Feature reference

The re-ranker reads a fixed set of feature columns per candidate protein/GO pair. This page answers, for the technician who deploys and operates the stack, what each column measures, who produces it, and whether that producer actually runs in the default export today. Where the API Core reference documents the code that reads these columns, this page documents the columns themselves.

The whole table below is generated at build time from protea_contracts.feature_docs.FEATURE_DOCS, the single feature registry shared across the PROTEA stack. It is never written by hand: edit a FeatureDoc in protea-contracts and this page follows on the next build. A drift lint, scripts/check_feature_docs.py, fails PROTEA CI if a declared feature loses its doc, if a doc names an undeclared column, or if a doc’s family disagrees with the canonical schema, so the reference cannot silently fall out of sync with the fingerprinted feature set.

How to read a row

Each column carries a status that tells the operator whether the value it sees in an export is a real signal or a placeholder:

PRODUCED:

A wired producer fills the column with a real value in the export. Some producers sit behind a performance flag that the canonical export enables; read the feature’s notes. A PRODUCED status means the producer runs, not that the value is non-trivial in a given deployment: a producer whose database source is empty (see interpro_*) still keeps its declared default, and the notes say so.

DECLARED_ABSENT:

The column is a first-class member of the schema and a producer exists, but no producer runs in the default export, so the export emits NaN and LightGBM reads it as missing. The six LAFA columns are in this state per ADR-D45.

POOL_INJECTED:

The PROTEA dump does not write the column at all; the lab’s pooled multi-manifest loader injects it as a per-source constant at stage time (for example plm_id and k_context). It is absent from the raw parquet dumps.

BROKEN:

The column is produced but a defect that can be pointed at in code or data makes it carry no signal. Used only where that defect is verifiable from the source tree.

The producer field names the exact callable (or the lab loader) that fills the column, so an operator can trace a suspect value back to the code path that wrote it. The notes field is where an operator learns the operational caveats: that the interpro_* columns keep their zero default whenever the InterPro GO-prediction table is unset for the deployment, and that plm_id/k_context never appear in a raw dump because the lab injects them.

The registry

There are 78 documented feature columns across 19 families.

Family knn

distance

Embedding distance from the query to the reference neighbour that voted this candidate term.

Family:

knn

Status:

PRODUCED

Producer:

protea.core._leaf_record_builder._LeafRecordBuilder.make_leaf_record (distance from protea.core._knn_transfer_runner._KnnTransferRunner)

Unit:

embedding-distance

Range:

non-negative; scale depends on the retrieval metric and PLM

Definition:

The PLM-embedding distance returned by the KNN retrieval step for the (query, reference) pair whose annotation contributed this candidate GO term. Smaller is a closer neighbour. Assembled onto the leaf record from the KNN runner; the metric is the one the retrieval index was built with for the active PLM.

vote_count

How many of the query’s KNN neighbours annotated this candidate GO term.

Family:

knn

Status:

PRODUCED

Producer:

protea.core._leaf_record_builder._LeafRecordBuilder._reranker_fields (runner.rr_vote_count)

Unit:

votes

Range:

bare count, 1..K

Definition:

Count of retrieved neighbours (within the K-neighbourhood) that carry this candidate term among their annotations. Defaults to 1 when no per-term tally is present.

k_position

Rank of the closest neighbour that voted this candidate term.

Family:

knn

Status:

PRODUCED

Producer:

protea.core._leaf_record_builder._LeafRecordBuilder._reranker_fields (runner.rr_k_position)

Unit:

rank

Range:

bare count, 1..K

Definition:

The 1-based position, in the distance-sorted neighbour list, of the nearest neighbour that annotated this candidate term. Lower means the term was proposed by a closer neighbour. Defaults to 1.

neighbor_distance_std

Spread of the query’s neighbour distances.

Family:

knn

Status:

PRODUCED

Producer:

protea.core._leaf_record_builder._LeafRecordBuilder._reranker_fields (runner.rr_distance_std)

Unit:

embedding-distance

Range:

non-negative

Definition:

Standard deviation of the distances of the query’s retrieved neighbours. A per-query quantity (same for every candidate of that query) describing how tight or diffuse the neighbourhood is. Defaults to 0.0.

neighbor_vote_fraction

Fraction of the K-neighbourhood that voted this candidate term.

