ESOL Reimplementation and Extension: Aqueous Solubility Prediction with Linear and Tree-Based Models

Abstract

This software release provides a reproducible reimplementation and extension of the ESOL aqueous-solubility prediction method. It compares the published ESOL equation, refitted linear models, Extra Trees, and histogram gradient boosting using AqSolDB and curated OChem data. The combined benchmark contains 10,000 retained measurements representing 9,560 molecular identity groups. Models use either the four original ESOL descriptors—calculated logP, molecular weight, rotatable-bond count, and aromatic proportion—or an expanded set of ten descriptors. Evaluation uses five matched molecular-group train/test splits, with nonlinear hyperparameter selection performed exclusively within grouped training folds. Extra Trees achieved the lowest mean test RMSE of 0.9852 log₁₀(mol/L), compared with 1.3376 for the ten-descriptor linear model. These results describe exploratory grouped random-holdout performance, not independent external validation. RDKit descriptors approximate the original Daylight implementation, and measurement conditions are not fully harmonized across the source datasets. The archive includes Python scripts, pinned dependencies, checksum-verified dataset download instructions, selected benchmark results, split record identifiers, provenance, and verification documentation. Raw source databases are not redistributed; full predictions and fitted models can be regenerated using the supplied workflow. Creator: Ngoc Bach Hoang. The creator’s original code, documentation, and research outputs are licensed under CC BY 4.0. Third-party rights remain unchanged. AI assistance is documented in the archive. This release has not undergone peer review. References: Original ESOL method: https://doi.org/10.1021/ci034243x AqSolDB: https://doi.org/10.1038/s41597-019-0151-1 Curated OChem study: https://doi.org/10.1038/s41597-024-03105-6 OChem data repository: https://doi.org/10.57745/CZVZIA