Chemically Informed and Leakage-Aware QSPR Modeling: The 4L-QSPR Framework Applied to Aqueous Solubility Prediction

Abstract Reliable quantitative structure-property relationship (QSPR) modeling requires validation protocols that measure chemical generalization rather than interpolation among closely related compounds. This contribution presents 4L-QSPR, a layered workflow for chemically informed and leakage-aware molecular property modeling. The framework separates endpoint-independent molecular grouping (Layer 1), supervised descriptor preselection (Layer 2), and nested group-aware model optimization using DOOIT2 (Layer 3, Dual Objectives Optimization with iterative Feature Pruning), with Layer 4 reserved for future deployment and updating. Layer 1 defines a reproducible structural hierarchy using a fixed hybrid distance that combines Morgan fingerprint Tanimoto distance with cosine distance in standardized RDKit physicochemical descriptor space. The representation uses radius-2, 2048-bit Morgan fingerprints and eight RDKit descriptors, with fixed weights of 0.80 and 0.20, respectively. No solubility values, supervised descriptor rankings, model residuals, or predictive metrics were used to construct or adjust Layer 1. Layer 2 performs endpoint-dependent descriptor cleaning and regressor-specific preselection, while Layer 3 confines hyperparameter optimization, recursive descriptor pruning, and final model selection to nested StratifiedGroupKFold validation using Layer 1 groups. The workflow was evaluated on a curated ESOL aqueous solubility dataset containing 1111 unique molecular identities. The final locked demonstration used the combined COSMO-PaDEL descriptor representation with LightGBM regression. Across three independent random seeds, the final model achieved MAE = 0.459 ± 0.004 logS units and R² = 0.897 ± 0.007 across seed-wise means. The representative seed-42 run gave outer-fold MAE = 0.460 ± 0.046 and R² = 0.895 ± 0.031, with 22.2 ± 5.8 selected descriptors. Recurrently selected COSMO-derived solvation terms and PaDEL structural descriptors supported chemically interpretable model behavior under stricter grouped validation. Scientific Contribution This work introduces a reproducible validation-oriented QSPR framework that explicitly separates endpoint-independent chemical grouping from supervised descriptor selection and nested model optimization. The ESOL case study shows that competitive and interpretable solubility models can be obtained under chemically grouped validation designed to reduce structural leakage.