Interpretable AI-driven pan-cancer pre-screening of actionable gene fusions from H&E slides: A prototype-guided mixture-of-experts framework to prioritize patients for RNA/DNA sequencing.

Abstract

e15011 Background: Actionable gene fusions such as ALK, ROS1, and NTRK define critical therapeutic subsets across solid tumors but suffer from the needle-in-a-haystack challenge due to extreme rarity in unselected populations. We hypothesized that those fusion-driven oncogenesis imprints subtle and detectable morphological patterns on routine H&E slides. We present a novel AI framework designed not only to predict fusion status but to enable a cost-effective pre-screening funnel. Methods: We developed PathMoE-Fusion, a prototype-guided mixture-of-experts framework to identify fusion-associated signals directly from routine hematoxylin and eosin (H&E)-stained whole-slide images. The mixture-of-experts architecture enables multiple specialized subnetworks (“experts”) to selectively capture distinct histomorphological patterns, which serve as representative and interpretable prototypes for profiling characteristic phenotypes. PathMoE-Fusion was trained and validated on a pan-cancer cohort of 6,273 patients across 25+ cancer types, including 1,553 fusion-positive cases confirmed by next generation sequencing assays in a CAP/CLIA-certified laboratory (including ALK: 884; RET: 307; ROS1: 255; and NTRK: 107). Model operating points were selected to support conservative pre-screening use, prioritizing high specificity and enrichment of fusion-positive cases. External validation was conducted on independent cohorts, including the TCGA Pan-Cancer Atlas, to assess generalizability across cancer types and domains. Results: PathMoE-Fusion, demonstrated robust pre-screening performance across multiple actionable gene fusion targets, achieving AUROC values ranging from 0.84 (ROS1) to 0.94 (NTRK). Crucially, when deployed as a pre-screening tool under a conservative operating point prioritizing high specificity, the model yielded positive predictive values of 80.2% for ALK and 70.2% for NTRK, indicating marked enrichment of fusion-positive cases relative to their baseline prevalence. In TCGA validation, PathMoE-Fusion showed improved pre-screening performance, with more favorable enrichment characteristics than baseline approaches. Conclusions: PathMoE-Fusion serves as a rapid, low-cost digital enrichment biomarker that transforms the search for rare gene fusions. By prioritizing specificity, it acts as a pre-screening filter to populate a high-yield queue for confirmatory DNA/RNA sequencing. This framework maximizes the cost-effectiveness of molecular diagnostics, ensuring that sequencing resources are concentrated on patients most likely to benefit from life-extending targeted therapies.