Multimodal foundation models in colorectal cancer: from prediction to trustworthy clinical insight.

Abstract

Colorectal cancer (CRC) is characterized by profound, multi-layered heterogeneity that limits the precision of conventional single-modality clinical tools. The emergence of multimodal foundation models (MFMs) represents a conceptual paradigm shift, moving beyond static biomarkers to capture the dynamic and evolving nature of CRC. MFMs integrate histopathology, radiology, multi-omics data (including the critical regulatory layer of epigenomics), and clinical variables into shared high-dimensional representational spaces. This integration enables improved prognostication, refined molecular subtyping, and in silico simulation of therapeutic perturbations within the tumor's functional landscape, thereby supporting rational and model-driven drug development. In this review, we synthesize the rapidly expanding body of CRC-specific MFM research and critically examine the unresolved challenges that currently limit clinical translation. We place particular emphasis on the transition from correlation to causal inference, the establishment of cross-population generalizability, and the resolution of key issues related to trustworthiness and clinical interpretability. Finally, we propose an actionable roadmap outlining regulatory, data governance, and translational requirements, including the lab-in-the-loop paradigm, necessary to position MFMs as a robust and equitable framework in clinical oncology.