Design a variant library

This improves an existing protein along an existing axis — stability, expression, mild activity tuning. It is not de novo design, and the distinction matters for what you should expect from the ranking.

Provide a wild type

Accepted: FASTA, SnapGene, GenBank, EMBL, raw DNA, or raw protein. Upload a file or paste text; the cap is 8 MiB.

The engine detects what it received. Two cases need a decision from you:

• A plasmid or multi-CDS record surfaces a CDS picker. The build refuses to guess which coding sequence you meant — that returns kind: needs_cds_choice with the options listed.
• Raw DNA with multiple stop codons triggers six-frame ORF discovery, and you pick the frame.

You can also press Identify via NCBI BLAST to resolve what the protein actually is, which returns UniProt accessions and AlphaFold links for the structure viewer. That step sends your sequence to NCBI.

Set five parameters

Understand what the ranking is

ESM-2 computes ΔLL for every position × 19 substitutions — how much more or less plausible the mutant is than the wild type, given the rest of the sequence. Simulated annealing then searches combinations, using cumulative ΣΔLL as fitness with penalties for stop codons and repeated positions.

Functional retention falls steeply with mutation count:

Published benchmark figures are on /api/benchmarks — a 2026-07-17 run over three public deep-mutational-scanning datasets and 7,761 variants gave a median Spearman correlation of 0.5165 and median top-decile precision of 0.386, using the 35M model. Those are the only figures that exist; there are none for the larger models.

Read the library

Six columns, with the wild type prepended as a reference row. A real run over the 720 nt EGFP CDS at the defaults:

RANK   MUTATIONS    FITNESS   GC%    TM (°C)   BP
WT     wild type    —         48.9   60.1      720
V0001  C49D,W58L    2.300     48.8   60.1      720
V0002  C49D,W58I    2.076     48.5   60.1      720
…
V0030  C49S,W58F    0.842     48.6   60.1      720

Fitness is the cumulative ΣΔLL described above, so it is comparable within one library and meaningless across two. The wild type has no fitness by definition — it is the baseline everything else is scored against.

Every row carries ESMFold and AlphaFold buttons that open a structure for that specific variant, and an expander for the full DNA and translation.

Filter the table

The results table takes filter expressions directly:

gc>50
tm>58
fitness>2
bp<800
W58L
C49

A bare mutation like W58L matches variants carrying it; C49 matches anything mutated at position 49. Numeric expressions compare the corresponding column.

Codon optimisation happens automatically

Each variant is reverse-translated against the host’s codon-usage table, then synonymously scrubbed of BsaI (GGTCTC), BsmBI (CGTCTC) and NotI (GCGGCCGC) sites — on both strands, because Type IIS enzymes cut either. That is what lets the insert survive Golden Gate assembly and vendor synthesis screening.

The export carries forward and reverse primers, their Tm and GC, and a suggested annealing temperature of min(Tm) − 5 °C. A WT pseudo-row is prepended so you can order the unmutated reference alongside as a control.

Export

csv, tsv, fasta (protein), fasta-dna, gb, xlsx or json. You can also edit a single variant’s DNA in place — the engine re-derives the translation, the mutation diff and the PCR metrics, and rejects an edit that breaks the frame or introduces a premature stop, because vendors will reject it too.

Close the loop

Screen the library, then log the outcomes — the measured value for each variant you tested. Round 2 re-ranks using a surrogate model trained on what you actually measured, rather than on the prior alone.

Interaction radar scans for epistasis between positions, which is the thing the zero-shot ranking is structurally blind to. Both are capped at 500 aa of protein.