what is a matched pair experiment

• Skip to secondary menu
• Skip to main content
• Skip to primary sidebar

Statistics By Jim

Making statistics intuitive

Matched Pairs Design: Uses & Examples

By Jim Frost Leave a Comment

What is a Matched Pairs Design?

A matched pairs design is an experimental design where researchers match pairs of participants by relevant characteristics. Then the researchers randomly assign one person from each pair to the treatment group and the other to the control group. This type of experiment is also known as a matching pairs design.

Statisticians recommend using this design to control for potential confounders that would otherwise bias the study’s results. The matched pairs experimental design is particularly advantageous for studies with limited sample sizes. When sample sizes are small, it can be challenging to achieve well-balanced groups through random assignment alone.

To conduct this type of experiment, researchers must identify the characteristics they’ll use to match the participants. Typically, these attributes include the potential confounders along with other relevant qualities such as age, gender, and race. Matching factors can incorporate medical history, lifestyle habits, and baseline measurements of the outcome of interest.

After identifying the criteria for the matched pairs design, researchers select pairs of participants with similar characteristics. Then they split each pair between the two experimental groups. For example, if a pair of participants match on the relevant variables, the researchers randomly assign one to the treatment group and the other to the control group.

This process creates two similar experimental groups. The goal is to reduce variability between groups relative to a typical between-subjects study.

Learn more about Experimental Designs , Random Assignment , and Control Groups .

Suppose a study evaluates the effectiveness of a new drug for treating hypertension. The researchers match participants on their age, gender, BMI, and baseline blood pressure and then randomly assign the members of each pair to receive the drug or a placebo.

This matched pairs design ensures that the treatment and control groups have similar characteristics at the beginning of the study. Notably, the process explicitly equalizes the factors the researchers know to affect hypertension. Consequently, if the mean blood pressures of the treatment and control groups differ at the end of the study, the researchers can confidently state that the drug caused the difference.

Advantages of a Matched Pairs Design

Helps researchers draw causal conclusions.

A matched pairs design helps researchers draw causal inferences by controlling for confounding variables. It helps ensure that the experimental groups are equivalent before the experiment. Hence, the experimental treatment likely caused the differences the researchers observed afterward.

Learn more about Confounding Variables and how they can bias the results.

Increases Statistical Power and Precision

Another advantage of this experimental design is that it helps increase the precision and statistical power of the study. By matching participants, the experimental design reduces the variability between groups, making it easier to detect a significant difference between them. This condition increases a hypothesis test’s ability to find an effect when it exists and produces a more precise estimate of the effect.

Learn more about Statistical Power and How Confidence Intervals Assess Precision .

Disadvantages of a Matched Pairs Design

2x dropouts.

With a matched pairs design, if one subject drops out of the study, the study must drop the other member of the pair. In other words, one dropout causes the study to lose two participants!

Matching Can Be Difficult

Researchers might find it challenging and time-consuming to find participants who match on all the characteristics. As the number of variables increases, the challenge of matching subjects for all of them also increases. This difficulty can increase the cost and logistical challenges of the study and limit the sample size.

Might Not Control All Confounders

This disadvantage is an extension of the previous one. If the outcome of interest is complex and involves many factors, a matched pairs design might not be able to match participants on all of them. When a design does not control a confounder, it can bias the results, making them untrustworthy.

In this case, researchers can use random assignment with a sufficiently large sample size. This approach requires larger samples, but it tends to produce equivalent experimental groups without requiring researchers to match subjects.

Despite these limitations, a matched pairs design is a valuable tool for conducting experiments. By carefully selecting and matching participants, researchers can use smaller sample sizes while increasing the statistical power of their study and obtain more precise estimates of the treatment effect.

Share this:

Reader Interactions

Comments and questions cancel reply.

Experimental Design: Types, Examples & Methods

Saul McLeod, PhD

Editor-in-Chief for Simply Psychology

BSc (Hons) Psychology, MRes, PhD, University of Manchester

Saul McLeod, PhD., is a qualified psychology teacher with over 18 years of experience in further and higher education. He has been published in peer-reviewed journals, including the Journal of Clinical Psychology.

Learn about our Editorial Process

Olivia Guy-Evans, MSc

Associate Editor for Simply Psychology

BSc (Hons) Psychology, MSc Psychology of Education

Olivia Guy-Evans is a writer and associate editor for Simply Psychology. She has previously worked in healthcare and educational sectors.

On This Page:

Experimental design refers to how participants are allocated to different groups in an experiment. Types of design include repeated measures, independent groups, and matched pairs designs.

Probably the most common way to design an experiment in psychology is to divide the participants into two groups, the experimental group and the control group, and then introduce a change to the experimental group, not the control group.

The researcher must decide how he/she will allocate their sample to the different experimental groups.  For example, if there are 10 participants, will all 10 participants participate in both groups (e.g., repeated measures), or will the participants be split in half and take part in only one group each?

Three types of experimental designs are commonly used:

1. Independent Measures

Independent measures design, also known as between-groups , is an experimental design where different participants are used in each condition of the independent variable.  This means that each condition of the experiment includes a different group of participants.

This should be done by random allocation, ensuring that each participant has an equal chance of being assigned to one group.

Independent measures involve using two separate groups of participants, one in each condition. For example:

• Con : More people are needed than with the repeated measures design (i.e., more time-consuming).
• Pro : Avoids order effects (such as practice or fatigue) as people participate in one condition only.  If a person is involved in several conditions, they may become bored, tired, and fed up by the time they come to the second condition or become wise to the requirements of the experiment!
• Con : Differences between participants in the groups may affect results, for example, variations in age, gender, or social background.  These differences are known as participant variables (i.e., a type of extraneous variable ).
• Control : After the participants have been recruited, they should be randomly assigned to their groups. This should ensure the groups are similar, on average (reducing participant variables).

2. Repeated Measures Design

Repeated Measures design is an experimental design where the same participants participate in each independent variable condition.  This means that each experiment condition includes the same group of participants.

Repeated Measures design is also known as within-groups or within-subjects design .

• Pro : As the same participants are used in each condition, participant variables (i.e., individual differences) are reduced.
• Con : There may be order effects. Order effects refer to the order of the conditions affecting the participants’ behavior.  Performance in the second condition may be better because the participants know what to do (i.e., practice effect).  Or their performance might be worse in the second condition because they are tired (i.e., fatigue effect). This limitation can be controlled using counterbalancing.
• Pro : Fewer people are needed as they participate in all conditions (i.e., saves time).
• Control : To combat order effects, the researcher counter-balances the order of the conditions for the participants.  Alternating the order in which participants perform in different conditions of an experiment.

Counterbalancing

Suppose we used a repeated measures design in which all of the participants first learned words in “loud noise” and then learned them in “no noise.”

We expect the participants to learn better in “no noise” because of order effects, such as practice. However, a researcher can control for order effects using counterbalancing.

The sample would be split into two groups: experimental (A) and control (B).  For example, group 1 does ‘A’ then ‘B,’ and group 2 does ‘B’ then ‘A.’ This is to eliminate order effects.

Although order effects occur for each participant, they balance each other out in the results because they occur equally in both groups.

3. Matched Pairs Design

A matched pairs design is an experimental design where pairs of participants are matched in terms of key variables, such as age or socioeconomic status. One member of each pair is then placed into the experimental group and the other member into the control group .

One member of each matched pair must be randomly assigned to the experimental group and the other to the control group.

• Con : If one participant drops out, you lose 2 PPs’ data.
• Pro : Reduces participant variables because the researcher has tried to pair up the participants so that each condition has people with similar abilities and characteristics.
• Con : Very time-consuming trying to find closely matched pairs.
• Pro : It avoids order effects, so counterbalancing is not necessary.
• Con : Impossible to match people exactly unless they are identical twins!
• Control : Members of each pair should be randomly assigned to conditions. However, this does not solve all these problems.

Experimental design refers to how participants are allocated to an experiment’s different conditions (or IV levels). There are three types:

1. Independent measures / between-groups : Different participants are used in each condition of the independent variable.

2. Repeated measures /within groups : The same participants take part in each condition of the independent variable.

3. Matched pairs : Each condition uses different participants, but they are matched in terms of important characteristics, e.g., gender, age, intelligence, etc.

Learning Check

Read about each of the experiments below. For each experiment, identify (1) which experimental design was used; and (2) why the researcher might have used that design.

1 . To compare the effectiveness of two different types of therapy for depression, depressed patients were assigned to receive either cognitive therapy or behavior therapy for a 12-week period.

The researchers attempted to ensure that the patients in the two groups had similar severity of depressed symptoms by administering a standardized test of depression to each participant, then pairing them according to the severity of their symptoms.

2 . To assess the difference in reading comprehension between 7 and 9-year-olds, a researcher recruited each group from a local primary school. They were given the same passage of text to read and then asked a series of questions to assess their understanding.

3 . To assess the effectiveness of two different ways of teaching reading, a group of 5-year-olds was recruited from a primary school. Their level of reading ability was assessed, and then they were taught using scheme one for 20 weeks.

At the end of this period, their reading was reassessed, and a reading improvement score was calculated. They were then taught using scheme two for a further 20 weeks, and another reading improvement score for this period was calculated. The reading improvement scores for each child were then compared.

4 . To assess the effect of the organization on recall, a researcher randomly assigned student volunteers to two conditions.

Condition one attempted to recall a list of words that were organized into meaningful categories; condition two attempted to recall the same words, randomly grouped on the page.

Experiment Terminology

Ecological validity.

The degree to which an investigation represents real-life experiences.

Experimenter effects

These are the ways that the experimenter can accidentally influence the participant through their appearance or behavior.

Demand characteristics

The clues in an experiment lead the participants to think they know what the researcher is looking for (e.g., the experimenter’s body language).

Independent variable (IV)

The variable the experimenter manipulates (i.e., changes) is assumed to have a direct effect on the dependent variable.

Dependent variable (DV)

Variable the experimenter measures. This is the outcome (i.e., the result) of a study.

Extraneous variables (EV)

All variables which are not independent variables but could affect the results (DV) of the experiment. Extraneous variables should be controlled where possible.

Confounding variables

Variable(s) that have affected the results (DV), apart from the IV. A confounding variable could be an extraneous variable that has not been controlled.

Random Allocation

Randomly allocating participants to independent variable conditions means that all participants should have an equal chance of taking part in each condition.

The principle of random allocation is to avoid bias in how the experiment is carried out and limit the effects of participant variables.

Order effects

Changes in participants’ performance due to their repeating the same or similar test more than once. Examples of order effects include:

(i) practice effect: an improvement in performance on a task due to repetition, for example, because of familiarity with the task;

(ii) fatigue effect: a decrease in performance of a task due to repetition, for example, because of boredom or tiredness.

Matched Pairs Design: Definition + Examples

A  matched pairs design  is an experimental design that is used when an experiment only has two treatment conditions. The subjects in the experiment are grouped together into pairs based on some variable they “match” on, such as age or gender. Then, within each pair, subjects are randomly assigned to different treatments

Example of a Matched Pairs Design

Suppose researchers want to know how a new diet affects weight loss compared to a standard diet. Since this experiment only has two treatment conditions (new diet and standard diet), they can use a matched pairs design.

They recruit 100 subjects, then group the subjects into 50 pairs based on their age and gender. For example:

• A 25-year-old male will be paired with another 25-year-old male, since they “match” in terms of age and gender.
• A 30-year-old female will be paired with another 30-year-old female since they also match on age and gender, and so on.

Then, within each pair, one subject will randomly be assigned to follow the new diet for 30 days and the other subject will be assigned to follow the standard diet for 30 days. At the end of the 30 days, researchers will measure the total weight loss for each subject.

Advantages & Disadvantages of a Matched Pairs Design

There are some notable advantages and some potential disadvantages of using a matched pairs design.

Advantages:

1. Controls for lurking variables .

A lurking variable is a variable that is not accounted for in an experiment that could potentially affect the outcomes of the experiment.

In the previous example, both age and gender can have a significant effect on weight loss. By matching subjects based on these two variables, we are eliminating the effect that these two variables could have on weight loss since we’re only comparing the weight loss between subjects who are identical in age and gender.

Thus, any difference in weight loss that we observe can be attributed to the diet, as opposed to age or gender.

2. Eliminates order effect . Order effect refers to differences in outcomes due to the order in which experimental materials are presented to subjects. By using a matched pairs design, you don’t have to worry about order effect since each subject only receives one treatment.

In our previous example, each subject in the experiment was only placed on one diet. If instead we made one subject use the standard diet for 30 days, then the new diet for 30 days, there could be an order effect due to the fact that the subject used one particular diet before the other.

Disadvantages:

1. Losing two subjects if one drops out. If one subject decides to drop out of the study, you actually lose two subjects since you no longer have a complete pair.

2. Time-consuming to find matches . It can be quite time-consuming to find subjects who match on certain variables, particularly if you use two or more variables. For example, it might not be hard to find 50 females to use as pairs, but it could be quite hard to find 50 female pairs in which each pair matches exactly on age.

3. Impossible to match subjects perfectly . No matter how hard researchers try, there will always be some variation within the subjects in each pair. The only way to match perfectly is to find identical twins who essentially share the same genetic code, which is actually why identical twins are often used in matched pairs studies.

Advantages of Using Ranges in a Matched Pairs Design

One way to make it slightly easier to find subjects that match is to use ranges for the variables you’re attempting to match on.

For example, instead of matching a 22-year-old with another 22-year old, researchers may instead create age ranges like 21-25, 26-30, 31-35, etc. so they can match one subject in the 21-25 age range with another subject in the 21-25 age range.

Using ranges has pros and cons. The obvious pro is that you can find matches more easily, but the con is that the subjects will match less precisely. For example, using the approach above it’s possible for a 21-year-old and a 25-year-old to be matched up, which is a rather notable difference in age. This is a trade-off that researchers must decide is worth or not in order to find pairs more easily.

A Simple Explanation of Internal Consistency

How to perform simple linear regression in stata, related posts, three-way anova: definition & example, two sample z-test: definition, formula, and example, one sample z-test: definition, formula, and example, how to find a confidence interval for a..., an introduction to the exponential distribution, an introduction to the uniform distribution, the breusch-pagan test: definition & example, population vs. sample: what’s the difference, introduction to multiple linear regression, dunn’s test for multiple comparisons.

Search site

What is a matched pairs design in behavioral science, what is a matched pairs design.

Matched pairs design is a research method used in experimental and quasi-experimental research to control for extraneous variables and reduce the influence of individual differences among participants. In this design, participants are paired based on similar characteristics or traits that are relevant to the study, such as age, gender, or socioeconomic status. Each pair is then randomly assigned to either the experimental group or the control group, ensuring that each group has a similar distribution of the matching variable. By controlling for these characteristics, researchers can more confidently attribute any observed differences in the outcome between the groups to the independent variable, rather than to individual differences or other confounding factors.

Examples of Matched Pairs Design

Drug treatment efficacy.

In a study evaluating the efficacy of a new drug for treating depression, participants could be matched on the severity of their symptoms at baseline. By pairing participants with similar depression scores, researchers can control for the initial severity of depression and better isolate the effects of the drug treatment.

Education Intervention

When assessing the impact of a new teaching method on student performance, researchers could match students based on their prior academic achievement. By pairing students with similar pre-intervention grades, the study can better account for individual differences in academic ability and more accurately measure the effect of the teaching method.

Behavioral Therapy

In a study examining the effectiveness of cognitive-behavioral therapy (CBT) for anxiety disorders, participants could be matched based on the type and severity of their anxiety symptoms. This matching would help control for the specific anxiety disorder and its severity, allowing for a more accurate assessment of CBT’s effectiveness.

Shortcomings and Criticisms of Matched Pairs Design

Difficulty in matching participants.

Finding suitable matches for all participants can be challenging, particularly in studies with a small sample size or with multiple matching variables. Incomplete or imperfect matching can introduce confounding factors and weaken the validity of the study’s conclusions.

Reduced Statistical Power

Because matched pairs design requires the formation of pairs, the effective sample size for statistical analyses is often smaller than in a completely randomized design. This can result in reduced statistical power, making it more difficult to detect significant effects.

Time and Resource Intensive

Matched pairs design can be more time-consuming and resource-intensive than other research designs, as it requires the collection of data on matching variables, the formation of suitable pairs, and additional data analysis techniques to account for the pairing structure.

Related Behavioral Science Terms

Belief perseverance, crystallized intelligence, extraneous variable, representative sample, factor analysis, egocentrism, stimulus generalization, reciprocal determinism, divergent thinking, convergent thinking, social environment, decision making, related articles.

Default Nudges: Fake Behavior Change

Here’s Why the Loop is Stupid

How behavioral science can be used to build the perfect brand

The Death Of Behavioral Economics

All Subjects

AP Statistics

Matched Pairs Design

Matched pairs design is a research method where each participant is paired with another participant who has similar characteristics, and then one member of the pair receives the treatment while the other serves as a control. This helps to eliminate confounding variables and increase the validity of the study.

Related terms

Control Group : In an experiment, the control group refers to the group that does not receive any treatment or intervention. It serves as a baseline for comparison to determine if there are any effects caused by the treatment.

Experimental Group : The experimental group in an experiment refers to the group that receives a specific treatment or intervention being studied. It is compared to the control group to evaluate whether or not there are significant differences between them.

Confounding Variables : Confounding variables are factors other than the independent variable that may influence or affect the dependent variable in an experiment. They can lead to inaccurate conclusions if not properly controlled for during data analysis.

" Matched Pairs Design " appears in:

Study guides ( 1 ).

• AP Statistics - 3.6 Selecting an Experimental Design

Subjects ( 3 )

• Data, Inference, and Decisions
• Experimental Design
• Honors Statistics

Practice Questions ( 5 )

• In a matched pairs design, each pair receives:
• What is a matched pairs design?
• What is the advantage of using a matched pairs design in an experiment?
• What is the key difference between a blocking design and a matched pairs design?
• In a matched pairs design, how are the two treatments assigned to the paired individuals?

© 2024 Fiveable Inc. All rights reserved.

Ap® and sat® are trademarks registered by the college board, which is not affiliated with, and does not endorse this website..

Matched Pairs Design: Definition + Examples

A  matched pairs design  is an experimental design that is used when an experiment only has two treatment conditions. The subjects in the experiment are grouped together into pairs based on some variable they “match” on, such as age or gender. Then, within each pair, subjects are randomly assigned to different treatments

Example of a Matched Pairs Design

Suppose researchers want to know how a new diet affects weight loss compared to a standard diet. Since this experiment only has two treatment conditions (new diet and standard diet), they can use a matched pairs design.

They recruit 100 subjects, then group the subjects into 50 pairs based on their age and gender. For example:

• A 25-year-old male will be paired with another 25-year-old male, since they “match” in terms of age and gender.
• A 30-year-old female will be paired with another 30-year-old female since they also match on age and gender, and so on.

Then, within each pair, one subject will randomly be assigned to follow the new diet for 30 days and the other subject will be assigned to follow the standard diet for 30 days. At the end of the 30 days, researchers will measure the total weight loss for each subject.

Advantages & Disadvantages of a Matched Pairs Design

There are some notable advantages and some potential disadvantages of using a matched pairs design.

Advantages:

1. Controls for lurking variables .

A is a variable that is not accounted for in an experiment that could potentially affect the outcomes of the experiment.

In the previous example, both age and gender can have a significant effect on weight loss. By matching subjects based on these two variables, we are eliminating the effect that these two variables could have on weight loss since we’re only comparing the weight loss between subjects who are identical in age and gender.

Thus, any difference in weight loss that we observe can be attributed to the diet, as opposed to age or gender.

2. Eliminates order effect . refers to differences in outcomes due to the order in which experimental materials are presented to subjects. By using a matched pairs design, you don’t have to worry about order effect since each subject only receives one treatment.

In our previous example, each subject in the experiment was only placed on one diet. If instead we made one subject use the standard diet for 30 days, then the new diet for 30 days, there could be an order effect due to the fact that the subject used one particular diet before the other.

Disadvantages:

2. Time-consuming to find matches . It can be quite time-consuming to find subjects who match on certain variables, particularly if you use two or more variables. For example, it might not be hard to find 50 females to use as pairs, but it could be quite hard to find 50 female pairs in which each pair matches exactly on age.

3. Impossible to match subjects perfectly . No matter how hard researchers try, there will always be some variation within the subjects in each pair. The only way to match perfectly is to find identical twins who essentially share the same genetic code, which is actually why identical twins are often used in matched pairs studies.

Advantages of Using Ranges in a Matched Pairs Design

One way to make it slightly easier to find subjects that match is to use ranges for the variables you’re attempting to match on.

For example, instead of matching a 22-year-old with another 22-year old, researchers may instead create age ranges like 21-25, 26-30, 31-35, etc. so they can match one subject in the 21-25 age range with another subject in the 21-25 age range.

Using ranges has pros and cons. The obvious pro is that you can find matches more easily, but the con is that the subjects will match less precisely. For example, using the approach above it’s possible for a 21-year-old and a 25-year-old to be matched up, which is a rather notable difference in age. This is a trade-off that researchers must decide is worth or not in order to find pairs more easily.

10.4 Matched or Paired Samples

When using a hypothesis test for matched or paired samples, the following characteristics should be present:

• Simple random sampling is used.
• Sample sizes are often small.
• Two measurements (samples) are drawn from the same pair of individuals or objects.
• Differences are calculated from the matched or paired samples.
• The differences form the sample that is used for the hypothesis test.
• Either the matched pairs have differences that come from a population that is normal or the number of differences is sufficiently large so that distribution of the sample mean of differences is approximately normal.

In a hypothesis test for matched or paired samples, subjects are matched in pairs and differences are calculated. The differences are the data. The population mean for the differences, μ d , is then tested using a Student's-t test for a single population mean with n – 1 degrees of freedom, where n is the number of differences.

Example 10.11

A study was conducted to investigate the effectiveness of hypnotism in reducing pain. Results for randomly selected subjects are shown in Table 10.11 . A lower score indicates less pain. The "before" value is matched to an "after" value and the differences are calculated. The differences have a normal distribution. Are the sensory measurements, on average, lower after hypnotism? Test at a 5% significance level.

Corresponding "before" and "after" values form matched pairs. (Calculate "after" – "before.")

The data for the test are the differences: {0.2, –4.1, –1.6, –1.8, –3.2, –2, –2.9, –9.6}

The sample mean and sample standard deviation of the differences are: x – d = –3.13 x – d = –3.13 and s d = 2.91 s d = 2.91 Verify these values.

Let μ d μ d be the population mean for the differences. We use the subscript d d to denote "differences."

Random variable: X ¯ d X ¯ d = the mean difference of the sensory measurements

H 0 : μ d ≥ 0

The null hypothesis is zero or positive, meaning that there is the same or more pain felt after hypnotism. That means the subject shows no improvement. μ d is the population mean of the differences.)

H a : μ d < 0

The alternative hypothesis is negative, meaning there is less pain felt after hypnotism. That means the subject shows improvement. The score should be lower after hypnotism, so the difference ought to be negative to indicate improvement.

Distribution for the test: The distribution is a Student's t with df = n – 1 = 8 – 1 = 7. Use t 7 . (Notice that the test is for a single population mean.)

Calculate the p -value using the Student's t-distribution: p -value = 0.0095

X ¯ d X ¯ d is the random variable for the differences.

The sample mean and sample standard deviation of the differences are:

x ¯ d x ¯ d = –3.13

s ¯ d s ¯ d = 2.91

Compare α and the p -value: α = 0.05 and p -value = 0.0095. α > p -value.

Make a decision: Since α > p -value, reject H 0 . This means that μ d < 0 and there is improvement.

Conclusion: At a 5% level of significance, from the sample data, there is sufficient evidence to conclude that the sensory measurements, on average, are lower after hypnotism. Hypnotism appears to be effective in reducing pain.

For the TI-83+ and TI-84 calculators, you can either calculate the differences ahead of time ( after - before ) and put the differences into a list or you can put the after data into a first list and the before data into a second list. Then go to a third list and arrow up to the name. Enter 1 st list name - 2 nd list name. The calculator will do the subtraction, and you will have the differences in the third list.

Using the TI-83, 83+, 84, 84+ Calculator

Use your list of differences as the data. Press STAT and arrow over to TESTS . Press 2:T-Test . Arrow over to Data and press ENTER . Arrow down and enter 0 for μ 0 μ 0 , the name of the list where you put the data, and 1 for Freq:. Arrow down to μ : and arrow over to < μ 0 μ 0 . Press ENTER . Arrow down to Calculate and press ENTER . The p -value is 0.0094, and the test statistic is -3.04. Do these instructions again except, arrow to Draw (instead of Calculate ). Press ENTER .

Try It 10.11

A study was conducted to investigate how effective a new diet was in lowering cholesterol. Results for the randomly selected subjects are shown in the table. The differences have a normal distribution. Are the subjects’ cholesterol levels lower on average after the diet? Test at the 5% level.

Example 10.12

A college football coach was interested in whether the college's strength development class increased his players' maximum lift (in pounds) on the bench press exercise. He asked four of his players to participate in a study. The amount of weight they could each lift was recorded before they took the strength development class. After completing the class, the amount of weight they could each lift was again measured. The data are as follows:

The coach wants to know if the strength development class makes his players stronger, on average. Record the differences data. Calculate the differences by subtracting the amount of weight lifted prior to the class from the weight lifted after completing the class. The data for the differences are: {90, 11, -8, -8}. Assume the differences have a normal distribution.

Using the differences data, calculate the sample mean and the sample standard deviation.

x ¯ d x ¯ d = 21.3, s d = 46.7

The data given here would indicate that the distribution is actually right-skewed. The difference 90 may be an extreme outlier? It is pulling the sample mean to be 21.3 (positive). The means of the other three data values are actually negative.

Using the difference data, this becomes a test of a single __________ (fill in the blank).

Define the random variable: X ¯ d X ¯ d mean difference in the maximum lift per player.

The distribution for the hypothesis test is t 3 .

H 0 : μ d ≤ 0, H a : μ d > 0

Calculate the p -value: The p -value is 0.2150

Decision: If the level of significance is 5%, the decision is not to reject the null hypothesis, because α < p -value.

What is the conclusion?

