sfMON: Predicting Drug Side Effect Frequencies Based on a Versatile Multi-objective Optimization Network

Introduction: Accurately predicting drug side effect (DSE) associations and their frequency profiles remains critical but unresolved in drug development. While computational models provide alternatives to costly experimental methods, current approaches have limitations: most support only binary predictions due to poor multitask design, and even advanced models struggle in cold-start scenarios for new drugs because they over-rely on historical data. This gap highlights the urgent need for computational tools that can rapidly identify DSE associations and frequencies early in drug development. Methods: We present sfMON, a multi-task framework that combines cross-attention and a multiobjective loss function to predict drug-side effect associations and frequencies. The model integrates molecular features, including GIN-based graph embeddings, Morgan fingerprints, and Mol2vec embeddings, via cross-attention, while side effects are encoded using BioBERT. Its multi-task architecture employs a multi-objective loss that combines binary cross-entropy (for association prediction) and mean squared error (for frequency estimation), enabling simultaneous optimization of both tasks. The complete dataset and source code for this study are available at https://github.com/yuliyi/sfMON. Results: Comparative benchmarking shows that sfMON outperforms state-of-the-art baselines, especially in cold-start scenarios. It achieves a statistically significant 4.3 percent reduction in root mean square error for drug side effect frequency estimation. Discussion: While sfMON improves drug side effect prediction, its strong reliance on regulatory sources dominated by common effects introduces a long-tailed bias, which limits accurate frequency prediction for rare side effects. Conclusion: sfMON’s effectiveness in dual-task prediction advances computational pharmacovigilance, offering substantial potential to accelerate drug repurposing and improve early-stage therapeutic safety.