Checkpoint Inhibitor Response Signatures in Single Cells

ISEF Category: Biomedical and Health Sciences

Subcategory: Immunology  ·  Difficulty: Advanced  ·  Setup: Home Setup  ·  Time: Full Year

The Hook

Some cancer patients get a huge benefit from checkpoint inhibitors, and others do not. The hard part is knowing who will respond before treatment starts. A single-cell immune atlas can act like a super-detailed map of each cell's gene activity. Your project is to find the pattern that separates the winners from the non-responders.

What Is It?

Single-cell RNA sequencing measures which genes are active in each cell, instead of averaging across a whole tumor. Think of it like reading thousands of tiny report cards instead of one class average. That lets you see which immune cells are present, what state they are in, and how those cells change after therapy.

Checkpoint inhibitors are drugs that take the brakes off the immune system. Some patients' T cells wake up and attack the tumor. Others stay quiet, even when they get the same drug. This project looks for a transcriptomic signature, a gene expression pattern, that can tell those two groups apart across several cancers.

You are not just asking which gene is highest. You are asking which combination of genes, cell states, or immune-cell ratios best tracks response. The strongest version of the project checks whether that pattern still works in a dataset you did not use for training.

Why This Is a Good Topic

This is a strong science fair topic because public datasets already exist, the question is clear, and the answer can be tested with real numbers. You can learn how to clean data, compare groups, build a classifier, and check whether your result holds up in a new cohort. The project also connects to a real medical problem, predicting who benefits from immunotherapy, so your analysis has real-world value.

Research Questions

• How does the immune-cell transcriptomic signature differ between responders and non-responders within melanoma?
• How does the same signature perform when you train on one cancer type and test on NSCLC or RCC?
• What is the effect of restricting the model to T cells, myeloid cells, or both on response prediction?
• To what extent do batch correction choices change the genes selected in the final signature?
• Does a small gene panel perform as well as the full transcriptomic signature on held-out cohorts?
• Which immune cell states contribute most to the score that separates responders from non-responders?

Basic Materials

• Laptop with 16 GB RAM or more.
• Stable internet connection.
• Free cloud storage or external drive with at least 100 GB available.
• Public single-cell RNA-seq datasets from GEO, cellxgene, or the Human Cell Atlas.
• Spreadsheet or notes app for tracking samples, metadata, and analysis decisions.

Advanced Materials

• University computing cluster or workstation with 64 GB RAM or more.
• Access to raw or processed single-cell RNA-seq count matrices.
• Matched clinical response metadata for checkpoint inhibitor treatment.
• Reference immune atlases for cell-type annotation.
• Versioned storage for large intermediate files and model outputs.

Software & Tools

• Python: Cleans data, trains models, and calculates response scores.
• R: Runs differential expression tests and makes summary plots.
• Scanpy: Processes and analyzes single-cell RNA-seq data in Python.
• Seurat: Integrates datasets, clusters cells, and compares cell states.
• Jupyter Notebook: Keeps code, notes, and figures in one place.

Experiment Steps

1. Define the responder label, cancer types, and cohort split before you touch the data.
2. Choose one training cohort and hold out at least one full cohort for final testing.
3. Harmonize cell labels and batch correction so site and platform differences do not drive the result.
4. Build the transcriptomic signature and compare it with simple baselines such as immune cell fraction or single-gene markers.
5. Test the signature on held-out cohorts and inspect which genes and cell states explain the score.

Common Pitfalls

• Mixing patients from the same study across train and test sets, which inflates accuracy.
• Letting batch effects from platform or site differences look like a response signal.
• Using too many rare cell types, which makes the signature unstable across cohorts.
• Skipping class balance checks, which can make a model look good by guessing the majority label.
• Treating one held-out cohort as proof, which hides failure on a different cancer type.

What Makes This Competitive

A strong version of this project does more than find a list of genes. You would test whether the signature survives across cancer types, sequencing platforms, and held-out cohorts. You can also compare a full gene score with a simpler cell-state model and show which one holds up under harder tests. That kind of design shows careful control over overfitting and clearer biological meaning.

Project Variations

• Focus on melanoma first, then test whether the same signature transfers to NSCLC and RCC.
• Build separate signatures from T cells, myeloid cells, and all immune cells to see which group predicts response best.
• Compare a full gene signature with a compact marker panel to test how many genes you really need.

Learn More

• NIH Gene Expression Omnibus: Search for public single-cell RNA-seq studies and treatment metadata.
• cellxgene: Browse annotated single-cell datasets and download immune atlas data.
• Human Cell Atlas: Find reference atlases and cell-type annotations for immune cells.
• PubMed: Search review articles on checkpoint inhibitor response biomarkers and tumor immunity.
• MIT OpenCourseWare: Use free lectures on statistics, machine learning, and data analysis.