At a 5% level of significance, from the sample data, there is not sufficient evidence to conclude that the strength development class helped to make the players stronger, on average.

Try It 10.12

A new prep class was designed to improve SAT test scores. Five students were selected at random. Their scores on two practice exams were recorded, one before the class and one after. The data recorded in Table 10.15 . Are the scores, on average, higher after the class? Test at a 5% level.

Example 10.13

Seven eighth graders at Kennedy Middle School measured how far they could push the shot-put with their dominant (writing) hand and their weaker (non-writing) hand. They thought that they could push equal distances with either hand. The data were collected and recorded in Table 10.16 .

Conduct a hypothesis test to determine whether the mean difference in distances between the children’s dominant versus weaker hands is significant.

Record the differences data. Calculate the differences by subtracting the distances with the weaker hand from the distances with the dominant hand. The data for the differences are: {2, 12, 7, –1, 2, 0, 4}. The differences have a normal distribution.

Using the differences data, calculate the sample mean and the sample standard deviation. x ¯ d x ¯ d = 3.71, s d s d = 4.5.

Random variable: X ¯ d X ¯ d = mean difference in the distances between the hands.

Distribution for the hypothesis test: t 6

H 0 : μ d = 0  H a : μ d ≠ 0

Calculate the p -value: The p -value is 0.0716 (using the data directly).

(test statistic = 2.18. p -value = 0.0719 using ( x ¯ d = 3.71 ,   s d = 4.5. ) ( x ¯ d = 3.71 ,   s d = 4.5. )

Decision: Assume α = 0.05. Since α < p -value, Do not reject H 0 .

Conclusion: At the 5% level of significance, from the sample data, there is not sufficient evidence to conclude that there is a difference in the children’s weaker and dominant hands to push the shot-put.

Try It 10.13

Five ball players think they can throw the same distance with their dominant hand (throwing) and off-hand (catching hand). The data were collected and recorded in Table 10.17 . Conduct a hypothesis test to determine whether the mean difference in distances between the dominant and off-hand is significant. Test at the 5% level.

This book may not be used in the training of large language models or otherwise be ingested into large language models or generative AI offerings without OpenStax's permission.

Want to cite, share, or modify this book? This book uses the Creative Commons Attribution License and you must attribute OpenStax.

Access for free at https://openstax.org/books/introductory-statistics/pages/1-introduction

• Authors: Barbara Illowsky, Susan Dean
• Publisher/website: OpenStax
• Book title: Introductory Statistics
• Publication date: Sep 19, 2013
• Location: Houston, Texas
• Book URL: https://openstax.org/books/introductory-statistics/pages/1-introduction
• Section URL: https://openstax.org/books/introductory-statistics/pages/10-4-matched-or-paired-samples

© Jun 23, 2022 OpenStax. Textbook content produced by OpenStax is licensed under a Creative Commons Attribution License . The OpenStax name, OpenStax logo, OpenStax book covers, OpenStax CNX name, and OpenStax CNX logo are not subject to the Creative Commons license and may not be reproduced without the prior and express written consent of Rice University.

Matched Pair Design Statistics: Enhancing Precision in Research

Matched pair design in statistics involves comparing two related groups. This method controls for variables that may affect the outcome.

Matched pair design, a vital component of experimental research, pairs similar subjects or groups together to minimize variability in the results. Researchers typically use this approach when they want to assess the effects of a specific treatment or intervention by comparing outcomes between the paired groups.

By matching subjects based on key characteristics, this design enhances the accuracy of attributing any observed differences to the factor under investigation rather than to extraneous variables. This statistical method is especially useful in small sample sizes and can be applied across various fields, from medicine to social sciences, providing insightful data for making informed decisions. The precise pairing process and controlled environment this design offers help scientists and researchers isolate the impact of their variables of interest.

Credit: www.simplypsychology.org

Introduction To Matched Pair Design

Delving into the realm of statistics , we often encounter experimental designs that aim to reduce variability and improve the reliability of results . One such approach is the Matched Pair Design . It’s a tactic researchers employ to compare two treatments while minimizing differences between the subjects. Let’s explore this design’s essence and its advantages in research.

Essence Of Matched Pair Design

The core concept of Matched Pair Design lies in its pairing mechanism . By matching subjects based on key characteristics, it ensures that each pair is as similar as possible . This similarity is crucial because it controls confounding variables , focusing solely on the treatment effects. In a Matched Pair Design, one subject from each pair receives one treatment , and the other subject receives another treatment . Researchers often use this design to tease apart the differences that the treatments may elicit.

Advantages In Research Settings

• Improves accuracy : By matching subjects closely, the design reduces the chance of external variables skewing the data .
• Controls variability : It minimizes effects of individual differences that are not related to the treatments being tested.
• Efficient use of data : Matched Pair Design tends to require fewer subjects to achieve statistical significance , making research more time and cost-effective .
• Flexible pairing : Researchers can match subjects on a multitude of characteristics, making it highly adaptable to various studies.

Key Principles Of Matched Pair Design

Understanding the key principles of Matched Pair Design can enhance the accuracy of statistical studies. This design pairs participants closely based on specific criteria. It helps reduce variability and draws clear conclusions on cause-effect relationships. Let’s explore the core principles behind this powerful statistical approach.

Creating Matched Pairs

To ensure credible results , creating matched pairs requires careful attention to detail:

• Identify key characteristics where participants are alike.
• Pair up participants to neutralize confounding variables .
• Use relevant data and metrics to match subjects correctly.

Creating pairs this way leads to more reliable comparisons .

Randomization And Control

Randomization and control play pivotal roles in Matched Pair Design:

Together, these principles strengthen the study’s integrity and its findings.

Implementing Matched Pair Design

If you want results you can trust, matched pair design statistics can be a game-changer. This method compares two groups that are alike in many ways. It helps find out if a certain change or treatment causes different results. Now, let’s learn how to put this powerful tool into action.

Selecting Variables For Matching

Choosing the right variables is key to a successful study . Think about the parts of your experiment that could affect the outcome. Here are some steps to select the best variables for matching:

• Identify core characteristics : Look for traits that could sway your findings, like age or gender.
• Analyze past data : Previous studies can show which variables matter most.
• Consider the treatment : Make sure variables match the treatment’s effects.

Pairing Techniques

Once you’ve selected your variables, the next step is pairing. Use these techniques to make pairs that are as close as a twin set:

• Random Pairing : This draws pairs randomly, aiming for a balanced mix.
• Exact Matching : Here, pairs are identical twins in the variables you chose.
• Rank Order : Sort participants, then pair top with top, bottom with bottom.

This table shows the methods side by side:

To make sure your matched pair design shines, always review your variables and pairing techniques. Get them right, and you’re on track for trustworthy, valuable results.

Credit: nap.nationalacademies.org

Analyzing Data From Matched Pair Experiments

Matched pair experiments shine in statistical analysis. They allow researchers to compare two sets of data that are linked. But once the experiment ends, the real work begins—analyzing the data. This piece delves into the intricate process of peeling back the layers of matched pair data.

Statistical Tests For Matched Pairs

Choosing the right statistical test is key. It ensures the data speaks the truth. Here are some tests often used:

• t-test for dependent samples : Checks if mean differences are statistically significant.
• Wilcoxon signed-rank test : Non-parametric alternative to the t-test.
• Sign test : Simpler non-parametric test based on direction of change between pairs.

To run a test, you’ll need:

• Data normality check : Ensures data fits the chosen test.
• Paired observations : Confirms data is properly matched.
• Significance level set : Defines the threshold for determining results.

Once these steps are completed, apply the test to interpret your findings.

Interpreting Results

Interpreting results unlocks the experiment’s value. Keep an eye out for these pointers:

Look for a p-value less than 0.05. This often means your findings are strong. A wide confidence interval suggests more data may be needed. A large effect size speaks to the difference’s importance.

Understanding these elements leads to credible conclusions from your matched pair experiment.

Case Studies

Matched pair design statistics offer a unique lens through which researchers can observe the effectiveness of interventions. This design reduces variability and increases the statistical power of the study. By looking at specific case studies, we can explore how this design enhances the reliability of results in various fields.

Matched Pair Design In Clinical Trials

Case studies in clinical trials show how matched pair design elevates research quality. Doctors use it to compare treatments. Patients with similar attributes like age and health status create pairs. One gets the new treatment. The other gets the standard or a placebo. This direct comparison often reveals which treatment works best.

• Better control for variables: Age, gender, and health level in pairs are similar.
• Clearer outcomes: Reduces outside influence on results.
• Increased reliability: Offers robust evidence for new treatments.

Innovative Applications In Social Sciences

Social scientists also leverage matched pair design. They study behaviors and attitudes. Participants with like characteristics form pairs. Researchers observe how different social factors affect them. This method isolates the factor’s effect and minimizes biases.

• Detailed insights: Delivers deeper understanding of social dynamics.
• Reduces bias: Matching on key attributes controls for confounding variables.
• Broader application: Adapts to various social science fields.

Credit: www.khanacademy.org

Challenges And Considerations

Using Matched Pair Design in statistics is smart. But it has its own set of puzzles to solve. What are these puzzles? Let’s dive in.

Limitations Of Matched Pair Design

Even the best tools have limits. Matched Pair Design is no different.

• Perfect matches are rare. Finding two things that are the same in every way is tough.
• Dropouts can skew results. If subjects leave, it messes up our matching.
• Not for large groups. When we have lots of subjects, this design gets hard.

This design shines in control, but these limitations need careful thought.

Ensuring Validity And Reliability

Validity and reliability are like the North Star in studies; they guide us to truth.

• Keep it tight. Stick to your plan to stop bias.
• Check your pairs. Make sure they really match up.
• Repeat the test. The same results over and again mean it’s reliable.

By focusing on these steps, we steer towards meaningful outcomes.

Frequently Asked Questions For Matched Pair Design Statistics

What is an example of a matched pair in statistics.

A matched pair in statistics occurs when researchers pair subjects based on similar characteristics before applying different treatments. For example, identical twins receiving different diets to study the effects on weight loss.

What Is The Statistical Advantage Of Using A Matched Pairs Design?

A matched pairs design boosts statistical power by reducing variability, ensuring that comparisons between conditions are more precise and require fewer subjects.

What Is A Matched Pair In Ap Stats?

A matched pair in AP Statistics refers to two related observations, often the same subject before and after a treatment, used for comparison in paired samples or experiments.

What Are The Strengths Of Matched Pairs Design?

Matched pairs design increases result accuracy by pairing similar subjects, reducing variability. It enables controlling for participant-specific variables, enhancing statistical power. This method minimizes the impact of confounding variables, leading to stronger, more reliable conclusions.

Understanding matched pair design statistics is crucial for precise data analysis. By recognizing the value of pairing subjects, researchers can significantly reduce variability. This method enhances the accuracy of statistical results, leading to more trustworthy conclusions. Embrace this approach to strengthen your research endeavors and glean meaningful insights from your data sets.

Leave a Reply Cancel reply

Your email address will not be published. Required fields are marked *

Save my name, email, and website in this browser for the next time I comment.

MOST POPULAR

Migrant Journey To Martha’s Vineyard Leads To Legal Battle Against Aviation Company, with Political Undercurrents

Orthodontics: A Key Component of a Beautiful and Healthy Smile

A confident smile is not just a sign of happiness; it's also an indicator of …

WYSL’s Bob Savage Responds to Pennysaver Vent

Geneseo, NY—Bob Savage, owner of WYSL Radio, recently issued a detailed response to criticism published …

Bucci and His Horse, Honorable Dillon, Have Sights Set on the Kentucky Derby

John Bucci and his prized horse, Honorable Dillon, prepare for a shot at the Kentucky …

Sign in to your account

Username or Email Address

Remember Me

Want to create or adapt books like this? Learn more about how Pressbooks supports open publishing practices.

Inference for Comparing Matched Pairs (HT for 2 Means, dependent samples)

More of the good stuff! We will need to know how to label the null and alternative hypothesis, calculate the test statistic, and then reach our conclusion using the critical value method or the p-value method.

The Test Statistic for a Test of Matched Pairs (2 Means from Dependent Samples):

[latex]t = \displaystyle \frac{\bar{x} - 0}{\frac{s}{\sqrt{n}}}[/latex]

What the different symbols mean:

[latex]n[/latex] is the sample size, or the number of pairs of data

[latex]df = n - 1[/latex] is the degrees of freedom

[latex]\mu_d[/latex] is the  mean value of the differences for the  population of all matched pairs of data

[latex]\bar{x}[/latex] is the sample mean of the computed differences for the paired sample data

[latex]s[/latex] is the sample standard deviation of the computed differences for the paired sample data

• [latex]\alpha[/latex] is the significance level , usually given within the problem, or if not given, we assume it to be 5% or 0.05

Assumptions when conducting a Test for Matched Pairs:

• The two samples or groups are dependent
• The matched pairs are a simple random sample
• The number of pairs of sample data is large ([latex]n  > 30[/latex]),  OR the pairs of values have differences from a population that is approximately normal.

Steps to conduct the Test for Matched Pairs:

• Identify all the symbols listed above (all the stuff that will go into the formulas). This includes [latex]n[/latex], [latex]df[/latex], [latex]\mu_d[/latex], [latex]\bar{x}[/latex], [latex]s[/latex], and [latex]\alpha[/latex]
• Identify the null and alternative hypotheses
• Calculate the test statistic, [latex]t = \displaystyle \frac{\bar{x} - 0}{\frac{s}{\sqrt{n}}}[/latex]
• Find the critical value(s) OR the p-value OR both
• Apply the Decision Rule
• Write up a conclusion for the test

Example 1: Global Warming and Climate Change [1]

In Michael Crichton’s book “The State of Fear,” a reference is made to reported temperatures declining in Punta Arenas, at a weather station in South America. The reference in the book indicates that the temperature decreases there discredit climate change. There is a danger, however, in using data from only one source and one time period when making statements that might have worldwide impact. Instead of using data from one location and one time, it might be better to look at trends from many stations and from multiple time periods. The table below shows collected temperature readings from 32 NASA-GISS stations based on a random sample of latitude-longitude coordinates. The table is a matched-pairs design, and the differences can be analyzed to determine if we have statistically convincing evidence of true global warming (on average). [NOTE: since we are talking about global warming , the implication is that temperatures would be rising , so the mean difference would be thought of as an increase for the alternative hypothesis.] You can get a copy of the table in Google Sheets format here .

Since we are being asked for convincing statistical evidence, a hypothesis test should be conducted. In this case, we are dealing with gains (differences) from pairs of data, the pre- and post-tests, so we will conduct a Test for Matched Pairs.

• [latex]n = 32[/latex] is the sample size, or the number of pairs of data
• [latex]df = n - 1 = 32 - 1 = 31[/latex] is the degrees of freedom
• You can either manually add up and divide by how many, or you can use the Excel or Sheets formula =average() and make sure the appropriate numbers are entered or selected
• You can also do the same for standard deviation; use the =stdev() formula in Excel or Sheets
• [latex]s = 0.296[/latex] is the sample standard deviation of the computed differences for the paired sample data
• [latex]H_{0}: \mu_d = 0[/latex]
• [latex]H_{A}: \mu_d  > 0[/latex]
• [latex]t = \displaystyle \frac{\bar{x} - 0}{\frac{s}{\sqrt{n}}} = \displaystyle \frac{0.35 - 0}{\frac{0.296}{\sqrt{32}}} = 6.689[/latex]
• Microsoft Excel : You don’t need to have the Data Analysis ToolPack installed for this. Since we already have the differences calculated and we have the mean and standard deviation on those differences (the gain column), we can use the regular t-distribution on those values, including the test statistic and the degrees of freedom. We can use the built-in T.DIST.RT function to help calculate it. The “RT” in the formula is for the “more than” problems. The function will be typed into an empty cell in Excel (either installed on your computer, or using the online version) as =T.DIST.RT(x,deg_freedom), where x is the [latex]t[/latex] test statistic we just calculated (but always entered as a positive value), and deg_freedom is the [latex]df[/latex] we calculated earlier. The “RT” in the formula is for the “more than” problems. Step 1 illustrates how we would enter =T.DIST.RT(6.689,31). Step 2 gives us 8.78E-08, which is scientific notation. This means we move the decimal to the left 8 spaces, and we have a bunch of zeros in front of the 878. This means our actual value, if we round to 4 places, would be 0.0000, which is the [latex]p-value[/latex].

• Google Sheets : You can also do this using the exact same built-in function within Google Sheets. We can use the built-in T.DIST.RT function to help calculate it. The function will be typed into an empty cell in Google Sheets as =T.DIST.RT(x,deg_freedom), where x is the [latex]t[/latex] test statistic we just calculated (but always entered as a positive value), and deg_freedom is the [latex]df[/latex] we calculated earlier. The “RT” in the formula is for the “more than” problems. Step 1 illustrates how we would enter =T.DIST.RT(6.689,31). Step 2 gives us 0.0000, which is the [latex]p-value[/latex].
• StatDisk : We can conduct this test using StatDisk, but slightly modified from the full process. Since we already have the mean and standard deviation on the differences (the gain), we can use the regular test for one mean. The nice thing about StatDisk is that it will also compute the test statistic. From the main menu above we click on Analysis, Hypothesis Testing, and then Mean One Sample (the calculated “gain” is like a single sample now). From there enter the 0.05 significance, along with the specific values as outlined in the picture below in Step 2. Notice the alternative hypothesis is the [latex]>[/latex] option. Enter the sample size, mean, and standard deviation. Now we click on Evaluate. If you check the values, the test statistic is reported in the Step 3 display, as well as the P-Value of 0.0000.

• Applying the Decision Rule: We now compare this to our significance level, which is 0.05. If the p-value is smaller or equal to the alpha level, we have enough evidence for our claim, otherwise we do not. Here, [latex]p-value = 0.0000[/latex], which is smaller than [latex]\alpha = 0.05[/latex], so we have enough evidence for the alternative hypothesis…but what does this mean?
• Conclusion: Because our p-value  of [latex]0.0000[/latex] is smaller than our [latex]\alpha[/latex] level of [latex]0.05[/latex], we reject [latex]H_{0}[/latex]. We have convincing statistical evidence of true global warming (on average).

Example 2: Summer Institute for Foreign Language Instruction [2]

At UA High School there is a summer institute to improve the skills of high school teachers of foreign languages. One summer institute hosted 20 French teachers for 4 weeks. At the beginning of the period, teachers were given a baseline exam covering Modern Language listening. After 4 weeks of immersion in French in and out of class, the exam was administered once again. The table below gives pretest and posttest scores. Do the results give convincing statistical evidence that the institute improved the teacher’s comprehension of spoken French? You can get a copy of the data table in Google Sheets format here .

• [latex]n = 20[/latex] is the sample size, or the number of pairs of data
• [latex]df = n - 1 = 20 - 1 = 19[/latex] is the degrees of freedom
• [latex]s = 2.893[/latex] is the sample standard deviation of the computed differences for the paired sample data
• [latex]t = \displaystyle \frac{\bar{x} - 0}{\frac{s}{\sqrt{n}}} = \displaystyle \frac{2.5 - 0}{\frac{2.893}{\sqrt{20}}} = 3.86[/latex]
• Microsoft Excel : You don’t need to have the Data Analysis ToolPack installed for this. Since we already have the differences calculated and we have the mean and standard deviation on those differences (the gain column), we can use the regular t-distribution on those values, including the test statistic and the degrees of freedom. We can use the built-in T.DIST.RT function to help calculate it. The “RT” in the formula is for the “more than” problems. The function will be typed into an empty cell in Excel (either installed on your computer, or using the online version) as =T.DIST.RT(x,deg_freedom), where x is the [latex]t[/latex] test statistic we just calculated (but always entered as a positive value), and deg_freedom is the [latex]df[/latex] we calculated earlier. The “RT” in the formula is for the “more than” problems. Step 1 illustrates how we would enter =T.DIST.RT(3.86,19). Step 2 gives us 0.000527, which is the [latex]p-value[/latex].
• Google Sheets : You can also do this using the exact same built-in function within Google Sheets. We can use the built-in T.DIST.RT function to help calculate it. The function will be typed into an empty cell in Google Sheets as =T.DIST.RT(x,deg_freedom), where x is the [latex]t[/latex] test statistic we just calculated (but always entered as a positive value), and deg_freedom is the [latex]df[/latex] we calculated earlier. The “RT” in the formula is for the “more than” problems. Step 1 illustrates how we would enter =T.DIST.RT(3.86,19). Step 2 gives us 0.000527, which is the [latex]p-value[/latex].
• StatDisk : We can conduct this test using StatDisk, but slightly modified from the full process. Since we already have the mean and standard deviation on the differences (the gain), we can use the regular test for one mean. The nice thing about StatDisk is that it will also compute the test statistic. From the main menu above we click on Analysis, Hypothesis Testing, and then Mean One Sample (the calculated “gain” is like a single sample now). From there enter the 0.05 significance, along with the specific values as outlined in the picture below in Step 2. Notice the alternative hypothesis is the [latex]>[/latex] option. Enter the sample size, mean, and standard deviation. Now we click on Evaluate. If you check the values, the test statistic is reported in the Step 3 display, as well as the P-Value of 0.00052.
• Applying the Decision Rule: We now compare this to our significance level, which is 0.05. If the p-value is smaller or equal to the alpha level, we have enough evidence for our claim, otherwise we do not. Here, [latex]p-value = 0.000527[/latex], which is smaller than [latex]\alpha = 0.05[/latex], so we have enough evidence for the alternative hypothesis…but what does this mean?
• Conclusion: Because our p-value  of [latex]0.000527[/latex] is smaller than our [latex]\alpha[/latex] level of [latex]0.05[/latex], we reject [latex]H_{0}[/latex]. We have convincing statistical evidence that the institute improved the teacher’s comprehension of spoken French.
• Adapted from the Skew The Script curriculum ( skewthescript.org ), licensed under CC BY-NC-Sa 4.0 ↵
• Adapted from The Introduction to the Practice of Statistics, 3rd Edition, by Moore & McCabe ↵

Basic Statistics Copyright © by Allyn Leon is licensed under a Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License , except where otherwise noted.

Share This Book

Reference Library

Collections

• See what's new
• All Resources
• Student Resources
• Assessment Resources
• Teaching Resources
• CPD Courses
• Livestreams

Study notes, videos, interactive activities and more!

Psychology news, insights and enrichment

Currated collections of free resources

Browse resources by topic

• All Psychology Resources

Resource Selections

Currated lists of resources

Matched Pairs

Matched pairs design is an experimental design where pairs of participants are matched in terms of key variables, such as age and IQ. One member of each pair is then placed into the experimental group and the other member into the control group.

• Share on Facebook
• Share on Twitter
• Share by Email

Types of Experiment: Overview

Study Notes

Experimental Design

Research methods - experimental method.

Quizzes & Activities

Example Answers for Research Methods: A Level Psychology, Paper 2, June 2019 (AQA)

Exam Support

Research Methods: MCQ Revision Test 1 for AQA A Level Psychology

Topic Videos

Example Answers for Research Methods: A Level Psychology, Paper 2, June 2018 (AQA)

Example answer for question 3 paper 1: a level psychology, june 2017 (aqa), model answer for question 11 paper 2: as psychology, june 2016 (aqa), a level psychology topic quiz - research methods, our subjects.

• › Criminology
• › Economics
• › Geography
• › Health & Social Care
• › Psychology
• › Sociology
• › Teaching & learning resources
• › Student revision workshops
• › Online student courses
• › CPD for teachers
• › Livestreams
• › Teaching jobs

Boston House, 214 High Street, Boston Spa, West Yorkshire, LS23 6AD Tel: 01937 848885

• › Contact us
• › Terms of use
• › Privacy & cookies

© 2002-2024 Tutor2u Limited. Company Reg no: 04489574. VAT reg no 816865400.

Matched Pairs Design vs Randomized Block Design

In a matched pairs design, treatment options are randomly assigned to pairs of similar participants, whereas in a randomized block design, treatment options are randomly assigned to groups of similar participants. The objective of both is to balance baseline confounding variables by distributing them evenly between the treatment and the control group.

Matched pairs design works in 2 steps:

• Divide participants into pairs by matching each participant with their closest pair regarding some confounding variable(s) like age or gender.
• Within each pair, randomly assign 1 participant to either the treatment or the control group (and the other will be automatically assigned to the other group).

Randomized block design works in 2 steps:

• Divide participants into several subgroups by putting together those who are similar regarding some confounding variable(s) like age or gender.
• Within each subgroup, randomly assign participants to either the treatment or the control group.

Here’s a figure that summarizes the difference between a matched pairs design and a randomized block design that are both trying to equalize the treatment and control groups with regards to gender and smoking status:

When working with a small sample, using simple randomization alone can produce, just by chance, unbalanced groups regarding the patients’ initial characteristics (for a detailed discussion see:   Purpose and Limitations of Random Assignment ). In these cases, ensuring equivalence between participants by using either a matched pairs design or a randomized block design will increase the statistical power and precision of the study.

Where randomized block design is better:

Matched pairs design may not be the best option in the following cases:

• If an eligible participant will have to wait a long time to be randomized because a suitable match is hard to find.
• If paired participants may not be similar regarding other important characteristics.
• If the subgroups have an odd number of participants. In this case, each will be left with 1 unpaired participant. Losing some participants this way can be problematic in cases where we are already working with a small sample, and/or very few participants are eligible for the study.

Where matched pairs design is better:

Matching is especially useful in cases where participants can be paired with themselves.

For instance, in order to study the effect of a new sunscreen, the new product can be applied to the right arm (the treatment group), and the left arm can be used as control.

Where a completely randomized design is better than both:

Neither matching nor blocking is necessary in studies with large sample sizes, since in these cases, simple randomization alone is enough to balance study groups.

• Friedman LM, Furberg CD, DeMets DL, Reboussin DM, Granger CB. Fundament als of Clinical Trials. 5th edition. Springer; 2015.
• Hulley SB, Cummings SR, Browner WS, Grady DG, Newman TB. Designing Clinical Research . 4th edition. LWW; 2013.

Further reading

• Randomized Block Design
• Matched Pairs Design
• Posttest-Only Control Group Design
• Pretest-Posttest Control Group Design

Want to create or adapt books like this? Learn more about how Pressbooks supports open publishing practices.

