Researchers at Harvard Medical School have developed COMPASS, an artificial intelligence model that significantly improves the prediction of which cancer patients will respond to immune checkpoint inhibitor therapies. The findings, published in Nature Medicine, address a critical gap in oncology where these life-extending drugs benefit only a fraction of recipients despite widespread clinical use. Immune checkpoint inhibitors work by removing molecular barriers that shield tumors from the immune system, transforming certain cancers into manageable conditions. However, response rates typically range from 10 to 40 percent depending on the malignancy. Existing predictive biomarkers frequently misclassify patients, leading to ineffective treatments, unnecessary side effects, and delayed progression management. Led by Marinka Zitnik, associate professor of biomedical informatics at HMS and associate faculty at the Kempner Institute, the COMPASS team engineered a concept bottleneck transformer architecture designed to analyze the activity of nearly 16,000 genes linked to immune cell states, tumor microenvironment interactions, and signaling pathways. Unlike conventional black-box algorithms, this structure generates human-interpretable outputs, explicitly detailing the biological rationale behind each prediction. This interpretability proved crucial for resolving anomalous cases, such as identifying why certain immune-desert tumors still responded to therapy or why inflamed tumors failed to react. The model was trained on genomic data from over 10,000 tumors across 33 cancer types sourced from The Cancer Genome Atlas, then fine-tuned using outcomes from 16 clinical trials covering seven cancer indications. In rigorous validation tests where individual trials were excluded from training, COMPASS consistently outperformed current best-practice methods by 8.5 to 10 percent. The accuracy advantage remained stable across varying cancer types, drug classes, sequencing platforms, and biopsy methods. If validated in prospective clinical settings, COMPASS could function as a clinical decision support tool, enabling oncologists to match patients with the most effective immunotherapies upfront. The system would also optimize clinical trial recruitment by enriching cohorts with high-probability responders, thereby reducing trial failure rates and accelerating therapeutic development. Furthermore, its transparent mechanistic insights may reveal novel immune pathways, guiding researchers toward next-generation drug targets. The development team plans to expand COMPASS by integrating electronic health records, comorbidity data, prior treatment histories, and single-cell sequencing profiles to further refine predictive precision. As the model moves toward real-world validation, it represents a significant advancement in precision oncology, leveraging artificial intelligence to transform immunotherapy from a trial-and-error approach into a reliably targeted intervention.