Research Summary

COMPASS AI Model Predicts Immunotherapy Response

Anthony Calabro, MA

Key Highlights

  • COMPASS predicted immune checkpoint inhibitor response from pretreatment bulk tumor RNA sequencing data using biologically grounded tumor-immune concepts.
  • Across 16 clinical cohorts, COMPASS improved average accuracy by 8.5% and area under the precision-recall curve by 15.7% compared with the next-best methods.
  • The model generalized to cancer types, immune checkpoint inhibitor regimens, and treatment targets excluded from training.
  • In a phase 2 urothelial carcinoma cohort, patients classified as responders had longer overall survival than those classified as nonresponders.

A pan-cancer artificial intelligence (AI) foundation model predicted responses to immune checkpoint inhibitors across multiple cancer types and treatment regimens, while providing interpretable insights into potential resistance mechanisms, according to a study published in Nature Medicine.

The model, called COMPASS, was developed to address the limited generalizability of existing immunotherapy biomarkers and prediction methods. COMPASS represented each tumor using 43 biologically grounded tumor immune microenvironment concepts and a cancer-type token derived from pretreatment transcriptomic data.

The researchers pretrained COMPASS using self-supervised contrastive learning on bulk RNA sequencing data from 10,184 tumors across 33 cancer types. The model encoded expression data for 15,672 protein-coding genes and projected them onto 132 curated gene signatures related to immune cell types, functional states, signaling pathways, and nonimmune tumor biology.

The AI model was subsequently evaluated using pretreatment RNA sequencing and clinical outcome data from 1,133 patients in 16 clinical cohorts. The cohorts included 7 cancer types and several immune checkpoint inhibitor regimens, including therapies targeting programmed cell death protein 1, programmed death ligand 1, and cytotoxic T-lymphocyte–associated protein 4, as well as combination regimens.

Study Findings

According to the study’s response definition, 346 patients (30.5%) had a complete or partial response and were classified as responders. The remaining 787 patients, or 69.5%, had stable or progressive disease and were classified as nonresponders.

In leave-one-cohort-out analyses, the partial fine-tuning and linear probing versions of COMPASS produced the strongest overall results. Compared with the second-best-performing methods, they improved average accuracy by 8.5%, area under the precision-recall curve by 15.7%, and Matthews correlation coefficient by 12.3%.

The researchers also evaluated whether models trained on 1 cohort could predict outcomes in another. Using the study’s definition of successful transfer—accuracy greater than the reference accuracy for the target cohort—COMPASS-LFT succeeded in 163 of 240 cohort-to-cohort evaluations, and COMPASS-PFT succeeded in 155. The strongest previously published comparator succeeded in 130 evaluations.

COMPASS also predicted responses in settings excluded from training. When lung adenocarcinoma cohorts were omitted from the training set, COMPASS-PFT achieved 76.5% accuracy in the held-out population. It achieved 70.8% accuracy for anti–CTLA-4 treatment when trained on cohorts receiving PD-1– or PD-L1–targeted therapy. When trained only on monotherapy cohorts, the model achieved 85.3% accuracy in patients receiving combination immunotherapy.

In a held-out phase 2 study of atezolizumab in metastatic urothelial carcinoma, patients classified as responders had a 1-year overall survival rate of 86%, compared with 40% among patients classified as nonresponders. The risk of death was higher in the predicted nonresponder group than in the predicted responder group, with a hazard ratio of 4.7 and a log-rank P value of 1.7 × 10⁻⁷.

Clinical Implications

According to the researchers, COMPASS may support indication selection, biomarker discovery, mechanistic hypothesis generation, and patient stratification in immunotherapy trials. Personalized response maps identified tumor-immune programs associated with response and resistance, including transforming growth factor β signaling, endothelial exclusion, CD4-positive T-cell dysfunction, and B-cell deficiency.

The study was limited by its reliance on bulk RNA sequencing and by incomplete adjustment for clinical covariates, as age, sex, and tumor stage were inconsistently annotated across cohorts. The absence of non–immune checkpoint inhibitor comparator groups also prevented the researchers from distinguishing predictive effects from prognostic effects.

The authors emphasized that COMPASS remains an exploratory tool requiring prospective clinical validation, assay validation, cross-platform calibration, and reproducible inference pipelines. They cautioned that predictions from COMPASS or similar models should not be used alone to deny immunotherapy.

Expert Commentary

“COMPASS links tumor transcriptomes to interpretable immune representations and supports biomarker discovery, mechanistic hypothesis generation, and patient stratification in immunotherapy trials,” the researchers concluded.

Reference
Shen W, Moon I, Nguyen TH, et al. Generalizable AI predicts immunotherapy outcomes across cancers and treatments. Nat Med. Published online July 3, 2026. doi:10.1038/s41591-026-04502-7