Harvard's COMPASS predicts immunotherapy response across cancers

For a decade the frustrating bottleneck around immune checkpoint inhibitors has not been whether they work but which patient will actually benefit, since response rates on this class of drugs sit somewhere between 10 and 40 percent depending on the cancer type. A team at Harvard Medical School published a model this week in Nature Medicine that tries to sharpen that guess, and unlike the current biomarker toolkit it is meant to work across cancers and across treatments in one shot.

The model is called COMPASS. According to the authors it was pre-trained on 10,184 tumors covering 33 cancer types from the Cancer Genome Atlas, then fine-tuned on 16 clinical trials spanning seven cancers where patients had been treated with anti-PD1, anti-PD-L1, anti-CTLA4, or combination therapies. In a leave-one-cohort-out evaluation the paper reports COMPASS beat 22 existing prediction methods by roughly 8.5 percent on accuracy, with a 15.7 percent lift on area under the precision-recall curve.

The architectural choice worth flagging is that COMPASS is a concept bottleneck transformer. Rather than mapping gene expression straight to a yes or no response, it routes about 16,000 genes through 44 human-readable immune concepts (immune cell states, tumor-microenvironment interactions, signaling pathways) and predicts from those. That gives an oncologist a rationale, not just a score. Senior author Marinka Zitnik, associate professor of biomedical informatics at Harvard Medical School, told reporters that identifying ICI responders is 'one of the central unsolved problems in oncology,' which is a fair way to describe the last decade of missed indications.

The honest caveat is that all of this is retrospective. Leave-one-cohort accuracy on published trials is a much easier bar than showing up in a real oncology clinic with a non-TCGA sequencing panel and actually changing a treatment decision. What the reporting does not tell you is how COMPASS performs on patients outside the academic-center populations that dominate TCGA, or how it holds up as newer ICI-plus combinations enter first-line care.

The forward look is more interesting than the headline. If the 44-concept layer generalizes the way the authors claim, COMPASS is not only a triage tool for the clinic; it is a hypothesis generator for drug discovery groups and a way for pharma sponsors to enrich ICI trial arms before enrolling patients.