AI Model “COMPASS” Predicts Immunotherapy Response in Cancer Patients

Researchers have developed COMPASS, a computational model designed to predict how patients with solid tumors will respond to immune checkpoint inhibitor (ICI) therapy. By analyzing transcriptomic data, the model maps tumor microenvironment features into a 44-dimensional space, providing a framework to identify which patients are most likely to benefit from immunotherapy. The findings, published in Nature, demonstrate that this interpretable, biology-grounded approach can outperform traditional biomarkers in specific clinical cohorts.

How does the COMPASS model function?

COMPASS operates using a three-tier architecture that mimics language modeling to interpret gene expression. According to the study, the model first uses a transformer-based gene language model (GLM) to encode individual gene activity. This data is then processed through a hierarchical projector that aggregates genes into functional concepts, such as specific immune cell types and signaling pathways. Finally, a classifier uses these concept representations to predict the likelihood of a patient’s response to treatment.

Unlike traditional “black-box” AI, COMPASS uses a “concept bottleneck” design. This means the model does not jump directly from raw genetic data to a prediction. Instead, it creates an intermediate layer of human-interpretable biological concepts. This transparency allows clinicians to see exactly which biological factors—such as the presence of cytotoxic T cells or specific immune checkpoints—are driving the model’s prediction for a particular patient.

What makes COMPASS different from existing biomarkers?

Standard biomarkers, such as Tumor Mutational Burden (TMB) or PD-L1 expression levels, often fail to capture the full complexity of the tumor microenvironment. While these metrics provide valuable snapshots, they frequently struggle with predictive accuracy across different cancer types and therapy regimens. COMPASS aims to address this by integrating pan-cancer data, allowing it to learn generalizable features of immune response that transcend individual tumor sites.

In benchmarking tests against 22 existing methods—including TIDE (Tumor Immune Dysfunction and Exclusion) and various immune signatures—COMPASS demonstrated competitive performance. By leveraging both parametric and non-parametric classifiers, the model can adapt to datasets of varying sizes, making it more robust in clinical settings where training data may be limited.

How is the model validated for clinical use?

To ensure the model’s reliability, the research team employed rigorous cross-cohort validation strategies. They utilized “leave-one-cohort-out” validation, where the model was trained on a broad range of immunotherapy-treated cohorts and then tested on an entirely held-out dataset. This approach is essential for assessing how well a model performs in real-world clinical environments where patient populations differ significantly from those in the initial training study.

Furthermore, the researchers generated “personalized response maps” for patients in the IMvigor210 cohort, which focuses on bladder cancer. These maps visualize how molecular features propagate through the model’s layers, offering a granular view of the biological reasoning behind a specific patient’s predicted outcome. This level of interpretability is a significant departure from standard genomic scoring, which often provides a single risk score without explaining the underlying biological drivers.

What are the limitations and future prospects?

While the results are promising, COMPASS is currently a computational tool based on retrospective analysis of deidentified datasets. As with all predictive models, its clinical utility depends on prospective validation in live clinical trials. The researchers noted that while the model handles cross-cancer heterogeneity well, its accuracy can be influenced by the quality and size of the input transcriptomic data.

Looking ahead, the integration of such models into clinical practice could help reduce the number of patients who undergo ineffective immunotherapy. By identifying non-responders early, clinicians may be better equipped to steer patients toward alternative treatments or combination therapies, potentially improving long-term survival outcomes. The code for the model and its processing pipeline are publicly available, allowing the broader research community to test and refine the tool for future oncology applications.

Key Takeaways

• Interpretable AI: COMPASS uses a concept-based architecture, allowing researchers to trace predictions back to specific biological markers like immune cell infiltration.
• Multi-modal Learning: The model was pretrained on 10,184 tumor samples from the TCGA dataset, enabling it to learn a robust, pan-cancer understanding of the tumor microenvironment.
• Improved Accuracy: By accounting for both gene identity and expression abundance, the model provides a more nuanced view of tumor biology compared to simple gene-signature scoring.
• Clinical Flexibility: The model supports multiple fine-tuning modes, ranging from zero-shot inference (NFT) for small datasets to full model adaptation (FFT) for larger clinical cohorts.