A new artificial intelligence model developed by Harvard Medical School researchers can predict which cancer patients will benefit from immunotherapy drugs with significantly greater accuracy than current methods, potentially transforming how doctors personalize cancer treatment. The tool, called COMPASS, analyzes nearly 16,000 genes to identify patients most likely to respond to immune checkpoint inhibitors (ICIs), a class of drugs that has revolutionized cancer care over the past decade.

Why Does It Matter That Some Patients Don't Respond to Immunotherapy?

Immune checkpoint inhibitors are remarkable drugs that work by removing the "invisibility cloak" cancer cells use to hide from the immune system. These drugs target proteins like PD-L1, PD-1, and CTLA-4 on tumor and immune cells, allowing the body's defenses to recognize and destroy cancer cells. For some patients, the results are life-changing. Former U.S. President Jimmy Carter, for example, survived nine years after a diagnosis of stage IV melanoma that had spread to his liver and brain, largely due to taking a PD-1 blocker called pembrolizumab.

However, this success story represents only a fraction of patients who receive these drugs. Clinical trials show that only 10 to 40 percent of patients experience meaningful benefit from ICIs, depending on their cancer type. The remaining patients not only face potential serious side effects but also lose valuable time receiving ineffective treatment while their cancers progress. This unpredictability has been a central challenge in oncology, leaving doctors and patients uncertain about whether to pursue this expensive and sometimes risky treatment path.

How Does COMPASS Work, and What Makes It Different?

COMPASS uses what researchers call a "concept bottleneck transformer architecture," which means it doesn't just spit out a prediction without explanation. Instead, it provides human-interpretable results, showing doctors and researchers the reasoning behind its predictions. This transparency is crucial because it allows clinicians to understand why the model thinks a particular patient will or won't respond.

The researchers trained COMPASS using data from 10,184 tumors across 33 different cancer types from the Cancer Genome Atlas, a public database containing genetic and molecular information from cancer patients. The model learned what patterns of gene activity correlated with patients who responded well to ICIs and those who didn't. They then fine-tuned this training using results from 16 clinical trials testing different ICI regimens across seven cancer types.

To test how well COMPASS actually works, the team removed individual clinical trials one at a time and asked the model to predict which patients in that missing trial would respond to immunotherapy. The results were impressive: COMPASS outperformed the best existing approaches for predicting ICI response by nearly 8.5 percent on average. This improvement held true across different cancer types, different ICI drugs, different gene sequencing methods, and different biopsy sites.

What Could This Mean for Cancer Patients and Doctors?

If these results hold up in future clinical trials, COMPASS could transform cancer care in several important ways. First, it could help oncologists decide which individual patients would benefit most from ICIs before they receive the drug, avoiding unnecessary treatment and side effects for those unlikely to respond. Second, it could make clinical trials more efficient by helping researchers enroll the best-matched participants, increasing the likelihood of meaningful responses and faster drug development. Third, because COMPASS explains its predictions, it could generate new insights about how the immune system fights cancer, potentially leading to entirely new drug targets and therapies.

The model's interpretability has already revealed surprising insights. For example, some patients with immune-inflamed tumors (tumors with lots of immune cells) didn't respond to ICIs, but COMPASS could explain why: their gene expression patterns suggested biological processes that actually impeded immune response. Conversely, some patients with immune-desert tumors (tumors with few immune cells) did respond, and their gene signatures suggested other types of immune activity at work. These discoveries challenge conventional thinking and could open new treatment avenues.

How to Prepare for Potential Future Use of AI-Guided Cancer Treatment

• Ask Your Doctor About Genetic Testing: If you or a loved one is considering immunotherapy, ask whether genetic or molecular testing is available to help predict response. As tools like COMPASS become validated and available, this information could become standard practice in cancer care.
• Understand Your Tumor's Characteristics: Learn whether your tumor is classified as immune-inflamed or immune-desert, and what that means for treatment options. This information can help you have more informed conversations with your oncology team about which therapies might work best.
• Discuss Clinical Trial Eligibility: If you're considering immunotherapy, ask your doctor whether you might be a good candidate for clinical trials testing new approaches. AI-guided patient selection could make these trials more likely to succeed and benefit participants.

Researchers at Harvard Medical School are already planning next steps to improve COMPASS further. They're exploring whether adding additional data could boost accuracy even more, including information from patients' electronic health records such as medical history, disease complications, and previous responses to other treatments. They're also considering incorporating data from single-cell sequencing, which could reveal the role of different cell populations in determining who responds to immunotherapy.

"ICIs are an exciting therapeutic modality that has transformed cancer treatment over the past decade by engaging the immune system to fight cancer cells and destroy them. By leveraging cutting-edge AI capabilities, we can identify who would be most likely to respond to a particular ICI before that patient receives the drug," said Marinka Zitnik, associate professor of biomedical informatics at Harvard Medical School.

Marinka Zitnik, Associate Professor of Biomedical Informatics, Harvard Medical School

The development of COMPASS represents a significant step toward precision medicine in cancer care. By combining artificial intelligence with deep biological knowledge, researchers are working to solve what has been one of oncology's most persistent puzzles: figuring out which patients will benefit from immunotherapy before they start treatment. While the model still needs validation in prospective clinical trials before it can be used in routine cancer care, the early results suggest that AI-guided prediction could soon help thousands of cancer patients receive the right treatment at the right time.