How Harvard Medical School's COMPASS AI Predicts Which Cancer Patients Will Respond to Immunotherapy

A major challenge in cancer immunotherapy is identifying which patients will actually benefit from immune checkpoint inhibitors (ICIs). Many patients receive costly treatments with significant side effects but no response. Now, researchers at Harvard Medical School, in collaboration with Roche Pharma, have developed COMPASS — a foundation AI model that predicts ICI response from tumor transcriptomic data with greater accuracy than current clinical biomarkers, while also explaining its reasoning .

How COMPASS predicts ICI response

COMPASS is a machine learning model that analyzes bulk RNA-seq data from a patient's tumor. It works in two stages :

• Pan-cancer pre-training: The model is first trained on 10,184 tumors across 33 cancer types from The Cancer Genome Atlas (TCGA) using self-supervised learning. This teaches it general patterns of tumor-immune biology without needing labeled outcomes .
• Clinical fine-tuning: The pre-trained model is then fine-tuned on 16 independent clinical cohorts covering 7 cancer types (melanoma, non-small cell lung cancer, gastric, bladder, kidney, glioblastoma, and head & neck cancer) and 6 different ICI regimens (including anti-PD-1, anti-PD-L1, anti-CTLA-4, and combinations) .

At prediction time, COMPASS takes a patient's tumor mRNA expression profile (in TPM units) plus their cancer type code and outputs two things :

• 44 concept scores representing immune cell states (e.g., CD8⁺ T cells, exhausted T cells, B cells), tumor-microenvironment interactions, and signaling pathways (e.g., IFN-γ, TGF-β).
• Predicted probabilities of response vs. non-response.

What makes COMPASS different from existing methods

Interpretable concept bottleneck architecture. Most deep learning models are black boxes — they produce a prediction without explanation. COMPASS uses a concept bottleneck design: it first maps raw gene expression onto 44 high-level, human-interpretable immune concepts, and then uses those concepts to make the final prediction . This means a clinician can see why a patient is predicted to respond — for example, high CD8⁺ T cell infiltration and strong IFN-γ pathway activation. Through SHAP analysis on these concept scores, researchers can identify which immune biology drives response or resistance in individual patients .

Pan-cancer generalizability. Many existing ICI prediction models are trained on a single cancer type or small cohort, limiting their usefulness. Because COMPASS is pre-trained on 33 cancer types and fine-tuned across diverse cohorts, it can generalize to new cancer types and ICI drugs not seen during fine-tuning . This is a significant advantage over single-cancer models.

Outperforms standard biomarkers. In a phase II urothelial cancer trial, COMPASS achieved a hazard ratio of 4.7 (p < 0.0001) for survival stratification, substantially outperforming PD-L1 expression and tumor mutational burden (TMB) . Across its evaluations, it outperformed 22 baseline methods, with increases of 8.5% in precision, 12.3% in MCC, and 15.7% in AUPRC over the best competing models .

Potential clinical implications

Better patient selection. COMPASS could help identify patients most likely to benefit from immunotherapy, sparing non-responders from ineffective treatments and unnecessary side effects .

Trial enrichment and design. The model can support clinical trial design by enabling more precise cohort selection, potentially reducing trial sizes and accelerating drug development .

Mechanistic insights into resistance. By revealing which immune pathways drive non-response, COMPASS could inform combination therapy strategies and identify novel resistance mechanisms .

Accessible tools. The COMPASS code, pre-trained models, and a web-based prediction tool are publicly available on GitHub and at immuno-compass.com .

Limitations to note

The COMPASS preprint (posted May 2025 on medRxiv) has not yet completed formal peer review, though it is published in PMC as an NIH-funded preprint . The model requires bulk RNA-seq data as input, which is not yet part of routine clinical workflows for all cancer patients. Independent prospective clinical validation studies have not yet been published, and the validation gap remains a critical challenge for AI models in this space .