Introduction
Deep learning-based models for anti-cancer drug response prediction typically face limitations of interpretability and dynamic learning in precision oncology and practical biomedical applications with ever-evolving data. Although existing methods employ different strategies to retain knowledge from previous tasks, they struggle within the vast space of drug and genomic data due to severe distribution shifts. Furthermore, privacy concerns in biomedical fields limit the applicability of data buffer-based approaches. We introduce the interpretable multi-contextual self-alignment (MCSA) framework composed of four cooperating modules: continual adversarial distillation (CAD), pluggable pharmacogenomic interpretable (PGI), stable interpretability regularizer (SIR), and prototype-aware interaction (PAI). Driven by designed multi-domain adversarial attack strategies, the CAD module captures contextual invariance in latent spaces by aligning representation contexts across tasks. The SIR leverages domain-prototype matching and feature constraints to align representations for interpretability and capture contextual invariance, while the PGI module identifies potential biomarkers via selective attention mechanisms. The PAI module simulates the complementary learning systems of human memory to enhance model stability. By progressively operating its modules, MCSA learns hierarchical context invariance and performs context-specific selfdistillation alignment. Experiments demonstrate that it outperforms SOTA methods, including a recent Nature approach, providing a unified interpretable framework for CDRP with superior resistance to forgetting and reliable interpretable insight.
Figure 1. The model architecture of MCSA, which is used for drug response prediction and precision oncology.