XaxiPiruli/protea-frozen-v226-2025-05-03

PROTEA frozen-data bundle — v226-2025-05-03

Frozen reference bundle produced by PROTEA for LAFA submission. Pairs with the ghcr.io/frapercan/protea/lafa:v0.7.0 inference container.

Training cutoff

• GOA UniProt release v226 dated 2025-05-03 (latest GOA release strictly before LAFA's earliest evaluation t0 of 2025-09-04).
• GO ontology release 2026-01-23 (superset of v226's GO terms; no annotation drops at load time).
• Reference proteins drawn from PROTEA's UniProt cache (574,627 canonical accessions, sequence cutoff ≤ UniProt release 2025_02).

Bundle contents

File	Size	Description
manifest.json	507 B	cutoff version + dates + source IDs + row counts + reranker provenance
reference_embeddings.parquet	1.1 GB	(accession, embedding[1024]) rows from ProtT5-XL
reference_annotations.parquet	21 MB	5,902,876 GO annotations rolled up to canonical accession
go_term_metadata.parquet	1.4 MB	48,251 GO term rows (go_term_id, go_id, aspect, name)
pca_state.npz	70 KB	mean (1024,) + components (16, 1024) for query PCA projection
anc2vec.npz	32 MB	Anc2Vec GO-term embedding dictionary (release 2020-10)
reranker/F.txt	994 KB	LightGBM booster for MFO (NK-MFO winner)
reranker/P.txt	778 KB	LightGBM booster for BPO (NK-BPO winner)
reranker/C.txt	994 KB	LightGBM booster for CCO (NK-CCO winner)
reranker/routing.json	929 B	per-aspect provenance + filename mapping

Method overview

PROTEA is a KNN-based protein function prediction system that augments nearest-neighbour retrieval with a per-aspect LightGBM re-ranker. For each query protein it (1) computes a single mean-pooled ProtT5 embedding, (2) finds the top-K most similar proteins in this frozen reference bank, (3) transfers GO term votes from those neighbours, (4) enriches each candidate term with a "v6" feature family (Anc2Vec ontology embeddings, neighbour centroids, query-PCA projection, optional taxonomy voters), and (5) scores each candidate with the aspect-specific LightGBM booster shipped under reranker/<aspect>.txt.

When no booster is available for a given aspect the candidates fall back to KNN distance ordering.

The boosters in this revision are the highest-Fmax single-tier specialisations from the v19-winners family (NK-MFO test_fmax=0.453, NK-BPO=0.467, NK-CCO=0.485). All three share feature_schema_sha=0d9b7219433f.

How to consume

# 1. Pull the inference image:
docker pull ghcr.io/frapercan/protea/lafa:v0.7.0

# 2. Download this dataset (single-shot snapshot):
huggingface-cli download \
    XaxiPiruli/protea-frozen-v226-2025-05-03 \
    --repo-type dataset \
    --local-dir /path/to/protea-frozen-v226-2025-05-03

# 3. Run a query batch (FASTA in, TSV out):
docker run --rm \
    -v /path/to/protea-frozen-v226-2025-05-03:/bundle \
    -v $HOME/.cache/huggingface:/hf-cache \
    -e HF_CACHE=/hf-cache \
    -v $PWD/queries.fasta:/queries.fasta \
    -v $PWD/predictions.tsv:/predictions.tsv \
    ghcr.io/frapercan/protea/lafa:v0.7.0 \
        --query_file /queries.fasta \
        --frozen_data_dir /bundle \
        --output_baseline /predictions.tsv \
        --aspect_separated

Output is a 3-column TSV: <query_accession> <go_id> <score> where score is the reranker output in [0, 1] for aspects covered by a booster, else 1 - cosine_distance.

Per-cutoff revisions

When LAFA extends its evaluation window past Mar_2026, PROTEA can publish a new protea-frozen-v<N>-<YYYY-MM-DD> revision (re-running the GOA loader + bundle exporter for the new cutoff). The same image (ghcr.io/frapercan/protea/lafa) accepts the new bundle without code changes; this lets LAFA test the recency effect of training-data freshness without re-engineering the inference path.

Reproducibility

• Inference container: ghcr.io/frapercan/protea/lafa:v0.7.0. Source: https://github.com/frapercan/PROTEA/tree/main/apps/lafa_container.
• Inference library: protea-method. All KNN, feature enrichment, and reranker scoring code lives here; the container is a thin shell that loads this bundle and calls protea_method.pipeline.predict.
• Bundle exporter: scripts/export_lafa_bundle.py.
• Method card: apps/lafa_container/METHOD_CARD.md.

Provenance pointers (manifest.json)

{
  "schema_version": "1",
  "cutoff_version": "v226",
  "cutoff_date": "2025-05-03",
  "embedding_config_id": "c0ae5b69-d6dc-41cf-a711-1739d3d2e170",
  "annotation_set_id": "5e1a7285-bb85-4816-88ab-2a32f7e96dd5",
  "ontology_snapshot_id": "35c3ad67-3002-47db-8f71-eeed69d22ad6",
  "n_refs": 574627,
  "n_annotations": 5902876,
  "n_go_terms": 48251,
  "embedding_dim": 1024,
  "feature_schema_sha": "0d9b7219433f",
  "reranker_routing": "per_aspect",
  "reranker_aspects": ["C", "F", "P"]
}

Authors

• Francisco Miguel Pérez Canales — PROTEA architecture, KNN + v6 feature pipeline, LAFA container.
• PhD co-supervisors: David Orellana-Martín (Universidad de Sevilla), Ana M. Rojas (CABD Sevilla).

Contact: frapercan1@gmail.com (primary), frapercan1@alum.us.es (academic).