PANTHER - Protein-Affinity for Nucleic Target-binding, Hybridization, and Energy Regression

Abstract

Although protein-RNA interactions are crucial for many biological processes, predicting their binding free energies (ΔG) is a challenging task due to limited available experimental data and the complexity of these interactions. To address this issue, we developed a machine learning-based model designed to predict energy-based scores for protein-RNA complexes, called PANTHER score. By applying a local-to-global approach, the here proposed methodology can be subdivided into four steps: (1) we derived 87,117 pairwise local interaction energies out of 331,744 obtained from molecular dynamics simulations for a training set composed by 46 curated protein-RNA complexes; (2) we trained ML models derived from pairwise interaction features to predict the local interaction energies without performing MD runs; (3) we integrated the predicted local interaction energies with our here proposed local-to-global methodology, to calculate the model-specific PANTHER score; (4) we test the model-specific PANTHER score on a test set of 7 complexes (5) we further exposed all the models to an external stress set which includes 110 complexes with experimental ΔG allowing for final selection of the optimal model for implementation in the PANTHER scoring pipeline. Among all the multiple regression models developed here and evaluated on the test set, Random Forest Regression exhibited the highest predictive performance as a model-specific PANTHER score, with a Pearson correlation coefficient of (r) of 0.80 and mean absolute error (MAE) of 1.79 kcal/mol. Furthermore, the Random Forest Regression model maintained strong predictive capabilities on the stress set as well with (r) of 0.64 and MAE of 1.63 kcal/mol. Benchmarking against existing tools on the stress test set, the PANTHER score demonstrated superior accuracy and reliability. This study highlights the effectiveness of machine learning in addressing data limitations through innovative strategies, positioning here proposed PANTHER score as a valuable tool for predicting protein-RNA binding affinities in biomolecular research and drug discovery.