8.1 Inference for Two Dependent Samples (Matched Pairs)

Learning Objectives

By the end of this chapter, the student should be able to:

• Classify hypothesis tests by type
• Conduct and interpret hypothesis tests for two population means, population standard deviations known
• Conduct and interpret hypothesis tests for two population means, population standard deviations unknown
• Conduct and interpret hypothesis tests for matched or paired samples
• Conduct and interpret hypothesis tests for two population proportions

Studies often compare two groups. For example, maybe researchers are interested in the effect aspirin has in preventing heart attacks.  One group is given aspirin and the other a placebo , and the heart attack rate is studied over several years.  Other studies may compare various diet and exercise programs.  Politicians compare the proportion of individuals from different income brackets who might vote for them. Students are interested in whether SAT or GRE preparatory courses really help raise their scores.

You have learned to conduct inference on single means and single proportions .  We know that the first step is deciding what type of data we are working with.  For quantitative data we are focused on means, while for categorical we are focused on proportions.  In this chapter we will compare two means or two proportions to each other.  The general procedure is still the same, just expanded.  With two sample analysis it is good to know what the formulas look like and where they come from, however you will probably lean heavily on technology in preforming the calculations.  

To compare two means we are obviously working with two groups, but first we need to think about the relationship between them. The groups are classified either as independent or dependent.  I ndependent samples consist of two samples that have no relationship, that is, sample values selected from one population are not related in any way to sample values selected from the other population.  Dependent samples consist of two groups that have some sort of identifiable relationship.

Two Dependent Samples (Matched Pairs)

Two samples that are dependent typically come from a matched pairs experimental design. The parameter tested using matched pairs is the population mean difference .  When using inference techniques for matched or paired samples, the following characteristics should be present:

• Simple random sampling is used.
• Sample sizes are often small.
• Two measurements (samples) are drawn from the same pair of (or two extremely similar) individuals or objects.
• Differences are calculated from the matched or paired samples.
• The differences form the sample that is used for analysis.

Confidence intervals may be calculated on their own for two samples but often, especially in the case of matched pairs, we first want to formally check to see if a difference exists with a hypothesis test.  If we do find a statistically significant difference then we may estimate it with a CI after the fact.

Hypothesis Tests for the Mean difference

In a hypothesis test for matched or paired samples, subjects are matched in pairs and differences are calculated, and the population mean difference, μ d , is our parameter of interest.  Although it is possible to test for a certain magnitude of effect, we are most often just looking for a general effect.  Our hypothesis would then look like:

H o : μ d =0

H a : μ d (<, >, ≠) 0

The steps are the same as we are familiar with, but it is tested using a Student’s-t test for a single population mean with n – 1 degrees of freedom, with the test statistic:

A study was conducted to investigate the effectiveness of hypnotism in reducing pain. Results for randomly selected subjects are shown in the figure below. A lower score indicates less pain. The “before” value is matched to an “after” value and the differences are calculated. The differences have a normal distribution. Are the sensory measurements, on average, lower after hypnotism? Test at a 5% significance level.

A study was conducted to investigate how effective a new diet was in lowering cholesterol. Results for the randomly selected subjects are shown in the table. The differences have a normal distribution. Are the subjects’ cholesterol levels lower on average after the diet? Test at the 5% level.

Confidence Intervals for the Mean difference

If we are using the t distribution, the error bound for the population mean difference is:

• use df = n – 1 degrees of freedom, where n is the number of pairs
• s d =  standard deviation of the differences.

A college football coach was interested in whether the college’s strength development class increased his players’ maximum lift (in pounds) on the bench press exercise. He asked four of his players to participate in a study. The amount of weight they could each lift was recorded before they took the strength development class. After completing the class, the amount of weight they could each lift was again measured. The data are as follows:

The coach wants to know if the strength development class makes his players stronger, on average.

Using the differences data, calculate the sample mean and the sample standard deviation.

Using the difference data, this becomes a test of a single __________ (fill in the blank).

Calculate the p -value:

What is the conclusion?

A new prep class was designed to improve SAT test scores. Five students were selected at random. Their scores on two practice exams were recorded, one before the class and one after. The data recorded in the figure below. Are the scores, on average, higher after the class? Test at a 5% level.

Image Credits

Figure 8.1: Ali Inay (2015). “Brunching with Friends.” Public domain. Retrieved from https://unsplash.com/photos/y3aP9oo9Pjc

Figure 8.3: Kindred Grey via Virginia Tech (2020). “Figure 8.3” CC BY-SA 4.0. Retrieved from https://commons.wikimedia.org/wiki/File:Figure_8.3.png . Adaptation of Figure 5.39 from OpenStax Introductory Statistics (2013) (CC BY 4.0). Retrieved from https://openstax.org/books/statistics/pages/5-practice

Figure 8.6: Kindred Grey via Virginia Tech (2020). “Figure 8.6” CC BY-SA 4.0. Retrieved from https://commons.wikimedia.org/wiki/File:Figure_8.6.png . Adaptation of Figure 5.39 from OpenStax Introductory Statistics (2013) (CC BY 4.0). Retrieved from https://openstax.org/books/statistics/pages/5-practice

An inactive treatment that has no real effect on the explanatory variable

The facet of statistics dealing with using a sample to generalize (or infer) about the population

The arithmetic mean, or average of a population

The number of individuals that have a characteristic we are interested in divided by the total number in the population

Numerical data with a mathematical context

Data that describes qualities, or puts individuals into categories

The occurrence of one event has no effect on the probability of the occurrence of another event

Very similar individuals (or even the same individual) receive two different two treatments (or treatment vs. control) then the difference in results are compared

The mean of the differences in a matched pairs design

The probability distribution of a statistic at a given sample size

The value that is calculated from a sample used to estimate an unknown population parameter

Significant Statistics Copyright © 2020 by John Morgan Russell, OpenStaxCollege, OpenIntro is licensed under a Creative Commons Attribution-ShareAlike 4.0 International License , except where otherwise noted.

Share This Book

LEARN STATISTICS EASILY

Learn Data Analysis Now!

What is: Matched Pairs

What is matched pairs.

Matched pairs refer to a statistical technique used primarily in the context of hypothesis testing and experimental design. This method involves pairing subjects or experimental units based on specific characteristics or criteria to control for confounding variables. By ensuring that each pair is as similar as possible, researchers can isolate the effect of the treatment or intervention being studied, leading to more reliable and valid results. Matched pairs are particularly useful in studies where random assignment may not be feasible or ethical, allowing for a more controlled comparison between groups.

Ad description. Lorem ipsum dolor sit amet, consectetur adipiscing elit.

Application of Matched Pairs in Research

In research, matched pairs are commonly employed in various fields, including psychology, medicine, and social sciences. For instance, in clinical trials, patients may be matched based on age, gender, or baseline health status before being assigned to different treatment groups. This approach minimizes the variability that could skew results, ensuring that any observed differences in outcomes can be attributed to the treatment itself rather than extraneous factors. The use of matched pairs enhances the internal validity of a study, making it a preferred method for many researchers.

Types of Matched Pairs Designs

There are several types of matched pairs designs, including complete and incomplete matching. In a complete matched pairs design, every participant in one group is paired with a participant in another group, ensuring a one-to-one correspondence. In contrast, an incomplete matching design may involve pairing only a subset of participants, which can be useful when dealing with larger populations or when certain characteristics are more critical than others. The choice of design often depends on the research question, available data, and the specific characteristics being controlled for.

Statistical Analysis of Matched Pairs

The analysis of matched pairs typically involves the use of paired statistical tests, such as the paired t-test or the Wilcoxon signed-rank test. The paired t-test is appropriate when the differences between pairs are normally distributed, allowing researchers to determine if there is a statistically significant difference in means between the two groups. On the other hand, the Wilcoxon signed-rank test is a non-parametric alternative that can be used when the normality assumption is violated. These tests provide valuable insights into the effectiveness of interventions or treatments by comparing outcomes within matched pairs.

Advantages of Using Matched Pairs

One of the primary advantages of using matched pairs is the reduction of variability, which enhances the precision of estimates. By controlling for confounding variables, researchers can draw more accurate conclusions about the effects of treatments or interventions. Additionally, matched pairs designs often require smaller sample sizes compared to completely randomized designs, making them more efficient in terms of resources and time. This efficiency is particularly beneficial in fields where data collection is costly or logistically challenging.

Challenges in Matched Pairs Design

Despite their advantages, matched pairs designs also present several challenges. One significant challenge is the difficulty in finding suitable matches for all participants, which can lead to incomplete data or biased results if not handled properly. Furthermore, the process of matching can introduce its own biases if the criteria used are not carefully considered. Researchers must also be cautious about over-matching, which can limit the generalizability of the findings. Addressing these challenges requires careful planning and a thorough understanding of the underlying assumptions of the matched pairs approach.

Matched Pairs vs. Independent Samples

When comparing matched pairs to independent samples, it is essential to recognize the fundamental differences in their design and analysis. Independent samples involve two separate groups that are not related, while matched pairs consist of related groups where each pair is linked by specific characteristics. The choice between these two designs often depends on the research question and the nature of the data. Matched pairs are generally more powerful for detecting differences because they control for individual variability, whereas independent samples may require larger sample sizes to achieve similar levels of statistical power.

Real-World Examples of Matched Pairs

Real-world applications of matched pairs can be found in various studies, such as clinical trials assessing the efficacy of new medications. For example, researchers may pair patients based on their pre-treatment health status and then assign one patient in each pair to receive the new medication while the other receives a placebo. This design allows for a direct comparison of outcomes, providing robust evidence regarding the medication’s effectiveness. Other examples include educational interventions where students are matched based on prior academic performance to evaluate the impact of new teaching methods.

Conclusion on Matched Pairs Methodology

In summary, matched pairs represent a powerful methodology in statistics and data analysis, offering researchers a means to control for confounding variables and enhance the validity of their findings. By carefully designing studies that utilize matched pairs, researchers can gain deeper insights into the effects of treatments and interventions, ultimately contributing to the advancement of knowledge in various fields. The strategic application of matched pairs can lead to more accurate conclusions and inform evidence-based practices across disciplines.

What is a Matched Pairs Design and what are some examples of it?

Table of Contents

A Matched Pairs Design is a type of research design in which two sets of data are compared, with each set consisting of pairs of similar subjects. In this design, the pairs are matched based on certain characteristics, such as age, gender, or other relevant factors. This ensures that any differences observed between the two groups can be attributed to the intervention or treatment being studied, rather than individual differences.

Some examples of Matched Pairs Design include comparing the effectiveness of two different medications on a group of patients with similar medical conditions or comparing the impact of two teaching methods on students with similar academic backgrounds. In both cases, the pairs would be matched to control for any confounding variables and to enhance the validity of the study results. This type of design is commonly used in medical and educational research, as well as in other fields where it is important to control for individual differences when examining the effects of a specific intervention or treatment.

A  matched pairs design  is an experimental design that is used when an experiment only has two treatment conditions. The subjects in the experiment are grouped together into pairs based on some variable they “match” on, such as age or gender. Then, within each pair, subjects are randomly assigned to different treatments

Example of a Matched Pairs Design

Suppose researchers want to know how a new diet affects weight loss compared to a standard diet. Since this experiment only has two treatment conditions (new diet and standard diet), they can use a matched pairs design.

They recruit 100 subjects, then group the subjects into 50 pairs based on their age and gender. For example:

• A 25-year-old male will be paired with another 25-year-old male, since they “match” in terms of age and gender.
• A 30-year-old female will be paired with another 30-year-old female since they also match on age and gender, and so on.

Then, within each pair, one subject will randomly be assigned to follow the new diet for 30 days and the other subject will be assigned to follow the standard diet for 30 days. At the end of the 30 days, researchers will measure the total weight loss for each subject.

Advantages & Disadvantages of a Matched Pairs Design

There are some notable advantages and some potential disadvantages of using a matched pairs design.

Advantages:

1. Controls for lurking variables .

A is a variable that is not accounted for in an experiment that could potentially affect the outcomes of the experiment.

In the previous example, both age and gender can have a significant effect on weight loss. By matching subjects based on these two variables, we are eliminating the effect that these two variables could have on weight loss since we’re only comparing the weight loss between subjects who are identical in age and gender.

Thus, any difference in weight loss that we observe can be attributed to the diet, as opposed to age or gender.

2. Eliminates order effect . refers to differences in outcomes due to the order in which experimental materials are presented to subjects. By using a matched pairs design, you don’t have to worry about order effect since each subject only receives one treatment.

In our previous example, each subject in the experiment was only placed on one diet. If instead we made one subject use the standard diet for 30 days, then the new diet for 30 days, there could be an order effect due to the fact that the subject used one particular diet before the other.

Disadvantages:

2. Time-consuming to find matches . It can be quite time-consuming to find subjects who match on certain variables, particularly if you use two or more variables. For example, it might not be hard to find 50 females to use as pairs, but it could be quite hard to find 50 female pairs in which each pair matches exactly on age.

3. Impossible to match subjects perfectly . No matter how hard researchers try, there will always be some variation within the subjects in each pair. The only way to match perfectly is to find identical twins who essentially share the same genetic code, which is actually why identical twins are often used in matched pairs studies.

Advantages of Using Ranges in a Matched Pairs Design

One way to make it slightly easier to find subjects that match is to use ranges for the variables you’re attempting to match on.

For example, instead of matching a 22-year-old with another 22-year old, researchers may instead create age ranges like 21-25, 26-30, 31-35, etc. so they can match one subject in the 21-25 age range with another subject in the 21-25 age range.

Using ranges has pros and cons. The obvious pro is that you can find matches more easily, but the con is that the subjects will match less precisely. For example, using the approach above it’s possible for a 21-year-old and a 25-year-old to be matched up, which is a rather notable difference in age. This is a trade-off that researchers must decide is worth or not in order to find pairs more easily.

Related terms:

• How to perfrom a matched pairs design
• Matched Group Design
• MATCHED-GROUP DESIGN
• How can I list all matched instances of a specific value in Excel?
• How To Create a Pairs Plot in Python
• # How to Create and Interpret Pairs Plots in R
• WORKSPACE DESIGN
• TOOL DESIGN
• OPTIMAL DESIGN
• ONE-GROUP PRE-POST DESIGN

• Voxco Online
• Voxco Panel Management
• Voxco Panel Portal
• Voxco Audience
• Voxco Mobile Offline
• Voxco Dialer Cloud
• Voxco Dialer On-premise
• Voxco TCPA Connect
• Voxco Analytics
• Voxco Text & Sentiment Analysis

• 40+ question types
• Drag-and-drop interface
• Skip logic and branching
• Multi-lingual survey
• Text piping
• Question library
• CSS customization
• White-label surveys
• Customizable ‘Thank You’ page
• Customizable survey theme
• Reminder send-outs
• Survey rewards
• Social media
• Website surveys
• Correlation analysis
• Cross-tabulation analysis
• Trend analysis
• Real-time dashboard
• Customizable report
• Email address validation
• Recaptcha validation
• SSL security

Take a peek at our powerful survey features to design surveys that scale discoveries.

Download feature sheet.

• Hospitality
• Academic Research
• Customer Experience
• Employee Experience
• Product Experience
• Market Research
• Social Research
• Data Analysis

Explore Voxco 

Need to map Voxco’s features & offerings? We can help!

Watch a Demo 

Download Brochures 

Get a Quote

• NPS Calculator
• CES Calculator
• A/B Testing Calculator
• Margin of Error Calculator
• Sample Size Calculator
• CX Strategy & Management Hub
• Market Research Hub
• Patient Experience Hub
• Employee Experience Hub
• NPS Knowledge Hub
• Market Research Guide
• Customer Experience Guide
• Survey Research Guides
• Survey Template Library
• Webinars and Events
• Feature Sheets
• Try a sample survey
• Professional Services

Get exclusive insights into research trends and best practices from top experts! Access Voxco’s ‘State of Research Report 2024 edition’ .

We’ve been avid users of the Voxco platform now for over 20 years. It gives us the flexibility to routinely enhance our survey toolkit and provides our clients with a more robust dataset and story to tell their clients.

VP Innovation & Strategic Partnerships, The Logit Group

• Client Stories
• Voxco Reviews
• Why Voxco Research?
• Careers at Voxco
• Vulnerabilities and Ethical Hacking

Explore Regional Offices

• Survey Software The world’s leading omnichannel survey software
• Online Survey Tools Create sophisticated surveys with ease.
• Mobile Offline Conduct efficient field surveys.
• Text Analysis
• Close The Loop
• Automated Translations
• NPS Dashboard
• CATI Manage high volume phone surveys efficiently
• Cloud/On-premise Dialer TCPA compliant Cloud on-premise dialer
• IVR Survey Software Boost productivity with automated call workflows.
• Analytics Analyze survey data with visual dashboards
• Panel Manager Nurture a loyal community of respondents.
• Survey Portal Best-in-class user friendly survey portal.
• Voxco Audience Conduct targeted sample research in hours.
• Predictive Analytics
• Customer 360
• Customer Loyalty
• Fraud & Risk Management
• AI/ML Enablement Services
• Credit Underwriting

Find the best survey software for you! (Along with a checklist to compare platforms)

Get Buyer’s Guide

• 100+ question types
• SMS surveys
• Financial Services
• Banking & Financial Services
• Retail Solution
• Risk Management
• Customer Lifecycle Solutions
• Net Promoter Score
• Customer Behaviour Analytics
• Customer Segmentation
• Data Unification

Explore Voxco 

Watch a Demo 

Download Brochures 

• CX Strategy & Management Hub
• The Voxco Guide to Customer Experience
• Professional services
• Blogs & White papers
• Case Studies

Find the best customer experience platform

Uncover customer pain points, analyze feedback and run successful CX programs with the best CX platform for your team.

Get the Guide Now

VP Innovation & Strategic Partnerships, The Logit Group

• Why Voxco Intelligence?
• Our clients
• Client stories
• Featuresheets

Matched Pairs Experimental Design

• October 5, 2021

SHARE THE ARTICLE ON

What is a Matched Pairs Experimental Design?

A matched pairs design is a type of experimental design wherein study participants are matched based on key variables, or shared characteristics, relevant to the topic of the study. Then, one member of each pair is placed into the control group while the other is placed in the experimental group. Participants are assigned to each group using random criteria, so as to avoid any potential bias.

See Voxco survey software in action with a Free demo.

When is the Matched Pairs Experimental Design Used

The matched pairs experimental design is most beneficial for studies that have small sample sizes. This is because it is harder to obtain balanced groups when using small sample sizes, even with the use of random assignment. 

Studies that employ smaller sample sizes generally have financial constraints or time constraints, making it unfeasible to have a larger sample size. With the use of the matched pairs design, researchers can improve the comparability of their study participants despite their smaller sample size, increasing the validity of the cause-and-effect relationship identified in the experiment. 

Additionally, matched pairs design can only be used when there are two treatment conditions so that one person from each pair can be assigned the first treatment and the other can be assigned the second treatment. 

How does a matched pair design function?

In this design, members are brought together because of a particular attribute or factors applicable to the concentrate and afterward split into various circumstances. A member will then be allotted to the control group in each pair, and the other member will be assigned to the trial group. The strategies are then equivalent to the free groups’ plan. Each group just encounters one degree of IV. The mean consequences of the matches would be analyzed after the trial.

Wondering what will be the cost of conducting survey research using Voxco?

Example of a Matched Pairs Experimental Design

Let’s take a look at the following example of matched pairs design in order to understand this experimental design better: 

Researchers want to find out how a new diet affects weight gain among underweight subjects. This experiment only has two treatment conditions, the new diet and the standard diet, hence the matched pairs design can be used. For this study, the researchers recruited 200 subjects which will be grouped into 100 pairs based on shared characteristics such as age, gender, weight, height, lifestyle, and so on. For example:

• A 20-year-old female within the weight range of 40-50 kgs and the height range of 156-160 cms will be paired with another 20-year-old female that falls into the same weight and height categories. 
• A 30-year-old male within the weight range of 50-60 kgs and the height range of 176-180 cms will be paired with another 30-year-old male that falls into the same weight and height categories. 

Once all 100 pairs are made, a subject from each pair will be randomly assigned into the treatment group (will be administered the new diet for 2 months) while the other subject from the pair will be assigned to the control group (will be assigned to follow the standard diet for two months). At the end of the time time period of 2 months, researchers will measure the total weight gain for each subject.

There are a few outstanding benefits and a few expected disadvantages of utilizing a matched-pairs design.

• Controls for hiding factors.

A hiding variable is a variable that isn’t represented in an examination that might influence the results of the investigation.

In the past model, both age and orientation can altogether affect weight reduction. By matching subjects in light of these two factors, we are wiping out the impact that these two factors could have on weight reduction since we’re just looking at the weight reduction between subjects who are indistinguishable in age and orientation.

In this manner, any distinction in weight reduction that we notice can be credited to the eating routine, instead of old enough or orientation.

• Wipes out order impact .

 Order impact alludes to contrasts in results because of the order where trial materials are introduced to subjects. By utilizing a matched pair design, you don’t need to stress over order impact since each subject just gets one treatment.

In our past model, each subject in the examination was just put on one eating regimen. If we made one subject utilize the standard eating regimen for 30 days, then, at that point, the new eating regimen for 30 days, there could be a request impact because of the way that the subject utilized one specific eating routine before the other.

• Diminished demand attributes

]Another benefit of matched pairs is their diminished demand attributes. Because we test all members just a single time, members are more averse to figure the analysis’ objective. This might lessen the gamble that members will change a part of their way of behaving because of information on the examination speculation. Therefore, lessening demand attributes might expand the legitimacy of the research.

Explore Features in details

Need a map of our platform? Browse through all that we have to offer

Disadvantages

• Losing two subjects if one exists .

 On the off chance that one subject chooses to exit the review, you lose two subjects since you never again have a total pair.

• Tedious to find matches.

It may very well be very tedious to observe subjects who match specific factors, especially assuming you utilize at least two factors. For instance, it probably won’t be difficult to come by 50 females to use as matches, yet it very well may be very elusive for 50 female matches in which each pair matches precisely on age.

• Difficult to impeccably match subjects.

Regardless of how diligently analysts attempt, there will generally be some variety inside the subjects in each pair. The best way to match impeccably is to observe indistinguishable twins who share a similar hereditary code, which is really why indistinguishable twins are much of the time utilized in paired match studies.

What are disadvantages of cluster sampling?

A matched pairs design is an experimental design where participants are matched in pairs based on shared characteristics before they are assigned to groups; one participant from the pair is randomly assigned to the treatment group while the other is assigned to the control group.

The matched pairs design is best suited to studies that have small sample sizes where it is harder to obtain balanced groups by using random allocation alone. Additionally, this research design can only be used in studies with two treatment conditions.

 Some advantages of the matched pairs design are:

• Reduced participant variables
• No order effect

Some limitations of the matched pairs design are:

• Losing two subjects if one drops out
• Time-consuming to find matches
• Matches are never perfect

Explore Voxco Survey Software

+ Omnichannel Survey Software 

+ Online Survey Software 

+ CATI Survey Software 

+ IVR Survey Software 

+ Market Research Tool

+ Customer Experience Tool 

+ Product Experience Software 

+ Enterprise Survey Software 

PX Outlook for 2023: Why Patient Survey Software is Going to Be Vital

PATIENT EXPERIENCE PX Outlook for 2024: Why Patient Survey Software is Going to Be Vital Free Download: Enhance your Patient Experience program by using these

Causal Research

The use of Causal Research in making Business Decisions See what question types are possible with a sample survey! Try a Sample Survey Table of

User Persona: Concept, Types and Benefits

User Persona: Concept, Types and Benefits SHARE THE ARTICLE ON Share on facebook Share on twitter Share on linkedin Table of Contents What is a

ROI for CX transformation

How to quantify ROI for CX Transformation? SHARE THE ARTICLE ON Table of Contents Customer experience informs how your customers perceive their interaction and journey

Digital Customer Experience Vs. Customer Experience

Digital Customer Experience Vs. Customer Experience SHARE THE ARTICLE ON Table of Contents What is Digital Customer Experience? Digital customer experience, also known as Digital

Data Lifecycle Management – Why is it Important?

Data Lifecycle Management – Why is it Important? SHARE THE ARTICLE ON Table of Contents Introduction When you realize how much data organizations collect everyday,

We use cookies in our website to give you the best browsing experience and to tailor advertising. By continuing to use our website, you give us consent to the use of cookies. Read More

Paired T-Test (Matched Pair/Repeated Measure)

Comparing two samples/populations/groups/means/values.

Two-sample paired T-test is performed when two observations are made on each observational unit. There are situations where completely randomized trials do not provide better responses towards the research questions. The example in Table 7 provides a few examples of such in which the repeated measures on the same observational unit would produce a better result in support of the research questions.

Table 7. Examples for Paired/Matched Paired/Repeated Measure Experiments

Assume that a researcher is interested to compare the drivability of two similar vehicles from two different manufacturers. She hired 15 test drivers and collected the drivability performance data provided in Table 8.

Table 8. Paired T-Test Data

Two-sample paired T-Test can be applied as the data comes in pairs for this experimental situation. Analysis can be performed manually using the paired T-Test formula provided Equation 6.

In MS Excel, manual analysis using the paired T-Test formula is provided in Table 9.

Table 9. Manual Analysis Results for Paired T-Test Using MS Excel

Analysis using the statistical function in MS excel is provided in Figure 10.

Figure 10. Two Sample Paired T-Test Analysis Results Using MS Excel

Figure 11. Two Sample Paired T-Test Analysis Results Using Minitab

Statistical Interpretation of the Results

We do not reject the null hypothesis because the p -value (0.430) is larger than the level of significance (0.05). [ p -value is the observed probability of the null hypothesis to happen, which is calculated from the sample data using an appropriate method, two-sample paired T-Test in this case]

Contextual Conclusion

Statistically, vehicle 1 and vehicle 2 are same with respect to the drivability rating by the test drivers. [rewrite the accepted hypothesis for an eighth grader without using the statistical jargon such as the p-value, level of significance, etc.]

Test Your Knowledge

Two sample population proportion test.

Thank you for visiting nature.com. You are using a browser version with limited support for CSS. To obtain the best experience, we recommend you use a more up to date browser (or turn off compatibility mode in Internet Explorer). In the meantime, to ensure continued support, we are displaying the site without styles and JavaScript.

