Loading…
Conotoxin design framework pairing a ProGen2-small fine-tune on cone snail venom peptides with an autoencoder classifier that screens the samples.
No providers recorded yet. Browse all providers
About two thousand conotoxins have ever been catalogued, against the roughly one million estimated to exist across cone snail venoms. That is far too small a corpus to fit a generative sequence model from scratch, and the feature that matters most is exactly the kind a small corpus fails to teach: a cysteine framework whose paired residues sit dozens of positions apart and staple the 10-to-70-residue peptide into a rigid, protease-resistant fold. ConoDL's answer is to import the grammar of proteins from elsewhere: its generator starts from a checkpoint already trained on a billion ordinary protein sequences, and spends the scarce conotoxin data only on what makes conotoxins distinctive.
ConoDL is a two-part framework from Weiwei Xue's group at Chongqing University, with the State Key Laboratory Breeding Base of Marine Genetic Resources at the Third Institute of Oceanography, posted to bioRxiv in September 2024 and published in the Journal of Computer-Aided Molecular Design that December. ConoGen, the generative half, is a ProGen2-small checkpoint fine-tuned on mature conotoxin peptides; ConoPred, the screening half, is a Wasserstein autoencoder classifier that scores how conotoxin-like a sequence is. The pairing is what makes the framework usable: unconditional sampling from a small fine-tune produces plausible and implausible peptides in bulk, and a discriminator trained on the same positives narrows a million raw samples to a shortlist worth folding.
ConoGen fine-tunes progen2-small — 12 transformer layers, 16 attention heads, model dimension 1024, rotary embeddings and a 32-token amino acid vocabulary, roughly 151 million parameters. The training set began as 3,281 mature peptides from ConoServer, filtered to lengths of 10–70 residues, standard amino acids only, and at least four cysteines, leaving 2,310 sequences; fine-tuning ran 15 epochs on a single NVIDIA GTX 3080 with AdamW at a learning rate of 1e-4. ConoPred was trained in two stages: a Wasserstein autoencoder for 500 epochs, then a classifier for 100 epochs on those 2,310 conotoxins as positives against 13,941 Swiss-Prot sequences as negatives, using focal loss to handle the imbalance and reaching an AUC of 1.0 and AUPR of 0.998 on both splits. The preprint's architecture section instead describes a 36-layer, 8-head, 1.2-billion-parameter transformer — the configuration of the original ProGen — which the released checkpoint and its config file do not match.
Sampling produced 10,000 sequences per ten-minute run and one million in total, of which 761,371 were non-redundant. Mean pairwise similarity between generated sequences and the training set was 10.86 against 15.58 among the generated set itself, on a scale where identical sequences score 36.8 and unrelated ones -14.6. Thirty-four representatives spanning every scaffold class were folded with ColabFold, and 37 were carried into 500 ns AMBER ff14SB simulations; 31 held below 3 A Ca RMSD, with the larger excursions confined to long segments lacking disulfide bonds.
Cone snail venom is an established therapeutic source — ziconotide, a synthetic omega-conotoxin MVIIA, is approved for chronic pain — but reaching new scaffolds has meant purifying venom from wild-caught snails or chemically modifying known toxins. ConoDL gives peptide chemists and marine natural product groups a computational front end instead: a library of scaffold-diverse candidates, a probability score for triage, and enough sampling control to bias generation toward a target length or cysteine pattern. The released sequence set is reusable directly, without running the model.
ConoDL is one of the first generative models built specifically for conopeptides, and it makes a case that transfer learning from a general protein language model suffices in a domain with only a few thousand known members — a situation common across venom peptides, defensins and other disulfide-rich families. The evaluation is entirely computational: no generated peptide was synthesized or assayed, so the output is a hypothesis list rather than a set of validated toxins. ConoPred judges only whether a sequence looks like a conotoxin, not what it would bind, and the authors name target prediction as the missing piece. The fine-tuned weights, the pretrained base checkpoint and the million-sequence library are on Zenodo under CC BY 4.0, with the training and sampling code on GitHub.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.