PeptiTox

The disulfide-braced scaffold shared by scorpion toxins, defensins and conotoxins assembles its business end from residues drawn from opposite ends of the chain: the cysteines that staple the fold sit dozens of positions apart, and the surface that engages an ion channel exists only once the peptide is folded. A classifier reading the peptide as a string has to recover those contacts indirectly, from co-occurrence statistics over a few thousand training sequences. The contacts themselves are never part of its input.

PeptiTox supplies them explicitly. Each peptide is embedded residue by residue with ESM-2 and, in parallel, folded by ESMFold; the predicted backbone becomes a graph whose nodes are residues and whose edges join every pair of Cα atoms within 15 Å. Node features concatenate the language model embedding with geometric descriptors from the local backbone frame — bond and dihedral angles, radial-basis-expanded interatomic distances, side-chain directions — while edge features carry the relative orientation, distance and sequence separation of the residues they connect. A graph transformer passes messages over that structure and a pooled graph embedding drives the binary toxic/non-toxic call. Both pretrained components are used as released and frozen: the only weights PeptiTox owns are the graph network and its classification head.

The model comes from Yanling Wang's group at Qilu Institute of Technology in Jinan, with collaborators at Macao Polytechnic University, and was published in the Journal of Chemical Information and Modeling in July 2025. It is one of three peptide toxicity predictors of broadly this shape to appear from three unrelated groups within a few months of each other — alongside PLPTP, which stays sequence-only with a BiLSTM over ESM-2, and StrucToxNet, which pairs ProtT5 embeddings with an equivariant GNN over ESMFold structures. The architectures differ in substance; the convergence reflects how cheaply ESMFold now supplies peptide structures, not a shared lineage.

#Key Features

• Structure as an explicit input: Toxicity is decided over predicted 3D contacts, so residues adjacent only after folding are adjacent in the model's input too.
• Two frozen backbones, one small trained head: ESM-2 and ESMFold are used off the shelf, keeping the trained artifact a compact graph network rather than a fine-tuned language model.
• Joint node and edge updates: Every layer refreshes edge representations from the residues they link, so contact features stay in step with residue features instead of being fixed at graph construction.
• Context gating across the whole peptide: A pooled graph-level vector gates each residue's representation at every layer, giving local updates a view of the peptide as a whole.
• Trained against noised geometry: Backbone coordinates and node features are perturbed during training, so the classifier cannot lean on the exact structure ESMFold happened to emit.

#Technical Details

Sequence features come from the 650M-parameter esm2_t33_650M_UR50D checkpoint, taking the 1,280-dimensional layer-33 representation per residue; these join 184 geometric features for a 1,464-dimensional node input, alongside 450-dimensional edge features. Structures are predicted with esmfold_v1 and reduced to N, Cα, C, O and a side-chain centroid per residue. The encoder is four graph transformer layers of 128 hidden units with four attention heads, each combining edge-conditioned multi-head attention, a position-wise feed-forward block, an edge-update MLP and the context gate; global mean pooling feeds a classifier narrowing 128 to 64 to a single logit. Training uses binary cross-entropy with Adam at a learning rate of 1e-4, weight decay 1e-3, dropout 0.2, coordinate noise of 0.1, batches of 128 and a patience of 50 epochs. The released benchmark is a balanced set of 3,864 peptides, half toxic, evaluated by stratified five-fold cross-validation and scored on accuracy, F1, MCC, AUROC, AUPRC, sensitivity and specificity.

#Applications

The intended use is triage: given candidate peptides — designed by a generative model, mined from venom or antimicrobial libraries, or enumerated as mutants of a lead — rank them by predicted toxicity and spend synthesis budget on the survivors. Because structure is predicted rather than measured, the model applies to sequences that have never been made, which is the regime peptide design operates in. The cost is that every prediction runs ESMFold first, making PeptiTox heavier per sequence than sequence-only predictors and bounding its accuracy by how well ESMFold resolves short, flexible peptides.

#Impact

PeptiTox is a compact demonstration that predicted structure is now cheap enough to use routinely for peptide property prediction, not only for structure prediction itself — an argument its two contemporaries reach independently. Its practical reach is set by what was released: the repository carries the model definition, the feature-extraction scripts and the benchmark CSV, but no trained checkpoint, no inference entry point and no license file, and there is no web server or model-hub listing. Reproducing a prediction means rerunning the cross-validation training from the published code.