Systematic Evaluation of Graph Neural Networks for Ligand-Based Virtual Screening on ChEMBL Datasets

The performance of target-specific, ligand-based virtual screening models is strongly influenced by dataset characteristics, including data availability, class imbalance, and evaluation strategies. In this work, we perform a systematic evaluation of graph neural networks (GNNs) using a ChEMBL-derived dataset spanning 5,368 targets and 1.59 million activity records, capturing the long-tailed distributions and target-specific imbalances commonly observed in pharmaceutical data. Through a systematic evaluation of multiple GNN architectures, we identify guidelines for model selection: while the Graph Isomorphism Network (GIN) consistently outperforms others on datasets with >100 samples (a mean ROC-AUC up to 0.94), simpler architectures are more robust under extreme data scarcity. Critically, our comparative analysis of splitting strategies reveals that random sampling yields artificially optimistic performance due to structural overlaps, whereas similarity-aware clustering exposes a substantial generalization gap (AUC drop > 0.3), cautioning against prevailing evaluation practices. We further demonstrate that multi-task learning serves as an effective remedy for small-target instability, providing significant and consistent performance gains. To underscore its translational value, we deploy this comprehensive framework in a virtual screening campaign against Mcl-1, yielding a chemically optimized lead, C4 (Ki = 0.58 μM), with verified cellular efficacy. Our findings highlight the importance of task-aware benchmark design and offer a practical strategy for reliable GNN application in drug discovery.