Abstract Membrane proteins play pivotal roles in cellular signaling and transport, making them prime targets for therapeutic intervention. Therefore, rapid screening and accurate property prediction of these proteins are essential. Recently, we proposed a new metric, membrane contact probability (MCP), to characterize the membrane-contacting features of membrane proteins and to further develop the Protein Relative Accessibility Predictor (ProtRAP). However, these models relied on evolutionary information obtained from multiple sequence alignments (MSAs), which constrained the speed of predictions. In this study, we introduce a new transformer-based model, ProtRAP-LM, which leverages protein language model (pLM) embeddings as input features to rapidly and accurately predict MCP and relative accessibility for each residue within a given protein sequence. ProtRAP-LM demonstrates superior performance on a 184-protein test set compared to previous MSA-based models, achieving a speed-up of more than 300 times on a workstation equipped with an RTX 3080 Nvidia graphics processing unit (GPU). As a result, entire proteomes can be predicted within hours. This advancement enables us to provide more comprehensive annotations of membrane protein sequences at the proteome-wide scale, particularly for single-pass transmembrane proteins, membrane-anchored proteins, and β-sheet-containing membrane proteins, all of which have long been a challenge in the field. In the end, we provide a comprehensive list of membrane proteins for 48 living organisms, offering a rich resource for investigating the structure and function of these essential biomolecules in the future. An online computation server of ProtRAP-LM is available at http://www.songlab.cn/ProtRAP-LM/home/ with the source code at GitHub (https://github.com/ComputBiophys/ProtRAP-LM) and BioCode (https://ngdc.cncb.ac.cn/biocode/tool/7994).
Wang et al. (Sun,) studied this question.