Solvation Free Energy Descriptors from Density Functional Theory for Unified Retention Prediction of Flavonoids and Isoflavonoids

<title>Abstract</title> <p> Context: Quantitative structure–retention relationship (QSRR) models are typically confined to a single chemical scaffold. A unified model capable of predicting the chromatographic retention of compounds with distinct conjugation topologies using a minimal set of physically interpretable descriptors has remained elusive. Herein, a cross-scaffold QSRR model encompassing both flavones and isoflavones was developed using only two solvation free energy descriptors from density functional theory (DFT): <bold>E</bold> _{<bold>W</bold>} (water) and <bold>E</bold> _{<bold>M</bold>} (methanol). Multiple linear regression yielded <bold>t</bold> _{<bold>R</bold>}  = − 0.8486 <bold>E</bold> _{<bold>W</bold>} +0.9396 <bold>E</bold> _{<bold>M</bold>} +40.55 ( <bold>R²=0.928</bold> , <bold>Q²</bold> _{<bold>LOO</bold>} <bold>=0.872</bold> , training <bold>MARE = 2.80%</bold> ). The model discriminates flavone–isoflavone positional isomers and retains predictive capability for isoflavones outside the training set, with the neobavaisoflavone outlier demarcating the applicability domain of the IEFPCM-based descriptors. Methods A two-tier conformational sampling strategy was employed: relaxed potential energy surface (PES) scanning of the B-ring dihedral angle using the PM6 Hamiltonian identified the global minimum-energy basin, circumventing the local-minimum trapping that frequently compromises direct DFT optimization; each candidate conformation was subsequently refined at the B3LYP-D3/6-31G(d) level. Implicit solvation free energies were computed at the same level of theory using the IEFPCM continuum model, while explicit hydrogen-bond binding energies were evaluated at sites identified by molecular electrostatic potential (ESP) analysis, with RAHB theory guiding site selection. All DFT and semi-empirical calculations were carried out with Gaussian 16W; GaussView 6.0 was employed for molecular visualization and input preparation. ESP analysis and visualization were performed with Multiwfn (version 2026.6.2) and VMD 1.9.3. Multiple linear regression was performed using MATLAB R2025b. </p>