Knowing in advance whether a patient will respond to immunotherapy could spare them months of ineffective treatment and serious side effects — yet current biomarkers like PD-L1 expression and tumor mutational burden remain frustratingly imprecise. A foundation model that generalizes across cancer types and drug classes represents a meaningful step toward individualized treatment selection.
The model, called COMPASS, was trained on bulk tumor transcriptomes — genome-wide RNA expression profiles obtained from standard tumor biopsies — and learns to predict immunotherapy response without being retrained for each cancer type or specific agent. Rather than relying on a single biomarker, it integrates the full transcriptomic landscape of a tumor, capturing complex gene expression patterns that correlate with immune engagement. The pan-cancer design is particularly notable: COMPASS demonstrated predictive performance across multiple cancer types and distinct immunotherapy regimens within a unified framework, published in Nature Medicine.
What makes this finding worth watching is the architectural choice to use a foundation model approach — pre-training on large, diverse data before fine-tuning — which mirrors advances in large language models but applied to oncogenomics. This positions COMPASS within a rapidly maturing class of biological AI tools, including similar transcriptome-based predictors explored for chemotherapy response. The practical implication is substantial: bulk RNA sequencing is already performed at many major cancer centers, meaning the computational layer could potentially be added without new biopsy infrastructure. However, critical limitations must be acknowledged. Validation cohort sizes, prospective versus retrospective design, and real-world clinical integration remain open questions from the excerpt alone. Foundation models in oncology also carry risks of batch effects and training data biases that can inflate apparent generalizability. This is a genuinely promising signal, but confirmatory prospective trials across diverse patient populations will be essential before COMPASS influences treatment decisions at the bedside.