Enhancing Graph Neural Networks for Multi-target Activity Prediction Through Multi-task Learning and Knowledge Distillation

• Arthur Cerveira UFPel
• Frederico Kremer UFPel
• Gabriel A. Gomes UFPel
• Ulisses B. Corrêa UFPel

Resumo

The early phases of the drug discovery pipeline increasingly rely on machine learning to identify active compounds and prioritize candidates for experimental testing. Quantitative structure-activity relationship (QSAR) models are widely employed to predict molecular activity for biological targets. While traditional QSAR models are often trained independently for each target, recent studies have shown that multi-task learning can improve predictive performance by exploiting shared patterns across related targets. In this work, we investigate the application of graph-based neural networks, specifically message-passing neural networks (MPNNs), in multi-task molecular activity prediction. Our methodology incorporates (1) knowledge distillation from pre-trained target-specific models to enrich multi-target activity datasets and (2) hierarchical clustering to group related tasks, allowing the model to better exploit inter-target similarities. We assess multi-task MPNNs performance relative to single-task learning, classical machine learning baselines, and feed-forward neural networks trained on molecular fingerprints. Our evaluation includes standard random splits as well as realistic lead optimization and hit identification benchmark splits, which better reflect practical drug discovery scenarios. Our results highlight the strengths and limitations of graph-based architectures in multi-task settings and provide insights into their practical utility for improving early-phase drug discovery workflows.