Deep learning has rapidly reshaped structural biology, enabling researchers to predict, analyse and design macromolecular structures at unprecedented scale. In venom research, these tools offer exciting opportunities to annotate toxins, explore structure–function relationships, guide construct and mutagenesis design, model toxin–target interactions, and inspire the design of venom-derived molecules with altered specificity or therapeutic potential.
However, predicted and designed structures can also be misleading. Computational methods will often return visually plausible models even when the underlying sequence information, evolutionary signal, biological context or training-set support is insufficient. Venom biology is a particularly useful lens through which to examine these tools, because venom-derived molecules often combine many of the features that remain challenging for structure prediction and design: disulfide-rich scaffolds, post-translational modifications, proteolytic processing, cross-species molecular recognition and interactions with dynamic membrane-protein targets such as ion channels and GPCRs.
In this talk, I will illustrate practical ways that deep learning-based structural models can be incorporated into venom-to-drug discovery workflows, while highlighting common pitfalls, limitations and best practices. Particular emphasis will be placed on interpreting confidence metrics, assessing predicted interfaces, understanding the limits of protein and peptide design, and treating computational models as structural hypotheses rather than experimental evidence.