Can Protein Structure Prediction Methods Capture Alternative Conformations of Membrane Proteins?

In this study, we benchmarked 12 representative protein structure methods on alternative conformation prediction in a curated dataset that consists of Inward-facing and Outward-facing states of Membrane Proteins (IOMemP) from the Protein Data Bank (PDB). The methods were benchmarked using the same input multiple sequence alignments (MSAs). Additionally, we tested 2 AlphaFold-based methods that manipulate MSAs to predict alternative conformations. Our dataset IOMemP and benchmark results could promote the development of alternative conformation prediction.

The curation of the dataset

The details can be found here.

The benchmarked methods and their configurations

MI mfDCA PSICOV CCMpred plmDCA RaptorX-Contact ResPRE trRosetta RaptorX-3DModeling ESM RoseTTAFold AlphaFold AF-Depth AF-Cluster

All methods are executed with default settings if there is no specific statement.

As for AF-cluster, AF-depth, and AlphaFold (v2.1.2), the mini modifications for their localization are provided in the corresponding patch files.

Note that AF-cluster default uses Alphafold (v2.2.0) as its prediction engine.

The calculation of the composite confidence score of IF and OF structures

We add an alignment function to AF2Rank (renamed as AF2Rank_homo_template) for the homologous template assignment, considering the different sequences lying between IF and OF structures.

The benchmark on AlphaFold input with MSAs, IF and OF structures, or their combinations

We also make AF2Rank_homo_template accept multiple templates and predict multiple times for models 1-5, so that we can AlphaFold predictions under different input configurations.

Reference

Tengyu Xie and Jing Huang. Can Protein Structure Prediction Methods Capture Alternative Conformations of Membrane Proteins? bioRxiv 2023.08.04.552045; doi: https://doi.org/10.1101/2023.08.04.552045.