• View all journals
• Explore content
• About the journal
• Publish with us
• Sign up for alerts
• Open access
• Published: 23 August 2024

Mapping spatial organization and genetic cell-state regulators to target immune evasion in ovarian cancer

• Christine Yiwen Yeh   ORCID: orcid.org/0000-0003-4217-8555 1 , 2 , 3   na1 ,
• Karmen Aguirre 1 , 4 , 5   na1 ,
• Olivia Laveroni   ORCID: orcid.org/0009-0008-4416-0475 1   na1 ,
• Subin Kim   ORCID: orcid.org/0000-0001-9339-5758 1 ,
• Aihui Wang   ORCID: orcid.org/0000-0002-4497-7441 6 ,
• Brooke Liang   ORCID: orcid.org/0000-0002-8823-2804 6 ,
• Xiaoming Zhang 6 ,
• Lucy M. Han 7 ,
• Raeline Valbuena   ORCID: orcid.org/0000-0001-7877-4151 1 ,
• Michael C. Bassik   ORCID: orcid.org/0000-0001-5185-8427 1 ,
• Young-Min Kim 1 ,
• Sylvia K. Plevritis   ORCID: orcid.org/0000-0003-2796-9180 2 , 8 ,
• Michael P. Snyder   ORCID: orcid.org/0000-0003-0784-7987 1 ,
• Brooke E. Howitt   ORCID: orcid.org/0000-0002-0309-6680 6 &
• Livnat Jerby   ORCID: orcid.org/0000-0002-4037-386X 1 , 5 , 9  

Nature Immunology ( 2024 ) Cite this article

23 Altmetric

Metrics details

• Genetic techniques
• Translational immunology
• Tumour immunology

The drivers of immune evasion are not entirely clear, limiting the success of cancer immunotherapies. Here we applied single-cell spatial and perturbational transcriptomics to delineate immune evasion in high-grade serous tubo-ovarian cancer. To this end, we first mapped the spatial organization of high-grade serous tubo-ovarian cancer by profiling more than 2.5 million cells in situ in 130 tumors from 94 patients. This revealed a malignant cell state that reflects tumor genetics and is predictive of T cell and natural killer cell infiltration levels and response to immune checkpoint blockade. We then performed Perturb-seq screens and identified genetic perturbations—including knockout of PTPN1 and ACTR8 —that trigger this malignant cell state. Finally, we show that these perturbations, as well as a PTPN1/PTPN2 inhibitor, sensitize ovarian cancer cells to T cell and natural killer cell cytotoxicity, as predicted. This study thus identifies ways to study and target immune evasion by linking genetic variation, cell-state regulators and spatial biology.

Similar content being viewed by others

Single-cell analysis reveals the stromal dynamics and tumor-specific characteristics in the microenvironment of ovarian cancer

Immune cell landscapes are associated with high-grade serous ovarian cancer survival

Single-cell tumor-immune microenvironment of BRCA1/2 mutated high-grade serous ovarian cancer

Multicellular dysregulation has an important function in the initiation and progression of a wide range of diseases, including cancer, where tumor development and accompanying immune responses depend on (and shape) the location of different cell-type populations, tissue properties and organization 1 , 2 , 3 . Cellular and animal models have been instrumental in identifying central immune suppressors 4 , 5 , 6 and have resulted in major breakthroughs in cancer treatment. However, many patients with cancer do not respond to current immunotherapies 7 , 8 , 9 , resulting, at least in part, from two central gaps. First, in contrast to the study of cancer genetics, in which genome sequencing of tumors across large and diverse patient populations has provided a foundation to study the genetic basis of cancer and develop targeted therapies, we still lack equivalent maps for tumor tissue organization to study the inherently spatial processes of multicellular dynamics and immune exclusion in patients. Second, identifying the regulators of cell states and reciprocal intercellular interactions poses additional challenges and requires functional interrogation across a larger search space of combinatorial gene–environment perturbations.

In tubo-ovarian high-grade serous carcinoma (HGSC)—the most common and aggressive form of ovarian cancer 10 —these gaps are pronounced. HGSC is often diagnosed at advanced stages, and is prone to chemoresistance, resulting in a 5-year survival rate below 50% 10 . HGSC genetics has been thoroughly characterized 11 , 12 , 13 , 14 , demonstrating nearly ubiquitous TP53 mutations and massive copy number alterations (CNAs), with some tumors also presenting with BRCA1/ BRCA 2 mutations and homologous recombination deficiency. Abundant tumor-infiltrating lymphocytes (TILs) are a robust prognostic marker of positive clinical outcomes in patients with HGSC 15 , 16 . Yet, while single-cell studies have provided important resources and insight into HGSC cell biology 17 , 18 , the molecular and cellular modalities that promote or suppress immune recruitment and infiltration in HGSC are unclear, and existing immunotherapies continue to have limited success in treating HGSC 19 , 20 .

Here, we provide a molecular map of HGSC spatial organization, that, in conjunction with data-driven experimental design and high-content CRISPR screens, enabled us to systematically uncover molecular and cellular regulators of HGSC tumor immunity, as well as genetic and pharmacological perturbations that affect it.

Single-cell spatial transcriptomic mapping of HGSC

To spatially map HGSC in the setting of metastatic disease, we applied in situ imaging with high-plex RNA detection at single-cell resolution to 130 HGSC tumors from a total of 94 patients to generate 202 tissue profiles, yielding a total of 2,598,277 high-quality single-cell spatial transcriptomes (Fig. 1a and Supplementary Table 1a,b ). Tumor tissues were obtained from the adnexa (ovaries/fallopian tube, n  = 84), and/or omentum ( n  = 46), with 36 patient-matched pairs of adnexal and omental tumors. All tumor tissue profiles were obtained from debulking surgeries in either the treatment-naive or the neoadjuvant chemotherapy-treated setting (Fig. 1b and Supplementary Table 2a,b ). Associated patient clinical data including treatment history (for example, PARP inhibitor, bevacizumab and immune checkpoint blockade (ICB)) and survival outcomes are also available (Fig. 1a,b , Supplementary Table 2a,b and Methods ). Eight patients in this cohort received ICB treatment, but in all cases ICB treatment was after sample collection. For 40 patients, we also obtained DNA sequencing data spanning a 648-gene panel (Fig. 1b , Supplementary Table 1b and Methods ), focused on actionable single-nucleotide variations, somatic CNAs, chromosomal rearrangements and tumor mutational burden (TMB), providing a basis to link tissue structure and somatic genetic aberrations.

a , Cohort and data overview. Created with BioRender.com . b , Summary of clinical history, tumor genetics and ST profiles per patient. Each column represents 1 of 94 patients. NACT, neoadjuvant chemotherapy; Beva, bevacizumab; PARPi, PARP inhibitor. c , Uniform manifold approximation and projection (UMAP) embedding of high-confidence spatial single-cell transcriptomes from the different datasets. n denotes number of cells within each cell-type annotation in the Discovery dataset. d , Representative tumor tissue ST images (11 of 202) with cells plotted in situ and colored based on cell-type annotations. e , Co-embedding of spatial single-cell transcriptomes from this study with six publicly available HGSC scRNA-seq datasets 17 , 24 , 25 , 26 , 27 , 28 , 29 . f , Cell-type composition ( y axis) per tissue profile ( x axis) from this study and in six publicly available scRNA-seq HGSC datasets 17 , 24 , 25 , 26 , 27 , 28 , 29 . g , Pairwise colocalization analysis: the number of tissue profiles ( x axis) where each pair of cell types ( y axis) shows significantly (BH FDR < 0.05, hypergeometric test) higher (red), lower (blue) or expected (gray) colocalization frequencies compared to those expected by random. h , log 2 colocalization quotient (CLQ) of T/NK cells with fibroblasts (CLQ T/NK cell→fibroblast , blue) and T/NK cells with malignant cells (CLQ T/NK cell→malignant, green, x axis) in ST tissue profiles from the Discovery dataset ( n  = 87 CLQ pairs, P  = 4.31 × 10 −10 , paired Wilcoxon rank-sum test). Light gray lines connect paired values in each ST tissue core (black dots). In the box plots, the middle line denotes the median, box edges indicate the 25th and 75th percentiles, and whiskers extend to the most extreme points that do not exceed ±1.5 times the interquartile range (IQR); further outliers (minima and maxima) are marked individually as black points beyond the whiskers; **** P  < 1 × 10 −4 , paired Wilcoxon rank-sum test. NS, not significant.

The spatial data were collected using three spatial transcriptomics (ST) platforms, allowing rigorous cross-platform validation of these recently developed technologies (Fig. 1a , Supplementary Fig. 1a–c and Supplementary Table 1b ). All three ST platforms used here provide detection of RNA molecules at subcellular resolution. As the spatial molecular imaging (SMI, also known as CosMx) 21 platform had the largest gene panels, we used SMI to generate most of the data in this study. Data were generated with the SMI platform in two rounds (Discovery and Test; Fig. 1a ). In the first round, we applied SMI to formalin-fixed and paraffin-embedded (FFPE) tissue microarrays (TMAs) to generate a Discovery dataset that spans 960 genes measured across 491,792 cells from 94 tumors. The discovery phase of the study was performed exclusively based on analyses of the Discovery dataset; all the spatiomolecular transcriptional programs identified in this study were derived from this dataset. In the second round, we applied SMI to another TMA from 34 additional (unseen) patients as well 4 whole-tissue sections (from patients included in the Discovery set), spanning 6,175 and 1,000 genes (Supplementary Table 1b and Supplementary Fig. 1a ), respectively, and a total number of 1,233,033 cells. The Test datasets were used only in the testing phase of the study to test if the key spatiomolecular findings generalize to unseen patients and larger fields of view (FOVs) within a tumor. For technical comparison and validation, in situ sequencing (ISS; via Xenium 22 ) was applied to profile 280 genes in an FFPE TMA of 32 tissue cores, and MERFISH 23 (multiplexed error-robust fluorescence in situ hybridization (ISH)) was applied to profile 140 genes in four fresh-frozen tissue sections ( Methods and Supplementary Table 1b ).

For robust data processing, we developed an analytical procedure that mitigates segmentation inaccuracies (Extended Data Fig. 1a–f and Methods ) and results in robust cell-type annotation through recursive clustering of the spatial single-cell gene expression profiles (Extended Data Fig. 2a,b and Methods ). Applying our pipeline to the Discovery dataset identified malignant cells ( n  = 314,191), T cells and natural killer (NK) cells ( n  = 28,676), B cells ( n  = 16,373), monocytes ( n  = 45,549), mast cells ( n  = 606), fibroblasts/stromal cells ( n  = 72,861) and endothelial cells ( n  = 13,536; Fig. 1c,d and Extended Data Fig. 2a,b ). The same procedure resulted in similar annotations of the Validation and Test datasets (Fig. 1c,d ). T/NK cells were then further stratified to NK ( n  = 6,897), CD4 + T ( n  = 6,040), CD8 + T ( n  = 8,439) and regulatory T cells (T reg cells; n  = 1,905) in the Discovery dataset (Extended Data Fig. 2c–g and Methods ), with similar T/NK stratification results obtained in the Test and Validation 1 datasets (Extended Data Fig. 2h,i and Methods ).

Cell-type annotations were validated in five ways. First, de novo cell-type signatures identified based on the annotated cells recapitulate well known cell-type markers ( Methods , Supplementary Table 3a,b and Extended Data Fig. 2a ). Second, cell-type annotations are aligned with matching hematoxylin and eosin (H&E) and immunohistochemical markers (Extended Data Fig. 3a–d ). Third, cell-type annotations were aligned both spatially (Extended Data Fig. 3e ) and compositionally (Extended Data Fig. 3f ) when examining biological and technical replicate-matched tissue samples profiled in the Discovery and Validation 1 datasets. Fourth, we generated a unified HGSC single-cell transcriptomic atlas by integrating the HGSC spatial data with six publicly available single-cell RNA-sequencing (scRNA-seq) datasets 17 , 24 , 25 , 26 , 27 , 28 , 29 (Fig. 1e,f and Supplementary Fig. 2a–c ). The unified co-embedding corroborates the cell-type annotations obtained independently based on the ST data alone (Fig. 1e , Supplementary Fig. 2a,b and Methods ). Fifth, using patient-matched CNA data, we examined for each gene in each cell type whether its expression is significantly associated with its copy number in the tumor (Benjamini–Hochberg false discovery rate (BH FDR) < 0.05, linear mixed-effects model (LMM); Methods ). In malignant cells, gene expression is patient specific and the expression of 42% of the genes matches their CNAs (that is, ‘in cis ’), compared to only 0–2% in the nonmalignant cell types (Supplementary Fig. 4a ). The genes that do not show in cis RNA-to-CNA associations in malignant cells are significantly more copy number stable (median = 2, range = 0–7) compared to those showing the association (median = 3, range = 0–20, P  = 5.04 × 10 −10 , one-sided Wilcoxon rank-sum test). In contrast to CNAs, a relatively small number of genes (0–4%) are associated with clinical and other genetic covariates in malignant and nonmalignant cell types (that is, age at diagnosis, disease stage, pathogenic BRCA1/ BRCA 2 mutational status and TMB; Supplementary Fig. 4a ).

Initial analyses of the data reveal heterogeneous tumor tissue compositions across patients (Fig. 1f ), yet a generalizable spatial organization at the macro level, where malignant cells and fibroblasts form spatially distinct compartments (which we refer to as the malignant and stromal compartments; Fig. 1g and Supplementary Fig. 3a,b ), such that T/NK cells preferentially localized in the stromal rather than the malignant compartment ( P  < 1 × 10 −4 , paired Wilcoxon rank-sum test; Fig. 1h , Supplementary Fig. 3c–g and Methods ). This macro-organization principle was first observed in the Discovery dataset (Fig. 1g,h and Supplementary Fig. 3c ) and subsequently validated in the Validation and Test datasets (Supplementary Fig. 3a,b,d–g ).

Patients with higher T/NK cell abundance had improved clinical outcomes ( P  = 2.03 × 10 −2 , univariate Cox regression, P  = 2.1 × 10 −4 , log-rank test), demonstrating that even a relatively small area within the tumor (Supplementary Fig. 1a ) is predictive of patient outcomes 5 years and even 8 years later.

T cell states reflect T cell tumor infiltration status

Using the Discovery dataset, we set out to map the immune cell-intrinsic and cell-extrinsic factors that mark immune infiltration and exclusion. First, taking an unbiased data-driven approach, we used unsupervised methods to embed and cluster the transcriptomes of cells from each immune cell type without any spatial information. The resulting embedding and clustering shows that immune cells residing in the malignant compartment are transcriptionally distinct from those that reside outside of it (Fig. 2a ). Next, we took a spatially supervised approach and identified for each of the five immune cell subtypes robustly represented in the data a tumor infiltration program (TIP), consisting of genes that are significantly (BH FDR < 0.05, LMM; Methods ) overexpressed or underexpressed in the pertaining immune cell subtype as a function of proximity to malignant cells (Fig. 2b,c , Extended Data Fig. 4a and Supplementary Table 4a ).

a , UMAP embedding of CD8 + T cells (Discovery dataset) derived from gene expression of all genes (top) or only T cell-specific genes (bottom). b , The association ( P value and effect size, LMM) of each gene (row) from the CD8 TIP with immune cell infiltration status, when considering CD8 + T cells and other immune cell types in the Discovery dataset (columns). c , Representative ST images from the Validation 1 dataset depicting the CD8 TIP identified in the Discovery dataset. P values denote per tissue core if the expression of the CD8 TIP is significantly higher in CD8 + T cells with a high (above median) versus low (below median) abundance of malignant cells within a radius of 30 μm (one-sided t -test). d , UMAP embedding of CD8 + T cells (Validation 1 dataset) from gene expression alone. e , Average gene expression ( z score) in fibroblasts (Discovery dataset) of the top 50 desmoplasia-associated genes (columns) in each tissue profile (rows, n  = 87). f , Representative tissue section (HGSC24, adnexa, Discovery dataset, 1 of 100), wherein the desmoplasia-associated genes capture stromal morphology on the per-cell level ( n  = 1,968 fibroblasts, P value = 7.23 × 10 −80 , one-sided Wilcoxon rank-sum test). These results were observed repeatedly across samples, as shown in e . g , Ligand–receptor interactions (lines) consisting of genes from the CD8 TIP and their respective ligand/receptor in the malignant compartment and stromal compartment. The arrows connect each gene to the cell type where it was found to mark the malignant or stromal compartment.

CD8 + T cell TIP (CD8 TIP) demonstrates that effector and exhausted CD8 + T cells more frequently colocalize with malignant cells ( P  = 3.24 × 10 −53 , LMM; Fig. 2b,c ), as also confirmed by annotating CD8 + T cells based on predefined signatures 30 (Fig. 2d and Extended Data Fig. 4b–e ). CD8 TIP upregulated genes include effector cytotoxicity genes as granzymes ( GZMA , GZMB and GZMH ) and perforin ( PRF1 ), chemokines ( CCL3 , CCL4 , CCL4L2 and CCL5 ), interferon gamma (IFN-γ; encoded by IFNG ), interferon signaling genes (for example, IFITM1 , IFNG , JAK1 and STAT1 ) and immune checkpoints ( CTLA4 , HAVCR2 , PDCD1 , TIGIT and LAG3 ), as well as KLRB1 (that is, CD161 ) and CXCR6 , which have been previously reported to suppress 31 or sustain 32 , 33 , 34 the cytotoxic function of exhausted CD8 + T cells, respectively. CD8 TIP downregulated genes include naive and memory T cell markers ( SELL , IL7R and CD44 ), the co-stimulatory gene CD28 , the granzyme encoded by GZMK and the chemokine receptor encoded by CXCR4 (Fig. 2b ). Extending CD8 TIP to the whole-transcriptome level based on scRNA-seq data 17 ( Methods and Supplementary Table 4b,c ) revealed the upregulation of other exhaustion-associated genes 30 , 35 (that is, ENTPD1 , BST2 , CD63 , MIR155HG , MYO7A and NDFIP2 ), and downregulation of additional naive T cell markers (for example, CCR7 , TCF7 , SATB1 and KLF2 ), with MALAT1 , KLF2 , CCR7 , GPR183 and TCF7 being the topmost downregulated genes in the extended CD8 TIP ( P  < 1 × 10 −16 , r s  > 0.23, Spearman correlation).

Testing these findings in the Test datasets, the CD8 TIP identified in the Discovery dataset was validated as an infiltration marker both in unseen patients and in whole-tissue sections ( P  = 4.22 × 10 −3 and P  < 1 × 10 −17 , LMM, Test datasets 1 and 2, respectively; Extended Data Fig. 4f,g ), and exhausted and effector CD8 + T cell subsets were enriched in proximity to malignant cells ( P  = 1.09 × 10 −5 and P  < 1 × 10 −17 , hypergeometric tests, for Test datasets 1 and 2, respectively; Extended Data Fig. 4h ).

CD8 TIP is not associated with age at diagnosis, disease stage (III or IV), pathogenic BRCA1/ BRCA 2 mutations, or TMB, but does show lower expression levels in samples after neoadjuvant treatment ( P  < 4.42 × 10 −3 , LMM, also when controlling for malignant cell abundance).

To investigate the role of the stroma in preferentially colocalizing with naive and memory T cells compared to effector and exhausted T cells, we integrated the ST data with sample-matched H&E stains independently annotated by a gynecologic pathologist (Supplementary Fig. 5a–e ), revealing two fibroblast subsets, one marking normal stroma and the other marking desmoplasia (that is, a neoplasia-associated alteration in fibroblasts and extracellular matrix with distinct tissue morphology 36 , 37 , 38 , 39 , 40 ; Fig. 2e,f , Supplementary Fig. 5a–g and Supplementary Table 5a ). As expected 41 , 42 , desmoplastic fibroblasts not only overexpress collagen fibril organization and extracellular matrix genes ( P  < 1 × 10 −2 , permutation test; Supplementary Fig. 5b and Supplementary Table 5b ), but also upregulate CXCL12 (the cognate ligand of CXCR4 ; Fig. 2e ) and were associated with niches enriched with T/NK cells ( P  < 1 × 10 −4 , LMM).

To link these findings to paracrine signaling, we compiled 2,678 ligand–receptor pairs based on three public resources 43 , 44 , 45 (Supplementary Table 5c and Methods ). Focusing on CD8 + T cells, we identified 24 ligand–receptor pairs that mark the interactions of CD8 + T cells with other cells in the malignant or stromal compartment ( Methods ). The resulting network (Fig. 2g and Supplementary Table 5d ) manifests suppressive ligand–receptor interactions in the malignant compartment (for example, CD80/CD86–CTLA4, CD8 + T cell–monocyte; TIM3–LAGLS9, CD8 + T cell–malignant cell) and CD8 + T cell-mediated chemoattraction of other immune cells via CCL2 and CCL5 . Colocalization of CXCR6–CXCL16 (CD8 + T cell–malignant cell) and CXCR4–CXCL12 (CD8 + T cell–fibroblasts) mark chemoattraction of infiltrating and stroma-residing CD8 + T cells, respectively (Fig. 2g ; BH FDR < 1 × 10 −10 , LMM; Supplementary Fig. 5j–l ). Of note, TCF1 (encoded by TCF7 and downregulated in the extended CD8 TIP) has been shown to directly repress CXCR6 expression 34 (upregulated in the CD8 TIP), suggesting that this central regulator of naive and resting T cells 46 represses CD8 + T cell chemoattraction to CXCL16 expressed by the malignant cells.

A malignant cell state marks and predicts T/NK cell abundance

Mapping the spatial distributions of T/NK cells within the malignant compartment revealed that TILs preferentially colocalize with a subset of malignant cells ( Methods , Fig. 3a–c , Extended Data Fig. 5a and Supplementary Table 6a ). Although malignant cell states are highly patient specific (Supplementary Fig. 4b ) and vary also within patients (Supplementary Fig. 4c–f ), we found that the connection between TIL location and malignant cell gene expression appeared repeatedly across the heterogeneous tumors in our cohort and external cohorts (Fig. 3 and Extended Data Figs. 5 and 6 ).

a , Heat map of genes in the M TIL (malignant transcriptional program that robustly marks the presence of TILs) program (Discovery dataset). Average expression of the top 104 M TIL genes (rows) across spatial frames (columns). b , Top gene sets enriched in M TIL based on Gene Ontology (GO) enrichment analysis. c , M TIL spatial distributions in six representative tumor tissue profiles (6 of 100, Discovery dataset). P values denote if M TIL expression is significantly (one-sided t -test) higher in frames with high-versus-low T/NK abundance, defined based on the median level in the respective tissue section. Matching cumulative analysis is provided in Extended Data Fig. 5e . d , M TIL expression in each malignant cell (Discovery dataset, n  = 297,960 cells), stratified based on the relative abundance of T/NK cells in their surroundings (top) and the presence of T/NK cells at different distances (bottom). e , ROC curves obtained for cross-validated SVM classifier using M TIL expression in malignant cells (Discovery dataset) to predict T/NK cell levels, at the sample, spatial frame and single-cell levels. f , M TIL spatial distributions in a representative region from one (of four) whole-tissue section (HGSC1, adnexa, Test 2 dataset; M TIL expression in TIL-high versus TIL-low niches, P  = 2.87 × 10 −107 , one-sided Wilcoxon rank-sum test). A full view of the whole-tissue section is provided in Extended Data Fig. 6g . g , Mean M TIL expression in malignant cells in each FOV (Test 2 dataset, n  = 878 FOVs), stratified based on the relative abundance of T/NK cells in each FOV. In d and g , in the box plots, the middle line denotes the median, box edges indicate the 25th and 75th percentiles, and whiskers extend to the most extreme points that do not exceed ±1.5 times the IQR; further outliers are marked individually with circles (minima/maxima). P values of group comparisons are derived from a one-sided Student’s t -test.

Formulating these findings, we used the Discovery dataset to identify a malignant transcriptional program that robustly marks the presence of TILs, abbreviated as the malignant TIL (M TIL ) program (Fig. 3a , Extended Data Fig. 5a,b and Supplementary Table 6a ). The program consists of 100 upregulated and 100 downregulated genes whose expression in malignant cells is significantly (BH FDR < 0.05, LMM) positively (M TIL -up) and negatively (M TIL -down) correlated with and predictive of T/NK cell infiltration (Fig. 3d,e ). M TIL overall expression in malignant cells ( Methods ) reflects both inter-sample and intra-sample variation in T/NK cell levels (Fig. 3d, e ), irrespective of anatomical site ( P  < 1 × 10 −30 , LMM; Extended Data Fig. 5c ). M TIL continuously increases as a function of T/NK cell abundance and proximity, also when stratifying the T/NK population into its respective cell subtypes (Extended Data Fig. 5d,e ). M TIL is associated and predictive of T/NK cell levels both in the Validation (Extended Data Fig. 6a,b,f ) and Test datasets, generalizing to unseen patients (Extended Data Fig. 6c,f ) and whole-tissue sections (Fig. 3f,g , Extended Data Fig. 6d,f,g and Supplementary Fig. 6 ). Likewise, an independent scRNA-seq dataset 47 demonstrates that M TIL expression in malignant cells is highest in tumors annotated as ‘infiltrated’, moderate in tumors annotated as ‘excluded’ and lowest in tumors annotated as ‘immune desert’ (Extended Data Fig. 6e ).

