The post introduces ESMFold2 and the ESM-C family as an open release from the Chan Zuckerberg Biohub, with day-one availability through Tamarind's web interface and API. ESMFold2 targets protein structure prediction, while ESM-C provides protein language-model representations.
The article highlights ESMFold2's reported results on FoldBench, including antibody-antigen and protein-protein complex prediction. It also describes ESMFold2-Fast as a lower-latency variant intended to make repeated structure evaluation more practical.
The Biohub release also reports de novo binder design against several disease-related targets using minibinder and scFv formats. The post summarizes reported hit rates and a PD-L1 example, while attributing those outcomes to the authors' experimental work.
ESM-C, also called Cambrian, is presented as a protein language model trained on a large sequence collection. The article describes its use for embeddings in tasks such as property prediction, similarity search, clustering, variant-effect analysis, and generative workflows.
The release can be run through Tamarind's web, API, agent, and pipeline interfaces, where teams can combine it with other structure and design methods.