G-protein-coupled receptors (GPCRs) make up a large and diverse group of membrane receptors that play crucial roles in cellular signal transduction. Their involvement in various physiological processes makes them attractive targets for therapeutic intervention. (1) Despite the success of small molecules and biologics in modulating GPCR activity, these approaches often face limitations such as off-target effects. (2) Peptide-based therapeutics offer a promising alternative due to their remarkable pharmacodynamic properties: high affinity, selectivity, and potency, (3,4) though their clinical use is challenged by limited in vivo stability and complex synthetic requirements. These limitations can be mitigated through strategies such as incorporating non-natural amino acids or chemical modifications (5) to enhance stability and synthetic feasibility.
Conventional methods for developing peptide drugs involve leveraging natural peptides and hormones found in the human body, as well as creating peptides that mimic these hormones. (6−8) Peptides can also be sourced from natural products, including those derived from bacteria, fungi, and plants. (9−13) Another approach involves the use of phage display techniques to discover and optimize peptide candidates. (14)
Recently, AI-based peptide rational design has gathered increasing attention. (15−17) EvoBind (18,19) employs in silico-directed evolution to design peptides and then evaluates the designs using confidence scores from the folding model AlphaFold-Multimer. (20) EvoPlay (21) improves it by replacing directed evolution with Monte Carlo tree search (MCTS) and reinforcement learning policy, (22) leading to more data-efficient peptide design. Additionally, PepMLM (23) uses a protein language model (PLM) to design peptide binders. Approaches for designing mini-proteins may also be applicable to peptide designs, such as AlphaFold-2 hallucination, (24) RFDiffusion-MPNN, (24) AlphaProteo, (25) and BindCraft. (26)
Although AI-driven peptide design approaches have successfully generated high-affinity binders, targeting GPCRs presents unique challenges due to their dynamic conformational landscape and functional plasticity as shown in GPCRmd. (27) GPCRs can adopt multiple active and inactive states, each stabilized by distinct ligand interactions, which determine whether a peptide functions as an agonist or antagonist. Effective GPCR-targeting peptides must not only achieve high binding affinity but also selectively stabilize specific receptor conformations, thereby modulating the energy landscape to favor the desired functional outcomes. Therefore, integrating GPCR conformational states and state-specific interactions into peptide design strategies is critical for achieving precise modulation of the receptor activity. Existing approaches for peptide design targeting GPCRs, such as HelixDiff, (28) typically define the structure of the GPCR protein used in the generation model. However, these methods do not allow for precise control over the GPCR’s functional state (e.g., active or inactive). This limitation hinders the ability to design peptide agonists and antagonists that accurately correspond to specific receptor states.
In this study, we present a computational method for the state-specific peptide design targeting GPCRs. The workflow begins with the generation of a diverse pool of candidate peptides using a peptide design module (29,30). A state-specified folding model is then employed to define the active or inactive states of GPCR targets, predict the structures of the GPCR–peptide complexes, and assign a confidence score to each complex. Complexes that do not align with the specified GPCR state, such as those with the binding pocket located within the membrane or those exhibiting a significant RMSD in the state-specific TM domain compared to the expected state, are filtered out. The remaining complexes are then ranked based on the integrated application of multiple confidence metrics. This ranking process allows for the prioritization of peptide candidates with the highest likelihood of success for further validation through in silico and experimental methods. To refine the folding model for GPCR–peptide interactions, we fine-tuned HelixFold-Multimer (HF-Multimer), (31) a reproduction of AlphaFold-Multimer (AF-Multimer) (20) on GPCR–peptide (32) and protein–peptide (33) data sets. This refinement resulted in improved predictive accuracy of GPCR–peptide complexes. Furthermore, we incorporated state-specific modeling techniques inspired by AlphaFold-Multistate (AF-Multistate), (34) enabling precise differentiation between the active and inactive states of GPCRs. These advancements facilitate the rational design of peptide agonists and antagonists specifically tailored to target specific GPCR receptors. The resulting model, named HelixFold-Multistate (HF-Multistate), provides a robust computational tool for advancing the design of state-specific peptide therapeutics targeting GPCRs.
