Pan-Cancer AI Model Targets Immunotherapy Challenge, as BioCOMPASS Tries to Put Biomarkers at the Core of Prediction

One of the toughest problems in cancer immunotherapy is not whether the drugs work, but in whom they will work. Immune checkpoint inhibitors such as PD-1 and CTLA-4 have enabled long-term tumor control in some patients, but many others still show no clear response after bearing side effects and time costs. If responses can be predicted more accurately before treatment, clinical strategies could potentially be adjusted earlier, and trial design could become more precise.

FirstWord Pharma reported that a pan-cancer AI tool called BioCOMPASS has improved the accuracy of predicting immunotherapy response. According to its research preprint, BioCOMPASS is an extended version of the existing COMPASS model. Its focus is not only on genomic or transcriptomic signals themselves, but on incorporating clinical biomarkers and treatment types into the model’s interpretation.

The specific task targeted by this model is to predict whether cancer patients are likely to respond after receiving immune checkpoint inhibitors. Public information shows that BioCOMPASS supports treatment indicators including PD-1, CTLA-4, and combination therapies, and uses multiple classes of input features, including 62 immune cell features, 42 CTLA4 and PD1 pathway activity scores, and auxiliary biomarkers such as TIDE and IPRES that are associated with immune escape or treatment resistance.

The research team stated in the preprint that BioCOMPASS uses designs such as treatment gating, concept alignment, pathway consistency, and auxiliary learning to help the model interpret patient data in a way that is closer to the biological context of immunotherapy. In other words, it does not simply look for statistical associations in the data, but tries to place “which type of immunotherapy was received” and “what signals the tumor immune microenvironment presents” within the same predictive framework.

In validation, the research report noted that in leave-one-cohort-out testing across 8 cohorts, BioCOMPASS increased accuracy from COMPASS’s 63.10% to 70.00%. The authors also used strategies such as leave-one-cancer-type-out and leave-one-treatment-out to test the model’s generalizability when facing unseen cancer types or treatment types. These designs can more closely approximate real-world application scenarios than a single data split, but the results currently still come from a preprint and are not equivalent to prospective clinical validation.

The limitations are therefore also quite clear. Immunotherapy response is affected by tumor type, prior treatments, sample processing, sequencing platforms, and response definitions; biases between different datasets may allow a model to perform well in research data but decline once placed into hospital workflows. To become a tool usable for treatment selection, BioCOMPASS still needs to demonstrate stability in independent, multicenter, prospective data, and explain how its outputs should be used together with existing clinical criteria such as PD-L1 expression and tumor mutational burden.

The significance of this progress lies in shifting the focus of biomedical AI from the abstract idea of “more accurate prediction” toward a more concrete clinical question: how to connect immunobiology, treatment mechanisms, and patient data into testable judgments. BioCOMPASS’s publicly available code also helps external researchers reproduce and examine the method. However, before regulation, clinical responsibility, and patient stratification standards are clarified, it is more like a promising research platform than a diagnostic and treatment tool that can directly replace physicians’ judgment.