PathGeneLink: Constructing Evidence-Grounded Gene–Histomorphology Associations toward Morphology-Grounded Evaluation of Virtual Molecular Predictions

Virtual spatial-transcriptomics models that infer gene expression from H&E are typically evaluated using gene-wise correlation and related measures comparing predicted with measured expression. These metrics quantify molecular agreement, but they do not test whether a predicted molecular state is consistent with the H&E-visible morphology expected from prior biological knowledge. We introduce PathGeneLink, an evidence-grounded human-AI pipeline for constructing directional gene-expression-to-histomorphology relationships for morphology-grounded evaluation. PathGeneLink first generates candidate relationships between gene functional classes, expression states, and named histomorphology, grounds them in biomedical literature, verifies the supporting evidence, and subjects them to clinician review. Of 333 assumptions reviewed by clinicians across 33 morphology features, 256 were retained. PathGeneLink then links these reviewed assumptions to individual genes by checking whether the gene’s own evidence supports the biological function described in the assumption. We validate this gene-specific linking step against a blinded expert-clinician review of 379 gene-assumption pairs spanning 12 genes. PathGeneLink reached 88.9% agreement with the clinician, with precision 0.77 and recall 0.83 (Cohen’s kappa = 0.72, 95% CI 0.64-0.80). PathGeneLink outperformed matched LLM baselines on every metric, with the clinician-reviewed biological assumption driving most of the gain and gene evidence adding a further improvement. These results support PathGeneLink as a scalable pipeline for constructing evidence-grounded gene-histomorphology knowledge, with the resulting knowledge intended for future morphology-grounded evaluation of virtual molecular predictions.