Harvard AI model predicts which cancer patients will respond to immunotherapy

Researchers at Harvard Medical School have developed an artificial intelligence model called COMPASS that can predict which cancer patients will respond to immune checkpoint inhibitors, a class of drugs that works in only a fraction of those who receive them. The study was published July 3 in Nature Medicine.medicalxpress+1

Addressing Oncology's Central Unsolved Problem

Immune checkpoint inhibitors have transformed cancer treatment since the first was approved by the FDA in 2011, but clinical trials show only 10% to 40% of patients respond depending on their cancer type. Nonresponders risk serious side effects while their cancers progress unchecked.pubmed.ncbi.nlm.nih+1

COMPASS analyzes the activity of nearly 16,000 genes with known roles in immune cell states, tumor-microenvironment interactions, and signaling pathways. The model was trained on data from 10,184 tumors across 33 cancer types drawn from the Cancer Genome Atlas, then fine-tuned using results from 16 clinical trials testing different checkpoint inhibitor regimens on seven cancer types.medicalxpress+1

"Understanding who will respond to ICIs is not a minor knowledge gap," said senior author Marinka Zitnik, associate professor of biomedical informatics at HMS and associate faculty at Harvard's Kempner Institute. "It is one of the central unsolved problems in oncology."medicalxpress

Outperforming Existing Methods

In testing, COMPASS outperformed the best existing approach for predicting checkpoint inhibitor response by nearly 10% on average, with an 8.5% increase in precision, a 12.3% improvement in Matthews correlation coefficient, and a 15.7% gain in area under the precision-recall curve. The accuracy held across different cancer types, drugs, gene transcript sequencing platforms, and biopsy sites.pubmed.ncbi.nlm.nih+1

Unlike black-box AI systems, COMPASS uses a concept bottleneck transformer architecture that delivers human-interpretable results. It decodes tumors into 44 biologically grounded immune concepts, allowing clinicians to understand the rationale behind each prediction. This interpretability helped explain outlier cases — for instance, why some patients with immune-inflamed tumors still failed to respond, and why others with so-called immune-desert tumors benefited from treatment.medicalxpress+1

Clinical Implications

If validated in prospective trials, COMPASS could serve as a clinical decision aid, helping oncologists match patients to therapies and improving enrollment in clinical trials by identifying participants most likely to show a meaningful response. Zitnik and her colleagues plan to test whether incorporating electronic health records and single-cell sequencing data could further improve accuracy.medicalxpress