Why choose between prediction power and mechanistic insight in QSRR modeling? A case study of artificial membrane...

Quantitative structure-retention relationship (QSRR) modeling of immobilized artificial membrane (IAM) chromatography was performed on a chemically diverse set of 2003 analytes to determine whether mechanistically interpretable, descriptor-based models can match the predictive performance of advanced machine-learning approaches. Chromatographic hydrophobicity indices (CHI_IAM) were measured under standardized fast-gradient IAM conditions and modeled using three molecular representation strategies: physicochemical descriptors, ECFP4 fingerprints, and a Graph Isomorphism Network (GIN) operating on molecular graphs. Among the descriptor- and fingerprint-based methods, the Gradient Boosting Regressor gave the best performance. The descriptor-based model achieved a test R² of 0.817, MAE of 4.43, and RMSE of 6.46 CHI_IAM units, whereas the GIN model showed only a marginal numerical improvement (R² = 0.828, MAE = 4.27, RMSE = 6.25). In contrast, fingerprint-based modeling performed worse (R² = 0.771, MAE = 5.04, RMSE = 7.22). Statistical analysis of paired prediction errors showed no significant difference between the descriptor-based and GIN models, while the fingerprint-based model was the least accurate of the three. The observed differences among the best models were smaller than the long-term reproducibility of the IAM protocol. Applicability-domain analysis further showed that approximately 94-99% of compounds fell within the interpolation region. SHAP analysis identified lipophilicity, ionization, charge, and polarity as the major determinants of IAM retention, with cationic character enhancing retention and high polarity reducing it. These results demonstrate that descriptor-based QSRR modeling can achieve near-deep learning performance while preserving mechanistic interpretability, providing a practical and physically grounded strategy for IAM retention prediction.