Prediction of pre- and postfusion conformations of class I fusion proteins with AlphaFold2

Abstract

Viruses such as coronaviruses or filoviruses use their surface glycoproteins (GPs) to attach to the host cell, triggering the fusion of the viral membrane with the endosome membrane. Epitopes on the viral GP are major targets for antibody-mediated recognition and neutralization. During the fusion process, the GP undergoes conformational changes triggered by fluctuations in environmental pH. Structural states are typically classified into three distinct conformations: prefusion, intermediate, and postfusion. These conformations serve as essential templates for prediction of conformational epitopes and structure-based vaccine design. Despite their importance, many viral GP structures remain absent from the Protein Data Bank (PDB). Fortunately, recent breakthroughs in computational structure prediction have greatly enhanced the accuracy and accessibility of protein modeling. In this study, we utilized AlphaFold2-Multimer (AF2-M), version 2.3, to predict various GP structural conformations and observed that the overall frequency of predictions in the postfusion conformation is low. Therefore, we hypothesized that adapting the AF2-M protocol is necessary to enrich for specific conformations, thereby enabling the prediction of both pre- and postfusion conformations. AF2-M requires only the input sequence and internally generates multiple sequence alignments (MSAs) and optional templates before applying its pretrained model weights. We tested the use of template data to enrich pre- or postfusion conformations and demonstrated that our approach significantly increases the prediction frequency of class I fusion protein structures in both conformations, with the template dataset playing a crucial role in guiding modeling towards the intended state. Furthermore, we showed that the lack of correlation between pLDDT and TM-scores suggests that low pLDDT values may obscure the presence of valid alternative conformations.