Conditional masked language model that designs and optimizes conotoxin peptides for a named receptor subtype, with wet-lab validated hits.
No providers recorded yet. Browse all providers
α-conotoxin ImI is twelve residues long and stapled by two disulfide bonds, and decades of mutagenesis have barely improved on it. Replacing every position in its two loops with each of the other nineteen amino acids yields a thousand variants, of which physics-based screening finds sixty-five that bind the α7 nicotinic acetylcholine receptor more tightly — a 6.5% hit rate. Freezing the three residues already known to be critical raises the yield to 253, but only by refusing to look anywhere interesting: constrain the scaffold and the hit rate climbs while the accessible sequence space collapses.
CreoPep answers that trade by making the target part of the input. Every training example carries a subtype label drawn from 53 receptor subtypes and a potency label, alongside a wild-type conotoxin and a handful of auxiliary peptides sharing its label, so one checkpoint can be pointed at eight different conotoxin–receptor systems by changing a token rather than by refitting. It was built by Rilei Yu's group in the Key Laboratory of Marine Drugs at Ocean University of China, with the Molecular Horizons institute at the University of Wollongong contributing the electrophysiology, and posted to arXiv in May 2025 before appearing in Communications Chemistry in January 2026.
The generative mechanism borrows its schedule from diffusion. In place of the fixed 15% corruption rate that ProtTrans-style masked language models train on, progressive masking grows the number of masked tokens over training; generation reverses the schedule, predicting one masked residue at a time until the peptide is complete.
The backbone is ProtBert's masked language modeling framework. Its BERT encoder is frozen and a second trainable BERT encoder, without an embedding layer, is stacked on top; auxiliary conotoxins are encoded separately through a Xavier-initialized multilayer perceptron, and the two feature streams are concatenated and reduced by a 1×1 convolutional block before a decoder produces the output. Training data came from ConoServer — 2,088 conotoxins with measured targets and IC50 values, binned as high potency at ≤1000 nM — across seven receptor families and 53 subtypes. On its validation set the final model reaches 76.63% Top-1 and 97.22% Top-5 accuracy on subtype, 93.36% on potency, and above 95% Top-1 on receptor target. Generating a thousand mutants for each of eight systems, OmegaFold pLDDT ranged from 58.93 for the Nav1.2 system to 70.43 for α7, against a training-set average of 71.46, while TM-scores below 0.5 and Hamming distances above 0.3 show the outputs are not near-copies.
The immediate use is lead optimization for peptide therapeutics against ion channels and receptors, a class with a marketed precedent in ω-conotoxin MVIIA for chronic pain. A user supplies a wild-type peptide and a target subtype token and receives ranked mutants; the same checkpoint then annotates uncharacterized conotoxins with a likely receptor and potency. Four Gradio interfaces on HuggingFace Spaces run each task in a browser, and the BERT vocabulary can be extended with tokens for non-canonical residues.
CreoPep is one of the few peptide generators to carry its designs through synthesis and assay. Of thirteen ImI-derived candidates, seven blocked human α7 nAChR by more than 50% at 10 µM, and two reached submicromolar potency — CP_α7_1 at 405.1 nM and CP_α7_6 at 504.6 nM. CP_α7_6 is the more striking result: it breaks the conserved S-X-P motif and lacks ImI's disulfide connectivity, reaching comparable potency through an entirely different interaction network. A second campaign on the harder GID/α4β2 system produced three actives from ten candidates. The limits are real: only one variant marginally exceeded wild-type ImI, the FoldX filter driving the augmentation loop scored below 70% accuracy on the ImI/α7 and Vc1.1/α9α10 systems, and the label vocabulary confines the model to conotoxin-like peptides. Alongside ConoDL, it makes conotoxin design a subfield with reproducible baselines.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.