Histopathology-guided multimodal pathomics identifies a PLEKHA6-positive epithelial state associated with phosphoinositide metabolism in esophageal cancer.

Esophageal cancer is biologically heterogeneous, and conventional clinicopathological staging incompletely captures variation in patient outcomes. This study aimed to determine whether histopathology-derived features associated with phosphoinositide metabolism could support exploratory prognostic stratification and identify biologically interpretable epithelial states. Diagnostic whole-slide images and bulk transcriptomic data from The Cancer Genome Atlas Esophageal Carcinoma (TCGA-ESCA) cohort were integrated with single-cell RNA-sequencing data. Tissue-rich image tiles underwent manual quality control, deep features were extracted using an ImageNet-pretrained ResNet-50 model, and phosphoinositide metabolism activity was quantified by single-sample gene set enrichment analysis (ssGSEA). Associated image features were evaluated using machine-learning survival models, followed by clinicopathological adjustment and multimodal molecular characterization. The selected random survival forest plus gradient boosting machine model stratified overall survival in the training and internal validation cohorts, with concordance indices of 0.762 and 0.723, respectively. The standardized pathology-derived risk score remained associated with overall survival after adjustment for age, sex, histological subtype, and pathological stage (hazard ratio, 2.37; 95% confidence interval, 1.60-3.50). However, calibration was imperfect, and performance estimates may be optimistic because the internal validation cohort informed model selection. Most initial pathology-phosphoinositide metabolism associations were attenuated after adjustment for histological subtype. Multimodal analyses prioritized PLEKHA6 as a candidate pathology-associated epithelial-state marker rather than a subtype-independent prognostic biomarker. PLEKHA6-positive epithelial cells exhibited inflammatory and microenvironment-related transcriptional programs, predicted midkine- and macrophage migration inhibitory factor-related communication features, and stronger inferred copy-number variation-like signals. Higher tumor-level PLEKHA6 expression was also associated with distinct inferred immune, metabolic, and predicted drug-response profiles, although malignant-cell identity was not established. Because the analyzed tiles were not derived from pathologist-annotated malignant regions, the image signal represents composite diagnostic-slide tissue context. These findings identify an exploratory pathology-derived prognostic signal and a PLEKHA6-associated epithelial state in esophageal cancer that require independent, subtype-specific, pathological, and experimental validation.