1
Optional settings
Use YAML file input
Upload a ESMFold2 YAML configuration file instead of filling out the UI. Include any MSA paths your chains need - directly-uploaded YAML is used as-is (MSAs are not auto-computed for direct upload).
Number of Seeds
Number of jobs to submit. Each job uses a random seed by default.
Seed (optional)
Set a specific seed for reproducibility, or leave blank for a random seed.
Model details
ESMFold2 (full): 48 pair folding layers + MSA encoder, best accuracy. ESMFold2-Fast: 24 pair folding layers, no MSA encoder - ~1.7x faster, much less memory, slight accuracy trade-off. Choose the model above the optional settings.
Output Format
File format used for predicted structures.
Recycling loops
Number of folding-trunk recycling loops. Leave blank for the default (ESMFold2 Full: 10, the paper config). More loops can improve accuracy at higher runtime cost (sequential, so no extra GPU memory).
Blank = default.
Diffusion steps
Number of diffusion sampling steps for the structure module. Leave blank for the default (ESMFold2 Full: 68, the paper config). More steps can improve accuracy at higher runtime cost.
Blank = default.
Use Multiple Sequence Alignment
When disabled, ESMFold2 will not use MSA data for prediction.
MSA Max Depth
Cap the number of MSA sequences passed to the model (top of the a3m, taken in file order - typically the most-informative as ColabFold ranks them first). Set to 0 for no cap.
Recommended: 1024. Use 0 to disable cap.
Custom A3M Files
Upload custom a3m (multiple sequence alignment) files to use instead of generating MSAs automatically. Each a3m file should correspond to a protein chain in your complex. MSA generation is skipped for chains you provide an a3m for.
Save raw model tensors
Include per-sample PAE, distogram, and atom-coordinate tensors as .pt files. Useful for downstream re-analysis. Off by default - these files can be 100 MB to several GB per sample for multi-chain complexes.
All sequences must be non empty.
Structure Prediction