COMPASS AI Predicts Who Responds to Cancer Immunotherapy

Cancer immunotherapy has produced some of the most striking recoveries modern oncology has ever recorded, yet it remains maddeningly unpredictable. A drug that dissolves one patient's tumors will do nothing for another whose disease looks, on paper, almost identical. This week a research team at Harvard Medical School put forward an artificial intelligence model built to close that gap. Named COMPASS, the system reads the gene activity inside a tumor and estimates whether a patient is likely to respond to immune checkpoint inhibitors, the class of drugs that has redrawn the map of cancer treatment over the past decade. The work was published in Nature Medicine on July 3, 2026, and it arrives with an unusually careful set of claims about what the model can and cannot yet do.

What makes COMPASS worth paying attention to is not only that it predicts more accurately than the tools clinicians currently lean on, but that it tries to explain itself. Instead of returning a bare score, it points to the biological programs behind its verdict. In a field where so many machine learning results feel like oracles handing down numbers, that shift toward interpretability may matter as much as the accuracy gain itself.

Why It Matters

Immune checkpoint inhibitors, or ICIs, work by stripping away a disguise. Tumor cells drape themselves in proteins such as PD-L1, PD-1, and CTLA-4 that act like an invisibility cloak, telling the immune system to stand down. ICIs block that signal, and the body's own T cells are freed to recognize the cancer and destroy it. For some patients this has turned a death sentence into a manageable, even curable, condition.

The trouble is arithmetic. Depending on the cancer type, only about 10 to 40 percent of patients benefit from these drugs. The rest gain nothing while still exposed to serious side effects, and they lose precious months on a treatment that was never going to work as their disease advances. Existing biomarkers offer rough guidance. Tumors thick with infiltrating immune cells, so-called immune-inflamed tumors, tend to respond, while barren "immune deserts" tend not to. But a large share of patients defy that logic entirely, which is why senior author Marinka Zitnik, an associate professor of biomedical informatics at Harvard Medical School, describes the challenge plainly. Knowing who will respond, she says, "is one of the central unsolved problems in oncology."

How COMPASS Works

COMPASS was trained on 10,184 tumors spanning 33 cancer types, drawn from The Cancer Genome Atlas, the vast public repository of sequenced tumor and matched-normal tissue. From that data the model learned which patterns of gene activity separate responders from nonresponders across a wide range of checkpoint drugs. It reads the behavior of nearly 16,000 genes tied to immune cell states, interactions within the tumor microenvironment, and the signaling pathways that govern them.

The architecture is the interesting part. COMPASS uses what its designers call a concept bottleneck transformer. Rather than compressing everything into an opaque internal representation and emitting a black-box prediction, the model routes its reasoning through a layer of human-interpretable immune concepts. Each of those concepts corresponds to a recognizable piece of tumor biology, so when COMPASS decides a patient is unlikely to respond, a clinician can trace that decision back to the specific immune programs that drove it. The researchers then fine-tuned the model on results from 16 clinical trials covering seven cancer types, testing different checkpoint regimens against real outcomes.

What the Results Reveal

To measure how well COMPASS generalizes, the team used a deliberately harsh test. They removed each clinical trial from the fine-tuning set one at a time, then asked the model to predict responders and nonresponders in the trial it had never seen. Across those held-out cohorts, COMPASS outperformed the best existing approach by 8.5 percent on average, and the improvement held across different cancer types, different drugs, different sequencing platforms, and different biopsy sites. That consistency is what separates a genuinely generalizable tool from one that has merely memorized a handful of datasets.

Because its reasoning is legible, COMPASS could also account for the puzzling cases that undermine simpler biomarkers. Some nonresponders whose tumors looked immune-inflamed turned out to carry gene expression tied to processes that quietly suppressed the immune attack. Conversely, some responders with apparent immune deserts showed signatures pointing to other forms of immune activity that standard measures miss entirely. In other words, the model did not just flag the outliers; it offered a biological story for why they broke the rules, and those stories are testable.

