DeepADR: multimodal prediction of adverse drug reaction frequency by integrating early-stage drug discovery information via Kolmogorov-Arnold networks.

Adverse drug reactions (ADRs) are a major cause of clinical trial failure and postmarket withdrawal, posing significant risks to public health and impeding drug development. While computational methods offer an alternative to costly preclinical testing, existing models often fail with novel compounds by requiring pre-existing information such as drug-ADR associations or by inadequately integrating diverse data sources. Here, we introduce DeepADR, a multimodal deep learning framework for predicting both the occurrence and frequency of ADRs using early-stage, readily available data. DeepADR integrates chemical structures and biological target profiles with semantic representations of ADR terms derived from a large language model (LLMs). These heterogeneous parameters are fused using a Kolmogorov-Arnold Network (KAN), which enhances the modeling of complex, nonlinear relationships among modalities to improve predictive performance. Our model outperforms existing methods in predicting both ADR occurrence and frequency, demonstrating robust generalization to new chemical entities. DeepADR showed consistently better performance than other models across both classification and regression tasks. By effectively integrating chemical, biological, and semantic datasets, DeepADR provides a powerful, scalable tool for the early-stage safety assessment and candidate prioritization. This framework not only facilitates the prioritization of safer drug candidates but also offers a methodology for predicting the toxicity of other hazardous materials, holding significant promise for advancing public health.