MSFCL: Drug Combination Risk Level Prediction Based on Multi-Source Feature Fusion and Contrastive Learning

Abstract

Accurate assessment of drug combination risk levels is crucial for guiding rational clinical medication and avoiding adverse reactions. However, most existing methods are limited to binary classification, which fails to quantify distinctions between risk levels and struggles with imbalanced data distribution and insufficient semantic alignment of heterogeneous features. To address these challenges, we propose MSFCL, a drug combination risk level prediction based on multisource feature fusion and contrastive learning. MSFCL integrates molecular structural features extracted by TrimNet with high-order topological relationships captured via a graph convolutional network. To enhance feature robustness, we fuse Morgan fingerprint similarity matrices with identity matrix-based prior constraints. To tackle data imbalance issues, we design an adaptive gradient-noise hybrid perturbation strategy to dynamically balance gradient direction guidance and Gaussian noise injection, enabling contrastive learning without requiring data augmentation. In addition, we implement multihead attention mechanisms and residual connections to improve multisource feature alignment while label smoothing and focal loss functions sharpen the training objectives. Extensive experiments on three benchmark data sets demonstrated that MSFCL outperformed baseline methods across all evaluation metrics. Specifically, on the DDInter data set, MSFCL achieved an average improvement of 9.84% in accuracy, 14.97% in macro-F1, 11.91% in macro-recall, and 12.94% in macro-precision. MSFCL also demonstrated superior generalization in multiclass classification tasks on the DrugBank and MDF-SA-DDI data sets.