Billions of protein sequences have been catalogued, but the biology behind most of them still remains to be characterized. Today we are releasing a world model of protein biology built to change that, giving researchers an accessible system for accelerating prediction, design, and discovery to drive new breakthroughs in curing disease.
We introduce ESMC,ESMFold2, and the ESM Atlas together to offer a unified, open set of tools for exploring and understanding proteins at evolutionary scale, opening a new frontier in the generative design of protein matter.
Build novel AI applications on a unified representation of protein space
From therapeutic protein engineering to fundamental insights into protein biology across life, ESMC delivers digital representations of proteins that match or exceed the capabilities of previous generation models while running significantly faster.
A next-generation protein language model trained on sequence data spanning the diversity of life, ESMC learns the rules of protein biology from billions of sequences and exposes them as a shared computational language. The result is a step change in performance and scalability, now available under an MIT license.
Predict structures for unknown sequences in a fraction of the time
ESMFold2 represents the next generation of ESMFold and is a state-of-the-art model that defines a new frontier for speed and fidelity. Researchers can now predict full atomic structures, including side chains, directly from a single sequence, with no multiple sequence alignments required.
Built on ESMC's learned representation, ESMFold2 also powers our open-source, lab-validated protocol for designing binders, giving researchers a more efficient path from target to functional candidate.
Discover functional relationships through the lens of ESMC
Identify functional relationships, compare proteins across evolution, and generate hypotheses for uncharacterized sequences across 6.8 billion proteins and 1.1 billion predicted structures.
Rather than organizing proteins by sequence homology or curated labels, ESM Atlas organizes them by features learned by ESMC and extracted via sparse autoencoders (SAEs), capturing interpretable biological signals such as binding activity and functional motifs, and surfacing patterns learned directly from training on billions of protein sequences.
Using a design approach based on inverting ESMFold2, our research team generated binders across both minibinder and antibody modalities for diverse therapeutic targets, produced large candidate sets in silico, and experimentally validated a subset.
ESM-designed proteins, confirmed in the lab to bind PD-L1, a key checkpoint target in cancer immunotherapy.
Throughout all targets, we recovered binders with nanomolar affinities and strong specificity. Characterized binders consistently bound their intended targets on the surface of cells.
Together, these tools make it possible for the scientific community to explore, understand, discover, and design proteins at an unprecedented scale and speed, turning the world model into a programmable substrate for rapid biological discovery and engineering.
Biohub's mission is to cure or prevent disease, and accelerating science is how we get there. That is why we are releasing all three as open source, and we invite you to join us on this mission.
Join the Slack community to hear updates, engage with our team, meet other researchers, and share your work.
Tom Sercu, VP AI Engineering and the Biohub Research Team
P.S. You are receiving this message because you are a registered EvolutionaryScale Forge user. We’re sharing a first look at our newest offerings as part of our broader mission to unite frontier AI and frontier biology to power your next breakthrough. You can unsubscribe at any time using the link below.
Chan Zuckerberg Biohub, Inc., 2682 Middlefield Road, Suite i, Redwood City, CA, 94063
