FerrGAT: Multi-Task Pre-Training of Graph Attention Networks for Low-Data Molecular Activity Prediction

Predicting molecular bioactivity in low-data regimes remains a central challenge in computational drug discovery, where labeled compounds for specialized tasks are scarce while related datasets are abundant. Here, we propose FerrGAT, a graph attention network framework that addresses this challenge through domain-relevant multi-task pre-training and dual-channel molecular representation learning. FerrGAT first pre-trains a shared GAT encoder on three mechanistically related tasks—antioxidant activity (GST inhibition, 245 compounds from ChEMBL target CHEMBL2095173; 83 active/162 inactive), kinase inhibition (3000 compounds spanning AXL, EGFR and VEGFR2 from ChEMBL; 2240 active/760 inactive), and cellular toxicity (7265 compounds from the Tox21 NR-AhR endpoint; 309 active/6956 inactive)—then transfers the learned representations to a target task via differential learning rate fine-tuning. The architecture fuses atom-level graph features from multi-head attention message passing with global physicochemical descriptors through a learned projection and provides built-in interpretability via attention weight visualization at the atomic level. We evaluated FerrGAT on ferroptosis inhibitor prediction as a representative low-data molecular classification task (1052 compounds from ChEMBL targets GPX4 and HMOX1 combined with 63 literature- and FerrDb-curated ferroptosis modulators; 409 active/643 inactive). In 5-fold cross-validation, FerrGAT achieved an AUC of 0.906, outperforming Morgan fingerprint baselines, including Random Forest (0.877), XGBoost (0.871), and an SVM (0.873). Ablation studies confirmed that domain-relevant pre-training improved AUC by 2.5% over training from scratch, while pre-training on unrelated tasks degraded performance, highlighting the importance of task-domain alignment. Applied to virtual screening of 30 FDA-approved tyrosine kinase inhibitors, the model identified Bemcentinib (AXL inhibitor, score = 0.918) as a top candidate, validated by AutoDock Vina molecular docking (−8.51 kcal/mol) and independent experimental evidence, including lipid peroxidation assays, Western blot, and cellular thermal shift analysis. These results demonstrate that domain-aware transfer learning with graph attention networks provides an effective and interpretable framework for molecular property prediction in data-limited scenarios.