StrucToxNet

Peptides fold. They adopt α-helices and β-sheets and assemble into larger arrangements, and that conformation governs their biological function and physicochemical behaviour. The computational screens that flag toxic candidates before synthesis — the ToxinPred series, CSM-Toxin, ToxIBTL, CAPTP — nonetheless read the primary sequence alone, so the conformational half of the signal never reaches the classifier.

StrucToxNet restores it by pushing each peptide through two frozen pretrained models at once. ESMFold, via the released esmfold_v1 checkpoint, predicts a structure with no multiple sequence alignment, and the ProtT5-XL-U50 encoder from the ProtTrans suite produces a 1024-dimensional context-aware embedding for every residue, used without additional fine-tuning. Neither backbone is retrained. The only weights fit for this task belong to an E(3)-equivariant graph neural network over a residue graph built from the predicted coordinates, plus its classifier.

The work comes from Shihu Jiao and colleagues at the University of Electronic Science and Technology of China and the University of Tsukuba, published in BMC Biology in July 2025. Peptide toxicity has drawn several language-model-based predictors in close succession, and they are separate models: PLPTP fine-tunes an ESM-2 backbone into a bidirectional LSTM from sequence alone, while PeptiTox, from an unrelated group, also pairs language model embeddings with ESMFold structures in a graph network. StrucToxNet is the one that keeps both backbones frozen and makes the geometry explicit through equivariant message passing over Cα coordinates.

#Key Features

• Two frozen backbones, one trained head: ProtT5-XL-U50 supplies residue embeddings and esmfold_v1 supplies coordinates, both used exactly as released; only the graph network and its prediction module are fit.
• Structure as a Cα contact graph: Residues whose Cα atoms fall within 8 Å become graph neighbours, and their pairwise distances are expanded through a Gaussian radial basis function into edge features.
• Equivariant message passing: Each layer updates node embeddings and 3D coordinates together, leaving the representation unchanged by translation, rotation, reflection and node reordering, without spherical harmonics.
• Geometric descriptors beside the embeddings: A 45-dimensional per-residue vector covering secondary structure, solvent accessibility, local geometry, residue orientation, sequence and spatial position, virtual surface area and contact count.
• Contrastive term for class imbalance: Following CAPTP, a contrastive objective balances a training set where non-toxic peptides outnumber toxic ones by roughly two and a half to one.

#Technical Details

Training and evaluation data are taken from the earlier CAPTP work: 1,818 toxic and 4,569 non-toxic peptides for training, with an independent test set of 320 toxic and 806 non-toxic. Both splits are drawn from that same curated collection, so the independent set is a held-out partition rather than an external cohort. The network stacks three equivariant graph convolutional layers at hidden dimension 256, max-pools the final node embeddings into a peptide-level vector, and maps it through an MLP with a sigmoid output. Training ran for up to 200 epochs on an NVIDIA A100, with fivefold cross-validation on the training split used only to search hyperparameters.

On the independent test set StrucToxNet reaches 93.18% balanced accuracy, 0.968 AUC, 91.56% sensitivity, 94.79% specificity and 0.852 MCC, against 91.59% and 0.811 for CAPTP, the strongest sequence-only baseline. The ablation shows where the gain originates: replacing the ProtT5 embeddings with one-hot encoding drops balanced accuracy to 88.88% and MCC to 0.755, whereas removing the structural leg costs comparatively little (92.80% and 0.845), and dropping the contrastive term 91.49% and 0.830. Retraining after discarding peptides whose predicted structures score below pLDDT 60 lifts balanced accuracy to 94.98% and AUC to 0.977.

#Applications

The intended use is early triage of peptide drug candidates: run a library of designed or natural sequences through the model and rank them by predicted toxicity before committing to synthesis and assays. That suits groups mining venoms, food proteins or marine organisms for bioactive peptides, and teams designing antimicrobial or anticancer peptides who need a safety filter beside an activity filter. Because structures are generated on the fly, the pipeline runs on sequences with none determined experimentally — which is nearly all of them.

#Impact

StrucToxNet shows that geometry can be added to a peptide property predictor at low cost: no new pretraining, two released checkpoints, and an equivariant head small enough to train on a single GPU. The ablation is the honest framing — the language model carries most of the discriminative load, and structure supplies a smaller increment worth only as much as the predicted conformations are, with short and flexible peptides named by the authors as the weak case. Further limits are stated plainly: the framework handles linear peptides only, with no support for cyclic peptides or post-translational modifications, and it emits a single binary call rather than distinguishing hemolytic, immunotoxic or neurotoxic mechanisms. Code, the trained checkpoint and the peptide datasets are all published, on GitHub and Zenodo.