Multi-state Structure Prediction of G Protein-Coupled Receptor Proteins via Prompting on AlphaFold

G protein-coupled receptors (GPCRs), an essential family of transmembrane proteins, widely participate in signal transduction in organisms and have long been recognized as a major class of therapeutic targets. In structure-based drug design, high-resolution structures of GPCRs in both active and inactive states are essential for designing agonists and antagonists, respectively. However, obtaining experimental GPCR structures is costly, while homology modeling and artificial-intelligence-driven approaches including AlphaFold 2 and AlphaFold 3 often show reduced accuracy for active-state conformations. To address the above limitations, we propose PromptGPCR, an AlphaFold-based inference framework aiming to predict highly accurate structures of GPCRs in both active and inactive states. We provide AlphaFold-Multimer and AlphaFold 3 with biological sequences based on knowledge of structural biology as prompts to guide the models in the multi-state prediction task. Experimental results demonstrate that PromptGPCR can accurately predict active and inactive structures compared to baselines, suggesting its ability to provide structural hypotheses where state-resolved experimental structures are unavailable. Furthermore, PromptGPCR exhibits higher success rates in molecular docking than baselines, and this indicates that the predictions of PromptGPCR may have a certain degree of usability in downstream application scenarios involving specific GPCR conformational states.