A protein can bind a toxin in a lab assay without functioning as an antivenom. That distinction is central here: the release makes a sizable AI-designed protein campaign inspectable, while clearly stopping short of claiming therapeutic efficacy.
Key Findings
- Claude generated 318,484 candidate designs, and 211 were tested by Adaptyv Bio and Twist Bioscience. Seventeen met the study’s criteria for binding, giving researchers data to examine both successful designs and the broader selection process.
- The binders target four of five neurotoxins: seven bind Cn2, four AaH II, four AaH I, and two LqhIII; none met the criteria for Ts1. The uneven results make this useful for studying where the approach worked and where it did not.
- The release includes sequences, assay results and kinetics, sensorgrams, structures, computational scores, and supporting tables. This makes it possible to compare predictions with experiments rather than relying on a headline success count.
- Cn2 competition experiments suggest some designs bind at or near the site used by the antibody LR, but neither cellular neutralization nor efficacy in animals has been established.
Who It’s For—and the Trade-off
A strong fit for protein-design researchers who want experimental data to analyze, benchmark, or build on. It is not evidence of a ready-to-use antivenom: the results concern binding to recombinant toxins, and the release does not demonstrate neutralization in cells or protection in animals.
