The AI that matters most this month is not writing anyone's emails. It is reading tumors and telling doctors which patients a cancer drug will actually help.

By Shashi Bellamkonda · July 7, 2026

reading the gene activity inside a tumor and predicting whether a patient will respond to immunotherapy is the kind of work analytical AI was built for, and the kind that rarely makes a keynote. On July 3, researchers at Harvard Medical School published a model called COMPASS that does exactly that. It predicts response to immune checkpoint inhibitors, the drugs that strip the immune-suppressing cloak off cancer cells so the body can attack them, and it does so more accurately than any prior approach across sixteen clinical cohorts (Harvard Medical School, 2026).

These drugs help only 10 to 40 percent of the patients who receive them, depending on the cancer (Harvard Medical School, 2026). The rest lose months to a treatment that will not work while their disease advances. Telling those two groups apart before the first dose is one of oncology's hard, unglamorous problems, and it is a pattern-finding problem, the sort a model can work through faster than a person reading one tumor at a time.

Analytical AI keeps solving the problems that matter and losing the headline

Generative AI takes the oxygen. Chat assistants, image tools, and copilots pull the budgets and the coverage, while the models quietly reshaping medicine, climate work, and drug discovery sit a layer down, doing the heavier lifting on data too large and too fragmented for a human to hold in their head.

COMPASS is one of those. It reads nearly 16,000 genes' worth of activity in a tumor sample and finds the signatures that separate responders from nonresponders (Harvard Medical School, 2026). No human oncologist reads 16,000 genes across thousands of past patients and holds the correlations in memory. The model does, and hands back a prediction with a reason attached.

I have made this argument before, using a laboratory diagnostics platform as the proof point. COMPASS is a cleaner one, because of how the team built its reasoning to be read.

It shows its work, and that is the part clinicians needed

Most high-accuracy models give a score and no story. A clinician, or a regulator, is left to trust a number with no way to check the logic. COMPASS uses a design its authors call a concept bottleneck transformer, which routes gene activity through 44 biologically grounded immune concepts before it predicts anything, so a doctor sees the immune cell states and signaling pathways behind each call (Nature Medicine, 2026).

That interpretability did real diagnostic work. The model explained its own outliers, showing why some patients with immune-rich tumors still failed to respond, and why some with immune-poor tumors did (Harvard Medical School, 2026). A black box cannot generate a new hypothesis about the immune system. This one can, which is how a prediction tool becomes a research instrument.

Open source and commercial services are not in tension here

COMPASS shipped on GitHub under an open license, installable in one command, trained on the Cancer Genome Atlas, a public database of 10,184 tumor samples across 33 cancer types (Harvard Medical School, 2026). Any research team can download it and run it today.

That openness and a healthy commercial market can coexist. The Harvard team is careful that COMPASS is not a cleared clinical product, and that its predictions have to hold up in prospective trials before a doctor leans on them (Harvard Medical School, 2026). Getting from published research code to something a hospital can run means validation on local data, regulatory clearance, monitoring, and support. A company that does that work, on top of an open foundation, is offering something worth paying for. Several of COMPASS's own funders, including Roche and AstraZeneca, build precisely those commercial layers, and two of the paper's authors now work at Roche (Harvard Medical School, 2026).

An open model that raises the floor on accuracy and interpretability makes the commercial layer more honest, not less viable. The vendor competes on the parts patients and regulators care about, safety, validation, and evidence, rather than on owning the algorithm.