An AI model just got better at predicting who cancer immunotherapy will actually help

Only a fraction of cancer patients respond to checkpoint inhibitors, so a model that flags likely non-responders before treatment starts could spare people from harsh side effects with little chance of benefit. What makes COMPASS notable is that it does not just spit out a probability, it gives clinicians an interpretable rationale, which is the piece that usually keeps AI tools out of real hospitals. Trained across 33 cancer types and 16 clinical trials, this is the kind of unglamorous, high leverage medical AI that rarely makes headlines but could reshape treatment decisions faster than flashier drug discovery announcements.