Prediction of metastatic potential of heterogeneous pancreatic ductal adenocarcinoma through gradient-based algorithms

Abstract

Background

Pancreatic ductal adenocarcinoma carries a dismal prognosis, with five-year survival below 10 % due to late presentation, aggressive nature, and profound intratumor heterogeneity. Existing prognostic models fail to account for the dynamic interplay between subclonal evolution and the tumor microenvironment, limiting their clinical utility for risk stratification and therapy guidance.

Methods

An in silico, multiscale framework was developed, combining an agent-based model of subclone proliferation, migration, quiescence, and apoptosis with continuum reaction–diffusion equations for oxygen, nutrients, extracellular matrix, and chemoattractants. Four clinically relevant microenvironmental scenarios were simulated to generate spatiogenetic signatures of subclonal adaptation. These signatures were then sequentially input into three gradient-boosting classifiers—CatBoost, LightGBM, and XGBoost—to predict each subclone's metastatic potential.

Results

All hybrid pipelines demonstrated robust discrimination of high-risk subclones, with XGBoost achieving the highest sensitivity (92 %) and specificity (89 %) in cross-validation. The model accurately recapitulated PDAC's known clinical features, such as hypoxia-driven invasive fronts and desmoplastic stroma–associated resistance. Importantly, it uncovered novel biomarker signatures—combinations of genetic mutations and microenvironmental factors—that correlated with early metastatic seeding in simulated cohorts.

Conclusion

This work introduces a novel multiscale hybrid framework that is original both in its formulation integrating agent-based, continuum, and immune dynamics and in its application, where simulated biological signatures are used to train gradient boosting models for accurate prediction of metastatic potential.