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

Area of Science:

• Computational chemistry and bioinformatics
• Pharmacology and drug safety
• Machine learning in healthcare

Background:

• Accurate drug combination risk assessment is vital for safe clinical practice.
• Existing methods often use binary classification, failing to distinguish risk levels and handle imbalanced data.
• Heterogeneous features in drug combination data present challenges for semantic alignment.

Purpose of the Study:

• To develop a robust method for predicting drug combination risk levels.
• To address limitations of existing binary classification methods, including data imbalance and feature alignment.
• To propose MSFCL (Multisource Feature Fusion and Contrastive Learning) for enhanced risk prediction.

Main Methods:

• Integrating molecular structural features (TrimNet) with graph convolutional networks for topological relationships.
• Fusing Morgan fingerprint similarity with prior constraints for feature robustness.
• Employing adaptive gradient-noise hybrid perturbation for contrastive learning on imbalanced data.
• Utilizing multihead attention, residual connections, label smoothing, and focal loss for feature alignment and objective sharpening.

Main Results:

• MSFCL significantly outperformed baseline methods across all evaluation metrics on three benchmark datasets.
• Achieved substantial improvements in accuracy (9.84%), macro-F1 (14.97%), macro-recall (11.91%), and macro-precision (12.94%) on the DDInter dataset.
• Demonstrated superior generalization capabilities in multiclass classification tasks on DrugBank and MDF-SA-DDI datasets.

Conclusions:

• MSFCL provides an effective solution for multiclass drug combination risk level prediction.
• The proposed method successfully addresses data imbalance and feature alignment issues.
• MSFCL offers a promising tool for guiding rational clinical medication and enhancing drug safety.