Neurotox: Deep learning decodes conserved hallmarks of neurotoxicity across venomous species

Neurotoxic proteins drive the most pathophysiological effects of animal envenomation, yet it remains unclear whether neurotoxicity is encoded directly within the protein sequence or emerges from higher-order structure binding and interactions with their target receptor. To address this, we developed Neurotox, a sequence-based deep learning framework trained on 200,000 curated protein sequences, with balanced representation of neurotoxic and non-neurotoxic proteins across taxa, achieving high classification accuracy (96%) with strong performance on unseen toxin families. We further introduced a controlled sequence-representation warping strategy that selectively perturbs neurotoxicity-relevant features, inducing a systematic loss of predicted neurotoxicity while preserving primary sequence identity. Structural modeling using AlphaFold 3 showed that, for most top-ranked toxins, warping disrupted {beta}-sheet architectures and reduced interface precision, with all top candidates showing highly significant effects (p < 0.0001). These structural changes were accompanied by recurrent cysteine-centered substitutions, implicating disruption of conserved disulfide frameworks. A single exception retained its global fold (C RMSD = 2.8 [A]), maintained low PAE, high pLDDT, and high pDockQ scores, and preserved a close arginine-glutamate contact (Arg53-Glu75), yet still exhibited marked attenuation of predicted neurotoxicity. These results suggest that neurotoxicity arises from distributed sequence features that shape secondary-structure organization and receptor interaction, rather than from isolated contact residues alone. TeaserDeep learning suggests the identification of neurotoxicity hallmarks directly from amino acid sequences across diverse species and toxin families.