Precomputed alignments for every reviewed human protein

Reuse precomputed MSAs for proteins and complexes.

Inspect alignment artifacts for one human protein, or prepare native model inputs from one protein, multiple protein chains, sequences, or an existing model file. GreenFold prepares the reusable MSA evidence; structure prediction still runs in your model's own environment.

One protein

Explore or download an MSA

Inspect paired and unpaired A3Ms, retained STO sources, sequence variants, and alignment statistics.

One protein or multiple chains

Prepare input for a model

Build a single-chain or multi-chain input from UniProt IDs, sequences and entities, or an existing model file.

One UniProt ID Multiple protein chains FASTA or entities Existing model file

Start here

Prepare for a specific model

Choose the model whose native input you want to build, then describe one protein or a multi-chain complex.

Don't need a model-specific bundle? Direct download MSA (A3M) →

API access

Use GreenFold through the API

Both options are generated from the same OpenAPI contract. Choose the interface that fits your task.

How to use GreenFold

From model choice to native inference

GreenFold prepares reusable alignment evidence for the model you selected. The original prediction software and its model weights remain unchanged.

01 Choose a model Select AlphaFold 2, AlphaFold 3, Boltz, OpenFold3, or Protenix.
02 Describe the target Provide the protein or complex information using that model's supported input.
03 Prepare native inputs GreenFold resolves indexed MSAs, normalizes pairing, and stages model-specific files.
04 Run the original model Download the prepared bundle or runner and execute it in the model's own environment.
20,420 reviewed human proteins Complete reviewed-human UniProtKB export frozen on 2025-10-21.
507 days sequential search time Proteome-scale JackHMMER work completed in the offline layer.
≈97,000 allocated CPU-hours Eight CPU threads allocated per precomputed database search.

What the service adds

More than an A3M download

GreenFold preserves how the alignments were produced, understands sequence edits, and translates the same evidence into the different pairing and input contracts used by modern structure-prediction models.

01

AF3-compatible JackHMMER provenance

The indexed corpus comes from the AF3 protein-search route across UniRef90, MGnify, small BFD, and the UniProt paired-search channel. MMseqs2 is not substituted for JackHMMER in this primary precomputation.

02

A complete, frozen reviewed-human snapshot

The searches span four documented AF3-compatible channels: UniRef90, MGnify, small BFD, and the UniProt paired-search source. The current artifact index retains these channel identities but does not encode a database-release field for each artifact.

03

Search compute ready before the request

For an indexed target, the user retrieves and adapts existing evidence instead of waiting for four database searches to run again. Online work begins with artifact resolution, transformation, and model-specific packaging.

04

Variant-aware alignment rewriting

GreenFold recognizes substitutions, deletions, truncations, and short insertions, updates the requested protein sequence, and rewrites the corresponding query-relative MSA columns. Related constructs can therefore reuse the indexed search rather than launch a new one.

05

Model-aware pairing for protein–protein interaction prediction

A paired.a3m download is not universally prediction-ready because each model coordinates homologs differently. GreenFold prepares pairing according to the selected model's input contract before native inference, rather than handing every predictor the same file and assuming it will interpret the rows correctly. For example, Boltz requires per-chain CSV alignments whose shared non-negative keys coordinate homologous rows across chains; GreenFold constructs this representation from the indexed paired and unpaired A3Ms.

Documentation

User-facing guides for API usage and model workflows.