Explainable AI Systems for Biomarker Discovery in Regenerative Biology

Abstract

Biomarker discovery in regenerative biology -- the identification of molecular signals that reliably indicate tissue regeneration capacity, therapeutic response, or regenerative failure -- is a critical enabling step for personalised regenerative medicine, enabling patient stratification, treatment monitoring, and outcome prediction. Machine learning models applied to multi-omics data (transcriptomics, proteomics, metabolomics, epigenomics) have demonstrated substantial power in identifying biomarker candidates that outperform single-molecule clinical markers in predicting regenerative outcomes. However, the clinical translation of ML-derived biomarker panels requires not only predictive accuracy but biological interpretability: clinicians and regulatory bodies require understanding of the biological mechanisms underlying biomarker associations, and unexplained black-box predictions are insufficient for clinical decision support in regenerative medicine. This paper proposes the Explainable Regenerative Biomarker AI (ERBA) framework, a methodology for biologically interpretable ML-based biomarker discovery in regenerative biology, integrating four explainability- centred components: a multi-omics biomarker selector (MOBS) that identifies predictive biomarker panels from integrated multi-omics data; a biological pathway explainer (BPE) that maps discovered biomarkers to regeneration-relevant biological pathways; a patient stratification engine (PSE) that clusters patients by regeneration capacity based on biomarker profiles; and a biomarker validation ranker (BVR) that prioritises candidates for experimental validation based on both predictive power and biological plausibility. ERBA is applied to three regenerative biology datasets: skeletal muscle regeneration post-injury (n = 284 patients), liver fibrosis regression (n = 312 patients), and cartilage repair post-ACI therapy (n = 196 patients). ERBA achieves biomarker panel AUC-ROC = 0.924 (SD = 0.028) for regenerative outcome prediction -- substantially outperforming single-omics ML baselines (0.798, SD = 0.038) -- while providing pathway-level biological explanations for 94.2% of top-ranked biomarkers. Twelve novel biomarker candidates identified by ERBA are validated in independent cohorts (n = 3 independent validation studies). The study contributes the ERBA specification, the RegBiomark benchmark, and twelve experimentally validated novel regenerative biomarker candidates.