What Comes Next

The Harvard team is careful about the boundary between promise and proof. COMPASS was validated on historical patient data, which is the right place to start but not the same as guiding a live treatment decision. The results would need to hold up in a prospective clinical trial, one where the model's predictions are made before patients receive their drugs and then checked against what actually happens. That is a multi-year undertaking, and the authors say so.

If it clears that bar, the payoff runs in three directions. Patients could be spared treatments unlikely to help them, sharpening the practice of personalized medicine. Trials of new therapies could enroll more efficiently by enriching for likely responders. And because the model surfaces the biological programs behind resistance, it hands researchers a set of concrete leads for new drug targets, aimed at the very patients current drugs leave behind. Zitnik's group has also connected COMPASS to the broader effort at Harvard's Kempner Institute to understand natural and artificial intelligence together, a reminder that the model is a research instrument as much as a clinical one.

Closing Thoughts

There is a quiet lesson threaded through this work. For years the debate about AI in medicine has framed accuracy and explainability as a trade-off, as if a model had to choose between being right and being understood. COMPASS suggests that framing may be too pessimistic. By building interpretability into its architecture rather than bolting an explanation on afterward, it managed to be both more accurate and more transparent than the tools it aims to replace.

Whether COMPASS earns a place in the clinic will ultimately be decided by trials, not by benchmarks, and that verdict is years away. But the direction it points feels right. The most useful medical AI may not be the system that simply knows the answer, but the one that can show a clinician the reasoning and let a human weigh it. For a disease as varied and stubborn as cancer, that kind of partnership between judgment and computation is probably the only version worth trusting.

한국어 요약

하버드 의과대학 연구진이 암 면역항암제(면역관문억제제, ICI)에 어떤 환자가 반응할지를 예측하는 인공지능 모델 'COMPASS'를 개발해 2026년 7월 3일 국제학술지 네이처 메디신에 발표했다. 면역관문억제제는 지난 10년간 암 치료의 판도를 바꿨지만 암종에 따라 환자의 10~40%에게만 효과가 나타나며, 나머지 환자는 부작용만 감수한 채 귀중한 시간을 잃는다. COMPASS는 종양의 유전자 발현 패턴을 읽어 반응 가능성을 예측하고, 단순한 점수가 아니라 그 판단의 근거가 된 면역 생물학적 프로그램까지 함께 제시한다.

연구진은 암 유전체 지도(TCGA)의 33개 암종·10,184개 종양 데이터로 모델을 학습시키고, 7개 암종을 다룬 16건의 임상시험 결과로 미세조정했다. 핵심은 '개념 병목 트랜스포머'라는 구조로, 예측이 불투명한 내부 계산이 아니라 사람이 이해할 수 있는 면역 개념들을 거치도록 설계돼 임상의가 결정 근거를 추적할 수 있다. 한 번도 학습하지 않은 임상시험을 하나씩 떼어내 검증한 결과, COMPASS는 기존 최고 방법보다 평균 8.5% 더 정확했고, 암종·약물·검사 플랫폼이 달라져도 성능이 일관되게 유지됐다.

연구책임자인 마린카 지트닉 교수는 반응 예측이 "종양학의 미해결 핵심 과제 중 하나"라고 말한다. 다만 이번 성과는 과거 환자 데이터로 검증한 단계이며, 실제 진료에 쓰이려면 예측을 먼저 내리고 결과와 대조하는 전향적 임상시험을 통과해야 한다. 검증에 성공한다면 맞춤형 치료, 효율적인 임상시험 등록, 새로운 약물 표적 발굴로 이어질 수 있다. 정확성과 설명 가능성을 동시에 잡으려 한 이 접근은, 답만 내놓는 AI가 아니라 근거를 보여주고 사람의 판단을 돕는 의료 AI의 방향을 시사한다. (참고: Nature Medicine, Shen W, Moon I, Nguyen TH, et al., 2026)