The function of many protein families, such as membrane-bound transporters and G protein-coupled receptors, are tightly coupled to transitions between major conformational states. Despite recent breakthroughs in deep learning methods for protein structure prediction, generalizable methods for accurately predicting these conformational states remain a challenge. We develop an effective algorithm for predicting multiple conformations of membrane proteins. Our method applies a negative entropy loss on the predicted auxiliary distogram of AlphaFold2, which prevents the AlphaFold model from focusing on one conformation and promotes the exploration of alternative conformations. We empirically evaluate our model on a set of 37 membrane proteins and achieve a >80% success rate of predicting two distinct conformations (each within 2.0 Å RMSD to experimental structures). Subsequent analyses reveal that our method is able to predict more than two major conformations in multiple cases, some of which have not been experimentally determined. In addition, we find that, among predicted structures within a conformational cluster, structural variations at a given residue position show correlation with experimental B-factor values. This suggests the variations observed in predictions might provide biologically useful information. We also developed a method for assigning scores to each predicted structure, which facilitates the selection of structures that may be closest to discrete conformations and allows for more informed interpretation of predictions. Our work provides an effective method to accurately predict alternative conformations of proteins at a low cost. We have released code for our method at https://github.com/fenglaboratory/egf.
Wu et al. (Sun,) studied this question.