Generalizable AI predicts immunotherapy outcomes across cancers and treatments

2026-07-03 · Source: Machine learning : nature.com subject feeds · Field: Science & Research — Artificial Intelligence & Machine Learning, Health & Medical Research, Life Sciences & Biology · Depth: Expert, extended

Summary

COMPASS, a pan-cancer foundation model, accurately predicts immunotherapy response from bulk tumor transcriptomes using a concept bottleneck transformer. Trained on 10,184 tumors across 33 cancer types, COMPASS encodes gene expression into 44 biologically grounded immune concepts. It outperforms 22 existing methods across 16 clinical cohorts spanning seven cancers and six immune checkpoint inhibitors (ICIs), achieving an 8.5% improvement in accuracy and a 15.7% increase in area under the precision-recall curve. The model generalizes to unseen cancer types and treatments, and in survival analyses, patients predicted as responders had significantly longer overall survival (hazard ratio = 4.7, P < 0.0001). Personalized response maps offer mechanistic insights, identifying programs like TGFβ signaling and B cell deficiency in non-responders.

Key takeaway

For AI Scientists and Research Scientists developing precision oncology tools, COMPASS offers a robust, interpretable framework for predicting immunotherapy outcomes. You should integrate concept bottleneck models and multi-stage fine-tuning. This enhances generalizability and mechanistic insight, especially when working with diverse cancer types or limited clinical data. This approach can improve patient stratification and biomarker discovery in future trials.

Key insights

COMPASS uses a concept bottleneck transformer to predict ICI response from transcriptomes, offering interpretable mechanistic insights.

Principles

• Biologically grounded concepts enhance model interpretability.
• Pan-cancer pretraining improves generalizability across therapies.
• Multi-stage fine-tuning adapts models to small cohorts.

Method

COMPASS pretrains on 10,184 TCGA transcriptomes using self-supervised contrastive learning, then fine-tunes on clinical cohorts with parameter-efficient strategies (e.g., partial fine-tuning, linear probing) to predict ICI response.

In practice

• Generate personalized response maps for individual patients.
• Identify specific resistance mechanisms (e.g., TGFβ signaling).
• Stratify patients for clinical trial design.

Topics

• Immunotherapy Response Prediction
• Concept Bottleneck Models
• Pan-Cancer AI
• Transcriptomics
• Tumor Microenvironment
• Precision Oncology

Code references

• mims-harvard/COMPASS-web
• mims-harvard/COMPASS

Best for: AI Scientist, Research Scientist, Machine Learning Engineer

Related on AIssential

• Predicting Immune Biomarkers with MultiModal Mixture-of-Expert Pathology Foundation Models Empowers Precision Oncology — MixTIME integrates multimodal pathology foundation models via a Mixture-of-Experts architecture to predict immune biomarkers for precision oncology.
• Gene dependency-informed inference of response to targeted cancer therapies — FORGE co-models drug response and gene dependency to predict cancer therapy benefit, improving mechanistic understanding and prediction.
• F Cancer — AI's impact on cancer is limited by technosolutionism, market failures, and insufficient biological understanding.
• A deep learning–based digital biopsy for predicting early recurrence in gastric cancer — The RSA model combines deep learning with clinical data to predict gastric cancer recurrence.
• Reusability report: Meta-learning for antigen-specific T cell receptor binder identification — PanPep generalizes well to unseen antigens with few known binders under specific negative sampling.
• Immunotherapy drug target identification using machine learning and patient-derived tumour explant validation — MIDAS, a multimodal GNN, effectively identifies novel immuno-oncology targets by integrating diverse biological data.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Machine learning : nature.com subject feeds.