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

Area of Science:

• Pharmacology
• Artificial Intelligence
• Computational Chemistry

Background:

• Current deep learning models for drug discovery are often single-task and single-modality, limiting their predictive power.
• These limitations overlook crucial factors in central nervous system (CNS) drug discovery, including blood-brain barrier (BBB) penetration and potential neurotoxicity.

Purpose of the Study:

• To develop a multimodal and multitask deep learning framework for simultaneous prediction of drug-target affinity (DTA), BBB permeability, and neurotoxicity.
• To enhance the representational capacity and cross-modal feature integration in CNS drug discovery models.

Main Methods:

• Proposed a multimodal and multitask learning framework (NeuMTL) incorporating mutual attention and attention pooling modules.
• Utilized early and late fusion strategies for improved interpretability and feature integration.
• Introduced NeuGradBalancer, a novel optimization strategy to manage gradient conflicts and ensure balanced task learning.

Main Results:

• NeuMTL achieved superior performance with low Mean Squared Errors (MSEs) for DTA prediction (0.124, 0.112, 0.412).
• Demonstrated high classification accuracies for BBB permeability and neurotoxicity prediction (0.912, 0.961, 0.972, 0.929).
• The model showed strong interpretability and outperformed existing methods on multisource datasets.

Conclusions:

• NeuMTL effectively enhances the safety and efficacy assessment of CNS drug candidates.
• The framework accelerates CNS drug discovery, as evidenced by its successful application in identifying potential autism therapeutics like AdipoRon.