Stokes Shift Prediction for Solvated Organic Chromophores: A Machine Learning Study of Generalization and Reliability

The Stokes shift is a key photophysical parameter for fluorescent materials, yet systematic assessment of machine learning (ML) generalization and prediction reliability for this property is still limited. In this work, five supervised regressors (Ridge, Lasso, Random Forest, XGBoost, LightGBM) are trained on the Deep4Chem database to predict the Stokes shift of solvated organic chromophores from combined chromophore and solvent descriptors, and evaluated with nested cross-validation under random and chromophore-wise grouped splits. Tree-based ensembles outperform linear baselines, and random splits yield substantially better apparent accuracy than grouped splits, indicating that conventional random splits overestimate extrapolation to unseen chromophores. External validation on the independent JUETAL dataset shows that tuned models achieve only moderate accuracy overall, with metrics closer to the grouped-split than the random-split regime. To quantify prediction reliability, a distance-based applicability domain (AD) is constructed, using a k-nearest-neighbors (k-NN) density metric. Imposing an error threshold corresponding to RMSE ≤ 20 nm identifies an in-domain subset comprising roughly 11% of the external molecules, within which average performance improves to R² ≈ 0.9 and MAE ≈ 10 nm. SHAP analysis further highlights nitrogen-containing conjugated motifs, tertiary amines, ring-fusion topology, and polarity-related descriptors as key determinants of the Stokes shift. Together, these results adapt established QSAR concepts of split strategy and AD to Stokes-shift prediction and delineate a narrow but highly reliable prediction domain for solvated organic chromophores.