Family:

knn

Status:

PRODUCED

Producer:

protea.core._leaf_record_builder._LeafRecordBuilder._reranker_fields (vote_count / runner.k_limit)

Unit:

fraction

Range:

0.0..1.0

Definition:

``vote_count`` divided by the retrieval neighbourhood size ``runner.k_limit``. A normalised consensus strength: 1.0 means every neighbour in the K-neighbourhood annotated this term.

neighbor_min_distance

Distance of the closest neighbour that voted this candidate term.

Family:

knn

Status:

PRODUCED

Producer:

protea.core._leaf_record_builder._LeafRecordBuilder._reranker_fields (runner.rr_vote_min_d)

Unit:

embedding-distance

Range:

non-negative

Definition:

Minimum over the voting neighbours of their query distance, for this candidate term. Falls back to the row’s own ``distance`` when no per-term minimum is recorded.

neighbor_mean_distance

Mean distance of the neighbours that voted this candidate term.

Family:

knn

Status:

PRODUCED

Producer:

protea.core._leaf_record_builder._LeafRecordBuilder._reranker_fields (runner.rr_vote_sum_d / vote_count)

Unit:

embedding-distance

Range:

non-negative

Definition:

Sum of the voting neighbours’ distances divided by ``vote_count`` (clamped to at least 1) for this candidate term. Falls back to the row’s own ``distance`` when no per-term sum is recorded.

Family alignment_nw

identity_nw

Sequence identity of the global (Needleman-Wunsch) alignment of query and reference.

Family:

alignment_nw

Status:

PRODUCED

Producer:

protea.core.feature_engineering.compute_alignment (parasail NW), assembled by protea.core._leaf_record_builder._LeafRecordBuilder._alignment_fields

Unit:

fraction

Range:

0.0..1.0

Definition:

Fraction of identical positions (matches / alignment length) in the parasail Needleman-Wunsch global alignment of the query and reference sequences.

Notes:

Populated when the ``compute_alignments`` export flag is enabled; with the flag off the column is emitted null (LightGBM reads it as missing).

similarity_nw

Sequence similarity of the global (Needleman-Wunsch) alignment.

Family:

alignment_nw

Status:

PRODUCED

Producer:

protea.core.feature_engineering.compute_alignment (parasail NW), assembled by protea.core._leaf_record_builder._LeafRecordBuilder._alignment_fields

Unit:

fraction

Range:

0.0..1.0

Definition:

Fraction of similar positions (parasail comparison line characters ``|`` or ``:``) over the alignment length of the NW global alignment. Similarity counts conservative substitutions, so it is >= identity.

Notes:

Populated when the ``compute_alignments`` export flag is enabled; null otherwise.

alignment_score_nw

Raw score of the global (Needleman-Wunsch) alignment.

Family:

alignment_nw

Status:

PRODUCED

Producer:

protea.core.feature_engineering.compute_alignment (parasail NW), assembled by protea.core._leaf_record_builder._LeafRecordBuilder._alignment_fields

Unit:

alignment score

Range:

unbounded (matrix-dependent)

Definition:

The parasail substitution-matrix score of the NW global alignment of query and reference. An unnormalised score that grows with alignment length and similarity.

Notes:

Populated when the ``compute_alignments`` export flag is enabled; null otherwise.

gaps_pct_nw

Gap percentage of the global (Needleman-Wunsch) alignment.

Family:

alignment_nw

Status:

PRODUCED

Producer:

protea.core.feature_engineering.compute_alignment (parasail NW), assembled by protea.core._leaf_record_builder._LeafRecordBuilder._alignment_fields

Unit:

fraction

Range:

0.0..1.0

Definition:

Fraction of the NW alignment columns that are gaps in either sequence.

Notes:

Populated when the ``compute_alignments`` export flag is enabled; null otherwise.

alignment_length_nw

Length of the global (Needleman-Wunsch) alignment.

Family:

alignment_nw

Status:

PRODUCED

Producer:

protea.core.feature_engineering.compute_alignment (parasail NW), assembled by protea.core._leaf_record_builder._LeafRecordBuilder._alignment_fields

Unit:

residues

Range:

bare count, non-negative

Definition:

Number of columns in the NW global alignment (matched positions plus gaps).

Notes:

Populated when the ``compute_alignments`` export flag is enabled; null otherwise.

