Multi-model Prediction and Comparative Study of Protein Folding Rates Based on LASSO Feature Selection

Accurate prediction of protein folding rates is essential for understanding protein dynamics. However, existing structural parameters show limited generalizability across two-state and multi-state proteins and are affected by multicollinearity. Here, LASSO regression was applied to a training set of 145 proteins to identify six structural features (L_eff, LRO, ACO, Sheet, H_nc, and P_β) from 12 structural parameters. These features were used to construct four machine learning models (RR, RF, SVM, and SSA-BP). Model performance was evaluated on an independent test set of 29 proteins, with bootstrap resampling used to assess performance uncertainty. On the full test set, RF achieved the lowest RMSE and MAE, whereas SSA-BP showed the highest correlation coefficient (R =0.9282). Folding-type-specific analysis revealed distinct model performance patterns between two-state and multi-state proteins. SHAP analysis further showed that the relative contributions of structural features varied across models and folding types, with L_eff, LRO, and ACO making substantial contributions. Overall, the proposed framework integrates feature selection, predictive modeling, and model interpretation, providing an interpretable approach for protein folding rate prediction.