ESMFold2
#28 in Wetenschop- & Bio-KIBiohub · v2 · 9× · tolest 30. Juni 2026
ESMFold2 is an Open Source AI model from Biohub for predicting and designing 3D structures of proteins and biomolecular complexes (proteins, DNA, RNA, ligands) directly from sequence data. It builds on the sequence representations of the language model ESMC (6 billion parameters, trained on approximately 2.8 billion sequences) and uses a diffusion-based, "looped" Transformer architecture to generate atomically resolved structures. According to the provider, the model achieves state-of-the-art results in protein-protein and antibody-antigen interaction prediction and is additionally used for designing de novo protein binders (minibinders, scFvs). It is available as an Open-Weight model under MIT license and can be used locally, via Hugging Face, or through the hosted Biohub Platform API.