Figure 1. State-specific peptide design pipeline. (a) Generation of the candidate peptides through the peptide design modules. (b) State-specific folding model (HelixFold-Multistate), an optimized version tailored for GPCR–peptide complexes, to select promising peptides. (c) Wet-lab experimental validation.
We validated that the GPCR–peptide structures predicted by HelixFold-Multistate are more effective at preserving the correct activation states of GPCRs while providing more accurate GPCR–peptide interaction poses for both active and inactive states of GPCR targets. Furthermore, the structural confidence scores derived from the HelixFold-Multistate model demonstrated a strong correlation with experimental affinity data for both peptide agonists and antagonists. We tested the state-specific peptide design pipeline by designing both agonists and antagonists for three GPCR targets, Apelin Receptor (APJR) and Glucagon-Like Peptide-1 Receptor (GLP-1R) and Growth Hormone Secretagogue Receptor (GHSR) using in silico and experimental validation. Designing antagonistic peptides for GPCRs is likely more challenging than agonists as most existing peptide therapeutics for GPCRs are agonists. (4) Additionally, while the GPCRdb database (32) contains few antagonist peptides, most GPCR-targeting antagonists are small molecules. The pipeline successfully generated de novo high-affinity agonists (<100 nM) and moderate-affinity antagonists, highlighting the precision and flexibility of the proposed methodology. Notably, the agonist peptides targeting APJR exhibit a high affinity, with an EC 50< 10 nM. The antagonist peptides targeting GLP-1R have an IC 50 of 874 nM.
We first evaluated the structural prediction accuracy and screening effectiveness of the optimized state-specific folding model, HF-Multistate, which serves as the most critical component of our peptide design pipeline. Following this, we assessed the pipeline’s performance by designing peptides for three GPCR targets, aiming to evaluate its ability to generate target-specific agonists and antagonists.
We evaluated the performance of different folding models from two perspectives: the accuracy of GPCR–peptide interface predictions and the structural accuracy of GPCR states, particularly in key regions. Our optimized state-specific folding model, HelixFold-Multistate (HF-Multistate), was benchmarked against AlphaFold-Multimer (AF-Multimer), (20) HelixFold-Multimer (HF-Multimer), (31) and AlphaFold-Multistate (AF-Multistate). (20) It is worth noting that AF-Multimer and AF-Multistate utilize an ensemble of five models, each executed five times with different random seeds for each prediction. In contrast, HF-Multimer and HF-Multistate achieve competitive results using a single model executed five times with varying seeds.
We curated a recent data set of agonist and antagonist peptide–GPCR structures from GPCRdb, (32) applying cutoffs for both release date and GPCR type. This process guarantees no overlap in GPCR types between the evaluation and training sets. Further details on the evaluation set are available in Supporting Information Section A. We employed multiple evaluation metrics, including DockQ, (35) interface root-mean-square deviation (iRMS, a component of DockQ), % correct (DockQ > 0.8), % correct (DockQ > 0.49), and % correct (iRMS < 2.0), to comprehensively assess the performance of various folding models in predicting GPCR–peptide interfaces. DockQ is a composite metric ranging from 0 to 1 that evaluates the quality of predicted protein–protein or protein–peptide interfaces by combining interface RMSD (iRMSD), ligand RMSD (lRMSD), and the fraction of native contacts (F nat). Higher scores indicate better interface accuracy. iRMS (interface RMSD) measures the deviation between the Cα atoms of interface residues in the predicted and experimental structures, focusing specifically on the local binding interface. All evaluation metrics were computed with respect to the experimentally resolved GPCR–peptide complex structures. The model was conditioned on the specified receptor state, which was derived from the annotations in the experimental data. As shown in Figure 2a–c, HF-Multistate demonstrates a top-tier performance in predicting GPCR–peptide interactions, showing a noticeable advantage over AF-Multistate. While AF-Multistate has been demonstrated to accurately predict the activation states of monomeric GPCRs, (34) it faces limitations in predicting interactions within GPCR–peptide complexes. The comparison between HF-Multistate and AF-Multistate highlights the importance of incorporating GPCR–peptide complex structural data for fine-tuning, as such data can improve predictive accuracy, particularly in modeling state-specific interactions.
