NeuMTL: A Unified Multimodal Framework for Multi-Task Prediction in CNS Drug Discovery

Deep learning has made significant progress in drug discovery. However, most existing models are single-task and single-modality, which not only limits their representational capacity but also tends to overlook critical factors in central nervous system (CNS) drug discovery, such as blood-brain barrier (BBB) permeability and neurotoxicity. To overcome the limitations of single-task models, we propose a multimodal and multitask learning framework that simultaneously predicts drug-target affinity (DTA), BBB permeability, and neurotoxicity. The framework employs mutual attention and attention pooling modules, along with early and late fusion strategies, to enhance interpretability and cross-modal feature integration. Moreover, a novel optimization strategy, NeuGradBalancer, mitigates gradient conflicts and ensures balanced learning across tasks. NeuMTL demonstrates superior overall performance, achieving low MSEs (0.124, 0.112, 0.412) in DTA prediction and high accuracies (0.912, 0.961, 0.972, 0.929) in BBB and neurotoxicity classification tasks, indicating its strong capability to enhance CNS drug screening safety and efficacy. Comprehensive experiments on multisource data sets demonstrate the superior performance and strong interpretability of the proposed NeuMTL model. Moreover, its application in autism drug screening identifies promising candidates like AdipoRon, underscoring its potential to accelerate CNS drug discovery.