Right Drug, Wrong Patient: The AI Out to End Cancer's Costliest Mistake

Immunotherapy can be close to a cure, but only for a lucky minority, and today almost no one can tell you in advance whether you are one of them. A new AI is trying to read that answer off your tumour before the first dose, and to do it for many cancers at once.

Picture the worst coin toss of your life. Heads, a drug wakes up your immune system and clears a cancer that should have killed you. Tails, you get months of harsh side effects, the tumour keeps growing, and the window to try something else quietly closes.

For most people offered immunotherapy, that is roughly the choice. And the strange part: the doctors flipping the coin with you usually cannot see which way it will land.

What patients want is a weather forecast. Not a promise, just an honest read before they bet months of their life: is this drug likely to work for me, yes or no. A model published in early July gets closer to that forecast than anything before, and it does the thing this field keeps failing at: it works across many cancers, not just one.

Your immune system is good at killing things that look wrong, which is exactly what a cancer cell is. So why does cancer grow at all? Because tumours learn to hide, flashing a molecular "do not attack" sign that switches off the immune cells sent to destroy them.

Immunotherapy takes the brakes off. The most common kind, a checkpoint inhibitor, blocks that "do not attack" signal so the immune system can see the tumour again. The checkpoints are proteins with names like PD-1, PD-L1 and CTLA-4. Picture them as an invisibility cloak the cancer wears. The drug pulls the cloak away.

When it works, it can work spectacularly. The first of these drugs were approved in 2011. One, pembrolizumab, is widely credited with keeping former US president Jimmy Carter alive for nine years after a melanoma, a skin cancer, had spread to his liver and brain. For the right person, this is the difference between a death sentence and a normal life.

Here is the catch that has haunted cancer medicine for a decade. Only about 10 to 40 percent of patients respond, depending on the cancer. The rest take the drug and get little or nothing.

That failure is not free. Checkpoint inhibitors can turn the immune system loose on healthy organs, with side effects from a rash to serious damage. They are expensive. And every month on a treatment that was never going to work is a month the cancer grows and other options slip away.

Doctors do have some clues. Tumours crowded with immune cells, called "hot" tumours, tend to respond better than "cold" ones with few immune cells inside. But the clues are crude, and many patients defy them: hot tumours that stay silent, cold ones that light up.

The new model is called COMPASS, built by researchers at Harvard Medical School and colleagues and published in Nature Medicine on 3 July 2026. Instead of leaning on one or two crude markers, it reads the tumour's gene activity: which of its genes are switched on, and how loudly.

That readout, around 16,000 genes' worth, carries much of the story: how the immune system is behaving around the tumour, how the cancer is signalling, what state the neighbouring cells are in. COMPASS learns which patterns separate responders from non-responders.

One design choice matters more than it sounds. Most medical AI is a black box: an answer with no reasons, which doctors rightly distrust. COMPASS is built to show its work, routing each prediction through biological ideas a specialist can name, so it can point to why a patient is likely to respond or not. That even let the team explain some odd cases, like a "hot" tumour that stayed silent.

Plenty of models can predict response for one cancer, tuned to one disease and one drug. They fall apart the moment you point them at a different cancer, a different drug, or samples processed in another lab. A pattern learned on melanoma may mean nothing in lung cancer.

Generalising is the whole point of COMPASS, and the reason it matters. It was trained on 10,184 tumours across 33 cancer types from a large public database, then tested against 16 real clinical trials covering seven cancers and six immunotherapy drugs. The team even hid an entire trial, asked the model to predict its patients cold, then checked the answers. On average it beat the best previous method by about 8.5 percent, and it held up across cancers, drugs and lab equipment.

A few percent may not sound like a revolution. But in a field where most tools break when conditions change, a model that stays useful across cancers, drugs and labs is rare.

If COMPASS holds up, the payoff is blunt and human. Spare the likely non-responders a punishing, costly treatment and move them to something else while there is still time. Get the likely responders to the drug that could save them sooner. It could also make drug trials smarter.

Now the part that must be said plainly. COMPASS has been tested on records of patients already treated, looking backwards. That is a fair test, but not the real one. It predicts, it does not yet prescribe.

None of that makes it small. The honest label today: promising, not proven.

Immunotherapy releases the brakes on your own immune system so it can attack the cancer. The most common type, checkpoint inhibitors, blocks the "do not attack" signals that tumours use to hide, such as PD-1 and PD-L1. When it works it can be close to a cure, but it only helps roughly 10 to 40 percent of patients, depending on the cancer. Part of the reason is that some tumours are full of immune cells the drug can unleash, while others are "cold" and have almost none to work with. Predicting in advance which patient is which has been one of oncology's hardest problems.

COMPASS predicts whether a specific patient is likely to respond to immunotherapy before they start it. It does this by reading the tumour's gene activity, the pattern of which genes are switched on, across roughly 16,000 genes tied to the immune system and the cancer's behaviour. Unlike most medical AI, it is built to explain its reasoning in biological terms a specialist can check, rather than handing over a black-box score. In testing across many cancers and drugs it beat the best previous method by about 8.5 percent on average. Importantly, it predicts which treatment to choose; it does not detect cancer in the first place.

Not yet. COMPASS has only been tested on data from patients who were already treated, looking backwards to see whether it would have called their outcomes correctly. That is encouraging, but the real test is a prospective clinical trial, where the model's predictions actually help guide treatment and researchers measure whether patients do better. The study team says this validation still needs to happen before COMPASS could be used as a decision aid in the clinic. For now it is a powerful research result and a signpost, not a tool your oncologist can order.

Related from Frontier Signal: our deep dive on the robot that can perform surgery without a surgeon. Frontier Signal explains frontier technology in plain English. This is general information, not medical advice.