Generalizable AI predicts immunotherapy outcomes across cancers and treatments

Can AI Decode the Complexity of Cancer Immunity?

Why do some patients achieve long-term remission with immunotherapy while others see no benefit at all? Immune checkpoint inhibitors (ICIs) have fundamentally changed oncology. Yet, their success is highly unpredictable. Researchers have long sought a reliable way to identify which patients will respond. Current biomarkers often fail to generalize across different cancer types or treatment regimens.

A new study introduces COMPASS, a foundation model designed to bridge this gap. By using tumor gene data, the model predicts whether a patient will respond to immunotherapy. Crucially, it attempts to explain why by linking predictions to specific biological processes. These include immune cell activity or tumor signaling.

The search for a universal predictor

The central question is whether a single model can capture the shared principles of immune response across human cancer. Immunotherapy works by "releasing the brakes" on the immune system. This allows T cells—the body's specialized killer cells—to attack tumor cells. However, the tumor microenvironment (the local ecosystem of cells and signals surrounding a tumor) is incredibly complex. It varies wildly between a patient with lung cancer and a patient with melanoma.

The authors investigate whether a model can move beyond simple, single-marker predictors. They want to understand the deeper, functional state of a patient's immune system. Their goal is to create a tool that adapts to new drugs and unseen cancer types. It does this by focusing on the underlying biological "concepts" that drive immunity.

Cracks in the biomarker foundation

Until now, clinical practice has relied on a handful of validated biomarkers. One is tumor mutational burden (TMB), which measures genetic mutations in a tumor. Another is PD-L1 expression, a marker of a specific checkpoint that drugs can block.

However, the authors note that these markers are often insufficient. Many high-TMB tumors remain resistant to treatment. Conversely, some low-TMB tumors respond robustly. Existing transcriptomic signatures (patterns of gene activity) also tend to be highly specialized. A signature that works well for melanoma might fail completely in lung or bladder cancer. This lack of generalizability creates a massive hurdle for drug development. Models trained on one specific clinical trial often fail to predict outcomes in the next.

Mapping the immune landscape

To solve this, the researchers developed COMPASS using a "concept bottleneck" architecture. Neural networks often jump directly from raw gene measurements to a prediction. Instead, the authors forced the model to pass information through a middle layer. This layer contains 44 biologically grounded "concepts." These concepts represent high-level biological themes like "exhausted T cells" or "interferon-gamma signaling." They act as a translator between raw data and clinical outcomes [Figure 1a].

The investigation began with a massive pre-training phase. The authors used self-supervised learning—a method where the model learns patterns from unlabeled data—on 10,184 tumors. These samples came from 33 different cancer types [Figure 1b]. By teaching the model the "language" of tumor transcriptomes (the complete set of RNA molecules in a cell), they built a foundation of general biological knowledge.

Following pre-training, the researchers tested the model's flexibility through various fine-tuning strategies [Figure 1e]. They evaluated COMPASS against 22 existing baseline methods across 16 independent clinical cohorts [Figure 2a]. To ensure the results were robust, they used a "leave-one-cohort-out" strategy. This means they trained the model on a set of cancers and then tested it on an entirely unseen cohort. This simulates real-world clinical deployment [Figure 2b].

Superior accuracy and biological insight

The results show that COMPASS significantly outperforms established methods. The study finds that COMPASS increases precision by 8.5%. It also improves the Matthews correlation coefficient (a metric for accuracy that accounts for class imbalance) by 12.3%. Most notably, it increases the area under the precision-recall curve (a measure of how well a model finds true positives) by 15.7% [Figure 2c]. These improvements are measured against the second-best performing baseline models.

The model also demonstrated remarkable generalizability. COMPASS achieved 83.7% accuracy in predicting responses for stomach adenocarcinoma. This was a cancer type the model had not seen during training [Figure 3a]. It also successfully transferred knowledge across different drug targets. It achieved 76.1% accuracy for anti-CTLA4 treatments after being trained primarily on anti-PD1/PD-L1 data [Figure 3c].

Beyond mere prediction, the model provided a window into why certain patients fail treatment. In a survival analysis of a Phase II trial for urothelial cancer, the authors found that COMPASS-identified responders lived significantly longer. They showed a hazard ratio of 4.7 compared to non-responders [Figure 6a]. This means the predicted responders faced nearly five times lower risk of death.

By analyzing "concept scores," researchers identified distinct resistance mechanisms. Some patients had "inflamed" tumors (meaning they had plenty of immune cells) but still failed to respond. These non-responders were often driven by specific mechanisms. These included TGF-$\beta$ signaling or vascular exclusion. In these cases, the tumor's environment physically or chemically blocks immune cells from working [Figure 6f].

From prediction to personalized medicine

If COMPASS can generalize across clinical data, the implications for oncology are profound. First, it could streamline early-phase clinical trials. Researchers can use multi-stage fine-tuning to adapt the model. They can take a model trained on broad cancer data and refine it for a specific drug or a rare cancer type .

This helps select the right patients for enrollment more accurately.

Second, the model offers a path toward "reverse translation." Clinicians could use personalized response maps to see why a drug failed . This could point directly to the next logical step. For example, a clinician might choose a combination therapy that blocks the TGF-$\beta$ pathway alongside the primary immunotherapy.

The authors acknowledge a key limitation. The model relies on bulk RNA-seq, which averages signals from all cells in a biopsy. This might miss subtle signals from rare immune cell populations. A logical next step would be to integrate COMPASS with spatial transcriptomics. This technique maps gene activity to specific locations within the tumor. Such data could confirm if the "exclusion" mechanisms identified by the model match the physical distance between immune cells and tumor cells.

Figures from the paper