Towards Mechanistic Biomarkers Discovery

Computational biomarkers derived directly from high-dimensional omics data often show limited reproducibility, interpretability, and generalizability. This study evaluates a mechanism-based strategy that integrates curated signaling models with gene-expression data through extreme currents (ECs), minimal steady-state subpathways derived from stoichiometric network models. BioModels models were converted to ECs using PoCaB, ECs were mapped to ENTREZ gene identifiers, and expression of genes within each EC was summarized by the first principal component to generate pathway-informed quantitative features. Features across models were combined and evaluated using Elastic Net, Sparse Group Lasso, gradient boosting, pathway-model boosting, and stacking. Predictive performance and feature-selection stability were assessed using repeated 10 × 10 cross-validation in a breast cancer survival dataset and a prostate cancer case-control dataset. In breast cancer, EC-based methods achieved predictive performance similar to conventional gene- and pathway-based representations. The key advantage was therefore not higher accuracy but greater mechanistic interpretability: each EC feature remains linked to a defined steady-state subpathway in a curated signaling model. Gradient boosting and pathway-model boosting selected features more consistently but produced less sparse models, whereas Sparse Group Lasso and stacking yielded smaller signatures with lower selection stability. In prostate cancer, classification was near-perfect across methods, again making interpretability, stability, and sparsity more informative than marginal differences in accuracy. These results support EC features as a biologically structured and more mechanistically interpretable representation that can preserve predictive performance comparable to conventional alternatives while exposing a practical trade-off between sparse signatures and stable feature selection.