Deep Learning in Structure-Activity Relationship Modeling

Abstract

Structure-activity relationship (SAR) modelling is a cornerstone of computational drug discovery, enabling the prediction of biological activity from molecular structure and accelerating lead optimisation. Deep learning architectures - particularly graph neural networks (GNNs), molecular transformers, and convolutional neural networks operating on molecular images -- have emerged as powerful alternatives to classical QSAR methods, capturing complex non-linear structure-activity relationships without manual feature engineering. This study systematically benchmarked five deep learning architectures -- Directed Message Passing Neural Network (D-MPNN/Chemprop), Graph Attention Network (GAT), Molecular Transformer (BERT-based, ChemBERTa), DeepChem CNN (ECFP input), and a multi-task Graph Isomorphism Network (GIN) -- against three classical baselines (Random Forest/ECFP4, XGBoost/ECFP4, Support Vector Machine) across eight diverse MoleculeNet benchmark datasets spanning toxicity prediction (Tox21, ClinTox), ADMET profiling (BBBP, BACE, HIV), and activity classification (SIDER, MUV, FreeSolv). On the primary benchmark (Tox21, 12 tasks, scaffold split), D-MPNN achieved the highest mean ROC-AUC (0.854 +- 0.018), outperforming Random Forest (0.741 +- 0.024) by 15.2% and ChemBERTa (0.831 +- 0.021) by 2.8%. GIN multi-task learning provided the greatest benefit for low-data endpoints (MUV dataset), improving mean ROC-AUC by 0.142 over single-task GNN. Transfer learning from ChEMBL pre-training improved ChemBERTa performance on all eight datasets (mean delta-AUC +0.038 +- 0.011). These results provide an evidence-based architecture selection framework for SAR modelling in pharmaceutical research and identify the conditions under which deep learning most substantially outperforms classical QSAR approaches.