Multiple-sequence alignment to expose co-evolution. None = fastest. MMseqs2 = fast default. JackHMMER = slowest, most sensitive. Better MSA → better structure/complex accuracy, especially for conserved proteins; sparse families may gain little.

Searches PDB for homologous templates (requires MMseqs2). Injects structural priors that can boost accuracy and domain placement. Helpful for difficult folds or multi-domain targets; may bias toward the template if over-similar.

Refinement passes through the network. More passes usually reduce clashes and improve local geometry with diminishing returns. Common: 3–6; difficult targets: 8–12; beyond that rarely helps but costs compute.

Sampling steps for the generator. More steps → better convergence/diversity, more compute. Quick screens: ~100-300. Balanced: ~300-600. Stringent finals: ~600+

Sets the random seed for sampling. Keep fixed to reproduce a result; change to explore diverse solutions under identical settings.

Number of independent structures Chai-1 will generate. You’ll get n different CIF structures.