Protein pKa Prediction by Tree-Based Machine Learning

Abstract

Protonation states of ionizable protein residuesmodulate many essential biological processes. For correct modelingand understanding of these processes, it is crucial to accuratelydetermine their pKavalues. Here, we present four tree-basedmachine learning models for protein pKaprediction. The fourmodels, Random Forest, Extra Trees, eXtreme Gradient Boosting(XGBoost), and Light Gradient Boosting Machine (LightGBM),were trained on three experimental PDB and pKadatasets, two ofwhich included a notable portion of internal residues. We observedsimilar performance among the four machine learning algorithms.The best model trained on the largest dataset performs 37% betterthan the widely used empirical pKaprediction tool PROPKA and15% better than the published result from the pKapredictionmethod DelPhiPKa. The overall root-mean-square error (RMSE) for this model is 0.69, with surface and buried RMSE values being0.56 and 0.78, respectively, considering six residue types (Asp, Glu, His, Lys, Cys, and Tyr), and 0.63 when considering Asp, Glu,His, and Lys only. We provide pKapredictions for proteins in human proteome from the AlphaFold Protein Structure Database andobserved that 1% of Asp/Glu/Lys residues have highly shifted pKavalues close to the physiological pH.