Immune checkpoint inhibitors have changed the treatment of many cancers by helping the immune system recognize and attack tumor cells. While these therapies can be remarkably effective, only a portion of patients benefit from them. Finding reliable ways to predict who will respond remains one of the biggest challenges in cancer care.

Researchers from Harvard Medical School have developed an artificial intelligence model called COMPASS that could help address this problem. Rather than relying on a single biomarker, the model analyzes RNA sequencing data from tumor samples to identify complex patterns of gene activity linked to successful immunotherapy responses.

COMPASS was trained using gene expression data from more than 10,000 tumors representing 33 different cancer types. Instead of treating every gene independently, the model organizes gene activity into 44 biologically meaningful immune concepts. These concepts describe immune cell states, communication between tumors and the surrounding tissue, and important signaling pathways that influence how cancers interact with the immune system.

Concept bottleneck foundation model for interpretable prediction of immunotherapy response

a, Transfer of immuno-oncology knowledge via hierarchical concept learning. b, Self-supervised pretraining on pan-cancer transcriptomes. c, Fine-tuning on clinical cohorts for explainable prediction of immunotherapy outcomes. d, Architecture of the COMPASS model. The model comprises three components: (1) a transformer-based gene encoder that transforms expression profiles into context-aware representations; (2) a hierarchical concept projector that progressively aggregates these into multi-scale TIME concepts; and (3) a task-specific classifier that outputs predictions from concept-level features. e, Flexible fine-tuning strategies for clinical adaptation. 

When evaluated across multiple clinical datasets, COMPASS consistently outperformed 22 existing prediction methods. It achieved higher accuracy in identifying patients who would benefit from immune checkpoint inhibitors and also worked well for cancer types and treatments that were not included during its final training. This ability to generalize makes the model particularly promising for use in a wide range of clinical settings.

The researchers also found that patients predicted by COMPASS to respond to treatment experienced significantly longer overall survival than those predicted not to respond. Beyond making predictions, the model provides insight into why a patient may or may not respond by highlighting biological processes associated with treatment resistance. For example, it can identify immune signaling pathways and immune cell changes that may prevent successful therapy even in tumors that appear to have active immune responses.

An important advantage of COMPASS is its interpretability. Many artificial intelligence models function as black boxes, making it difficult to understand how they reach their conclusions. COMPASS instead connects its predictions to recognizable biological concepts, allowing researchers and clinicians to better understand the underlying mechanisms driving treatment outcomes.

Although additional clinical validation is needed before the model becomes part of routine patient care, COMPASS represents an important step toward more personalized cancer treatment. By combining RNA sequencing with biologically informed artificial intelligence, the approach may help physicians identify the patients most likely to benefit from immunotherapy while also uncovering new targets for future therapies.

Availability – The Python implementation of immunotherapy response prediction models as well as COMPASS is available on GitHub at https://github.com/mims-harvard/COMPASS.