Multimodal Feature Fusion for Bone Toxicity Prediction and Local Platform

Area of Science:

• Computational toxicology
• Drug safety
• Bone metabolism

Background:

• Drug-induced osteotoxicity poses significant risks in clinical practice and drug development.
• Existing machine learning models struggle to capture complex structure-toxicity relationships.

Purpose of the Study:

• To develop a novel multimodal predictive model for accurate osteotoxicity prediction.
• To address limitations of traditional methods in capturing molecular structure-toxicity interactions.

Main Methods:

• Curated a dedicated bone toxicity dataset.
• Proposed BTP-MFFGNN, a multimodal model integrating molecular fingerprints and graph-based features.
• Employed a tailored graph neural network with attention mechanisms and adaptive gating fusion.

Main Results:

• BTP-MFFGNN achieved significant improvements in osteotoxicity prediction.
• Achieved an accuracy (ACC) of 0.85 and an area under the curve (AUC) of 0.92.
• Demonstrated 13% and 8% increases in ACC and AUC, respectively, over previous best models.

Conclusions:

• BTP-MFFGNN provides an effective computational framework for bone toxicity assessment.
• The model offers valuable support for safer drug discovery and mechanism-driven toxicology.
• An accessible platform, OsteoToxPred, was developed for practical application.