BUILD_RECIPE.md

Reproducing the 70%-identity representative FASTAs

The canonical implementation is dev/data/process_full_corpus.py. All Python environments are managed by uv; MMseqs2 is invoked as a pinned external binary and its version is recorded in each clustering receipt.

1. Pin and download the three source arms

Source	Frozen input
UniRef90	UniRef release 2023_02 archive, 211,819,312,677 bytes, MD5 353681f464572bb199fa032f714d4669
MGnify	Protein DB 2023_02 mgy_clusters.fa.gz, 83,473,342,442 bytes, MD5 332d36d2a943bdb769237a03e050ed03
OMG/IMG	All 959 Parquet objects in tattabio/OMG, 1,253,813,127,320 bytes total, pinned by path, size, and LFS SHA-256 in dev/data/omg_upstream_shards.tsv

2. Normalize and quality-filter

For every record, in order:

1. remove whitespace and a terminal *;
2. uppercase the sequence;
3. map unsupported amino-acid characters to X;
4. reject sequences shorter than 60 residues;
5. reject sequences with more than 20% non-canonical residues; and
6. in the OMG payload, keep numeric JGI/IMG accessions and remove ERZ MGnify- origin rows so MGnify is not represented twice.

The measured funnel was:

Source	Original records	Short	>20% ambiguous	Wrong source	Quality eligible
UniRef90	170,669,877	4,755,787	29,797	0	165,884,293
MGnify	729,215,663	117,372,608	26,021	0	611,817,034
OMG/IMG	3,280,269,924	145,672,826	0	1,048,850,563	2,085,746,535

3. Global exact deduplication

Accepted records are partitioned into 256 leading-byte SHA-256 buckets. Exact normalized sequences are collapsed globally while preserving every (sha256, source, source_id) membership in Parquet. A source-specific FASTA view is then emitted for every source in which the sequence occurred.

The source views contain 165,884,293 UniRef90, 611,788,129 MGnify, and 963,673,186 OMG/IMG unique sequences. Their global union contains 1,661,993,387 exact unique sequences; source-view counts are larger because one digest can belong to multiple source arms.

4. Source-specific diversity clustering

Each source view is clustered independently with MMseqs2 Linclust:

mmseqs linclust sequences clusters tmp \
  --min-seq-id 0.70 \
  -c 0.80 \
  --cov-mode 1 \
  --cluster-mode 2 \
  --threads 64

--cov-mode 1 applies the 80% coverage requirement to the shorter sequence. createtsv emits the representative/member digest pairs and result2repseq plus result2flat --use-fasta-header emits the representative FASTA.

5. Reproduction commands

uv sync --frozen
PIPE=dev/data/process_full_corpus.py
ROOT=/absolute/path/to/protein-corpus
MMSEQS=/absolute/path/to/mmseqs

uv run --frozen python "$PIPE" download \
  --data-root "$ROOT" \
  --omg-manifest dev/data/omg_upstream_shards.tsv \
  --download-workers 8

uv run --frozen python "$PIPE" normalize --source uniref90 \
  --input "$ROOT/raw/uniref90_2023_02/uniref2023_02.tar.gz" \
  --output "$ROOT/normalized/uniref90"

uv run --frozen python "$PIPE" normalize --source mgnify \
  --input "$ROOT/raw/mgnify_2023_02/mgy_clusters.fa.gz" \
  --output "$ROOT/normalized/mgnify"

# Expand all 959 OMG inputs from dev/data/omg_upstream_shards.tsv, then:
uv run --frozen python "$PIPE" normalize --source omg_img \
  "${omg_inputs[@]}" --output "$ROOT/normalized/omg_img"

uv run --frozen python "$PIPE" deduplicate \
  --input "$ROOT/normalized/uniref90" \
  --input "$ROOT/normalized/mgnify" \
  --input "$ROOT/normalized/omg_img" \
  --output "$ROOT/deduplicated"

for source in uniref90 mgnify omg_img; do
  uv run --frozen python "$PIPE" cluster \
    --dedup-root "$ROOT/deduplicated" \
    --source "$source" \
    --output "$ROOT/clusters/$source" \
    --mmseqs "$MMSEQS" \
    --threads 64
done

Every output directory is create-once. The builder refuses to overwrite an existing normalization, deduplication, or clustering directory.

6. Boundary of this RAW release

This release stops immediately after source-specific 70%-identity clustering. It does not perform evaluation homology exclusion, exact evaluation exclusion, cross-source representative ownership, length 32–16,384 filtering, validation selection, or Parquet packing. Those steps produce the final LuminBench-Nano-ESMC release.