De novo designed proteins (from language models) and naturally occurring proteins borrow heavily from other functionally similar sequences

Summary

Proteins designed de novo from language models interpolate the design space of their training data, mirroring how naturally occurring proteins also borrow heavily form each other. Brian Naughton in a series of twitter posts noted that esmGFP (generated using (ESM)3) has 58% sequence identity to any naturally occurring GFP, but 74% when considering mixing and matching across protein sequences (https://gistpreview.github.io/?81c106aa6e8519bcc77acdbde939a2fa/ESM3_GFP_alignment.html). Likewise, the CRISPR enzyme openCRISPR (generated using fine-tuned (ProGen2)) from (1) can be recreated to 98% completion using three Streptococcus sequences (https://gistpreview.github.io/?f6684bb2fefa838bfff5082465df2ff0).

Figures

Ref (1)

Ref Brian Naughton’s Twitter; red regions have two or more consecutive novel residues; the long loop in the top right is not conserved

1.

Ruffolo JA, Nayfach S, Gallagher J, Bhatnagar A, Beazer J, Hussain R, et al. Design of highly functional genome editors by modelling CRISPR–Cas sequences. Nature. 2025;645(8080):518–25. Available from: https://doi.org/10.1038/s41586-025-09298-z