Survival-Relevant Directional Pathology–Omics Discordance from Frozen Whole-Slide Foundation Embeddings

KuanTing Wu

Keywords: computational pathology, multimodal foundation models, multi-omics, survival analysis, precision oncology

TL;DR: Directional pathology–omics discordance from frozen whole-slide foundation embeddings identifies survival-relevant residual biology beyond measured molecular assays in LUAD and TCGA-KIRC.

Abstract: Multimodal life-science foundation models are often evaluated by direct endpoint prediction: can a learned image representation recover a molecular or clinical label? We study a complementary use case for frozen whole-slide embeddings. After generating an out-of-fold morphology-derived molecular readout, we ask whether the direction of disagreement between the image-derived score and the measured molecular assay retains survival information. Using fixed TITAN–CONCH whole-slide embeddings, we define a signed image–molecular residual, adjust it only for outcome-free technical quality variables, and decompose it into positive and negative directional components. In an anonymized in-house resected NSCLC/LUAD cohort and in public TCGA-KIRC, locked directional discordance terms remain associated with recurrence-free or overall survival after adjustment for the observed molecular target and available clinicopathologic covariates. These retrospective findings do not imply that pathology can replace molecular testing. Instead, they support directional cross-modal residual analysis as a lightweight way to evaluate and interpret biomedical foundation embeddings.

Email Sharing: We authorize the sharing of all author emails with Program Chairs.

Data Release: We authorize the release of our submission and author names to the public in the event of acceptance.

Submission Number: 110