Biohub's AI designs lab-proven antibodies in days

ESM is an open-source biological "world model" suite from Biohub that lets researchers computationally predict structures and design functional protein binders and antibodies from scratch on a computer. It condenses what used to be months of wet-lab candidate screening into a few days of compute.

The system consists of three parts: ESMC, a foundation protein language model trained on roughly 2.8 billion sequences across all forms of life; ESMFold2, the structure prediction and de novo design engine; and ESM Atlas, a catalog spanning 6.8 billion sequences and 1.1 billion predicted structures. All three components are free and open for researchers.

How it works

Unlike AlphaFold's geometry-heavy pipeline that typically relies on external multiple sequence alignments (MSAs), ESMFold2 operates directly on representations learned by the ESMC language model. By training on billions of raw sequences, ESMC internalizes the folding and interaction rules of biology without needing explicit structural supervision during pretraining.

Architecturally, ESMFold2 uses a looped transformer. Instead of adding raw parameter count, it feeds representations through the same set of parameters across multiple recurrent passes, backpropagating gradients through the entire loop during training. Each pass refines the 3D atomic coordinates based on the previous round.

This delivers two key advantages. First, you can scale compute at inference time: tough targets simply get more loops without retraining. Second, relying on recurrence rather than bloat avoids overfitting on the limited pool of experimentally resolved protein structures in public databases.

Real-world performance and lab validation

The candidate search workflow takes two steps. First, candidate generation explores ESMFold2's joint sequence-structure space to generate tens of thousands of designs in roughly two days. Second, scoring and ranking evaluate binding confidence and structural stability in under a day. Both phases scale effortlessly across GPUs.

For high-throughput screening, ESMFold2-Fast folds a 1024-length protein in just 9.4 seconds. That speed enables hundreds of thousands of candidate simulations per day.

Biohub validated designs in wet labs against five clinical cancer and immunology targets: EGFR, PD-L1, CTLA-4, CD45, and PDGFRβ. Across targets, compact de novo minibinders achieved hit rates between 36% and 88%, averaging 70% when allocating more inference compute. Harder single-chain antibody variable fragments (scFvs) reached 15% to 29% hit rates, nearly doubling from 12% to 21% under higher compute.

Binding affinities reached nanomolar levels. A de novo scFv designed against PD-L1 showed 4.3 nM affinity and successfully restored T-cell signaling in cellular assays. Cryo-EM imaging of an EGFR minibinder confirmed that the physical binder matched the model's predicted geometric orientation down to the atomic level.

Why it matters

While AlphaFold proved AI can predict how existing proteins fold, ESM reaches for the harder task: designing entirely new functional proteins that biology never created. The cryo-EM structural match proves ESMFold2 is generating legitimate solutions rather than memorizing existing databases.

For academic labs, biotech startups, and researchers working on neglected diseases or underfunded targets, this changes the economics of drug discovery. You no longer need millions of dollars in automated physical screening infrastructure to find an initial therapeutic lead; you just need a cluster and a few days of runtime.