AI Designs State-Specific Peptides That Target GPCRs as Agonists or Antagonists
A new AI pipeline named HelixFold-Multistate designs peptides that selectively activate or block key drug-target receptors with nanomolar potency.
Summary
Researchers at Baidu and Zonsen PepLib developed a computational pipeline called HelixFold-Multistate to design peptides that target G protein-coupled receptors (GPCRs) in their active or inactive conformational states. By combining a peptide generation module with a fine-tuned structural folding model, the system can distinguish between agonist- and antagonist-favoring receptor conformations. The model was validated against real experimental structures and then applied to three therapeutic GPCR targets. It successfully generated agonist peptides for the Apelin Receptor (APJR) with EC50 below 10 nM, agonist and antagonist peptides for the Growth Hormone Secretagogue Receptor (GHSR), and a competitive inhibitor for the GLP-1 Receptor with an IC50 of 874 nM — demonstrating both design precision and experimental reproducibility.
Detailed Summary
G protein-coupled receptors (GPCRs) are among the most important drug targets in medicine, regulating heart rate, hormone secretion, metabolism, neurological signaling, and more. Most existing GPCR drugs are small molecules, but peptides offer superior selectivity and potency. A longstanding challenge is that GPCRs exist in multiple conformational states — active and inactive — and a peptide's function as an agonist (activator) or antagonist (blocker) depends on which state it stabilizes. Until now, most AI peptide design tools ignored this distinction.
A team from Baidu's NLP division and Zonsen PepLib Biotech developed HelixFold-Multistate (HF-Multistate), a fine-tuned structural folding model that incorporates state-specific information about GPCR conformations. The model was trained on GPCR–peptide complex structures from the GPCRdb database and optimized to accurately predict both the peptide–receptor binding interface and the functional conformation of key transmembrane (TM) helices, particularly TM3 and TM6 in class A GPCRs and TM3, TM6, and TM7 in class B GPCRs, which are critical for activation.
Benchmarked against AlphaFold-Multimer, HelixFold-Multimer, and AlphaFold-Multistate, HF-Multistate achieved top performance on DockQ scores (interface accuracy) and lowest RMSD across key TM domains. Notably, it outperformed AlphaFold-Multistate on peptide–GPCR interaction prediction despite using a single model with five random seeds rather than an ensemble of five models run five times — making it computationally more efficient. Crucially, it correctly distinguished active from inactive receptor states in cases where AF-Multimer failed entirely.
The pipeline was applied to three GPCR targets. For APJR (Apelin Receptor), both agonist and antagonist peptides were generated and validated experimentally; agonists achieved EC50 values below 10 nM, ranking among the most potent reported. For GHSR (Growth Hormone Secretagogue Receptor), agonist and antagonist peptides were also successfully designed. For GLP-1R (Glucagon-Like Peptide-1 Receptor), a competitive inhibitor peptide was identified with an IC50 of 874 nM. Structural confidence scores from HF-Multistate correlated strongly with experimental binding affinity data across both agonist and antagonist categories, validating the model's utility as a screening tool.
The pipeline addresses a gap in the field: while agonist peptide design has seen progress, antagonist peptide design for GPCRs is substantially harder and clinically underexplored. This framework offers a generalizable approach for designing state-selective peptide therapeutics for any GPCR with known structural data, with potential implications for metabolic disease, cardiovascular conditions, and endocrine disorders.
Key Findings
- HF-Multistate outperformed AF-Multimer and AF-Multistate on GPCR–peptide interface prediction using DockQ and iRMS metrics.
- APJR agonist peptides achieved EC50 below 10 nM, among the highest potencies reported for this receptor.
- A GLP-1R competitive inhibitor peptide was designed with IC50 of 874 nM, rare for a peptide antagonist at this target.
- The model correctly distinguished active vs. inactive GPCR states where standard AlphaFold-Multimer failed.
- Structural confidence scores from HF-Multistate correlated with experimental binding affinity for both agonists and antagonists.
Methodology
The study used a two-stage pipeline: a peptide generation module produced candidate sequences, then HF-Multistate (a fine-tuned version of HelixFold-Multimer incorporating state-specific TM domain modeling) predicted complex structures and scored them. Evaluation used DockQ, iRMS, and TM-RMSD metrics on a held-out set of recent GPCR–peptide crystal structures from GPCRdb, followed by wet-lab validation including binding affinity assays for three GPCR targets.
Study Limitations
The pipeline was demonstrated on only three GPCR targets, limiting generalizability claims. Antagonist peptide affinity (IC50 ~874 nM for GLP-1R) remains moderate and may require further optimization for clinical utility. The method relies on available experimental GPCR structures; receptors lacking high-quality active and inactive state structures may not benefit equally.
Enjoyed this summary?
Get the latest longevity research delivered to your inbox every week.