Learn/Predict a structure
Recycles, samples and seeds
Three knobs mean the same thing on every model that has them. Recycles feed the prediction back into the model until it stops changing. Samples draw several structures and rank them by confidence. A seed fixes the randomness so a run can be repeated, or varied on purpose.
Recycles
| Metric | 1BRS | 1 recycle | 10 recycles |
|---|---|---|---|
| pLDDT | 95.38 | 96.11 | |
| pTM | 0.9649 | 0.9772 | |
| ipTM | 0.9549 | 0.9746 | |
| Ranking score | 0.9569 | 0.9751 | |
| gPDE | 0.3109 | 0.2963 |
Protenix on barnase and barstar after one pass and after its default ten, against the crystal structure. A small, well-folded complex converges early; the extra passes buy little. Large proteins and hard complexes keep improving for longer, with diminishing returns.
Samples
| Metric | 1BRS | Sample 1 | Sample 2 | Sample 3 | Sample 4 | Sample 5 |
|---|
Boltz-2 drawing five structures of the same complex from five starting noises. Where they agree the answer is settled; where they spread, the model is unsure and more draws, or more seeds, are how to find out. The best by confidence is the one the lab shows.
The same setting, a different argument on each model
Each model names these settings its own way in its API. The Fold lab shows one name and sends the model its own argument; the table says which model takes which.
| Setting | Model and argument |
|---|---|
| RecyclesHow many times the model refines its own prediction. More is slower and usually a little better. | AlphaFold 2 num_recyclesBoltz-2 recycling_stepsESMFold 2 num_loopsESMFold 2, fast num_loopsChai-1 num_trunk_recyclesProtenix n_cycleOpenDDE n_cycleBoltz-1 recycling_stepsMol-1 n_cycle |
| SamplesHow many predictions to draw; the best by confidence is kept. | AlphaFold 2 num_modelsBoltz-2 diffusion_samplesESMFold 2 num_diffusion_samplesESMFold 2, fast num_diffusion_samplesChai-1 num_diffn_samplesOpenFold 3 num_diffusion_samplesProtenix n_sampleOpenDDE n_sampleBoltz-1 diffusion_samplesMol-1 n_sample |
| SeedFixes the randomness so a run can be repeated. | Boltz-2 seedESMFold 2 seedESMFold 2, fast seedChai-1 seedOpenFold 3 seedProtenix seedOpenDDE seedBoltz-1 seedMol-1 seed |
Start at the defaults. Raise samples and seeds for a complex whose ipTM sits between 0.5 and 0.8, and recycles for a large protein that has not settled.
Read more: Recycling and convergence · Seeds and sampling