COMPASS Adds Another Layer of AI Questions to Immunotherapy Trials

Immune checkpoint inhibitors have reshaped the treatment outlook for some cancers, yet they have always come with a difficult reality: with the same drug, some people achieve durable remission, while others bear only side effects and the cost of time. The COMPASS model recently published by Harvard Medical School and collaborators is an attempt to determine, before treatment begins, which side a patient is more likely to fall on based on gene expression signals in the tumor.

The study was published in Nature Medicine. According to the paper’s abstract, COMPASS is a pan-cancer foundation model that uses large-scale tumor transcriptome data and, through a “concept bottleneck” architecture, links gene activity to more interpretable immune-related concepts before predicting a patient’s response to immune checkpoint inhibitors. The research team said the model was first trained on 10,184 tumors across 33 cancer types, then evaluated in 16 clinical cohorts covering 7 cancer types and 6 immune checkpoint inhibitors.

Compared with 22 benchmark methods, the paper reports that COMPASS improved average accuracy by 8.5% and area under the precision-recall curve by 15.7%; patients classified by the model as responders also showed better separation in overall survival. These results make COMPASS not just a classifier for a single cancer type, but more like a patient-stratification framework that can be tested across different cancer types and different drug contexts.

Yet that is also precisely why clinical trial design becomes more complicated. If researchers incorporate this kind of model into immunotherapy trials, they must decide whether it is used to screen enrollment, stratify randomization, explore biomarkers, or serve only as a post hoc analysis tool; each position involves different statistical designs, risks of failure, and ethical questions. If the model wrongly excludes potential beneficiaries, or directs patients who are unlikely to benefit toward treatment, the consequences are not merely bias in an algorithmic score.

The official code repository shows that COMPASS can output response and non-response predictions from gene expression tables, and can also extract features at the gene, gene-set, and 44-concept levels. The project download page lists datasets, pretrained models, an all-immune-checkpoint-inhibitor model fine-tuned on 1,133 patients, and model files for drugs including atezolizumab, ipilimumab, nivolumab, and pembrolizumab. These open materials help external researchers reproduce and extend the work, but some clinical cohorts still require applications through controlled databases such as EGA and dbGaP, so data accessibility is not uniform.

For ordinary patients, the most important boundary is this: COMPASS remains a research tool, not medical advice that can directly determine treatment. The project website also clearly states that its use is for research and education. Even if retrospective data show that the model can distinguish groups more likely to benefit, real clinical adoption will still require prospective trials demonstrating that it can improve decision-making, rather than merely reorganizing signals in existing data.

The larger scientific question is generalizability. Tumor transcriptomes can be affected by sample handling, sequencing platform, tumor purity, and differences among patient populations; response to immunotherapy is also not determined only by tumor cells, but also involves the microenvironment, prior treatments, comorbidities, and drug combinations. COMPASS compresses these complex factors into more readable immune concepts, which is an attractive direction, but it also requires researchers to re-examine whether the model remains reliable before each use in a new cancer type, new drug, or new population.