A team screened 689 peptides from cone-snail venom with a protein language model, and one of them turned out to block a nerve receptor that has resisted good drugs for decades. They then optimized that hit into a molecule with nanomolar potency and solved its structure sitting inside the receptor. The AI-picked candidate and the microscope landed on the same peptide.
The work, published in Acta Pharmaceutica Sinica B ↗, targets the alpha-7 nicotinic acetylcholine receptor. That receptor sits on nerve and immune cells and is one of the switches nicotine flips. Because of where it acts, it has been chased for years as a target for Alzheimer's disease, schizophrenia, and the inflammation that the vagus nerve normally reins in. The trouble has always been finding molecules precise enough to hit alpha-7 and not the many closely related nicotinic receptors, such as the alpha-4-beta-2 subtype, sitting next to it.
Cone snails, the marine genus Conus, are a long-standing source of that kind of selectivity. Their venoms are mixes of small peptides, each shaped to grip one ion channel or receptor with high selectivity, and one of them, ziconotide, sold as Prialt, is an approved painkiller that blocks N-type calcium channels. The catch is that most conopeptides are knotted together by multiple disulfide bridges, which makes them hard to synthesize, hard to model, and scarce in the databases that machine-learning tools train on. There simply are not enough solved examples to learn from.
Working around thin data
The group got around the data problem by narrowing the search. Instead of the hardest cases, they focused on disulfide-poor conopeptides, the ones held together by fewer sulfur bridges, and built a screening model on ESM-2, the protein language framework from Meta AI. ESM-2 is trained on millions of protein sequences drawn from UniProt, the same class of tool behind recent structure predictors, and it can rank candidate peptides without needing a large task-specific training set.
That screen of 689 disulfide-poor conopeptides surfaced a single antagonist, which the authors named SS1. A round of structure-activity work, swapping and tuning individual residues, produced an optimized version, [ΔQP,S8R]SS1, that blocked the receptor at a concentration of 49.2 nanomoles per liter. In plain terms, it works at the low doses that real drugs need, with better selectivity and stability than the starting peptide.
The potency was checked with patch-clamp electrophysiology, the standard way to watch a single receptor open and close. The confirmation step is what separates this from a pure computational exercise. Using cryo-electron microscopy, the team resolved the structure of alpha-7 bound to the peptide at 3.3 angstroms, sharp enough to see how it fits. The peptide holds on through a mix of hydrogen bonds, hydrophobic contacts, and links to the sugar chains on the receptor's surface, and it appears to lock the receptor in a blend of its closed and desensitized shapes rather than its open, signaling one.
What it is and is not
This is a discovery-and-structure paper, not a drug. There are no animal results and no patients. An antagonist that shuts alpha-7 down is a research tool and a starting scaffold more than a finished therapy, since several of the diseases attached to this receptor want it switched on, not off. What the peptide does clearly is give medicinal chemists a defined handle on a receptor that has been hard to drug.
The method is the part that carries over. A language model trained on generic protein data, pointed at a deliberately tractable slice of a hard peptide family, produced a hit that held up under a microscope. That recipe, pick the part of a scaffold family the data can support and verify the winner structurally, travels to other venom-derived peptides and other receptors.
Neither SS1 nor its optimized form has a card on peptidemodel yet, which is normal for a molecule this new. The receptor itself sits on the neuroprotective ↗ shelf, alongside the other peptides being built against nervous-system targets.