Sign in

Coban Brooks

@cobanbrooks.bsky.social
51 followers 305 following 9 posts

Protein engineer/bioML/PhD student. Building robot scientists @ Romero Lab/Duke Univ.

PostsRepliesMedia
Reposted by Coban Brooks
Kevin K. Yang 楊凱筌 @kevinkaichuang.bsky.social · 20/08/2026
Binder design is nice and all, but here we have three agents sharing data and autonomously controlling a lab to design enzymes with shifted substrate scopes and high activity! @cobanbrooks.bsky.social @pascalnotin.bsky.social @philromero.bsky.social www.biorxiv.org/content/10.6...
0205
Coban Brooks @cobanbrooks.bsky.social · 20/08/2026
You can read more here: www.biorxiv.org/content/10.6...
biorxiv.org
Learning protein function through autonomous experimental interaction
Biological AI learns primarily from existing observations, but many questions cannot be answered from available data alone. Here we show that AI can instead acquire knowledge by acting directly on bio...
000
Coban Brooks @cobanbrooks.bsky.social · 20/08/2026
We let it operate like this for a month, during which it discovered enzymes with significantly shifted substrate specificities that persisted under purified enzyme kinetics experiments we performed afterwards. [7/n]
110
Coban Brooks @cobanbrooks.bsky.social · 20/08/2026
We built a robot lab capable of assembling genes, expressing them, and performing enzyme assays, and then a computational agent that could learn from the corresponding data and propose new sequences to test. [6/n]
110
Coban Brooks @cobanbrooks.bsky.social · 20/08/2026
We propose a way out of this regime: instead of being observational, biological AI should be interactive. It should be able to perturb biological systems, measure the response, and learn from feedback. [5/n]
100
Coban Brooks @cobanbrooks.bsky.social · 20/08/2026
Oftentimes, what these models view as fitness — how much a given sequence looks like a naturally-observed one — is completely orthogonal to the fitness protein engineers care about, which is concerned with how the protein actually functions in real life! [4/n]
100
Coban Brooks @cobanbrooks.bsky.social · 20/08/2026
This means that the current paradigm in bioAI is largely observational - these datasets are static, and even the best models fail when tasked to engineer proteins with functions not well represented in these datasets. [3/n]
100
Coban Brooks @cobanbrooks.bsky.social · 20/08/2026
Biological AI has gotten really good at a lot of things: protein structure prediction, sequence generation, designing biomolecules with new properties. It’s not uncommon that the cool new protein model is in some way built on top of some sort of large-scale pretraining. [2/n]
100
Coban Brooks @cobanbrooks.bsky.social · 20/08/2026
I’m really excited to share our new preprint, out today, where we built a fully autonomous system for enzyme engineering — integrating a self-driving lab with a generative protein language model that worked together over a month to engineer substrate specificity in glycoside hydrolase enzymes🤖 🧬
100
Coban Brooks @cobanbrooks.bsky.social · 29/11/2025
Excited to be at #EurIPS + @workshopmlsb.bsky.social next week in Copenhagen talking about our work building self-driving labs to engineer new proteins. If you’re into lab automation, active learning, Bayesian opt, or proteins in general let’s meet!
010