NTxPred2

Neurotoxins are among the most potent bioactive molecules known, and the ones that matter for drug development are not only snake and cone snail venoms. Any engineered peptide or recombinant protein intended for human use, and any novel protein entering the food supply through a genetically modified crop, carries some risk of acting on the central nervous system. Establishing that risk experimentally means animal testing — slow, expensive, and increasingly discouraged — hence the interest in predicting neurotoxicity from sequence.

NTxPred2 comes from the Raghava group at IIIT-Delhi, which released the original NTxPred in 2017. That first tool, like most of its contemporaries, used a single model for both short neurotoxic peptides and full-length neurotoxins. NTxPred2's central design decision is to abandon that: peptides and proteins get separate models, trained on separate datasets, because a classifier fitted to one length regime degrades badly on the other. The paper demonstrates this directly with a cross-dataset evaluation, applying each model to the other's data. It appeared as a bioRxiv preprint in March 2025 and in Protein Science in July 2025.

The models are built on ESM-2. Earlier neurotoxicity work relied on composition and evolutionary features; NTxPred2 replaces them with protein language model representations, either fine-tuning ESM-2 end to end or feeding its embeddings to a tree ensemble — the same recipe the group applied to anticancer proteins in AntiCP3.

#Key Features

• Separate peptide and protein models: The peptide model covers sequences of 7 to 50 residues, the protein model sequences of 51 and above, with a third combined model for mixed-length input.
• Fine-tuned ESM-2 checkpoint released: The peptide classifier is an esm2_t30_150M_UR50D fine-tune with a linear classification head, published on HuggingFace in safetensors format under GPL-3.0 behind a contact-sharing gate.
• Embedding plus tree ensemble for proteins: Longer sequences are scored by an Extra Trees classifier over ESM-2 embeddings reduced by recursive feature elimination, which outperformed end-to-end fine-tuning in this regime.
• Scan and design modules: Sliding windows localize neurotoxic regions inside a protein, and a design module generates single-residue mutants — or the full mutational library — and scores each.
• Multiple delivery routes: A web server, a pip install ntxpred2 package, a conda environment, a GPL-3.0 GitHub repository, and a Zenodo archive carrying the code and the curated dataset.

#Technical Details

Sequences come from reviewed UniProt entries: 5,173 candidate neurotoxic peptides retrieved by keyword, and 566,303 non-toxic sequences retrieved with a NOT toxin NOT neurotoxin filter. After removing non-canonical residues and duplicates, the peptide set holds 877 neurotoxic peptides, balanced by 877 length-matched non-toxic peptides drawn from CD-HIT clusters at 60% identity. The protein set reduces 3,755 neurotoxic proteins to 775 non-redundant sequences with CD-HIT at 40% identity, matched by 775 non-toxic proteins. Each set is split 80:20, with the 20% independent portion withheld from training, testing, model selection, and hyperparameter tuning.

Composition and binary profile features with classical classifiers reach an independent AUC of 0.97 on peptides but only 0.85 on proteins, rising to 0.89 when PSSM evolutionary profiles are added. Classical models over language model embeddings do better on proteins: ProtBERT embeddings with logistic regression give 0.963 on peptides, while esm2-t30 embeddings with Extra Trees give 0.934 on proteins and 0.949 on the combined set. Fine-tuned ESM-2 classifiers were evaluated at four scales; esm2-t30 was best on peptides at an independent AUC of 0.984 and MCC 0.898, and esm2-t6 best on proteins at 0.912. The final deployed configurations are the esm2-t30 fine-tune for peptides, an Extra Trees model over 48 RFE-selected esm2-t30 embedding dimensions for proteins (AUC 0.937, MCC 0.731), and an Extra Trees model over 84 SVC-L1-selected dimensions for the combined case (AUC 0.954, MCC 0.780). On the same independent sets, the best NTxPred model reached 0.922 on peptides, 0.715 on proteins, and 0.783 combined.

#Applications

The tool is aimed at safety assessment. Peptide therapeutic programmes can filter neurotoxic candidates out of a design library before synthesis, and food-safety workflows can screen novel crop proteins without an animal study. The scan module answers a different question — given a protein of interest, which region carries the neurotoxic signal — useful in venom proteomics and in engineering toxin scaffolds for therapeutic delivery, while the design module supports iterative detoxification of a lead sequence.

#Impact

NTxPred2's main contribution is methodological restraint applied well: it establishes empirically that peptide and protein neurotoxicity need separate models, then reports which representation wins in each regime rather than declaring language models the answer everywhere. The datasets are small by foundation model standards — 1,754 peptides and 1,550 proteins — and the negative classes are defined by UniProt keyword exclusion rather than by demonstrated absence of neurotoxicity, so the metrics are best read as an upper bound. The HuggingFace checkpoint covers the peptide model only and is gated behind a contact-sharing agreement; the protein and combined models ship in the standalone package and the Zenodo archive.