🧵Today is an exciting day! We have just announced ESM3, the design of a new green fluorescent protein, and the launch of our company. Here's a short thread about the scaling and infrastructure work involved in making ESM3 happen! 🧵

read the license before using ESM-3; github.com/evolutionarysc…. you're welcome, pharmas and biotechs.

What if we could universally recombine, insert, delete, or invert any two pieces of DNA? In back-to-back nature papers, we report the discovery of bridge RNAs and 3 atomic structures of the first natural RNA-guided recombinase - a new mechanism for programmable genome design

I think it's a relevant metric, it's just not very hard. If someone could take a GFP and make it 10000x brighter than any known ones with 192 shots I'd be far more impressed, even if it was at 3 mutations. ESM3 excites me most because it is a relatively high success rate model

I’m really excited to share our new pre-print “Functional protein mining with conformal guarantees”, with Seyone Chithrananda Anastasios Nikolas Angelopoulos Peter Hyungjun Yoon Michael Jordan @doudna_lab biorxiv.org/content/10.110… 🧵

Thomas Wolf I'm surprised the MuMath-Code paper wasn't more discussed on Twitter. It is awesome! Super dense in information. Only reference I found was: x.com/gm8xx8/status/… who incidentally tweets about a lot of really interesting papers that don't get much attention.

Does more compute equate with greater risk? What is our track record at predicting what risks emerge with scale? I don't have much time anymore to write, but this felt important enough to write down some thoughts on. arxiv.org/pdf/2407.05694…

👇👇 for biotech

Then we repeated the experiment, but we dehydrated the tissue first. The architecture was fine. We left it at room temperature for a year. Still fine. Ran over it with a car, had a former Astros starting pitcher throw a fastball at it, fired a shotgun at it. Still fine. (15/18)

📈 One standard graph in all bio deep learning papers should be: max similarity to anything in training set vs performance. (Reviewers shouldn't have to guess if there might be overfitting issues).

This last chapter "Horoscopes" is only 3 pages but I keep thinking it should be at the start of the book, because it is about scientific and social context of statistics. Anyway here it is:

Complex numbers are weird and counter-intuitive. For instance, did you know that

Short, interesting antibody design paper using available tools (EvoDiff, ColabFold, HADDOCK, OpenMM). No wet-lab validation, but interesting to see the pipelines people are putting together. Appears to be related to silico.bio biorxiv.org/content/10.110…

I wrote about some recent studies identifying widespread gene-environment interactions (GxE) acting on common traits, but through an unexpected mechanism. A few highlights 🧵:

A central challenge of AI-driven biotech is that in drug discovery a fully validated medicine is worth a million times more than a million early-stage ideas that only have little validation. This is unfortunately at odds with the AI domain where we almost always attempt to

Thrilled to announce a preprint describing collaborative work between our group and James Fraser (A$AP J), @BShoichet, and Adam Renslo. The paper shows a practical application of the Thompson Sampling method we published earlier this year. biorxiv.org/content/10.110…

“Shotgun” is a way better name than bootstrap, it isn’t even close.

What if a [MASK] was all you needed? ModernBERT is great, but we couldn't stop wondering if it could be greater than previous encoders in different ways. Maybe we don't need task-specific heads? Maybe it can do all sort of tasks with only its generative head? Spoilers: Yes

Has nominative determinism gone too far (h/t Benjamin)