Gene-set enrichment analyses demonstrate the connection between M TIL and immune evasion 48 , 49 , 50 , 51 , 52 , 53 . M TIL -up includes chemokines (for example, CCL5 , CXCL10 , CXCL9 and CXCL16 the cognate ligand to CXCR6 ), and oxidative stress genes (for example, GPX3 and SOD2 ; Fig. 3a,b ), and is enriched with multiple immune response genes, including antigen presentation (for example, B2M , CIITA and HLA-A/ HLA- B/ HLA- C ), interferon gamma response genes (for example, IDO1 , IFI27 , IFIH1 , OAS1/ OAS 2/ OAS 3 , JAK1 and STAT1 ) and cell adhesion molecules (for example, ICAM1 , ITGAV and ITGB2 ; BH FDR = 1.91 × 10 −9 , 2.86 × 10 −10 , 4.59 × 10 −2 , respectively, hypergeometric test; Fig. 3b and Supplementary Table 6b ). M TIL -up also includes immune suppression genes, most notable is LGALS9 , encoding galectin 9—the ligand of the immune checkpoint TIM3 (that is, HAVCR2 ), which is upregulated in the infiltrating T/NK cells (Fig. 2g ). However, there is no significant correlation between M TIL and the expression of exhaustion signatures in the nearby T cells ( r s  < 0.046, P  > 0.05, Spearman correlation, Discovery dataset; Supplementary Table 6c ). M TIL -down reflects diverse processes including Wnt signaling (for example, CTNNB1 , FZD3/ FZD 4/ FZD 6 , SMO , FGFR2 and WNT7A ), epigenetic regulation ( DNMT3A and HDAC1/ HDAC 11/ HDAC 4/ HDAC 5 ), insulin signaling (for example, IGFR1 and IGFBP5 ) and cell differentiation (for example, BMP7 , BMPR1A , ETV4 , FGFR1/ FGFR 2 , FYN , S100A4 , SMAD4 and SMO ; BH FDR < 0.05, hypergeometric test; Fig. 3b and Supplementary Table 6b ). Comparing the M TIL program to 13 malignant signatures previously identified in a comprehensive HGSC scRNA-seq study 17 , we found that 156 of the 200 genes in the M TIL program were not included in any of these signatures (Supplementary Table 6d ).

In our cohort, M TIL expression is not associated with patient age at diagnosis, disease stage, neoadjuvant chemotherapy, TMB or sample anatomical site (Supplementary Fig. 7 and Supplementary Table 7a–c ), but it is moderately positively associated with pathogenic BRCA1/ BRCA 2 mutations ( P  = 0.0423, LMM). M TIL expression is associated with improved overall survival, both in our cohort ( n  = 54, P  = 0.011, Cox regression, based on mean expression of malignant cells in adnexal tumors, while controlling for age at diagnosis, disease stage and treatment history; Fig. 4a,b , Supplementary Table 8a–d and Methods ), and in an external HGSC cohort ( n  = 111, P  = 0.024, Cox regression, controlling for patient age and stage; Fig. 4c ).

a , Hazard ratios (HRs) estimated for each predictor from multivariate Cox proportional hazards models of overall survival in the HGSC cohort of this study (Discovery and Test 1 datasets, n  = 30 and 54 patients for genomic and non-genomic features, respectively). Bars indicate 95% confidence intervals ( Methods ). Arrowheads indicate that the 95% interval extends beyond the HR limits shown in the x axis. * P value < 0.05, multivariate Cox proportional hazards models. b , Kaplan–Meier curves and numbers-at-risk table of overall survival in patients with HGSC (Discovery and Test 1 datasets); patients stratified by average M TIL expression (left), and T/NK cell density (right) in adnexal tumors. c , Kaplan–Meier curves and numbers-at-risk tables of ICB PFS probability (melanoma 56 , left; NSCLC 57 , middle) and overall survival (external HGSC cohort, right) with patients stratified by tumor M TIL expression. d , M TIL expression is significantly higher in patients with HER2-negative breast cancer with pCR (pathogenic clinical response) versus without pCR in two arms of the I-SPY2 clinical trial (durvalumab + olaparib ( n  = 71 patients) 58 and pembrolizumab + paclitaxel ( n  = 67 patients) 59 ). In the box plots, the middle line denotes the median, box edges indicate the 25th and 75th percentiles, and whiskers extend to the most extreme points that do not exceed ±1.5 times the IQR; further outliers are marked individually (minima/maxima). P values derived from one-sided Student’s t -test. e , T/NK cell levels estimated from bulk transcriptomics and TMB (mut/kB) as predictors of ICB responses in the datasets shown in c . Kaplan–Meier curves and numbers-at-risk tables of ICB PFS probability in patients with melanoma 56 stratified by T/NK cell levels (1) and TMB (2), and in patients with NSCLC 57 stratified by T/NK cell levels (3), and of overall survival in patients with HGSC stratified by T/NK cell levels (4). In b , c and e , P values were calculated from the Wald statistic of covariate-controlled Cox proportional hazards regression models. The log-rank P value was derived from comparing discretized predictors (high = top quartile versus low = bottom quartile).

Testifying to its generalizability, M TIL expression in malignant cells is predictive of T/NK cells in the vicinity of the malignant cells also in a previously published SMI dataset collected from five patients with non-small cell lung cancer (NSCLC) 21 (area under the receiver operating characteristic (AUROC) curve > 0.71; Extended Data Fig. 6h,i ). Likewise, M TIL expression was significantly correlated with the expression of a T/NK cell signature in external bulk gene expression datasets, both in HGSC ( n  = 578, r s  = 0.72, P  < 2.2 × 10 −16 , Spearman correlation) and other cancer types ( r s  > 0.66, P  < 1 × 10 −9 , Spearman correlation; Supplementary Fig. 8a ). However, although M TIL is strongly supported by bulk gene expression data as a marker of T/NK cell levels, attempting to rediscover the M TIL signature based on the covariation structure of HGSC bulk gene expression data across patients resulted in poor performances (area under the precision recall curve (AUPRC) of 0.202 and 0.306 for the prediction of M TIL -up and M TIL -down genes, respectively; Supplementary Fig. 8b ), underscoring the need for single-cell ST studies.

To expand beyond the 960-gene panel in the Discovery dataset, we identified 200 additional genes that were significantly coexpressed with the M TIL program based on the Test 1 dataset (Supplementary Table 6e and Methods ), and among these are RUNX1 , CCAAT/enhancer-binding protein-encoding genes ( CEBPB and CEBPD , which encode for subunits of the RUNX1 co-activation complex 54 , 55 ), and complement genes ( C1S and C3 ) as a part of the extended M TIL -up module, as well as the stem cell marker encoded by LGR5 , and the BAF complex subunit encoded by SMARCB4 , as a part of the extended M TIL -down module.

M TIL expression predicts clinical response to ICB

We hypothesized that higher M TIL expression may represent more immunogenic malignant cell states and could thus be predictive of clinical responses to ICB in HGSC and potentially other cancer types. As genomic datasets from ICB trials are currently not available in HGSC and given the generalizability of the M TIL program to other cancer types (Extended Data Fig. 6h,i and Supplementary Fig. 8a ), we tested this hypothesis in five external bulk gene expression datasets obtained from tumor samples of other cancer types before ICB treatment.

The M TIL program overall expression scores ( Methods ) were predictive of ICB response in four of the five cohorts that we tested. M TIL scores were predictive of ICB responses in the melanoma cohort 56 ( n  = 152, P  = 1.35 × 10 −3 , progression-free survival (PFS) Cox regression, controlling for patient sex, treatment status, TMB and anatomical site; patient age was not available, Fig. 4c ), NSCLC cohort 57 ( n  = 121, P  = 0.027, PFS Cox regression model, controlling for patient sex, age, smoking status and tumor histological type; Fig. 4c ) and in two independent arms of the I-SPY2 trial in HER2-negative breast cancer 58 , 59 (durvalumab/olaparib, n  = 71, P  = 2.52 × 10 −4 , one-sided t -test, AUROC = 0.65, and pembrolizumab/paclitaxel: n  = 69, P  = 6.37 × 10 −3 , one-sided t -test; AUROC = 0.70; Fig. 4d ). In the urothelial cancer cohort 60 , M TIL program scores and programmed death ligand 1 (PD-L1) levels were not predictive of clinical responses ( n  = 205, P  > 0.05, one-sided t -test). Collectively, M TIL predictive performances were comparable and, in several cases, superior to those of other ICB response biomarkers, including PD-L1 gene expression, TMB levels and estimates of TIL levels (Fig. 4e and Supplementary Fig. 8c ).

Lastly, we note that M TIL is not associated with TMB based on our data ( r s  = −0.077, P  = 0.67, Spearman correlation), as well as the melanoma and urothelial ICB cohorts, where TMB information was available ( r s  < 0.09, P  > 0.32, Spearman correlation). These findings suggest that M TIL is an orthogonal property that is not directly linked to TMB but is predictive of ICB response in several cancer types.

M TIL expression and T/NK cell abundance in the tumor as a function of CNAs

A key question is whether M TIL is merely reflecting the response of malignant cells to the presence of TILs and is thus a surrogate marker of TILs, or whether M TIL is regulated by other cell-intrinsic processes and is driving TIL infiltration and potentially also malignant cell susceptibility to TIL-mediated cytotoxicity, making it a causal predictive biomarker of ICB responses. Given the strong connection we observed between CNAs and malignant cell transcriptomes (Supplementary Fig. 4a ), we turned to examine the connection between M TIL expression and CNAs to probe at this question.

First, we note that M TIL inter-patient variation supersedes its intra-tumoral variation, as observed also after regressing out the impact of the tumor microenvironment compositions, or when considering only malignant cells in TIL-deprived environments (Fig. 5a ; P  < 1 × 10 −30 , analysis of variance (ANOVA) test). Moreover, when considering only malignant cells that are in TIL-deprived niches in the tumor, the M TIL score is still predictive of whether the entire tumor tissue core has high/low TIL levels (AUROC = 0.76, 0.89 and 0.83, when predicting which samples have TIL levels above the median, and 75th and 85th percentiles).

a , M TIL expression in adnexal malignant cells (Discovery and Test 2 datasets, n  = 264,825 cells) residing in tissue niches where T/NK cells were not detected, stratified by tissue profiles labeled by patient and dataset. Dashed brackets indicate adnexal malignant cells from patient-matched tissue profiles from Discovery and Test 2. b , M TIL expression in malignant cells (Discovery dataset), stratified by somatic copy number of six respective genes based on patient-matched bulk tumor genomic profile (LMM, n  = 40 patients). c , Top CNAs showing a significant (BH FDR < 0.05, LMM; Methods ) association with M TIL expression in malignant cells in the Discovery dataset. d , CNA–M TIL –TIL models. e , Deletion (red) of M TIL -up genes and amplification (light blue) of M TIL -down genes that are significantly (BH FDR < 0.05, one side t -test) associated with low T/NK cell levels (estimated based on gene expression of T/NK cell signatures; Methods ) in an independent TCGA HGSC cohort of 578 patients 12 . Exact P values are provided in Supplementary Table 9a . f , ELISA quantification of IFN-γ (1:100) and TNF (undiluted) in NK-92 supernatant treated with various concentrations of recombinant human BMP7. g , M TIL chemokines and the matching chemokine receptors in immune cell TIPs. h , Fold change of NK-92 cell migration of CXCR6 + NK-92 cells derived via CRISPR activation (Supplementary Fig. 9 ) versus control NK-92 cells (transduced with non-targeting control CRISPR activation guides, left) at varying CXCL16 concentrations. In the box plots in a and b , the middle line denotes the median, box edges indicate the 25th and 75th percentiles, and whiskers extend to the most extreme points that do not exceed ±1.5 times the IQR; further outliers are marked individually with circles (minima/maxima). In f and h , error bars represent the mean ± s.d. for f and mean ± s.e.m. for h ; comparisons are indicated via brackets; * P  < 0.05, ** P  < 0.01, *** P  < 0.001, **** P  < 0.0001 (ordinary one-way ANOVA); brackets that are not shown denote nonsignificant ( P  > 0.05) comparisons. Data from n  = 3 biological replicates were collected per condition in f and n  = 4 technical replicates were collected per condition in h .

Source data

Second, in addition to RNA–CNA associations in cis (Supplementary Fig. 4a ), M TIL expression strongly correlated with the copy number of multiple genes in our cohort ( Methods ). Among the positively correlated ones are interferon receptors IFNGR2 , IFNAR1 and IFNAR2 , as well as interferon regulatory factor 1 ( IRF1 ) and RUNX1 , and the top negatively correlated ones being TCF7L2 , FGFR2 and AXL ( P  < 5 × 10 −3 , LMM; Fig. 5b,c ).

Third, CNAs of M TIL genes are predictive of TIL abundance scores in an independent cohort from The Cancer Genome Atlas (TCGA) of 578 HGSC tumors 12 (AUROC = 0.82, on unseen test samples, support vector machine (SVM) model; Methods ), such that tumors with amplification of M TIL -down genes (for example, BMP7 , DNMT3A , FZD3 , MYL9 , SRC and TGFB2 ) or deletion of M TIL -up genes—including both chemokines ( CX3CL1 , CXCL10 , CXCL9 , CXCL16 and CCL5 ) and other genes (for example, ICAM1 , GPX3 and NR3C1 )—have significantly lower TIL abundance scores compared to tumors without these copy number changes (BH FDR < 5.0 × 10 −3 , one-sided t -test; Fig. 5d,e and Supplementary Table 9a ).

Mechanistically, the composite effect of CNAs (or other forms of genetic/epigenetic aberrations) in M TIL genes and regulators can lead to immune evasion through diverse mechanisms. To demonstrate this, we show that BMP7 —one of the topmost repressed genes in the M TIL program (Extended Data Fig. 5a ) that is amplified in TIL-deprived HGSC tumors (Fig. 5e )—suppresses IFN-γ and tumor necrosis factor (TNF) secretion in NK-92 cells (Fig. 5f ). Likewise, M TIL chemokines whose deletion was associated with low TIL levels in the TCGA data ( CXCL9 , CXCL10 , CXCL16 , CCL5 and CX3CL1 ; Fig. 5e ) recruited different subsets of TILs and other immune cells based on ligand–receptor colocalization analyses in the HGSC spatial data (Fig. 5g ). To functionally demonstrate this, we activated CXCR6 in NK-92 cells via CRISPR–dCas9 activation ( Methods and Supplementary Fig. 9a,b ) and showed a dose-dependent and CXCR6-dependent NK cell directional migration toward CXCL16 (Fig. 5h ).

M TIL sensitizes cancer cells to T/NK cell cytotoxicity

Given our findings and previous studies where genes and pathways represented in the M TIL program have been shown to have an important function in antitumor immune responses 61 , 62 , 63 , 64 , we hypothesized that M TIL reflects not only the response of malignant cells to TILs, but also an intrinsically regulated malignant cell state that impacts TIL-mediated tumor control. Supporting this hypothesis, a collection of previously published CRISPR screens 48 , 49 , 50 , 51 , 52 shows that M TIL -up is enriched with genes that sensitize cancer cells to immune-mediated selection pressures (including ICAM1 , JAK1 , NLRC5 , SOD2 and STAT1 ; P  = 1.82 × 10 −4 , hypergeometric test), while M TIL -down includes genes with desensitizing effects ( BCL2 , FGFR1 , HDAC1 , HDAC5 , ITGB5 and RELA ).

To functionally probe the M TIL program genes and examine their effect on ovarian cancer cell response to lymphocyte cytotoxicity, we performed high-content CRISPR knockout (KO) screens in ovarian cancer cells in monoculture and two types of co-cultures with cytotoxic T lymphocytes (CTLs), including one co-culture with T cell antigen receptor (TCR)-engineered CD8 + T cells and another co-culture with NK cells.

Instead of targeting only genes in the M TIL program itself, we devised a meta-analysis pipeline to identify program regulators based on available Perturb-seq datasets ( Methods ). Using three previously published Perturb-seq datasets 65 , 66 , 67 (Supplementary Fig. 10a ), we identified 43 and 104 perturbations that result in significantly higher and lower expression of the program, respectively (Fig. 6a–d , Supplementary Fig. 10b,c , Supplementary Table 10a and Methods ). Demonstrating the value of this approach, it revealed a wider and more diverse set of regulators, most of which are not included in the M TIL program itself or not included in the gene panel of the Discovery dataset (Supplementary Table 1b ). The positive regulators are enriched (BH FDR < 0.05, hypergeometric test) for genes involved in telomere maintenance (for example, CCT3/CCT4/CCT7 , RTEL1 ), transcriptional regulators (for example, DDX17 , IKZF3 , KLF1 , MTOR , SIRT7 , TP73 and XRCC6 ), protein metabolism (for example, CYC1 , SRP14 and SRP9 ) and cytokine signaling (for example, IRF1 , NUP85 and SEH1L ). The positive M TIL regulators also include the RUNX1 complex genes ( CBFA2T3 , CEBPA and CEBPE ), aligned with our findings that CNAs of RUNX1 are positively associated with the M TIL program (Fig. 5c ) and that RUNX1 is a part of the extended M TIL program. Negative M TIL regulators are enriched for chromatin organization (for example, DNMT1 , INO80 , TAF10 and WDR5 ), Wnt pathway, Myc targets and immune resistance genes 48 , 49 , 50 , 51 , 52 , 68 (BH FDR < 1 × 10 −3 , hypergeometric test). The top negative regulator identified here is PTPN1 , which is supported by both gene activation and inhibition screens (Fig. 6a,c ).

a , b , Differential M TIL expression (two-sided t -test comparing cells with the respective perturbation to cells with control sgRNAs) for M TIL altering perturbations identified in K562 (myelogenous leukemia) ( a ) and RPE1 (human retinal pigment epithelial) ( b ) cell lines Perturb-seq data 65 , 66 . c , d , Representative UMAP embeddings of M TIL altering perturbation: cells were labeled based on the sgRNA detected (top) and based on M TIL expression (bottom) in K562 ( c ) and RPE1 ( d ) cell lines. z denotes −log 10 ( P value), two-sided t -test, comparing M TIL expression in the perturbed versus control cells.

Based on these findings, we designed a pooled knockout screen of 74 M TIL genes and regulators (Supplementary Table 10a,b ) to test their function in ovarian cancer cells (TYK-nu cell line; Fig. 7a and Extended Data Figs. 7 and 8 ). Mapping fitness upon genetic perturbations under both adaptive and innate immune selection pressures (Fig. 7a,b ; BH FDR < 0.05, MAGeCK; Methods ) along with Perturb-seq scRNA-seq readouts in monoculture and co-culture with NK cells (Fig. 7a,c ), we identified perturbations that activate or repress the program and tracked subsequent effects of these perturbations on immune escape. In total, we profiled 18,585 high-quality single-cell transcriptomes, each assigned to an ovarian cancer cell with a single guide RNA (sgRNA) confidently identified, and a median of 4,251 genes detected per cell (Fig. 7c and Extended Data Fig. 9a ). Differentially expressed genes were identified for each gene knockout across the three conditions (Fisher’s method; Methods ), resulting in 74 gene ‘perturbation signatures’ ( Methods and Extended Data Fig. 9b,c ) that were then used to identify gene knockouts that significantly repressed or activated the M TIL program, denoted as ‘activators’ and ‘repressors’, respectively (Fig. 7d and Methods ).

a , Overview of experimental design. Created with BioRender.com . b , Ovarian cancer cell (TYK-nu) differential fitness (MAGeCK 82 ) under CD8 + T cell and NK cell selection pressures. c , UMAP of scRNA-seq profiles from Perturb-seq screen. Each dot corresponds to an ovarian cancer cell (TYK-nu) with 1 of the 232 guides confidently detected, cultured in monoculture or co-culture with NK cells at a 1:1 or a 2.5:1 effector-to-target ratio. d , Differential expression of M TIL genes (Fisher’s combined test; Methods ) when comparing ovarian cancer cells with the respective gene KO to those with NTC sgRNAs. e , Differential expression of M TIL -up genes upon different gene KOs under different conditions, shown for genes identified as M TIL repressors or activators. f – h , Gene KOs alter the cancer cell transcriptional response to NK cells. f , Differential expression of the gene KO signature in control ovarian cancer cells in monoculture versus co-culture (two-sided t -test). g , Gene KO signature expression in control ovarian cancer cells in monoculture versus co-culture; statistical significance per KO shown in f . h , UMAPs as in c with cells colored according to differential gene KO signature expression.

Validating our hypothesis and approach, the top perturbations activating the program— PTPN1 and ACTR8 KO—sensitize malignant cells to T/NK cell cytotoxicity (Fig. 7b,d,e and Extended Data Fig. 9b ), while the top perturbations that repress the program, IFNGR1 , IRF1 and STAT1 KOs, confer resistance to T cell-mediated killing (Fig. 7b,d,e and Extended Data Fig. 9b ). This further supports the causal link between CNAs of interferon signaling genes and M TIL expression (Fig. 5c ). Knockout of M TIL repressors ACTR8 , PTPN1 , FGFR1 , MAPK1 and MED12 was found to sensitize cancer cells to immune elimination also in previous in vivo CRISPR screens 48 , 49 , 50 , 51 , 52 . Demonstrating that the transcriptional response to TILs can be genetically rewired, our data also show that knockout of M TIL repressors ( ACTR8 , DNMT1 , FGFR1 , PTPN1 , MED12 and MIF ) mimics and amplifies the transcriptional responses to NK cells, while knockout of M TIL activators, as STAT1 , IFNGR1 , INTS2 , IRF1 , PARP12 and others, represses and counteracts the transcriptional response to NK cells (Fig. 7f–h and Extended Data Fig. 9c ).

To examine if the inhibition of M TIL repressors substantially impacts immune-based cancer cell elimination, we generated syngeneic PTPN1 and ACTR8 KO TYK-nu ovarian cancer lines (Extended Data Fig. 10a,b ). Both knockouts significantly sensitized the cancer cells to both NK cell-mediated cell death ( P  = 6.29 × 10 −14 and P  = 8.79 × 10 −15 for PTPN1 KO and ACTR8 KO, respectively; time-controlled LMM in comparison to non-targeting control (NTC)) and T cell-mediated cell death ( P  = 6.53 × 10 −34 and P  = 6.58 × 10 −25 for PTPN1 KO and ACTR8 KO, respectively; time-controlled LMM in comparison to NTC), as quantified over 16-h monitoring of caspase-3/caspase-7 activity in monoculture versus co-culture with NK-92 cells (Fig. 8a ) and co-culture with TCR-engineered CD8 + T cells (Fig. 8b ). Although ACTR8 is considered an essential gene (Supplementary Table 10c ), its knockout did not impact ovarian cancer cell viability (Fig. 8a and Extended Data Fig. 10b ). Focusing on NK cell-mediated killing, we tested the PTPN1/ PTPN 2 inhibitor ABBV-CLS-484 (refs. 50 , 69 ) in both the TYK-nu and OVCAR3 cell lines. The drug had minimal effect on cell viability in monoculture (Fig. 8c ) but led to a substantial increase in ovarian cancer cell killing by NK cells, as observed for both ovarian cancer cell lines across a range of doses ( P  = 5.10 × 10 −17 and P  = 1.61 × 10 −30 , for TYK-nu and OVCAR3, respectively; dose-controlled LMM; Fig. 8c ).

a , Fluorescent caspase-3/caspase-7 activity monitored in NTC and PTPN1 , ACTR8 and B2M KO syngeneic TYK-nu cell lines in monoculture and co-culture with NK-92 cells over 16 h. b , Fluorescent caspase-3/caspase-7 activity monitored in NTC, PTPN1 and ACTR8 KO syngeneic TYK-nu cell lines in monoculture and co-culture with TCR-specific CD8 + T cells over 16 h. In a and b , P values were derived from Satterthwaite’s ANOVA in time-controlled two-sided LMMs; n  = 3 technical replicates per experimental condition. c , PTPN1 / PTP N2 inhibitor (ABBV-CLS-484) increased NK-mediated cytotoxicity in TYK-nu (left) and OVCAR3 (right) ovarian cancer cell lines in a dose-dependent manner. P values were derived from Satterthwaite’s ANOVA in dose-controlled two-sided LMMs; n  = 3 technical replicates per experimental condition. For a – c , all data shown represent the mean + s.e.m. GCU, global counting unit.

Here we provide a comprehensive spatial mapping of HGSC tumors, revealing generalizable principles of tissue organization and lymphocyte infiltration within these aggressive and genetically unstable tumors. Our study demonstrates the connection between somatic genetic aberrations, malignant transcriptional dysregulation and immune evasion at the cellular and tissue levels, providing a new perspective to the barriers preventing the antitumor immune response in patients with HGSC and new leads to de-repress HGSC cancer immunogenicity.

Our study puts forward new frameworks to delineate complex multicellular processes and phenotypes through the lens of spatial organization. These include linking cell states to genetic variation across individuals and using existing Perturb-seq datasets to identify latent regulators of spatial cell states. We show that our data-driven approach provides a framework to uncover cell-state regulators, even when the transcriptional level of the regulator is not linked to the cell state of interest in the unperturbed state (for example, PTPN1 ). As available Perturb-seq datasets are still limited in their scope and diversity ( Supplementary Discussion ), it is likely that we are still not fully scanning the search space of cell-state regulators. In the case of the M TIL program, additional M TIL regulators beyond those identified here probably exist, as further suggested by our CNA analyses. As more Perturb-seq datasets, as the one generated here, become available across a more diverse range of cell types and conditions, it will be possible to use Perturb-seq data more effectively to extrapolate from one context to another with increasing accuracy 70 , 71 and, in the case of malignant cell states, using CNA-to-RNA and CNA-to-cell-state associations to further guide Perturb-seq experimental design.

The key findings from our study provide new leads and resources to study HGSC immune evasion toward new diagnostic and intervention strategies.

ICB and other immunotherapies have shown modest effects in tumors with low TIL levels at baseline 3 , 15 . Our findings demonstrate that this may be not only due to immune exclusion per se, but also due to malignant cell-intrinsic differences between TIL-rich and TIL-deprived tumors that protect malignant cells even in the presence of targeting CTLs. Supporting this model, we show that CTLs have a substantial effect on the cancer cell transcriptome (Fig. 7c ), such that perturbing malignant cells to prevent or enhance this malignant cell transcriptional response significantly impacts cancer cell susceptibility to CTL cytotoxicity, as we show in highly controlled co-cultures where spatial segregation is unlikely to have a major effect.

We show that stratifying patients based on such malignant cell-intrinsic features—whether through gene expression or CNAs—can help determine patient response to ICB and which aspects of the immune response are genetically dysregulated. Instead of a single immune evasion driver, we propose that immune evasion in HGSC is a result of the composite effects of multiple gene deletions and amplifications that dysregulate both well-established mechanisms (as interferon signaling and chemokine mediated recruitment) as well as specific genes and processes proposed by our analyses and data (for example, BMP7 and RUNX1 ). As ST data provide a single snapshot in time, we also note that M TIL -high areas that are deprived of TILs may mark situations where the snapshot is not representative of TIL location in the past or the probability that these areas will be infiltrated in the future. This may explain the improved ability of M TIL to predict clinical response to ICB compared to TIL levels in certain cohorts.

