Thank you very much ☺️

This work has been possible thanks to an incredible team effort w/ Alejandro, @marcguellc.bsky.social @noeliaferruz.bsky.social and the integra therapeutics team!

👀

But independent of this, what I find is that since Bill Degrado coined the term in the 80s, the field has evolved so much that it no longer captures today’s methods.

E.g in this (all-time favourite!) review, Baker and Po-ssu define de novo in a way that would exclude most AI methods today:

That’s an interesting paper, I must read 🙂

A bit off-tangent now, but I forgot to mention that Pubmed isn’t normalized, and all terms look as they are increasing exponientially, see for example banana 😄 (1/n)

Interestingly, diffusion models can naturally sample proteins with 0% seq identity!
In any case, last time I said we had generated de novo sequences with a pLM a bunch of reviewers avalanched toward my paper with knives and axes and I had to use “artificial”🤣

Before AI we reserved that term for proteins with no evolutionary history, i.e., designed from first principles, and consequently often had sequence identities ca. 0%. pLMs, unless steered, tend to naturally sample at >30-40% seqid and rarely explore novel folds.

Uncancelled! 🥳

Wohoooo 😍😍😍🚀🚀🚀