Family alignment_sw

identity_sw

Sequence identity of the local (Smith-Waterman) alignment of query and reference.

Family:

alignment_sw

Status:

PRODUCED

Producer:

protea.core.feature_engineering.compute_alignment (parasail SW), assembled by protea.core._leaf_record_builder._LeafRecordBuilder._alignment_fields

Unit:

fraction

Range:

0.0..1.0

Definition:

Fraction of identical positions (matches / alignment length) in the parasail Smith-Waterman local alignment of the query and reference sequences.

Notes:

Populated when the ``compute_alignments`` export flag is enabled; null otherwise.

similarity_sw

Sequence similarity of the local (Smith-Waterman) alignment.

Family:

alignment_sw

Status:

PRODUCED

Producer:

protea.core.feature_engineering.compute_alignment (parasail SW), assembled by protea.core._leaf_record_builder._LeafRecordBuilder._alignment_fields

Unit:

fraction

Range:

0.0..1.0

Definition:

Fraction of similar positions (comparison-line ``|`` or ``:``) over the alignment length of the SW local alignment.

Notes:

Populated when the ``compute_alignments`` export flag is enabled; null otherwise.

alignment_score_sw

Raw score of the local (Smith-Waterman) alignment.

Family:

alignment_sw

Status:

PRODUCED

Producer:

protea.core.feature_engineering.compute_alignment (parasail SW), assembled by protea.core._leaf_record_builder._LeafRecordBuilder._alignment_fields

Unit:

alignment score

Range:

unbounded (matrix-dependent)

Definition:

The parasail substitution-matrix score of the SW local alignment of query and reference. Unnormalised.

Notes:

Populated when the ``compute_alignments`` export flag is enabled; null otherwise.

gaps_pct_sw

Gap percentage of the local (Smith-Waterman) alignment.

Family:

alignment_sw

Status:

PRODUCED

Producer:

protea.core.feature_engineering.compute_alignment (parasail SW), assembled by protea.core._leaf_record_builder._LeafRecordBuilder._alignment_fields

Unit:

fraction

Range:

0.0..1.0

Definition:

Fraction of the SW alignment columns that are gaps in either sequence.

Notes:

Populated when the ``compute_alignments`` export flag is enabled; null otherwise.

alignment_length_sw

Length of the local (Smith-Waterman) alignment.

Family:

alignment_sw

Status:

PRODUCED

Producer:

protea.core.feature_engineering.compute_alignment (parasail SW), assembled by protea.core._leaf_record_builder._LeafRecordBuilder._alignment_fields

Unit:

residues

Range:

bare count, non-negative

Definition:

Number of columns in the SW local alignment (matched positions plus gaps).

Notes:

Populated when the ``compute_alignments`` export flag is enabled; null otherwise.

Family length

length_query

Amino-acid length of the query protein.

Family:

length

Status:

PRODUCED

Producer:

protea.core._leaf_record_builder._LeafRecordBuilder._alignment_fields (from pair features)

Unit:

residues

Range:

bare count, positive

Definition:

Residue count of the query sequence, carried on the pair-feature record.

length_ref

Amino-acid length of the reference (neighbour) protein.

Family:

length

Status:

PRODUCED

Producer:

protea.core._leaf_record_builder._LeafRecordBuilder._alignment_fields (from pair features)

Unit:

residues

Range:

bare count, positive

Definition:

Residue count of the reference sequence, carried on the pair-feature record.

Family taxonomy_pair

taxonomic_distance

Tree distance between the query’s and reference’s NCBI taxa.

Family:

taxonomy_pair

Status:

PRODUCED

Producer:

protea.core.feature_engineering.compute_taxonomy, assembled by protea.core._leaf_record_builder._LeafRecordBuilder._taxonomy_fields

Unit:

tree edges

Range:

bare count, non-negative; null when a taxon id is absent

Definition:

Distance between the query and reference organisms in the NCBI taxonomy tree, computed from their lineages (0 when identical). Null when either taxon id is missing.

Notes:

Populated when the ``compute_taxonomy`` export flag is enabled; null otherwise.

taxonomic_common_ancestors

Number of shared lineage nodes between the query and reference taxa.

