Event Overview
The full extent of protein biology is beyond what experimental approaches in the physical laboratory can characterize. Accurate digital representations of proteins could therefore accelerate discovery through virtual experiments. In our preprint, we describe ESMFold2, a model that processes ESMC’s language model representations through a looped architecture to predict atomic resolution protein structures directly from sequence. ESMFold2 combines a simplified folding module with a stable recurrence mechanism, allowing the model to benefit from deeper computation at inference time, and reveals a log linear relationship between language model training compute and downstream structure prediction accuracy. As a result, ESMFold2 achieves state-of-the-art performance across benchmarks for protein and biomolecular complex prediction. This includes antibody antigen interactions, a critical modality for the design of medicines, with further gains when MSA context or additional recurrent loops are provided at inference.
This session will cover the science behind the model, followed by a hands-on demo showing how to fold structures using ESMFold2, whether you’re working through the API, the Fold App, or running it locally.
Speakers
Zeming Lin
Zeming Lin is a principal research scientist on the AI research team at Biohub. He was formerly EvolutionaryScale and Meta with a PhD from NYU. His work spans PyTorch and the ESM series of protein language models.
Ishaan Mathu
Ishaan Mathur is a machine learning engineer on the AI Products team at Biohub. He develops the inference API platform which provides free, on-demand access to the ESM family of models. He has a background in ML infrastructure and MLOps.