Multimodal Feature Fusion for Bone Toxicity Prediction and Local Platform

Drug-induced osteotoxicity refers to the detrimental effects of certain drugs on bone metabolism, density, and structure, posing serious safety concerns in clinical practice, drug development, and environmental health. Although previous studies have attempted to use machine learning methods to predict osteotoxicity, traditional approaches often struggle to capture the complex relationships between molecular structure and toxicity. To address this issue, we curate a dedicated bone toxicity data set and propose a novel multimodal predictive model, termed BTP-MFFGNN, which integrates molecular fingerprints with graph-based features. By designing a graph neural network specifically tailored to address the complex interactions in osteotoxicity, along with advanced attention mechanisms and adaptive gating fusion strategies, our model can precisely capture the nonlinear relationship between molecular structure and toxicity, revealing the intricate molecular interactions in depth. Experimental results demonstrate that BTP-MFFGNN achieves significant improvements in osteotoxicity prediction, with an ACC of 0.85 and an AUC of 0.92, representing 13 and 8% increases, respectively, compared to the best previous model. To enable practical application, we develop a local platform named OsteoToxPred (refer to the demo at https://pzj-123456.github.io/), which supports SMILES input and delivers rapid, visualized predictions. Our work provides an effective computational framework for bone toxicity assessment and offers valuable support for safer drug discovery and mechanism-driven toxicology research. All of the codes are freely available online at https://github.com/zhaoqi106/BTP-MFFGNN.