How to select the best model from AlphaFold2 structures?

Abstract

AbstractAmong the methods for protein structure prediction, which is important in biological research, AlphaFold2 has demonstrated astonishing accuracy in the 14th Community Wide Experiment on the Critical Assessment of Techniques for Protein Structure Prediction (CASP14). The accuracy is close to the level of experimental structure determination. Furthermore, AlphaFold2 predicts three-dimensional structures and estimates the accuracy of the predicted structures. AlphaFold2 outputs two model accuracy estimation scores, pLDDT, and pTM, enabling the user to judge the reliability of the predicted structures. Original research of AlphaFold2 showed that those scores had good correlations to actual prediction accuracy. However, it was unclear whether we could select a structure close to the native structure when multiple structures are predicted for a single protein. In this study, we generated several hundred structures with different combinations of parameters for 500 proteins and verified the performance of the accuracy estimation scores of AlphaFold2. In addition, we compared those scores with existing accuracy estimation methods. As a result, pLDDT and pTM showed better performance than the existing accuracy estimation methods for AlphaFold2 structures. However, the estimation performance of relative accuracy of the scores was still insufficient, and the improvement would be needed for further utilization of AlphaFold2.