Structure prediction

ESMFold 2

Fast all-atom structure prediction for biomolecular complexes.

ESMFold2 predicts all-atom structures for protein complexes, DNA, RNA, ligands, modified residues, and covalent complexes through the released 6B-parameter model.

Documentation ↗ Input parameters ↗ Examples ↗ Home page ↗ Paper ↗ Source code ↗ License ↗

No jobs before this; it will run immediately.

Not logged in. Register an account to keep track of your jobs and results, and add presets and bookmarks. Email results to this address when complete:

Ready

Example preset

Choose an example to fill in its settings and input structures.

Input mode

Input

Job name Used to name the predicted mmCIF file.

Molecules

Add one box per unique ESMFold2 entity. Set its ID (or comma-separated IDs for identical copies), sequence or ligand, zero-indexed residue modifications, and optional protein/RNA MSA. An MSA is an absolute .a3m path on the compute node, a {"sequences": [...]} object whose first row is the query, or ESMFold2's serialized form; left blank, the chain folds from its sequence alone, which is ESMFold2's usual mode. Ligands may contain one or more comma-separated CCD_ codes.

Pocket conditioning (JSON, optional) Native ESMFold2 pocket object with binder_chain_id and contacts.

Distogram conditioning (JSON, optional) Native ESMFold2 distogram_conditioning array with chain_id and JSON matrix values.

Covalent bonds (JSON, optional) Native ESMFold2 covalent_bonds array using chain_id, residue index, and atom index fields for each endpoint.

StructurePredictionInput (JSON) JSON representation of ESMFold2's StructurePredictionInput. Sequence entries use type protein, dna, rna, or ligand; polymers take id and sequence, while ligands take id plus either smiles or a ccd list. Native pocket, distogram, covalent-bond, modification, and serialized MSA fields are passed through. A protein or RNA msa may also be an absolute .a3m path, which is read with MSA.from_a3m. In this form: Modification positions are zero-indexed, exactly as in the official ESMFold2 Python dataclass. Use the JSON-safe shape produced by serialize_structure_prediction_input.

StructurePredictionInput file (JSON) The JSON file defining this prediction, in ESMFold2's own StructurePredictionInput format. In this form: Choose a file. Its contents are read here and sent as the input document.

Inference

Recycling loops Recycling loops passed to ESMFold2InputBuilder.fold. The released API default is 20.

Diffusion sampling steps Diffusion sampling steps passed to ESMFold2InputBuilder.fold. The released API default is 200.

Diffusion samples Number of independently sampled structures. More than one produces one mmCIF per sample.

Random seed Seeds input preparation and diffusion sampling.

Language-model dropout Inference-time LM embedding dropout. The released folding API defaults to 0.3; set 0 to disable it.

Runtime

Pair-operation chunk size Chunk size for memory-heavy pair operations. The released model defaults to 64; smaller values use less memory. Set 0 to disable chunking.

Device Auto selects CUDA when available and otherwise CPU. The published 6B model is intended for GPU inference.

Model precision Use the checkpoint's published dtype, or explicitly cast model weights to BF16 or FP32.

Ready