Figure 2. Comparison of folding models for structural prediction accuracy in recent GPCR–peptide complexes from the PDB. (a–c) Comprehensive assessment of GPCR–peptide interface prediction accuracy using multiple evaluation metrics. (d–e) Assessment of structural accuracy in GPCR states through evaluation of the transmembrane (TM) domain in GPCR class A. (f–h) Assessment of the TM domain in GPCR class B. (i) Scatter plot comparison of TM6-RMSD between AF-Multimer and HF-Multimer, indicating HF-Multistate is more state-sensitive than AF-Multimer. CML1 active mode (PDB ID: 7YKD) is taken as an example to illustrate that AF-Multimer (gray) predicts the wrong activation state while HF-Multistate (light green) can predict the correct state. (j) Scatter plot comparison of iRMS between AF-Multistate and HF-Multistate for GPCR class B. CALCR active mode (PDB ID: 8F0K) is taken as an example. (i,j) Scatter plot comparison of iRMS between AF-Multistate and HF-Multistate for GPCR class A and class B families. (k) Illustration of the state-specific feature of HF-Multistate using SSR2. AF-Multimer model (left) cannot distinguish receptor state from active mode (7XAU) and inactive mode (7XNA), while HF-Multistate (right) can.
Although DockQ and related metrics are well suited for evaluating the accuracy of GPCR-peptide interfaces, they are less sensitive to structural precision in functionally important regions of the GPCR receptor, i.e., the transmembrane (TM) domain. (34) To address this limitation, we conducted a focused analysis of TM domain structural accuracy across two major GPCR families: class A (rhodopsin-like receptors) and class B (secretin-like receptors). Following the methodology outlined in a previous study, (36) we focused our analysis on TM3 and TM6 for class A GPCRs and on TM3, TM6, and TM7 for class B GPCRs since these regions are crucial for peptide hormone recognition and receptor activation. In class A GPCRs, TM3 and TM6 undergo state-dependent conformational changes: TM6 moves outward in active states to facilitate G-protein binding. At the same time, TM3 exhibits a cytosolic tilt and overall rotation with only minor changes. In class B GPCRs, TM3, TM6, and TM7 exhibit more intricate state-specific dynamics: TM6 undergoes outward shifts to drive receptor activation, TM3 maintains structural integrity through minor adjustments, and TM7’s extracellular region undergoes significant conformational changes to enable peptide hormone binding and activation. Across both classes, TM3 and TM6 function as key “macroswitches” in the receptor activation process. Figure 2d–e and f–h present a comparative analysis of the performance of multiple folding models on TM domains for class A and class B GPCRs, respectively, based on the root-mean-square deviation (RMSD) of key transmembrane (TM) domains. RMSD quantifies the structural deviation between predicted and experimental atomic coordinates, providing a measure of accuracy in modeling these functionally important regions. For class B GPCRs, the comparison of % Correct (TM RMSD < 2.0) is not very significant. Therefore, we included the distribution of all data points for clarity. AF-Multistate and HF-Multistate outperform AF-Multimer in structural prediction accuracy across the TM domains, with notable improvements in the critical TM3 and TM6, emphasizing the advantages of state-specific modeling. More importantly, HF-Multistate consistently achieves the highest precision across all key TM domains in terms of TM RMSD, highlighting HF-Multistate’s superior ability to incorporate the distinct activation states of GPCRs during peptide binding, providing a deeper understanding of their functional dynamics.
To demonstrate HF-Multistate’s improved modeling of functionally relevant GPCR regions over both AF-Multimer, we selected cases based on TM6 RMSD (Figure 2i). In the resulting scatter plots, each point represents a complex from the test set, with the x-axis and y-axis showing the same metric calculated by HF-Multistate and AF-Multimer, respective
