Design extremophilic enzyme variants

Input an enzyme sequence; the pipeline proposes phenotype-shifted variants that preserve the active site, then folds and scores them.

Job
Standard 20 AA + ambiguity codes (B, Z, X, U, O). Whitespace and FASTA headers are stripped. Best results ≤ 1022 aa (ESM-2 window); 1022–1500 aa is windowed with a warning; > 1500 aa is rejected.
Designs auto-span conservative → aggressive.
Design Aggressiveness

A single knob controls how far designs move from wild-type. The N designs above automatically span the range, so one query returns a conservative–to–aggressive spectrum. The active-site freeze and catalytic-RMSD gate are invariant at every level — aggressiveness moves only the mutable-surface budget, never the catalytic core.

Leave blank to use the coordinated per-level schedule. Any value here pins that lever across all designs. Hover a label for its definition.

Gibbs iterations are convergence-driven, not set here ⓘ

Target Phenotypes
Pipeline Components
Generator — Masked Language Model

Proposes substitutions on the mutable positions. The extremophilic adapter is the LoRA fine-tune from this project.

Structure Prediction (folding)

Folds both wild-type and designs. Folding both with the SAME method makes the active-site RMSD reflect the sequence change, not cross-method bias.

MSA / Conservation

Builds the alignment for sequence-weighted conservation scoring (directs where the generator is allowed to mutate).

Active-Site Annotation

Identifies catalytic / functional residues to freeze. Sequence- and annotation-based sources; combined across all selected.

Structural Homology (Foldseek)

Structural search — transfers active-site geometry from homologs and scores fold-compatibility of designs.

Structural Gate

Rejects designs whose backbone the inverse-folding model finds implausible; survivors are refolded and gated on catalytic-atom RMSD.

Scoring

Ranks designs by the per-phenotype classifier (this project's Stage-2 heads).