The Concept Bottleneck: COMPASS Maps Tumor Immune Biology Before It Predicts Who Responds to Checkpoint Inhibitors

Most patients treated with immune checkpoint inhibitors do not respond, and biomarkers like PD-L1 expression or tumor mutational burden generalize poorly across cancer types and drug regimens. A new pan-cancer foundation model published in Nature Medicine argues the problem is not missing data — it is missing structure.

COMPASS (Concept-bottleneck Pan-cancer Model for immunotherapy Outcome and mechanistic UnderStanding) predicts immunotherapy response from pretreatment bulk tumor transcriptomes by routing gene expression through 44 biologically grounded immune concepts before making a prediction. The work, Generalizable AI predicts immunotherapy outcomes across cancers and treatments, was led by collaborators across Harvard Medical School, Roche, and Zhejiang University.

A Foundation Model With a Readable Bottleneck

Starting from 15,672 protein-coding genes, COMPASS uses a transformer-based gene encoder and hierarchical concept projector to organize tumor transcriptomes into immune cell states, tumor–microenvironment interactions, stromal programs, and signaling pathways — TGFβ signaling, IFNγ activity, endothelial exclusion, and dozens more.

The model is pretrained on 10,184 tumors across 33 cancer types using self-supervised contrastive learning, then fine-tuned on clinical immunotherapy cohorts. That two-stage design lets COMPASS transfer knowledge from large unlabeled pan-cancer data into small labeled clinical datasets — the reality of translational oncology.

Benchmark Results

Evaluated on 1,133 patients from 16 clinical cohorts spanning seven cancer types and six immune checkpoint inhibitor regimens:

• Outperformed 22 baseline methods on average
• Improved accuracy by 8.5%, AUPRC by 15.7%, and MCC by 12.3% versus best competing methods
• Generalized to cancer types and treatments withheld during fine-tuning — including predicting combination therapy response when trained only on monotherapy cohorts
• In survival analyses, COMPASS-classified responders had hazard ratio 4.7 for overall survival (P < 0.0001) in held-out metastatic urothelial carcinoma data

Mechanism, Not Just Scores

COMPASS generates personalized response maps connecting gene expression to immune concepts. In immune-inflamed non-responders, the model highlights TGFβ signaling, endothelial exclusion, CD4+ T cell dysfunction, and B cell deficiency — resistance programs that broad immune phenotypes miss.

Marinka Zitnik told Inside Precision Medicine the architecture forces predictions through "an intermediate layer, or bottleneck, before making the final call." That bottleneck is biologically interpretable by design.

The team also reports multi-stage fine-tuning outperforms single-stage adaptation for early-phase trials with small drug-specific cohorts — relevant for indication selection before large efficacy readouts exist.

Clinical Translation Caveats

COMPASS remains a research tool. Prospective clinical validation, assay standardization, and multi-center testing are essential before clinic deployment. The authors explicitly state it should not be used as a standalone basis for clinical decisions or to deny patients immunotherapy.

Still, the paper marks a shift in how biotech AI is evaluated: not just AUC on a single cohort, but generalization across withheld cancers, drugs, and checkpoint targets — plus mechanistic maps that trial designers can actually use.

Sources

• Nature Medicine — Generalizable AI predicts immunotherapy outcomes across cancers and treatments (2026)
• COMPASS Project — Generalizable AI for Cancer Immunotherapy (July 2026)
• Inside Precision Medicine — Predicting Cancer Immunotherapy Response Better with COMPASS AI Model (2026)
• Medical Xpress — AI tool improves prediction of who will respond to cancer immunotherapy drugs (July 2026)