Multi-Seed Ensemble of Graph Attention Networks and Gradient Boosting for pIC50 Prediction: An Honest Out-of-Fold Benchmark on ChEMBL Data
Summary
No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.
Summarise with VIRAT Free to try with your Science Catalogs account.Abstract
Accurate prediction of compound potency (pIC50) is critical for prioritizing drug candidates in early-stage discovery.Wepresentamulti-seedensemblecombiningGraphAttentionNetworks(GAT),XGBoost, LightGBM, and CatBoost, validated on a curated subset of 828 compounds from ChEMBL v33. Unlike prior work reporting metrics on non-independent validation sets, all results here are computed strictly out-of-fold (OOF) across 5-fold cross-validation with 3 random seeds (15 total evaluations per model). The ensemble achievesR2=0.738,RMSE=0.597,andMAE=0.472 onclippedpIC50values[−0.5,4.0].We demonstratethatpreviouslypublishedbenchmarksclaimingR2=0.86onthissamedatasetarenot reproducible under rigorous OOF evaluation, and provide transparent analysis of model limitations. All code, data, and predictions are publicly available.
Fields of Study & Subjects
Computational Drug Discovery Methods
Machine Learning in Materials Science
Machine Learning in Healthcare
Publication details
- Journal / source
- ChemRxiv
- Publication date
- 1 April 2026
- Language
- EN
- Document type
- Preprint