Morphological Landscape Mapping Decodes Pathological Heterogeneity and Proteomic Programs in HCC

Intratumour heterogeneity (ITH) drives the clinical trajectory of HCC, yet routine pathology relies on global classifications that often mask local architectural diversity. We developed an unsupervised artificial intelligence framework to deconvolute HCC histology into a comprehensive lexicon of morphotypes, link them to in situ proteomic programs, and derive spatial biomarkers for precision prognostication. We developed Morphological Landscape Mapping (MLM), a deep-learning framework integrating multi-scale clustering with optimal-transport similarity. Applied to 1,448 whole-slide images across four independent HCC cohorts, MLM distilled the tumour landscape into 16 reproducible morphological phenotypes associated with distinct prognostic outcomes. Deep in situ proteomics (>8,000 unique proteins) linked MLM-derived morphotypes to specific molecular programs. Notably, an aggressive loss-of-adhesion morphotype exhibited hypoxia signalling activation and focal-adhesion downregulation, consistent with its fragmented architecture, whereas favourable trabecular morphotypes retained xenobiotic metabolic machinery. Building on these phenotypes, we developed the Morphotypic Spatial Entropy Index (MSEI) to quantify architectural complexity within the tumour ecosystem. In multivariable Cox models, MSEI remained an independent predictor of survival in two large HCC cohorts (adjusted HR 2.57, 95% CI: 1.66-3.98; and adjusted HR 2.21, 95% CI 1.28-3.81) after accounting for tumor stage and vascular invasion, and it further stratified risk within early-stage (BCLC 0-A) disease. MLM transforms routine H&E slides into quantitative, molecularly characterized maps of HCC heterogeneity. By coupling morphology with underlying biology and spatial organization, this framework provides a scalable, cost-effective foundation for morphology-guided precision oncology.