Pathomic biomarkers derived from whole-slide histopathology predict hepatocellular carcinoma progression and enable GPC3-linked tumour microenvironment stratification in a preclinical experimental oncology model.

UNLABELLED: Hepatocellular carcinoma (HCC) is the third leading cause of cancer-related mortality worldwide, with five-year survival below 20% for advanced-stage disease despite advances in experimental therapeutics including immune checkpoint inhibitors and GPC3-targeted agents. Objective, reproducible tools for quantifying tumour microenvironment (TME) architecture and linking tissue features to pharmacodynamically actionable biomarkers are a critical unmet need in preclinical oncology.

METHODS

We developed a fully automated computational pathology pipeline in R (v4.4.2) integrating Macenko stain normalisation, Otsu tissue segmentation, K-means clustering (k = 3) into necrotic/inflamed/healthy tissue classes, multi-scale feature extraction, and a random forest progression classifier. The pipeline was applied to whole-slide H&E images from a diethylnitrosamine (DEN)-induced mouse HCC model (n = 20) and prospectively validated against expert pathologist annotations. GPC3 immunohistochemistry served as independent biological validation of the computational necrotic fraction as a pharmacodynamic readout.

RESULTS

Tissue segmentation achieved pixel accuracy of 91.3% (Cohen's κ = 0.87; Dice = 0.893 ± 0.031). K-means cluster assignments agreed with expert scoring at κ = 0.83. Necrotic fraction increased from 8.4 ± 2.1% (controls) to 34.7 ± 6.3% (advanced HCC; p < 0.001) and correlated with serum GPC3 (Pearson r = 0.79, p < 0.001), independently validated by GPC3 IHC (ρ = 0.82, p < 0.001). The six-feature random forest classifier achieved AUC = 0.89 (sensitivity 84.6%, specificity 90.0%), outperforming single-feature morphometry (AUC = 0.71) and unaided expert assessment (AUC = 0.76). Tissue complexity score correlated with histopathological grade (ρ = 0.81, p < 0.001).

CONCLUSIONS

This validated, explainable pipeline delivers pathomic biomarkers for HCC staging, TME characterisation, and GPC3-correlated tumour burden quantification, providing a GPU-free, reproducible platform for preclinical evaluation of experimental therapeutics, immunotherapy stratification, and pharmacodynamic monitoring of GPC3-targeted agents.