We anticipate that the detailed mapping of HGSC tumors provided here will help inform the design of new interventions including T/NK cell engineering strategies to enhance T/NK cell infiltration. Our findings demonstrate that the stroma differentially retains or sequesters certain subsets of T/NK cells, but not others, providing new leads to activate or inhibit chemokine receptors as CXCR6 and CXCR4 to mobilize endogenous or engineered T/NK cells into the tumor. Our findings also underscore the need to map T cell clonality as a function of location at the micro level to examine whether T cells that reside in the malignant and stromal compartments are part of the same or different TCR clones, and dynamically track tumor-reactive T/NK cells to examine if these can egress back to the stroma to avoid or reverse exhaustion 72 .

Our data provide new leads to target HGSC resistance, including the inhibition of ACTR8 and PTPN1 . PTPN1 ’s protein product PTP1b is inactivated by oxidation 73 , which may explain M TIL activation under oxidative stress (as indicated by the upregulation of GPX3 and SOD2 ). PTP1b posttranscriptional regulation may also explain why it could not have been identified as an M TIL regulator with a standard approach of gene expression correlations. PTP1b is a negative regulator of insulin and leptin signaling 74 that has been an attractive drug target for treatment of type 2 diabetes and obesity 75 , 76 , 77 , 78 . PTPN1 KO mice have been shown to be viable and resistant to the development of obesity and diabetes, with more recent work demonstrating that PTPN1/PTPN2 inhibition is enhancing antitumor immune response, primarily via activation of CD8 + T cells 50 , 69 , 79 . Here we show that PTPN1 KO in ovarian cancer cells as well as its inhibition via ABBV-CLS-484 selectively sensitized ovarian cancer cells to both T cell-mediated and NK cell-mediated killing, providing rationale to include patients with HGSC in the ongoing phase I clinical trials ( NCT04777994 and NCT04417465 ) 80 , 81 .

Taken together, this integrative study provides a blueprint to functionally map and probe the molecular landscape of multicellular interplay in complex biological tissues and reveals spatial, molecular and genetic aspects of immune escape in HGSC, opening new avenues to activate targeted immune responses.

Human tumor specimen collection

All ST data other than the Validation 2 dataset were obtained from archival clinical FFPE tumor tissues, retrospectively procured from archival storage under an Institutional Review Board (IRB)-approved protocol (number 44615). In patients with both adnexal and omental tumors available for study, tumor blocks from both sites were selected by an expert gynecologic pathologist (B.E.H.) using histopathologic review of the associated H&E slides. HGSC diagnosis was confirmed in all cases. Tumor content as well as tissue quality and preservation were assessed for inclusion in the study. The Validation 2 dataset was obtained from fresh HGSC tumors that were collected at the time of surgery by Stanford Tissue Procurement Shared Resource facility with the appropriate written informed consent and institutional IRB approval (number 11977). Samples were flash frozen and stored at −80 °C until requested for this study. Samples were embedded in optimal cutting temperature (OCT) compound. Sections were generated using a cryostat and stained with H&E, which were reviewed by an expert gynecologic pathologist (B.E.H.) to confirm the diagnosis, quality and tumor content. Summary statistics of tissue sections, tumors and patients profiled are available in Supplementary Table 1a . Annotations at the patient level and tissue level are provided in Fig. 1b and Supplementary Table 2a,b , respectively.

Primary CD8 + T cells were isolated from de-identified blood samples received from the Stanford Blood Center (IRB number 13942).

Tumor tissue DNA sequencing

HGSC tumor sample selection for next-generation sequencing (NGS) was based on the assessment of overall tumor content by a board-certified expert pathologist (B.E.H). Solid tumor tissue was digested by proteinase K. Total nucleic acid was extracted from FFPE tissue sections using Chemagic 360 sample-specific extraction kits (PerkinElmer). Percentage tumor cellularity as a ratio of tumor to normal nuclei was verified against pathologist-derived assessment, with a minimum requirement of 20% tumor content. Macro-dissection was utilized as required to enrich specimens below the 20% threshold. Specimens that met the 20% threshold of tumor to normal nuclei were selected for DNA sequencing. DNA sequencing was subsequently performed via Tempus Labs according to the xT platform protocol 83 . Additional information about NGS data generation and processing is provided in Supplementary Information Section 1.9 .

ST data collection via SMI

The Discovery and Test datasets were generated using the CosMx SMI instrument according to the company’s protocols as described here and in greater detail in Supplementary Information Section 1.5 . In brief, CosMx Universal Cell Characterization RNA 960-gene panel and the CosMx Human 6K Discovery Panel were used (Supplementary Table 1b ), consisting of ISH probes. Each reporter set contains 16 readout rounds with four different fluorophores, creating a 64-bit barcode design with a Hamming distance of 4 (HD4) and a Hamming weight of 4 (HW4) to ensure low error rates. Probe fluorescence was detected at subcellular resolution via the CosMx SMI instrument, and the signal was aggregated to identify the specific RNA molecule measured in each location 21 .

SMI tissue preparation and RNA assay were performed as follows. Five-micron tissue sections were cut from FFPE TMA tissue blocks and adhered onto VWR Superfrost Plus Micro Slides (VWR, 48311-703) or Leica BOND Plus slides (Leica Biosystems, S21.2113.A). After sectioning, the tissue sections were air-dried overnight at room temperature. Tissue preparation was performed as described in the CosMx SMI Manual Slide Preparation Manual (MAN-10184-02). Briefly, the tissues underwent deparaffinization, heat-induced epitope retrieval for 15 min at 100 °C, and enzymatic permeabilization with 3 µg ml −1 digestion buffer for 30 min at 40 °C. Subsequently, a 0.0005% working concentration of fiducials were applied to the tissue, followed by post-fixation and blocking using NHS-acetate. Finally, an overnight hybridization was performed using the CosMx Universal Cell Characterization 960 plex RNA Panel or the CosMx Human 6K Discovery Panel of probes. The next day, the tissues were subjected to stringent washes to eliminate any unbound probes. The tissues were stained with CosMx Nuclear Stain, CosMx Hs CD298/B2M, CosMx Hs PanCK/CD45 and CosMx Hs CD3 nuclear and segmentation markers before loading onto the instrument. The slide and coverslip constitute the flow cell, which was placed within a fluidic manifold on the SMI instrument for morphological imaging and in situ analyte readout. Analysis run on the instrument was set up using the 60 s FOV pre-bleaching profile and segmentation profile for human tissue.

ST data collection via Xenium

The Validation 1 dataset was generated via 10x Genomics’ Xenium platform according to the company’s protocols as described here and in greater detail in Supplementary Information Section 1.6 . In brief, 10x Genomics’ Xenium ISS technology was used with the Xenium Human Breast Panel for multiplexed measurement of 280 genes (Supplementary Table 1b ). Xenium hybridization padlock probes were designed to contain two complementary sequences that hybridize to the target RNA 84 . Probes also contain a third sequence encoding for a gene-specific barcode such that once the paired ends of the probe bind to the target RNA and ligate, a circular DNA probe is generated for rolling circle amplification. Five-micron FFPE TMAs were sectioned onto a Xenium slide, deparaffinated, permeabilized and incubated with a Xenium probe for probe hybridization and barcode amplification, as described in detail in Supplementary Information Section 1.6 . Following washing and background fluorescence quenching 84 , slides were placed into an imaging cassette and loaded on the Xenium Analyzer instrument for morphological imaging and in situ analyte readout.

ST data collection via MERFISH

The Validation 2 dataset was generated via the Vizgen platform according to the company’s protocols as described here and in greater detail in Supplementary Information Section 1.7 . In brief, a custom 140-gene panel was designed with an additional set of 50 blank negative control barcodes based on the MERFISH design that incorporates combinatorial labeling with an error-robust encoding scheme to mitigate detection errors 85 . Four HGSC fresh-frozen tissue samples were preserved in OCT compound and stored at −80 °C before sectioning. Ten-micron tissue sections were cut from the fresh-frozen OCT tissue blocks and adhered onto MERSCOPE slides (Vizgen, 20400001). After sectioning, the tissue sections were fixed with 4% paraformaldehyde in 1× PBS for 15 min, washed three times with 1× PBS, and incubated overnight at 4 °C in 70% ethanol. As described in detail in Supplementary Information Section 1.7 , following tissue sample preparation process samples were loaded onto the MERSCOPE instrument (Vizgen, 10000001) for analyte readout and morphological imaging.

Cell segmentation

Cell segmentation was performed using a deep-learning-based segmentation image processing algorithm, Mesmer 86 , from the DeepCell platform on raw TIFF images. The cell segmentation algorithm was chosen after systematic comparisons with the Omnipose 87 algorithm as described in Supplementary Information Section 1.7 (Extended Data Fig. 1a–f ). The inputs for whole-cell segmentation for SMI images included immunofluorescence (IF) images of DAPI and CD298/B2M for nuclear and cell membrane detection, respectively. MERFISH whole-cell image segmentation was performed with DAPI and cell membrane stains (Vizgen stain boundary kit, 10400009). Nuclear segmentation was performed for ISS images wherein the input includes DAPI IF stain.

ST data quantification and processing

Preprocessed RNA in situ data include RNA transcripts confidently identified for each gene and their spatial coordinates. Given these data, each RNA transcript was aligned to the cell segmentation outputs described above based on its spatial coordinates. Cell count matrices, C , were generated by counting the number of RNA transcripts detected within the segmentation boundaries of each cell j for each gene i to yield \({c}_{i,j}\) for entry of C in each ST dataset. Cell counts were converted to transcripts per million (TPM) according to equation ( 1 ):

where G is the total number of genes in each ST dataset.

Expression levels were quantified as shown by equation ( 2 ):

Cells with fewer than 50, 20 and 5 genes detected in the SMI, Xenium and MERFISH data were excluded, as well as cells with exceptionally large volume (>441 μm 2 ).

Expression of a gene signature or set was computed by considering all the genes in the signature/set, with additional normalizations to filter technical variation, similarly to the procedure reported before 88 with some modifications as described in Supplementary Information Section 2.1 . For gene programs, the expression of the upregulated set minus the expression of the downregulated set was computed as the program expression.

The location of each cell was defined based on the location of its centroid. The r -neighborhood of a cell was defined as all the cells that reside at a distance of at most r μm from the cell. Spatial frames were defined by binning the tissue section FOV to squares with a size of 60 μm × 60 μm (that is, 3,600 μm 2 ), with a median number of 53 cells per frame.

As described in detail in Supplementary Information Section 2.2 , the cell-type annotation procedure was applied separately for each of the three spatial datasets via an initial cell-type assignment followed by an iterative subsampling procedure to obtain robust cell-type assignments with confidence levels. Cell-type signatures used for this purpose were derived from previous HGSC scRNA-seq datasets 26 , 27 , 29 , 89 .

Co-embedding ST and scRNA-seq datasets

A reference single-cell atlas was generated to examine consistency across spatial and scRNA-seq cohorts and validate cell-type annotations. The atlas includes spatial datasets collected here and six scRNA-seq HGSC cohorts 17 , 24 , 25 , 26 , 27 , 28 , 29 . Preprocessed gene expression matrices were downloaded from Synapse ( syn33521743 ) 17 , the Gene Expression Omnibus (GEO; GSE118828 , GSE173682 , GSE147082 and GSE154600 ) 24 , 26 , 25 , 28 and https://lambrechtslab.sites.vib.be/en/data-access/ 27 , 29 . Tumor samples derived from other anatomical sites, other than the adnexa or omentum, were removed to match the scope of this study. All nine datasets were co-embedded with reciprocal principal component analysis using the top 30 PCs fit on each dataset, using the Seurat R package version 5.1.0 implementation 90 , and then visualized with two-dimensional UMAP 91 . A detailed description of the co-embedding pipeline is available in Supplementary Information Section 2.3 .

Mixed-effects modeling

Mixed-effect models were used to capture codependencies and the hierarchical structure of the data, where covariates at different levels (for example, cell, spatial frame and sample) are sampled from different distributions. The lme4 (version 1.1-35.4) 92 and lmerTest (version 3.1–3) R packages 93 were used to fit the models using the standard restricted maximum-likelihood method, identify the latent variables that maximize the posterior probability and compute P values and sum of squares in type II ANOVA via the Satterthwaite degree of freedom method.

Identifying spatial gene expression programs

Immune infiltration programs (Supplementary Table 4a ) were identified by analyzing the Discovery dataset with the LMMs described above using the frame-level abundance of malignant cells as a measure of the infiltration level. To prevent impact of ambient RNA, only genes that had significantly higher expression levels (pairwise one-sided t -test P value < 1 × 10 −3 ) in respective immune cell types were considered, using pairwise t -tests when comparing the respective immune cell type to all other cell types. CD8 TIP was extended based on scRNA-seq data 17 (Supplementary Table 4b ). Analyzing the CD8 + T cells from this scRNA-seq cohort, the top 50 genes that were significantly correlated (BH FDR < 1 × 10 −10 , Spearman correlation) with CD8 TIP expression were identified (Supplementary Table 4b ). The M TIL program (Supplementary Table 6a ) was identified in a similar manner in the Discovery dataset, defining the presence of T/NK cells as a binary covariate at the frame level. P values were corrected for multiple hypotheses testing using the BH test, and topmost genes with FDR < 0.05 were reported.

Mapping ST data to clinical and genetic features

Mixed-effect models were used to compute the association between the expression of each gene in the different cell types and the patient-matched CNA measurements obtained at the bulk tumor level. Of the 626 genes with CNA measurements, 159 were also included in the Discovery dataset (SMI) panel. For each cell type and each of these 159 genes the following model was fit: tpm ~ (1 | patient) + cna + nact + sites, where tpm denotes the expression of the gene in cells from cell type k , cna denotes the copy number of the same gene (CNA in cis ), ‘nact’ denotes treatment status and ‘sites’ denote the anatomical site (Supplementary Fig. 4a ). To derive associations of clinical covariates, treatment status, tumor genomics and anatomical site, a similar model was fit (tpm ~ (1 | patients) + age + stage + nact + sites) on all 960 genes in the Discovery dataset (Supplementary Fig. 4a ). Here, ‘age’ denotes age at diagnosis (≤65 or >65 years), and ‘stage’ denotes disease stage (III or IV). Similarly, to examine the connection between CNAs and M TIL expression, all 626 genes with CNA were tested with an LMM, considering only malignant cells from the samples with genomic profiling, with M TIL expression as the dependent variable.

Spatial ligand–receptor network

A unified list of 2,678 unique ligand–receptor pairs was compiled based on three published ligand–receptor networks 43 , 44 , 45 (Supplementary Table 5c ). A CD8 + T cell-centered network (Supplementary Table 5d ) was defined as follows. Ligand–receptor pairs are included in the malignant compartment of the network if they consist of a gene upregulated in the CD8 TIP and: (1) a gene upregulated in the TIP of another immune cell type, or (2) a gene upregulated in the M TIL program. Likewise, ligand–receptor pairs are included in the stromal compartment of the network if they consist of a gene downregulated in the CD8 TIP and: (1) a gene downregulated in the TIP of another immune cell type, or (2) a gene that marks fibroblasts in T/NK cell high niches.

Survival analysis of the HGSC spatial cohort

Survival modeling and visualization was performed using the ‘survival’ (version 3.7-0) and ‘survminer’ (verion 0.4.9) R package 94 , 95 . The time-to-event ‘overall survival’ variable was constructed with the follow-up time (fu_time1; Supplementary Table 2a,b ) defined as days between diagnosis and last follow-up, and the patient status (dead or alive) at last follow-up (event; Supplementary Table 2a,b ). For all survival analyses of the HGSC spatial cohort, we combined the SMI TMA datasets from both Discovery and Test 1 datasets to sufficiently power the analyses. Multivariate and univariate Cox proportional hazards regression models were fit with clinical and treatment features, tumor genomic features and spatiomolecular features in adnexal tumors, and P values were computed via the Wald statistic test. Kaplan–Meier curves were plotted to visualize the predictability of predefined discretization of M TIL expression and T/NK cell density on patient overall survival, and P values were computed via the log-rank test. These analyses as well as additional confounding analyses are described in detail in Supplementary Information Sections 2.7 and 2.6 , respectively.

Survival and ICB response predictors

Confounder-controlled Cox proportional hazards regression models were used to quantify the prognostic value of a given marker in predicting ICB PFS in the ICB melanoma 56 and NSCLC 57 cohorts and in predicting overall survival in the International Cancer Genome Consortium (ICGC) Australian Ovarian Cancer Study cohort. In the breast and urothelial ICB cohorts, PFS was not available and thus categorical clinical response annotations were used to examine if the marker was associated with and predictive of response based on student’s two-sample t -tests and AUROCs. Further details on survival and ICB response predictions and comparison to other ICB response biomarkers are provided in Supplementary Information Section 2.8 .

CNA analyses of TCGA data

TCGA data of array-based gene expression (EXP-A) and copy number somatic mutations were downloaded from the ICGC. The TIL levels of each sample were computed as the expression of a T cell signature. Amplifications and deletions were defined as a copy number log-transformed value (‘segment_mean’) above or below 0.5 and −0.5, respectively. A one-sided t -test was performed per M TIL gene to examine if samples with deletion in M TIL -up gene loci or amplifications in M TIL -down gene loci have significantly lower TIL levels compared to samples without the respective amplification/deletion. Amplifications and deletions that were found to show a statistically significant association are reported in Supplementary Table 9a . SVM classifiers were generated to predict if a tumor has high (above median) TIL levels based on the CNA levels of all M TIL genes, using the ‘e1071’ (version 1.7–14) R package.

Quantifying IFN-γ and TNF secretion

NK-92 cells were treated with increasing concentrations of recombinant human BMP7 (Neuromics, PR27026). After 24 h, the supernatant was collected, centrifuged at 277 g for 10 min, transferred into a new 1.5 ml tube, and stored at −80 °C. The supernatants collected were diluted at a 1:100 ratio before the IFN-γ ELISA assay. The concentration of IFN-γ and TNF was determined using the ELISA MAX Deluxe Set Human IFN-γ (BioLegend, 430115) and ELISA MAX Deluxe Set Human TNF (BioLegend, 430215) according to the manufacturer’s instructions and analyzed on a Varioskan LUX Multimode Plate Reader (Thermo Fisher). A standard curve was fitted using GraphPad Prism’s nonlinear regression (curve fit) built-in analysis on a log–log axis. The analyte concentration was calculated based on the fitted line according to the manufacturer’s instructions.

Lentivirus production

Lenti-X 293T cells were cultured in cOPTI-MEM (opti-MEM, Gibco, 31985088), 1× GlutaMAX (Gibco, 35050061), 1 mM sodium pyruvate (Corning, 25-000-Cl), 5% FBS (Gibco, A3840302) and 1× non-essential amino acid (Corning, 25-025-CI). At ~90% confluency, cells were incubated with TransIT-Lenti (MirusBio, 6603) transfection mixture at 37 °C with 5% CO 2 . The transfection mixture included cOPTI-MEM supplemented with 14 µg of the respective transfer plasmid, 10 µg psPAX2 (Addgene, plasmid number 12260) and 4.33 µg pMD2.G (Addgene, plasmid number 12259). After 6 h of transfection, the medium was replaced with fresh cOPTI-MEM supplemented with 1× ViralBoost (Alstem Bio, VB100) and incubated for an additional 16 h. The supernatant was harvested 24 h and 48 h after transduction. Harvested viral supernatants were pooled and concentrated with Lenti-X Concentrator (Takara Bio, 631232) by centrifugation at 1,500 g for 45 min. Viral pellets were resuspended in medium at a volume 100 times smaller than the original volume and stored at −80 °C until retrieved for experiments.

CRISPR gene activation in NK-92 cells

Lentivirus for NK-92 transduction was produced as described above with the following vectors: dcas9-vp64-gfp (Addgene, plasmid number 61422), LentiMPH V2 (Addgene, plasmid number 89308) and a guide RNA (gRNA) backbone (Addgene, plasmid number 112925) cloned with non-targeting (5′-GGTCCATGGGTGGAGTTACG-3′) or CXCR6 (5′-GGATCTGAAGGACGGGAGT-3′) protospacer sequences. sgRNA protospacer oligonucleotides were purchased from IDT, annealed, and cloned into gRNA backbone digested with FastDigest BamHI (Thermo Fisher, FD0054). Briefly, NK-92 cells were transduced with dcas9-vp64-gfp at a multiplicity of infection (MOI) < 0.3. The GFP-positive cells were sorted using a Sony Biotechnology SH800S Cell Sorter, transduced with the LentiMPH V2 vector at an MOI < 0.3, and selected with 500 μg ml −1 hygromyin. dCas9-MPH NK-92 cells were transduced with sgRNA at an MOI < 0.3 and selected with 1 μg ml −1 puromycin. Flow cytometry was used to assess CXCR6 protein expression. NK-92 cells transduced with the NTC or CXCR6 sgRNA and were stained with Human CXCR6 antibody (BioLegend, 35600525; 1:200 dilution) as described in the flow cytometry analysis of TYK-nu cells section. Cells were analyzed using a Sony Biotechnology SH800S Cell Sorter. All data were analyzed using FlowJo version 10.10.0.

Transwell migration assay

NK-92 cells were harvested, washed with serum-free RPMI media (Gibco, 72400-047) and resuspended in 1 ml of RPMI containing 1% FBS (Gibco, A3840102). Cell numbers were determined by Countess 3 Automated Cell Counter (Invitrogen, AMQAX2000) and 2 × 10 5 cells in 200 μl were placed in each transwell insert (Corning, 3428). Around 600 μl of RPMI containing 1% FBS with various concentrations of recombinant human CXCL16 (PeproTech, 300-55) and untreated control was added to the bottom of a 24-well plate and incubated at 37 °C with 5% CO 2 . Four hours after incubation, top inserts were removed and the migrated cells at the bottom were collected and counted via flow cytometry (SH800S Cell Sorter) with cell counting beads (Precision Count Beads; BioLegend, 424902) for precise cell counts.

Perturb-seq meta-analyses and data-driven screen design

Publicly available Perturb-seq datasets 65 , 66 were used to identify M TIL regulators. For each dataset, counts were converted to TPM values, and two-sided t -tests were performed to identify differentially expressed genes for each perturbation in each one of the screens, comparing the cells with the perturbation to those with control sgRNAs. M TIL expression was computed, and a two-sided t -test was performed to examine if M TIL expression was significantly higher or lower in the cells with the perturbation compared to the control cells (with control sgRNAs). For perturbations that showed a significant effect on the M TIL expression (BH FDR < 0.05, t -test), hypergeometric tests were used to further confirm that the perturbation significantly represses or activates the M TIL genes, having opposite effects on the M TIL -up and M TIL -down gene subsets.

CRISPR KO library of ovarian cancer cells

Guide sequences were selected from the Human CRISPR Knockout Pooled Library (GeCKO v2) 96 . The pooled sgRNA library was purchased from GenScript in a plasmid format utilizing the pLentiGuide-Puro vector. In total, the library includes 232 sgRNAs targeting 74 genes with three guides per gene and ten non-targeting controls (Supplementary Table 10b ). Lentiviral stocks were obtained as described in ‘Lentivirus production’ by transfecting lentiCas9-Blast and the custom sgRNA lentiviral library into Lenti-X 293T cells (Takara, 632180).

TYK-nu ovarian cancer cell line (JRCB Cell Bank, JCRB0234.0) was used for the CRISPR screens. To obtain stable Cas9 expression in the TYK-nu cell line (TYK-nu Cas9 ), 100,000 wild-type TYK-nu cells were seeded in a 24-well plate (Corning, 3526) and incubated overnight. Cells were transduced with the lentiCas9-Blast lentivirus at an MOI of 0.2 with 8 µg ml −1 of polybrene (MilliporeSigma, TR-1003) and incubated overnight in 37 °C with 5% CO 2 . Transduced TYK-nu cells were then washed with DPBS, and selected over 10 days with 10 µg ml −1 of blasticidin (Invivogen, ant-bl-05). Successful transduction of Cas9 was validated via western blot (Extended Data Fig. 8a,c ) and flow cytometry analyses (Extended Data Fig. 8b,d ). The resulting TYK-nu Cas9 cells were then transduced at an MOI of 0.15 with the sgRNA lentiviral library and selected over 5 days with 0.5 µg ml −1 of puromycin (Invivogen, ant-pr-1).

Knockout efficiency was quantified via pMCB306 plasmid. The plasmid contains puromycin-T2A-EGFP with EF-1 alpha promoter and an EGFP-targeting sgRNA driven by a mU6 promoter. Following cell transduction with the pMCB306 plasmid, loss of GFP fluorescence indicates functional Cas9 activity, as cleavage of GFP by Cas9 results in loss of fluorescence, whereas intact GFP retains fluorescence. TYK-nu Cas9 cells were transduced at an MOI of 0.15 with pMCB306 virus and 8 µg ml −1 of polybrene and incubated overnight at 37 °C with 5% CO 2 . Transduced TYK-nu Cas9 cells were washed with DPBS, and selection was conducted over 5 days with 0.5 µg ml −1 of puromycin (Invivogen, ant-pr-1). Editing efficiency was validated via flow cytometry (Extended Data Fig. 8b ).

Flow cytometry analysis

Flow cytometry analysis was conducted to sort and analyze TYK-nu Cas9,B2M-KO cells and a TYK-nu Cas9 GFP transduced cell line. TYK-nu Cas9,B2M-KO cells were washed in 1× PBS and stained with Alexa Fluor 700 anti-human B2M antibody (BioLegend, 395708; 1:20 dilution) for 20 min. Additional cells were set aside to use as unstained controls and to adjust gating. The cells were washed twice in PBS with 1.5% FBS after staining and were filtered through a 35-µm cell strainer before analysis on an LSR II instrument. TYK-nu Cas9 GFP cells were prepared in the same way without the staining and analyzed on a Sony Biotechnology SH800S Cell Sorter. All plots were generated with FlowJo version 10.8.1.

