FGMA: A Functional Group-enhanced Deep Learning Method for Drug Toxicity Prediction

Introduction: Accurate drug toxicity prediction is vital for drug discovery, but existing atom-based methods are often computationally inefficient for complex molecules. This limitation creates a need for more efficient yet powerful predictive models. We introduce FGMA, a novel framework designed to address this challenge. Methods: FGMA integrates functional groups with molecular fingerprints for a multi-view molecular encoding. The framework extracts key functional groups, treating them as unified entities to reduce computational overhead. These are then combined with MACCS fingerprints using a crossattention mechanism to capture both local chemical motifs and global structural properties. Results: Experimental results demonstrate that FGMA achieves high accuracy and robustness in toxicity prediction tasks. In addition, interpretability analyses successfully identified how specific structural features and functional groups within drug molecules contribute to toxicity outcomes. Discussion: Our findings validate that combining functional groups with fingerprints is an effective strategy that balances computational efficiency and representational power. The model's interpretability offers valuable, actionable insights into structure-toxicity relationships, aiding in the design of safer compounds. Conclusion: FGMA offers an effective, efficient, and interpretable tool for drug toxicity analysis. It stands to accelerate the drug discovery pipeline by enabling faster and more insightful safety assessments of potential drug candidates.