BioMarkAdapt: A Dynamic, Continual, and Interpretable Framework for Adaptive Biomarker Discovery

The dynamic and multifaceted nature of diseases necessitates flexible and interpretable computational models for biomarker discovery. However, existing methods predominantly rely on static prediction paradigms, fail to adapt continuously to new tasks without forgetting prior knowledge, and lack transparent, task-aware explanations. To address these limitations, we introduce BioMarkAdapt, a novel framework for interpretable and continually adaptive biomarker discovery. Our framework is built upon three core innovations: (1) Dynamic Task-Aware Prediction, which uses a Task-Aware Modulation mechanism to specialize model reasoning for distinct clinical contexts; (2) Knowledge-Anchored Continual Learning, which leverages Gene Ontology to consolidate fundamental biological knowledge and mitigate catastrophic forgetting; and (3) Interpretable Evidence Tracing, which provides task-specific, traceable explanations linking predictions to relevant biological pathways. Extensive experiments on four benchmark datasets (MPO, HPO, GWAS, and CAFA2 wPPI) demonstrate that BioMarkAdapt achieves state-of-the-art performance, significantly outperforming prior methods (e.g., +1.47 Fmax and +2.03 AUC on MPO). Ablation studies confirm the contribution of each component, while sequential learning evaluations demonstrate effective knowledge retention (e.g., 92.3% of performance retained). Furthermore, BioMarkAdapt delivers biologically plausible explanations, with evidence weights exhibiting a Spearman correlation of 0.712 with ground-truth associations. Our work provides a robust, adaptable, and trustworthy framework for advancing precision medicine.