Feature-Adaptive Meta-Learning for Few-Shot Drug Synergy Prediction

Drug combination therapy plays an important role in personalized cancer treatment by improving therapeutic efficacy and reducing drug resistance through synergistic interactions. However, existing drug synergy prediction methods usually rely on sufficient labeled samples and struggle to capture cell-line-specific response patterns under sample-scarce conditions. To address this problem, this paper proposes FMBA-MetaSyn, a few-shot drug synergy prediction method based on feature-adaptive modulation and BOIL-inspired adaptation. FMBA-MetaSyn first extracts drug structural features from molecular graphs and biological features from cell-line gene expression profiles, and then integrates them into drug-drug-cell representations. A task-level feature modulation module is further designed to use limited labeled samples to adapt sample representations consistently. In addition, a lightweight high-level residual Adapter is updated in the inner loop while keeping the basic feature extractors stable, reducing overfitting in few-shot scenarios. Experimental results show that FMBA-MetaSyn achieves competitive performance in both synergy score regression and synergistic drug combination identification.