HerbGNN: A Directed Message Passing Neural Network Benchmark for Early Enrichment in Herbicide Virtual Screening under Extreme Class Imbalance

Virtual screening in agrochemical research presents distinct cheminformatics challenges that remain underrepresented in benchmarking studies, most notably the extreme class imbalance inherent to curated herbicide datasets and the lack of systematic comparisons between modern deep learning architectures and classical machine learning baselines under realistic screening conditions. Here, we introduce HerbGNN, a directed message passing neural network (D-MPNN) framework for herbicide activity prediction, and present a comprehensive benchmark against Random Forest and XGBoost. The models were evaluated on a dataset comprising 363 annotated herbicidal actives from the HRAC 2026 classification and 50,000 structurally diverse background molecules from ChEMBL, resulting in an approximate class imbalance ratio of 1:140. We show that conventional discrimination metrics such as AUROC can provide misleading estimates of model performance under such conditions and argue that the area under the precision-recall curve (AUPRC) and the enrichment factor at 1% (EF1%) offer more informative evaluation criteria for early enrichment tasks. To mitigate scaffold bias and data leakage, all models were assessed using scaffold-stratified cross-validation based on Bemis-Murcko decomposition with bootstrap confidence intervals. HerbGNN achieved a mean AUROC of 0.956 ± 0.05, a mean AUPRC of 0.573 ± 0.19, and a mean EF1% of 66.6, corresponding to a 41% improvement in early enrichment over Random Forest and an approximately 80-fold enrichment relative to random selection. Applicability domain analysis and atomic importance mapping further indicate that the model captures chemically meaningful patterns associated with herbicidal activity.