The prediction of protein structures has been significantly advanced by AI-driven methodologies, particularly AlphaFold2, which can often provide researchers with highly accurate predicted structures to facilitate subsequent studies. Membrane proteins are crucial for cellular communication and transport, serving as major targets for drug development. However, AlphaFold2 was not specifically trained to model membrane proteins or to account for the lipid membrane environment, which may lead to potential inaccuracies in the overall construction of transmembrane topology. These limitations necessitate specialized computational frameworks to enhance the predictions of membrane protein structures. Building on our prior work (ProtRAP) on the prediction of residue-level accessibilities, including RLA (Relative Lipid Accessibility) and RSA (Relative Solvent Accessibility), we present RAPFold—a modified workflow for enhanced prediction of membrane protein structures. RAPFold is based on OpenFold, which is a PyTorch implementation of AlphaFold2, and it integrates RLA and RSA to enhance the predictions of membrane protein structures. In the structure module, we construct a virtual lipid bilayer and orient the protein backbones within it. A loss value is generated based on relative accessibilities and inferred transmembrane topology, while the gradient is back-propagated to optimize the single representation, thereby refining the predicted structure. Compared to OpenFold, RAPFold significantly improves the prediction accuracy for certain membrane proteins, resulting in membrane protein structures that align more closely with the physicochemical properties of lipid membranes. This work establishes a computational framework for enhancing the structural prediction of membrane proteins. The prediction results will be stored in a newly developed database for reference in future studies on membrane proteins.
Kang et al. (2026) studied this question.