Cancer CD8 + T cell co-cultures

TYK-nu Cas9, NY-ESO-1+ cells were generated by transducing TYK-nu Cas9 cells to stably express the NY-ESO-1 antigen, as described in detail in Supplementary Information Section 3.1 . Primary human CD8 + T cells were isolated from human whole-blood buffy coat and transduced to express a NY-ESO-1 TCR (1G4) construct (generously provided by the laboratory of K. Wucherpfennig 97 , DFCI), as described in detail Supplementary Information Section 3.2 .

TYK-nu T cell co-cultures were performed in parallel with either NY-ESO-1 TCR + or wild-type CD8 + T cells from the same donor as a control, as follows. TYK-nu Cas9, NY-ESO-1+ cells were seeded into a clear-bottom, black-walled 96-well plate and incubated in 100 µl of T cell medium (Supplementary Information Section 3.2 ) per well overnight. CD8 + T cells were activated with Dynabeads for 72 h. Following magnetic removal of the Dynadeads, T cells were added to the TYK-nu Cas9, NY-ESO-1+ culture at varying effector-to-target (E:T) ratios to a total volume of 100 µl per well. The cells were co-cultured for 24 to 72 h.

TYK-nu cell viability and IFN-γ levels in the co-culture were measured to validate the cytotoxicity of the edited CD8 + T cells. For ELISA readouts, supernatants were collected from the 96-well plate at the end of each co-culture period, spun down at 400 g for 5 min and stored at −20 °C in single-use aliquots for subsequent ELISA assays (Extended Data Fig. 7h ). The supernatants collected were diluted at a ratio of 1:1,000 before the IFN-γ ELISA assay (BioLegend, 430104). For cell viability readouts, each well was washed twice with 200 µl of DPBS to remove the T cells at the end of each co-culture period. PrestoBlue cell viability dye (Thermo Scientific, A13261) was added to each well and incubated for 30 min before fluorescence plate reader reading (Tecan Infinite, M1000).

Cancer NK cell co-cultures

TYK-nu ovarian cancer cells (JRCB Cell Bank, JCRB0234.0) were cultured in EMEM (American Type Culture Collection, 30-2003) with 10% heat-inactivated FBS (Life Technologies, A3840102). NK-92 cells (American Type Culture Collection, CRL-2407) were cultured in RPMI 1640 Medium, GlutaMAX Supplement, HEPES medium (Gibco, 72400-047) with 10% heat-inactivated FBS (Life Technologies, A3840102), 200 U ml −1 recombinant human IL-2 (PeproTech, 200-02), 1 mM non-essential amino acids (Cytiva, SH30238), 1 mM sodium pyruvate (Cytiva, SH3023901) and 1% penicillin–streptomycin 100× solution (Cytiva, SV30010). For co-culture experiments, TYK-nu cells were seeded in a black-walled, clear-bottom 96-well plate (Greiner, 655090) and incubated overnight. NK-92 cells were added in varying E:T ratios, and the cells were incubated for 24 to 72 h (Extended Data Fig. 8e ). NK-92 cell line cytotoxicity was validated using PrestoBlue cell viability dye (Thermo Scientific, A13261) following the manufacturer’s protocol.

To validate the specificity of NK cell cytotoxicity in the co-culture experiments, a TYK-nu Cas9,B2M-KO cell line was generated by transducing TYK-nu Cas9 cells with B2M sgRNA lentivirus at an MOI of 0.15. Successful transduction was validated via flow cytometry and western blot analyses (Extended Data Fig. 8c,d ). All cell lines were routinely tested for mycoplasma using the Promokine PCR Mycoplasma Test Kit I/C (PromoKine, PK-CA91-1024).

CRISPR screen in cancer CD8 + T cell co-culture

TYK-nu Cas9, NY-ESO-1+ cells were transduced with the 232 sgRNA library at an MOI of 0.15 and selected with 0.5 µg ml −1 of puromycin over a period of 5 days. The resulting library cells were seeded in a 75-mm flask and allowed to adhere overnight to maintain >1,000× coverage. NY-ESO-1 TCR + CD8 T cells were added at a 5-to-1 E:T cell ratio. The TYK-nu library cells were grown in: (1) monoculture, (2) co-culture with wild-type CD8 + T cells and (3) co-culture with NY-ESO-1 TCR + CD8 T cells. In all three conditions, TYK-nu cells were incubated for 72 h, in either monoculture or co-culture, before being washed twice in 1× DPBS to remove the CD8 + T cells. TYK-nu library cells were snap frozen and stored at −80 °C before genomic DNA extraction and sgRNA amplification. As a second selection, 2 days after recovery, cells were grown under the same conditions again for 72 h and then allowed to recover again before collection and sequencing. All samples were sequenced on a MiSeq Micro V2 in a single-end run.

CRISPR and Perturb-seq screens in cancer NK co-culture

TYK-nu Cas9 cells were transduced with the sgRNA library at an MOI of 0.15 and selected with 0.5 µg ml −1 puromycin for 5 days. The first screen was performed for sgRNA and Perturb-seq readouts. TYK-nu library cells were seeded in a 75-mm dish (Corning, 353136) and allowed to adhere overnight. NK-92 cells were added at 1:1 and 2.5:1 E:T cell ratios. Perturb-seq readouts 67 , 98 were obtained from TYK-nu library cells grown for 48 h in monoculture and co-culture with NK-92 cells. After the completed growth timeline, TYK-nu library cells were washed twice with 10 ml 1× DPBS to remove the suspended NK-92 cells. Two replicates from each condition were put into a single-cell suspension according to the 10× ‘Single Cell Suspensions from Cultured Cell Lines for Single Cell RNA Sequencing’ protocol (10x Genomics, CG00054 Rev B). The libraries were prepared according to the Chromium Next GEM Single Cell 5′ Reagent Kits v2 (Dual Index) with Feature Barcode technology for CRISPR Screening protocol (10x Genomics, CG000510 Rev B). Equimolar amounts of indexed libraries were pooled and sequenced on a NextSeq 2000 P3 in a paired-end run. A subset of replicated TYK-nu library cells was allowed to recover for an additional day until confluency before being snap frozen and stored at −80 °C. Genomic DNA of the snap-frozen cells was extracted using the Quick-DNA Midiprep Plus Kit (Zymo Research, D4075). sgRNA amplification was performed following a previously published protocol 52 . Equimolar amounts of indexed libraries were pooled and sequenced on a MiSeq Nano V2 in a single-end run.

A second screen was performed for sgRNA sequencing. TYK-nu library cells were seeded in a six-well dish (Cole-Parmer, 0192770) and were allowed to adhere overnight. NK-92 cells were added at 2.5:1, 5:1 and 7.5:1 E:T ratios for 48 h. Library cells were allowed to recover for 3 days before being snap frozen and prepared for genomic DNA extraction as described above. Each experimental condition was performed in triplicates with >1,000× cells per sgRNA, resulting in 6 and 12 sequencing samples from the first and second screen, respectively.

CRISPR screen and Perturb-seq data analyses

Raw fastq files were processed using the cellranger pipeline (10x Genomics Cell Ranger 7.1.0), and counts were converted to TPM values. For each condition (monoculture, 1:1 co-culture, and 2.5:1 co-culture), data were analyzed to remove nonmalignant cells. Seurat R package was used for k -nearest neighbor clustering, resulting in a distinct NK cluster in the co-culture conditions, with expression of CD3E and NCAM1 . This cluster was removed, and only cancer cells with a detection of a single sgRNA were retained for downstream analyses. For each of the three conditions, differentially expressed genes were identified for each perturbation using a two-sided t -test comparing the cells with the perturbation to those with NTCs. Fisher’s test was used to combine the three P values. Hypergeometric tests were performed to examine if the upregulated or downregulated genes identified for each perturbation were enriched with M TIL -up or M TIL -down genes, or vice versa, and the combined P values (Fisher’s test) were reported as the final summary statistics.

MAGeCK algorithm (version 0.5.9.4) 82 was used to compute differential fitness effects in the cancer cells under the monoculture and co-culture conditions, either with the CD8 + T cells or with the NK cells. In brief, the sgRNA counts of the different samples were first median normalized to adjust for the effect of library sizes and read count distributions. Second, the variance of read counts was estimated by sharing information across the different sgRNAs, allowing to fit a negative binomial model to test whether sgRNA abundance differs significantly between treatments (that is, co-culture) and controls (that is, monoculture or co-culture with nonspecific T cells). Third, sgRNAs were ranked based on P values calculated from the negative binomial model, and an α-robust ranking aggregation algorithm was used to identify positively or negatively selected genes. The pairwise tests across the different screens were combined with Fisher’s statistic P values as the summary statistics.

Single-hit CRISPR knockout validation

PTPN1 and ACTR8 sgRNA sequences were taken from the library used in the CRISPR screen (Supplementary Table 10b ). Lentivirus production and transduction of PTPN1 and ACTR8 knockouts were generated as described in the lentiviral production section. TYK-nu Cas9,NY-ESO-1+ cells were transduced at an MOI < 0.5 and selected with puromycin for 5 days. Knockout efficiency was validated with Sanger sequencing using the Synthego ICE analysis tool (Synthego Performance Analysis, ICE Analysis. 2019. V3.0. Synthego (April 2024); Extended Data Fig. 10a,b ).

Apoptosis assay

TYK-nu cells (TYK-nu PTPN1KO , TYK-nu ACTR8KO , TYK-nu NTC ) were seeded at a density of 80,000 cells per ml in a 24-well dish and incubated overnight. The next day, TYK-nu cells were pretreated with NucView 488 caspase-3 substrate (Biotium, 10403) at a concentration of 1 μM for 30 min. NK-92 cells or NY-ESO-1 TCR + CD8 + T cells were added at a 4:1 or 1:1 E:T ratio, respectively. The plates were then placed into a Sartorious Incucyte S3 and imaged every hour for 16 h. Sixteen images were taken per well for each time point using a ×10 objective. Incucyte analysis software was used to compute integrated intensity of green fluorescence signal (global counting units × µm²/image).

PTPN1/PTPN2 inhibition

Wild-type TYK-nu cells were seeded at a density of 80,000 cells per ml in a 24-well dish and incubated overnight. The next day, NK-92 cells were added at a 2.5:1 E:T ratio and PTPN1/PTPN2 inhibitor ABBV-CLS-484 (MedChemExpress, HY-145923) was added at varying concentrations (4 μM, 8 μM and 16 μM). After a 48-h incubation, NK-92 cells were washed away, and cytotoxicity was measured using the PrestoBlue cell viability dye according to the manufacturer’s instructions. The same procedure was conducted in parallel with TYK-nu monoculture. Similarly, wild-type OVCAR3 cells were seeded at a density of 200,000 cells per ml. NK-92 cells were added at a 1.5-to-1 E:T ratio, and ABBV-CLS-484 was added at varying concentrations (0.5 μM, 1 μM, 2 μM, 4 μM, 8 μM and 16 μM). After a 24-h incubation, NK-92 cells were washed away, and cytotoxicity was measured using the PrestoBlue cell viability dye. The same procedure was conducted in parallel with OVCAR3 monoculture.

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.

Data availability

ST data, targeted genomics, de-identified clinical meta data and single-cell Perturb-seq data have been deposited in Zenodo via https://doi.org/10.5281/zenodo.12613839 (ref. 99 ) and the Single Cell Portal ( SCP2640 , SCP2641 , SCP2650 , SCP2644 , SCP2646 and SCP2707 ). Additional data and code to reproduce the figures are provided in the Zenodo repository. Data deposited in the SCP are also available for interactive web visualization. scRNA-seq studies with HGSC tumor samples were accessed from Synapse ( syn33521743 ), the GEO ( GSE118828 , GSE173682 , GSE147082 , GSE154600 and GSE146026 ) and https://lambrechtslab.sites.vib.be/en/data-access/ . An additional external validation dataset hosted on the European Genome-Phenome Archive ( EGAD00001006973 and EGAD00001006974 ) was made available for this study through a Data Access Agreement with Genentech. Ovarian cancer ‘Sequencing-based Gene Expression’ data and accompanying overall survival and clinical annotations of the Australian Ovarian Cancer Study cohort (OV-AU) were downloaded from the ICGC ( https://docs.icgc-argo.org/docs/data-access/data-download/ ). TCGA data of ‘array-based gene expression’ (EXP-A) and ‘copy number somatic mutations’ were also downloaded from the ICGC ( https://docs.icgc-argo.org/docs/data-access/data-download/ ). Processed gene expression and clinical data of ICB-treated patients with cancer were downloaded from Supplementary Tables in Liu et al. 56 , Zenodo via https://doi.org/10.5281/zenodo.7625516 (ref. 57 ), http://research-pub.gene.com/IMvigor210CoreBiologies/ and the GEO ( GSE173839 and GSE194040 ). Perturb-seq datasets used for the meta-analysis were downloaded from https://gwps.wi.mit.edu/ and the GEO ( GSE133344 ). Source data are provided with this paper.

Code availability

Code to reproduce results and figures presented in this study is provided as a GitHub repository ( https://github.com/Jerby-Lab/HGSC_SpatialPerturbational ).

Galon, J. & Bruni, D. Tumor immunology and tumor evolution: intertwined histories. Immunity 52 , 55–81 (2020).

Article   CAS   PubMed   Google Scholar  

Biswas, S. et al. IgA transcytosis and antigen recognition govern ovarian cancer immunity. Nature 591 , 464–470 (2021).

Article   CAS   PubMed   PubMed Central   Google Scholar  

Chen, D. S. & Mellman, I. Oncology meets immunology: the cancer–immunity cycle. Immunity 39 , 1–10 (2013).

Article   PubMed   Google Scholar  

Leach, D. R., Krummel, M. F. & Allison, J. P. Enhancement of antitumor immunity by CTLA-4 blockade. Science 271 , 1734–1736 (1996).

Chamoto, K., Yaguchi, T., Tajima, M. & Honjo, T. Insights from a 30-year journey: function, regulation and therapeutic modulation of PD1. Nat. Rev. Immunol. https://doi.org/10.1038/s41577-023-00867-9 (2023).

Anderson, A. C., Joller, N. & Kuchroo, V. K. Lag-3, Tim-3, and TIGIT: co-inhibitory receptors with specialized functions in immune regulation. Immunity 44 , 989–1004 (2016).

Hegde, P. S. & Chen, D. S. Top 10 challenges in cancer immunotherapy. Immunity 52 , 17–35 (2020).

Sharma, P., Hu-Lieskovan, S., Wargo, J. A. & Ribas, A. Primary, adaptive, and acquired resistance to cancer immunotherapy. Cell 168 , 707–723 (2017).

Ribas, A. & Wolchok, J. D. Cancer immunotherapy using checkpoint blockade. Science 359 , 1350–1355 (2018).

Torre, L. A. et al. Ovarian cancer statistics, 2018. CA Cancer J. Clin. 68 , 284–296 (2018).

Article   PubMed   PubMed Central   Google Scholar  

Macintyre, G. et al. Copy number signatures and mutational processes in ovarian carcinoma. Nat. Genet. 50 , 1262–1270 (2018).

Bell, D. et al. Integrated genomic analyses of ovarian carcinoma. Nature 474 , 609–615 (2011).

Article   CAS   Google Scholar  

Garsed, D. W. et al. The genomic and immune landscape of long-term survivors of high-grade serous ovarian cancer. Nat. Genet. 54 , 1853–1864 (2022).

Jiménez-Sánchez, A. et al. Unraveling tumor–immune heterogeneity in advanced ovarian cancer uncovers immunogenic effect of chemotherapy. Nat. Genet. 52 , 582–593 (2020).

Joyce, J. A. & Fearon, D. T. T cell exclusion, immune privilege, and the tumor microenvironment. Science 348 , 74–80 (2015).

Zhang, L. et al. Intratumoral T cells, recurrence, and survival in epithelial ovarian cancer. N. Engl. J. Med. 348 , 203–213 (2003).

Vázquez-García, I. et al. Ovarian cancer mutational processes drive site-specific immune evasion. Nature 612 , 778–786 (2022).

Zhang, A. W. et al. Interfaces of malignant and immunologic clonal dynamics in ovarian cancer. Cell 173 , 1755–1769 (2018).

Doo, D. W., Norian, L. A. & Arend, R. C. Checkpoint inhibitors in ovarian cancer: a review of preclinical data. Gynecol. Oncol. Rep. 29 , 48–54 (2019).

Kandalaft, L. E., Dangaj Laniti, D. & Coukos, G. Immunobiology of high-grade serous ovarian cancer: lessons for clinical translation. Nat. Rev. Cancer https://doi.org/10.1038/s41568-022-00503-z (2022).

He, S. et al. High-plex imaging of RNA and proteins at subcellular resolution in fixed tissue by spatial molecular imaging. Nat. Biotechnol. 40 , 1794–1806 (2022).

Ke, R. et al. In situ sequencing for RNA analysis in preserved tissue and cells. Nat. Methods 10 , 857–860 (2013).

Moffitt, J. R. & Zhuang, X. RNA imaging with multiplexed error-robust fluorescence in situ hybridization (MERFISH). Methods Enzymol. 572 , 1–49 (2016).

Shih, A. J. et al. Identification of grade and origin specific cell populations in serous epithelial ovarian cancer by single cell RNA-seq. PLoS ONE https://doi.org/10.1371/journal.pone.0206785 (2018).

Olalekan, S., Xie, B., Back, R., Eckart, H. & Basu, A. Characterizing the tumor microenvironment of metastatic ovarian cancer by single-cell transcriptomics. Cell Rep. 35 , 109165 (2021).

Regner, M. J. et al. A multi-omic single-cell landscape of human gynecologic malignancies. Mol. Cell 81 , 4924–4941 (2021).

Olbrecht, S. et al. High-grade serous tubo-ovarian cancer refined with single-cell RNA sequencing: specific cell subtypes influence survival and determine molecular subtype classification. Genome Med. 13 , 111 (2021).

Geistlinger, L. Multiomic analysis of subtype evolution and heterogeneity in high-grade serous ovarian carcinoma. Cancer Res. https://doi.org/10.1158/0008-5472.CAN-20-0521 (2020).

Qian, J. et al. A pan-cancer blueprint of the heterogeneous tumor microenvironment revealed by single-cell profiling. Cell Res. 30 , 745–762 (2020).

Zheng, L. et al. Pan-cancer single-cell landscape of tumor-infiltrating T cells. Science 374 , abe6474 (2021).

Sade-Feldman, M. et al. Defining T cell states associated with response to checkpoint immunotherapy in melanoma. Cell 175 , 998–1013 (2018).

Wu, V. H. et al. The GPCR–Gαs–PKA signaling axis promotes T cell dysfunction and cancer immunotherapy failure. Nat. Immunol. https://doi.org/10.1038/s41590-023-01529-7 (2023).

Di Pilato, M. et al. CXCR6 positions cytotoxic T cells to receive critical survival signals in the tumor microenvironment. Cell https://doi.org/10.1016/j.cell.2021.07.015 (2021).

Tooley, K. et al. Pan-cancer mapping of single CD8 + T cell profiles reveals a TCF1:CXCR6 axis regulating CD28 co-stimulation and anti-tumor immunity. Cell Rep. Med. https://doi.org/10.1016/j.xcrm.2024.101640 (2024).

Zheng, C. et al. Landscape of infiltrating T cells in liver cancer revealed by single-cell sequencing. Cell 169 , 1342–1356 (2017).

Casey, L. & Singh, N. Ovarian high-grade serous carcinoma: assessing pathology for site of origin, staging and post-neoadjuvant chemotherapy changes. Surg. Pathol. Clin. 12 , 515–528 (2019).

Layton, T. B. et al. Cellular census of human fibrosis defines functionally distinct stromal cell types and states. Nat. Commun. 11 , 2768 (2020).

Plikus, M. V. et al. Fibroblasts: origins, definitions, and functions in health and disease. Cell 184 , 3852–3872 (2021).

Lendahl, U., Muhl, L. & Betsholtz, C. Identification, discrimination and heterogeneity of fibroblasts. Nat. Commun. https://doi.org/10.1038/s41467-022-30633-9 (2022).

Buechler, M. B. et al. Cross-tissue organization of the fibroblast lineage. Nature https://doi.org/10.1038/s41586-021-03549-5 (2021).

Guerrero-Juarez, C. F. et al. Single-cell analysis reveals fibroblast heterogeneity and myeloid-derived adipocyte progenitors in murine skin wounds. Nat. Commun. 10 , 650 (2019).

Foster, D. S. et al. Integrated spatial multiomics reveals fibroblast fate during tissue repair. Proc. Natl Acad. Sci. USA 118 , e2110025118 (2021).

Ramilowski, J. A. et al. A draft network of ligand–receptor-mediated multicellular signalling in human. Nat. Commun. 6 , 7866 (2015).

Efremova, M., Vento-Tormo, M., Teichmann, S. A. & Vento-Tormo, R. CellPhoneDB: inferring cell–cell communication from combined expression of multi-subunit ligand–receptor complexes. Nat. Protoc. 15 , 1484–1506 (2020).

Browaeys, R., Saelens, W. & Saeys, Y. NicheNet: modeling intercellular communication by linking ligands to target genes. Nat. Methods 17 , 159–162 (2020).

Zhao, X., Shan, Q. & Xue, H.-H. TCF1 in T cell immunity: a broadened frontier. Nat. Rev. Immunol. https://doi.org/10.1038/s41577-021-00563-6 (2021).

Hornburg, M. et al. Single-cell dissection of cellular components and interactions shaping the tumor immune phenotypes in ovarian cancer. Cancer Cell 39 , 928–944 (2021).

Lawson, K. A. et al. Functional genomic landscape of cancer-intrinsic evasion of killing by T cells. Nature 586 , 120–126 (2020).

Patel, S. J. et al. Identification of essential genes for cancer immunotherapy. Nature 548 , 537–542 (2017).

Manguso, R. T. et al. In vivo CRISPR screening identifies Ptpn2 as a cancer immunotherapy target. Nature 547 , 413–418 (2017).

Griffin, G. K. et al. Epigenetic silencing by SETDB1 suppresses tumour intrinsic immunogenicity. Nature 595 , 309–314 (2021).

Joung, J. et al. CRISPR activation screen identifies BCL-2 proteins and B3GNT2 as drivers of cancer resistance to T cell-mediated cytotoxicity. Nat. Commun. 13 , 1606 (2022).

Zaretsky, J. M. et al. Mutations associated with acquired resistance to PD-1 blockade in melanoma. N. Engl. J. Med. 375 , 819–829 (2016).

Hu, N., Zou, L., Wang, C. & Song, G. RUNX1T1 function in cell fate. Stem Cell Res. Ther. 13 , 369 (2022).

Rossetti, S. & Sacchi, N. RUNX1: a microRNA hub in normal and malignant hematopoiesis. Int. J. Mol. Sci. 14 , 1566–1588 (2013).

Liu, D. et al. Integrative molecular and clinical modeling of clinical outcomes to PD1 blockade in patients with metastatic melanoma. Nat. Med. 25 , 1916–1927 (2019).

Ravi, A. et al. Genomic and transcriptomic analysis of checkpoint blockade response in advanced non-small cell lung cancer. Nat. Genet. 55 , 807–819 (2023).

Pusztai, L. et al. Durvalumab with olaparib and paclitaxel for high-risk HER2-negative stage II/III breast cancer: results from the adaptively randomized I-SPY2 trial. Cancer Cell 39 , 989–998 (2021).

Wolf, D. M. et al. Redefining breast cancer subtypes to guide treatment prioritization and maximize response: predictive biomarkers across 10 cancer therapies. Cancer Cell 40 , 609–623 (2022).

Mariathasan, S. et al. TGFβ attenuates tumour response to PD-L1 blockade by contributing to exclusion of T cells. Nature 554 , 544–548 (2018).

Jhunjhunwala, S., Hammer, C. & Delamarre, L. Antigen presentation in cancer: insights into tumour immunogenicity and immune evasion. Nat. Rev. Cancer 21 , 298–312 (2021).

Zhang, H. et al. Blocking Wnt/β-catenin signal amplifies anti-PD-1 therapeutic efficacy by inhibiting tumor growth, migration, and promoting immune infiltration in glioblastomas. Mol. Cancer Ther. 20 , 1305–1315 (2021).

Spranger, S., Bao, R. & Gajewski, T. F. Melanoma-intrinsic β-catenin signalling prevents anti-tumour immunity. Nature 523 , 231–235 (2015).

Roulois, D. et al. DNA-demethylating agents target colorectal cancer cells by inducing viral mimicry by endogenous transcripts. Cell 162 , 961–973 (2015).

Norman, T. M. et al. Exploring genetic interaction manifolds constructed from rich single-cell phenotypes. Science 365 , 786–793 (2019).

Replogle, J. M. et al. Mapping information-rich genotype-phenotype landscapes with genome-scale Perturb-seq. Cell 185 , 2559–2575 (2022).

Frangieh, C. J. et al. Multimodal pooled Perturb-CITE-seq screens in patient models define mechanisms of cancer immune evasion. Nat. Genet. 53 , 332–341 (2021).

Sheng, W. et al. LSD1 ablation stimulates anti-tumor immunity and enables checkpoint blockade. Cell 174 , 549–563 (2018).

Baumgartner, C. K. et al. The PTPN2/PTPN1 inhibitor ABBV-CLS-484 unleashes potent anti-tumour immunity. Nature 622 , 850–862 (2023).

Ji, Y., Lotfollahi, M., Wolf, F. A. & Theis, F. J. Machine learning for perturbational single-cell omics. Cell Syst. 12 , 522–537 (2021).

Roohani, Y., Huang, K. & Leskovec, J. Predicting transcriptional outcomes of novel multigene perturbations with GEARS. Nat. Biotechnol. 42 , 927–935 (2024).

Weber, E. W. et al. Transient rest restores functionality in exhausted CAR-T cells through epigenetic remodeling. Science 372 , eaba1786 (2021).

Salmeen, A. et al. Redox regulation of protein tyrosine phosphatase 1B involves a sulphenyl-amide intermediate. Nature 423 , 769–773 (2003).

Myers, M. P. et al. TYK2 and JAK2 are substrates of protein-tyrosine phosphatase 1B. J. Biol. Chem. 276 , 47771–47774 (2001).

