Computationally design de novo toxin-neutralizing proteins with deep learning

The proposal

Use deep-learning protein design (e.g., RFdiffusion + ProteinMPNN + AlphaFold2 filtering) to computationally design small proteins that bind and block conserved venom toxins, then manufacture them by microbial fermentation. Unlike antibodies discovered by immunizing animals or screening large libraries, these binders are generated in silico from a toxin's structure alone, targeting the family's conserved features for breadth.

Why it would work

De novo design sidesteps the biggest cost and consistency problems of current antivenom: no animal immunization, no large-library screening, recombinant microbial production with minimal batch-to-batch variation, and very low cost at scale. The designed proteins are tiny (~100 amino acids) and extremely heat-stable, which aids deep-tissue penetration and could ease cold-chain requirements. Crucially, they can be designed against toxins (like three-finger toxins) that are poorly immunogenic and therefore weakly covered by conventional antivenoms.

Evidence

Vázquez Torres et al. (Nature, 2025; Baker lab with DTU and Liverpool School of Tropical Medicine) de novo designed proteins binding short-chain and long-chain α-neurotoxins and cytotoxins from the 3FTx family. The neurotoxin binders had sub-nanomolar affinity and melting temperatures up to >95 °C, neutralized their targets in vitro at 1:1 ratios, and gave 100% protection in mice against lethal short-chain neurotoxin and α-cobratoxin challenges (including when given as rescue 15 minutes post-envenoming).

Implementation path

Design and validate binders for each dominant toxin family; combine complementary binders into broad cocktails; optimize the weaker binders (e.g., cytotoxin) to in vivo efficacy; progress through preclinical then clinical development. Near-term, the authors propose using them as "fortifying agents" to strengthen conventional antivenoms, especially for elapid bites.

Trade-offs and limitations

All results are preclinical. Each binder targets a single toxin or toxin class, so broad coverage requires designing and combining many. In vitro potency did not always translate: the cytotoxin binder did not reduce dermonecrotic lesions in vivo, needing further optimization. Work so far is limited to elapid 3FTx toxins; viper enzymes (SVMP/SVSP) are not yet addressed.