AI-Driven Prediction of Protein Structural Stability: A Machine Learning Approach for Accurate RMSD Estimation

Root Mean Square Deviation (RMSD) prediction of tertiary protein structure is very important for structural biology, drug design, and protein engineering. Conventional models, however, cannot ensure high accuracy due to the inherent complexity of protein flexibility and stability. In this study, RMSD prediction is enhanced by combining Machine Learning (ML) and Explainable Artificial Intelligence (XAI) through the innovation of a novel Compactness Factor and the logarithmic transformation. These modifications are aimed at improving model performance by capturing structural properties more effectively. In predictive power evaluation, we apply several ML models, including Random Forest Regression (RFR), K-Nearest Neighbors Regression (KNNR), Bootstrap Aggregating regression (BAGGINGR), CatBoost Regression (CBR), Decision Tree Regression (DTR), Extra Trees Regression (ETR), and Extreme Gradient Boosting Regression (XGBR). Among these, ETR is the best performing with R<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> of 0.713, adjusted R<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> of 0.712, Mean Squared Error (MSE) of 10.80, Root Mean Squared Error (RMSE) of 3.28, and Mean Absolute Error (MAE) of 2.17, which is an improvement over current research. The interpretability of the model is also enhanced using Shapley Additive Explanations (SHAP) and Local Interpretable Model-agnostic Explanations (LIME), which contribute significant features to predictions. The results indicate that a combination of feature engineering, ML models, and XAI techniques improves RMSD prediction, pushing computational protein research and biomedical applications forward.