Johnson, T. O., Ermolieff, J. & Jirousek, M. R. Protein tyrosine phosphatase 1B inhibitors for diabetes. Nat. Rev. Drug Discov. 1 , 696–709 (2002).

Zhang, Z.-Y. Drugging the undruggable: therapeutic potential of targeting protein tyrosine phosphatases. Acc. Chem. Res. 50 , 122–129 (2017).

Greisman, J. B. et al. Discovery and validation of the binding poses of allosteric fragment hits to protein tyrosine phosphatase 1b: from molecular dynamics simulations to X-ray crystallography. J. Chem. Inf. Model. 63 , 2644–2650 (2023).

Sharma, B. et al. Recent advance on PTP1B inhibitors and their biomedical applications. Eur. J. Med. Chem. 199 , 112376 (2020).

Wiede, F. et al. PTP1B is an intracellular checkpoint that limits T-cell and CAR T-cell antitumor immunity. Cancer Discov. 12 , 752–773 (2022).

Iracheta-Vellve, A. et al. Abstract 606: targeting the immune checkpoint PTPN2 with ABBV-CLS-484 inflames the tumor microenvironment and unleashes potent CD8 + T cell immunity. Cancer Res. 82 , 606 (2022).

Article   Google Scholar  

LoRusso, P. M. et al. Abstract CT257: first-in-human phase 1 studies of PTPN2/1 inhibitors ABBV-CLS-484 and ABBV-CLS-579 in locally advanced or metastatic tumors. Cancer Res. 83 , CT257 (2023).

Li, W. et al. Quality control, modeling, and visualization of CRISPR screens with MAGeCK-VISPR. Genome Biol. 16 , 281 (2015).

Beaubier, N. et al. Clinical validation of the tempus xT next-generation targeted oncology sequencing assay. Oncotarget 10 , 2384–2396 (2019).

Janesick, A. et al. High resolution mapping of the tumor microenvironment using integrated single-cell, spatial and in situ analysis. Nat. Commun. 14 , 8353 (2023).

Chen, K. H., Boettiger, A. N., Moffitt, J. R., Wang, S. & Zhuang, X. Spatially resolved, highly multiplexed RNA profiling in single cells. Science 348 , aaa6090 (2015).

Greenwald, N. F. et al. Whole-cell segmentation of tissue images with human-level performance using large-scale data annotation and deep learning. Nat. Biotechnol. 40 , 555–565 (2022).

Cutler, K. J. et al. Omnipose: a high-precision morphology-independent solution for bacterial cell segmentation. Nat. Methods 19 , 1438–1448 (2022).

Jerby-Arnon, L. et al. A cancer cell program promotes t cell exclusion and resistance to checkpoint blockade. Cell 175 , 984–997 (2018).

Izar, B. et al. A single-cell landscape of high-grade serous ovarian cancer. Nat. Med. 26 , 1271–1279 (2020).

Stuart, T. et al. Comprehensive integration of single-cell data. Cell 177 , 1888–1902 (2019).

McInnes, L., Healy, J., Saul, N. & Großberger, L. UMAP: uniform manifold approximation and projection. J. Open Source Softw. 3 , 861 (2018).

Bates, D., Mächler, M., Bolker, B. & Walker, S. Fitting linear mixed-effects models using lme4. J. Stat. Softw. 67 , 1–48 (2015).

Kuznetsova, A., Brockhoff, P. B. & Christensen, R. H. B. lmerTest package: tests in linear mixed effects models. J. Stat. Softw. https://doi.org/10.18637/jss.v082.i13 (2017).

Therneau, T. M. survival: survival analysis. R package version 3.5-5; https://CRAN.R-project.org/package=survival (2023).

Therneau, T. M. & Grambsch, P. M. Modeling Survival Data: Extending the Cox Model (Springer, 2000).

Sanjana, N. E., Shalem, O. & Zhang, F. Improved vectors and genome-wide libraries for CRISPR screening. Nat. Methods 11 , 783–784 (2014).

Mathewson, N. D. et al. Inhibitory CD161 receptor identified in glioma-infiltrating T cells by single-cell analysis. Cell 184 , 1281–1298 (2021).

Dixit, A. et al. Perturb-seq: dissecting molecular circuits with scalable single-cell RNA profiling of pooled genetic screens. Cell 167 , 1853–1866 (2016).

Yeh, C. Y. et al. Mapping spatial organization and genetic cell state regulators to target immune evasion in ovarian cancer (version 3). Zenodo https://doi.org/10.5281/zenodo.12613839 (2024).

Download references

Acknowledgements

We thank R. Villarin Akana, G. Schmitt, G. Allard and K. Spees for help with setting up assays and models related to this work; L. McGuire for help with artwork; K. Wucherpfennig (Dana-Farber Cancer Institute) for generously providing NY-ESO-1 constructs; Stanford Genomics Service Center (SFGF), N. Neff and the rest of the Stanford Chan Zuckerberg Biohub Center team for ongoing support. L.J. is a Chan Zuckerberg Biohub Investigator and an Allen Distinguished Investigator. L.J. holds a Career Award at the Scientific Interface from the Burroughs Wellcome Fund (BWF) and a Liz Tilberis Early Career Award from the Ovarian Cancer Research Alliance (OCRA). This study was supported by the BWF (1019508.01; to L.J.), National Human Genome Research Institute (NHGRI; U01HG012069; to L.J.), OCRA (889076; to L.J), Under One Umbrella, Stanford Women’s Cancer Center, Stanford Cancer Institute, a National Cancer Institute (NCI)-designated Comprehensive Cancer Center (251217; to B.E.H. and L.J.) and an NCI Center Support Grant (P30CA124435; to B.E.H.), as well as funds from the Departments of Pathology (to B.E.H.) and Genetics (to L.J.) at Stanford University and from the Chan Zuckerberg Biohub (to L.J.). The content is solely the responsibility of the authors and does not necessarily represent the official views of the NCI.

Author information

These authors contributed equally: Christine Yiwen Yeh, Karmen Aguirre, Olivia Laveroni.

Authors and Affiliations

Department of Genetics, Stanford University School of Medicine, Stanford, CA, USA

Christine Yiwen Yeh, Karmen Aguirre, Olivia Laveroni, Subin Kim, Raeline Valbuena, Michael C. Bassik, Young-Min Kim, Michael P. Snyder & Livnat Jerby

Department of Biomedical Data Science, Stanford University School of Medicine, Stanford, CA, USA

Christine Yiwen Yeh & Sylvia K. Plevritis

Department of Medicine, Stanford University School of Medicine, Stanford, CA, USA

Christine Yiwen Yeh

Cancer Biology Program, Stanford University, Stanford, CA, USA

Karmen Aguirre

Stanford Cancer Institute, Stanford University School of Medicine, Stanford, CA, USA

Karmen Aguirre & Livnat Jerby

Department of Pathology, Stanford University School of Medicine, Stanford, CA, USA

Aihui Wang, Brooke Liang, Xiaoming Zhang & Brooke E. Howitt

Department of Pathology, California Pacific Medical Center, San Francisco, CA, USA

Lucy M. Han

Department of Radiology, Stanford University School of Medicine, Stanford, CA, USA

Sylvia K. Plevritis

Chan Zuckerberg Biohub, San Francisco, CA, USA

Livnat Jerby

You can also search for this author in PubMed   Google Scholar

Contributions

C.Y.Y., K.A., O.L., B.E.H. and L.J. designed the study. C.Y.Y., K.A., O.L., B.E.H. and L.J. conducted the research and interpreted the results. K.A., O.L., S.K. and Y.-M.K. performed the experiments with L.J. supervision. A.W., B.L., X.Z., L.M.H. and B.E.H. collected, annotated and provided archival tissue samples for profiling. C.Y.Y. and L.J. performed the computational and statistical analyses. R.V., M.C.B. and S.K.P. provided additional support on data interpretation. M.P.S. supported spatial data collection. C.Y.Y., K.A., O.L., B.E.H. and L.J. wrote the paper. B.E.H. and L.J. obtained funding. L.J. supervised the study. All authors reviewed and approved the paper.

Corresponding authors

Correspondence to Brooke E. Howitt or Livnat Jerby .

Ethics declarations

Competing interests.

M.P.S. is a cofounder and scientific advisor of Personalis, SensOmics, Qbio, January AI, Fodsel, Filtricine, Protos, RTHM, Iollo, Marble Therapeutics and Mirvie. M.P.S. is a scientific advisor of Yuvan, Jupiter, Neuvivo, Swaza and Mitrix. M.C.B has outside interest in DEM Biopharma. The other authors declare no competing interests.

Peer review

Peer review information.

Nature Immunology thanks Denarda Dangaj Laniti, Jason Moffat and the other, anonymous, reviewer(s) for their contribution to the peer review of this work. Primary Handling Editor: Nick Bernard, in collaboration with the Nature Immunology team.

Additional information

Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Extended data

Extended data fig. 1 cell segmentation..

(a) Representative whole-cell segmentation performed for the Discovery dataset. Input data includes DAPI immunofluorescent (IF) and cell membrane stain. Cell boundaries represented as white contours. This is a single representative sample out of 100. Similar results were obtained for all 100 other samples. (b) Representative nuclear segmentation performed for Validation 1 dataset. Input data includes DAPI IF stain. Cell boundaries represented as white contours. This is a single representative sample out of 32 samples. Similar results were obtained for all other samples. (c) Representative comparison of Mesmer vs. Omnipose cell segmentation in a tissue profile (1 of 100) from the Discovery dataset. (d) Representative comparison of Mesmer Nuclear (left) and Mesmer Nuclear-with-Expansion (right) segmentation in a tissue profile (1 of 32) from Validation 1 dataset. (e) P values denoting if cell type confidence scores are significantly higher (one-sided Wilcoxon sum rank test) for whole cells segmented by Mesmer vs. Omnipose for each cell type in the Discovery dataset. (f) P values denoting if cell type confidence scores are significantly higher (one-sided Wilcoxon sum rank test) in cells segmented by Mesmer Nuclear vs. Mesmer Nuclear-with-Expansion for each cell type in Validation 1 dataset.

Extended Data Fig. 2 Cell type annotations of spatial transcriptomics.

(a) UMAP projection (matches (b) and Fig. 1c ) of high confidence ( Supplementary Methods ) spatial single cell transcriptomes (Discovery dataset); cells colored by overall expression of pre-defined cell type signatures (Supplementary Table 3 , a , Methods ). (b) UMAP projection of high confidence (top panel, matches (a) and Fig. 1c ) spatial single cell transcriptomes (Discovery dataset) to yield a reference map. UMAP projection of all cells in the Discovery dataset (bottom panel) onto the high confidence reference map. (c-d) UMAP embedding of single positive (that is, CD4 + or CD8 + ) T cell transcriptomes (Discovery dataset), cells colored by (c) CD8 and CD4 expression, and (d) expression of de novo CD8 (left) and CD4 (right) T cell expression signatures. (e) Projection of double negative T/NK cell transcriptomes (Discovery) onto CD8/CD4 T cell reference map in (b), with cells colored by overall expression of the de novo CD8 (left) and CD4 (right) T cell gene signatures (Supplementary Table 3 , b ). (f) UMAP embedding of CD4 T cell transcriptomes (Discovery dataset), cells colored by CD4 expression (left) and FOXP3 expression (right). (g) UMAP as in (f), with cells (Discovery dataset) colored based on de novo FOXP3 + CD4 T cell gene signature expression (left) and pre-defined regulatory T cell (Treg) signature expression ( Methods ; Supplementary Table 3 , a ). (h-i) UMAP embedding T/NK single cell transcriptomes in Validation 1 dataset, cells colored by (h) final T/NK cell subtype annotations, (i) detection of (from left to right): CD4 , CD8A/B , FOXP3 (regulatory T cell marker), and NCAMI (NK cell marker).

Extended Data Fig. 3 Cross-platform ST data validations.

(a) Hematoxylin & Eosin (H&E) staining (left), Immunofluorescence (middle) of in situ cell type annotations in the Discovery dataset (right) for four representative tissue FOV (4 of 100). (b-c,e-f) cells colored according to cell type legend in (a). (b) H&E staining (left), Immunohistochemistry (IHC) stain for CD163 (middle; monocyte marker) with corresponding in situ cell type annotations (right) in a representative tissue core (1 of 32) in Validation 1 dataset. (c) H&E (left), IHC stain for FOXP3 (middle, Treg marker), and corresponding in situ cell type annotations (right) in a representative tissue (1 of 32) FOV from Validation 1 dataset. (d) H&E stains of HGSC6 omentum tumor tissue (1 of 100) resolving morphology of plasma cells (black arrows) identified based on the Discovery tissue profile shown in panel (a)(iii). (e) H&E (left), in situ cell type annotations from Validation 1 dataset (middle), and Discovery dataset (right) from technical replicate pairs (2 of 39). Each dataset was processed and annotated separately. White box denotes region of tissue profiled by ISS via Xenium in the Validation 1 dataset that corresponds an adjacent region profiled by SMI in the Discovery dataset (same row). (f) Cell type proportion in technical replicates profiled both in the Discovery and Validation 1 datasets. Straight lines correspond to the linear regression fit; grey ribbons correspond to 95% confidence interval; r s denotes the Spearman correlation coefficient.

Extended Data Fig. 4 CD8 T cell states reflect CD8 T cell tumor infiltration levels.

(a) Size and overlap between the tumor infiltration programs (TIPs) identified in the Discovery dataset for the five different immune cell subsets, shown for the up-regulated (left) and down-regulated (right) subsets. (b) Stratification CD8 T cell subsets based on tumor infiltration status (Validation 1 dataset). (c-d) UMAP embedding of CD8 T cells (Discovery dataset) from gene expression only; cells colored by CD8 T cell states 30 (c), overall expression of predefined CD8 T cell signatures 30 . (d). (e) Stratification CD8 T cell subsets based on tumor infiltration status (Discovery dataset). (f) Expression of CD8 TIP in infiltrating vs. non-infiltrating CD8 T cells (Test 1 and Test 2 datasets); p- value derived from one-sided student’s t -test. (g) CD8 TIP expression marks infiltrating CD8 T cells in the Test datasets, shown in situ for a representative whole tissue section (HGSC2, Adnexa, 1 of 4); p- value derived from one-sided student’s t -test. (h) Abundance of malignant cells in a 30um radius of CD8 T cells in Test datasets, stratified by CD8 T cell subset.

Extended Data Fig. 5 M TIL marks T/NK infiltration at micro- and macro-scales.

(a) Statistical significance and effect size showing the association of each gene’s expression in malignant cells with T/NK cell levels, quantified via mixed effect models (two-sided) applied to the Discovery dataset ( Methods ). (b-d) In the Discovery dataset: M TIL expression in malignant cells as a function of (b) discretized T/NK cell levels across tissue profiles ( n  = 99 profiles, top) and spatial frames ( n  = 6699 frames, bottom), (c) T/NK cell levels in spatial frames ( n  = 6699 frames) across anatomical sites, (d) presence of T/NK cell subtypes in spatial frames ( n  = 6699 frames): CD4 T cells (left), CD8 T cells (middle), and NK cells (right); p- values derived from one-sided student’s t -test. (e) Cumulative probability analysis of fraction of T/NK cells in spatial frames stratified by M TIL expression in 6 representative tissue profiles (Discovery dataset) shown in Fig. 3c . In (b-d) Boxplots middle line: median; box edges: 25 th and 75 th percentiles; whiskers: most extreme points that do not exceed ± IQR x 1.5; further outliers are marked individually with circles (minima/maxima).

Extended Data Fig. 6 M TIL is predictive of T/NK infiltration.

(a-d) M TIL expression in malignant cells, stratified by discretized TIL levels in a malignant cell’s niche in (a) Validation 1 dataset, (b) Validation 2 dataset, (c) Test 1 dataset, and (d) Test 2 dataset. (e) M TIL expression in malignant cells, stratified by tissue immune subtyping in Hornburg et al scRNA-seq study 47 . In (a-e) p -value derived from one-sided t -tests. (f) M TIL expression in malignant cells as a predictor of T/NK cell levels. Predictive performances are quantified and visualized via the receiver operating characteristic (ROC) curves shown per ST dataset. Area under the ROC (AUROC) curve is reported in parentheses. (g) In situ M TIL expression marks T/NK cell levels shown in a representative whole tissue section (HGSC1, Adnexa, Test 2 dataset; M TIL expression is higher in TIL-high versus TIL-low niches, p  = 2.87 × 10 −107 , one-sided Wilcoxon rank sum test). A magnified version of region (1) is shown in Fig. 3f , region (2) is magnified in the right image. (h-i) M TIL predicts T/NK cell levels at the microenvironment level in an independent ST SMI data from NSCLC 21 . (h) ROCs depicting prediction performances in NSCLC when predicting the top 10%, 25%, and 50% most T/NK cell rich frames based on the M TIL expression in malignant cells. (i) M TIL expression in NSCLC malignant cells stratified by the level of T/NK cells in their vicinity (‘high’ and ‘low’ depict the top and bottom quartiles, respectively, and ‘moderate’ otherwise). p -value derived from one-sided mixed effect tests. In (a-e, i) boxplots middle line: median; box edges: 25 th and 75 th percentiles; whiskers: most extreme points that do not exceed ± IQR x 1.5; further outliers are marked individually with circles (minima/maxima).

Extended Data Fig. 7 Establishing an ex vivo model of TCR-dependent T cell cytotoxicity.

(a) Top: NY-ESO-1 [1G4] TCR lentiviral construct used to engineer primary human CD8 T cells. Bottom: NY-ESO-1 peptide with 1G4 epitope lentiviral construct used to edit TYK-nu Cas9 cells to express the 1G4 NY-ESO-1 antigen. A non-functional, extracellular domain of human growth factor receptor (NGFR) was used to detect and isolate NY-ESO-1 expressing cancer cells. Created with BioRender.com . (b) Representative flow cytometric analysis gated on the expression of the non-functional NGFR tag to quantify TYK-nu Cas9 cells transduced to express NY-ESO-1 antigen (TYK-nu Cas9,NY-ESO-1+ ). (c) qPCR quantification of CTAG1B mRNA expression in TYK-nu Cas9,NY-ESO-1+ cells relative to A375 melanoma cell line with endogenous CTAG1B expression. (d) Western blot of NY-ESO-1 expression from NY-ESO-1 transduced MDA-MB-231 Cas9, TYK-nu Cas9,NY-ESO-1+ , TYK-nu Cas9 , and A375 whole cell lysates. GAPDH was used as a loading control. Data shown in (d) is one representative experiment repeated three times with similar results. (e) Representative flow cytometric analysis of CD8 + T cells isolated from PBMC. (f) Representative flow cytometric analysis of NY-ESO-1 TCR transduced CD8 T cells. HA (α chain) and PC (β chain) double-positive CD8 + T cells were sorted (left). Cells were re-analyzed immediately after sorting to determine sorting quality (middle). Sorted HA + PC + CD8 + T cells that were frozen and thawed were re-sorted to determine population purity over time (right). (g) TCR-dependent cytotoxicity: NY-ESO-1 TCR expressing primary CD8 T cells were co-cultured with TYK-nu Cas9 cells or TYK-nu Cas9,NY-ESO-1+ cells at variable effector to target cell ratios (E:T). The percentage of dead cancer cells was calculated by normalizing to cancer cell monoculture conditions. (h) ELISA quantification of IFNγ secreted in the co-culture supernatant (1:1000). In (g) and (h): co-cultures were performed using n  = 3 technical replicates per condition and n  = 3 different T cell donors; comparisons are indicated with brackets; p -values **** p  < 1 × 10 −4 (two-way analysis of variance (ANOVA) with multiple comparisons for (g) and (h)); ‘ns’ denote non-significant ( p  > 0.05) comparisons. Exact p- values and raw blot images are provided with the Source Data. Data shown for (g) represent mean ± standard deviation and (h) mean ± s.e.m.

Extended Data Fig. 8 Establishing an in vitro cancer-NK model for CRISPR screens.

(a) Western blot of Cas9 protein from WT and Cas9 transduced whole cell lysates. Alpha tubulin measured as a loading control. (b) Representative flow cytometric analysis gated on GFP expression to measure Cas9 efficiency using pMCB306 plasmid. Loss of GFP denotes Cas9 activity ( Methods ). (c) Western blot of beta-2-microglobulin (B2M) from whole cell lysates of WT, Cas9, and B2M KO TYK-nu. GAPDH measured as a loading control. (d) B2M surface expression by flow cytometry in B2M wt and B2M KO Cas9 TYK-nu cells. (e) 24-to-72-hour time course cell viability in co-cultures of TYK-nu Cas9 and NK-92 cells at variable effector to target cell ratios. Percent killing was calculated by normalizing to monoculture conditions. Co-cultures were performed in 4 replicates per condition as shown. (f) 48-hour cell viability of B2M KO and B2M WT TYK-nu cell lines in co-culture with NK-92 cells. Percent killing was calculated by normalizing to the respective monoculture conditions. Data shown in (a) and (c) are one representative experiment repeated two or more times with similar results. In (e) and (f) , co-culture data is represented by mean ± s.e.m. for (e) and mean ± standard deviation for (f) with each experiment performed in n = 4 technical replicates; p -values **** represent p  < 1 × 10 −4 and * p  < 0.05 (two-way analysis of variance (ANOVA) with multiple comparisons); ‘ns’ shown denote non-significant ( p  > 0.05) comparisons. Exact p- values and raw blot images are provided with the Source Data. All statistical tests were conducted on GraphPad Prism version 10.2.3.

Extended Data Fig. 9 Perturb-seq screen in ovarian cancer identifies immune response regulators.

(a) Number of cells detected with sgRNAs targeting each gene in the CRISPR knockout (KO) library. (b) Gene expression of M TIL -up genes under different gene KOs. (c) Gene KOs mimic (top and second tows) and repress (third and bottom rows) transcriptional response to NK cells: Expression of KO gene signatures ( ACTR8 , MED12 , IRF1 , and STAT1 ) across ovarian cancer cells ( n  = 18,585 cells in each row) stratified based on culture condition and gene KO combination. Boxplots middle line: median; box edges: 25 th and 75 th percentiles; whiskers: most extreme points that do not exceed ± IQR x 1.5; minima and maxima are depicted by extreme ends of whiskers.

Extended Data Fig. 10 Validation of PTPN1 KO and ACTR8 KO in ovarian cancer cells.

(a-b) Discordance of base pairs corresponding to KO target genes (a) PTPN1 and (b) ACTR8 generated from Sanger sequencing using Synthego ICE Analysis tool (v3). Non-targeting control (NTC) depicted in grey. Alignment window for sequences depicted with dashed black bar; interference window for sequences depicted with solid black bar; start of guide sequence is depicted as a grey dotted line.

Supplementary information

Supplementary information.

Supplementary Figs. 1–10, Methods, Discussion, Tables (1–10) List and References.

Reporting Summary

Supplementary tables 1–10.

Supplementary Tables as a single workbook with multiple tabs.

Source Data Fig. 5

Statistical source data for Fig. 5f,h.

Source Data Fig. 8

Statistical source data for Fig. 8a–c.

Source Data Extended Data Fig. 7

Statistical source data for Extended Data Fig. 7c,g,h.

Unprocessed western blot for Extended Data Fig. 7d.

Source Data Extended Data Fig. 8

Statistical source data for Extended Data Fig. 8e,f.

Unprocessed western blots for Extended Data Fig. 8a.

Unprocessed western blots for Extended Data Fig. 8c.

Rights and permissions

Open Access This article is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License, which permits any non-commercial use, sharing, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if you modified the licensed material. You do not have permission under this licence to share adapted material derived from this article or parts of it. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by-nc-nd/4.0/ .

Reprints and permissions

About this article

Cite this article.

Yeh, C.Y., Aguirre, K., Laveroni, O. et al. Mapping spatial organization and genetic cell-state regulators to target immune evasion in ovarian cancer. Nat Immunol (2024). https://doi.org/10.1038/s41590-024-01943-5

Download citation

Received : 04 October 2023

Accepted : 25 July 2024

Published : 23 August 2024

DOI : https://doi.org/10.1038/s41590-024-01943-5

Share this article

Anyone you share the following link with will be able to read this content:

Sorry, a shareable link is not currently available for this article.

Provided by the Springer Nature SharedIt content-sharing initiative

Quick links

• Explore articles by subject
• Guide to authors
• Editorial policies

Sign up for the Nature Briefing: Cancer newsletter — what matters in cancer research, free to your inbox weekly.

Advertisement

Supported by

How A.I. Can Help Start Small Businesses

Entrepreneurs say use of artificial intelligence for a variety of tasks is accelerating the path to hiring and, ideally, profitability.

• Share full article

By Sydney Ember

Sean Ammirati has been teaching a class on entrepreneurship for more than a decade.

A professor at Carnegie Mellon University, Mr. Ammirati has groups of mostly graduate students start businesses from scratch over the course of the spring semester. Some of the start-ups that his 49 students created this year were classic examples of the form: a dating app for couples in long-distance relationships, a personalized fitness app.

But Mr. Ammirati also noticed something unusual.

“I have a pretty good sense how fast the progress that students should make in a semester should be,” he said. “In 14 years, I’ve never seen students make the kind of progress that they made this year.”

And he knew exactly why that was the case. For the first time, Mr. Ammirati had encouraged his students to use generative artificial intelligence as part of their process — “think of generative A.I as your co-founder,” he recalled telling them.

The students began sharing their ideas for use cases on a dedicated Slack channel. They used generative A.I. tools such as ChatGPT, GitHub Copilot and FlowiseAI to help them with tasks including marketing, coding, product development and recruitment of early customers.

By the end of the class in May, venture capitalists were descending on Carnegie Mellon’s campus in Pittsburgh.

“It felt to me like what I felt like in the mid-2000s, when cloud and mobile happened at the same time,” said Mr. Ammirati, who is himself an entrepreneur. Generative A.I., he believed, could similarly change innovation “by an order of magnitude.”

We are having trouble retrieving the article content.

Please enable JavaScript in your browser settings.

Thank you for your patience while we verify access. If you are in Reader mode please exit and  log into  your Times account, or  subscribe  for all of The Times.

Thank you for your patience while we verify access.

Already a subscriber?  Log in .

Want all of The Times?  Subscribe .