Family:

taxonomy_pair

Status:

PRODUCED

Producer:

protea.core.feature_engineering.compute_taxonomy, assembled by protea.core._leaf_record_builder._LeafRecordBuilder._taxonomy_fields

Unit:

count

Range:

bare count, non-negative

Definition:

Count of taxonomy nodes shared by the query and reference lineages (their common-ancestor path length). 1 when the taxa are the same, 0 when unrelated / unknown.

Notes:

Populated when the ``compute_taxonomy`` export flag is enabled.

taxonomic_relation

Categorical relationship of the query and reference taxa.

Family:

taxonomy_pair

Status:

PRODUCED

Producer:

protea.core.feature_engineering.compute_taxonomy, assembled by protea.core._leaf_record_builder._LeafRecordBuilder._taxonomy_fields

Range:

one of {same, ancestor, descendant, child, parent, root-only, unrelated}

Definition:

One of ``same``, ``ancestor``, ``descendant``, ``child``, ``parent``, ``root-only`` or ``unrelated``, derived from the two lineages. Categorical: the lab encodes it to a stable integer code for LightGBM.

Notes:

Categorical feature. Populated when the ``compute_taxonomy`` export flag is enabled.

Family taxonomy_voters

tax_voters_same_frac

Fraction of voting neighbours from the same organism as the query.

Family:

taxonomy_voters

Status:

PRODUCED

Producer:

protea.core._leaf_record_builder._LeafRecordBuilder._tax_consensus_fields (protea.core._leaf_record_builder._LeafRecordBuilder._tax_same_frac)

Unit:

fraction

Range:

0.0..1.0; NaN when taxonomy disabled

Definition:

Over the neighbours that voted this candidate term, the fraction whose taxon equals the query’s. NaN when taxonomy is disabled for the run.

tax_voters_close_frac

Fraction of voting neighbours taxonomically close to the query.

Family:

taxonomy_voters

Status:

PRODUCED

Producer:

protea.core._leaf_record_builder._LeafRecordBuilder._tax_consensus_fields (protea.core._leaf_record_builder._LeafRecordBuilder._tax_close_frac)

Unit:

fraction

Range:

0.0..1.0; NaN when taxonomy disabled

Definition:

Over the neighbours that voted this candidate term, the fraction whose taxon is taxonomically close to the query’s (near in the lineage tree). NaN when taxonomy is disabled.

tax_voters_mean_common_ancestors

Mean shared-lineage depth between the query and its voting neighbours.

Family:

taxonomy_voters

Status:

PRODUCED

Producer:

protea.core._leaf_record_builder._LeafRecordBuilder._tax_consensus_fields (protea.core._leaf_record_builder._LeafRecordBuilder._tax_ca_mean)

Unit:

count

Range:

non-negative; NaN when taxonomy disabled

Definition:

Average, over the neighbours that voted this candidate term, of the count of taxonomy nodes shared with the query. NaN when taxonomy is disabled.

Family go_context

go_term_frequency

How common the candidate GO term is across the reference pool.

Family:

go_context

Status:

PRODUCED

Producer:

protea.core._leaf_record_builder._LeafRecordBuilder._reranker_fields (runner.go_term_freq)

Unit:

count

Range:

bare count, non-negative

Definition:

Corpus frequency of the candidate term over the reference annotation pool, used as a base-rate prior so a ubiquitous term is not treated like a specific one. Read from the runner’s precomputed term-frequency map; defaults to 0 for an unseen term.

ref_annotation_density

How many annotations the voting reference protein carries.

Family:

go_context

Status:

PRODUCED

Producer:

protea.core._leaf_record_builder._LeafRecordBuilder._reranker_fields (runner.ref_ann_density)

Unit:

count

Range:

bare count, non-negative

Definition:

Annotation count of the reference protein that supplied this candidate, a proxy for how densely studied that reference is. Read from the runner’s per-reference density map; defaults to 0.

Family anc2vec_neighbor

anc2vec_neighbor_cos

Cosine of the candidate term to the neighbours’ semantic centroid.

Family:

anc2vec_neighbor

Status:

PRODUCED

Producer:

protea.core._leaf_record_builder._LeafRecordBuilder._anc2vec_features / protea.core._leaf_record_builder._LeafRecordBuilder._anc2vec_fields

Unit:

cosine

Range:

-1.0..1.0; NaN when no embedding

Definition:

Cosine similarity between the candidate GO term’s Anc2Vec embedding (GO release 2020-10-06 pretrained) and the unit centroid of the voting neighbours’ embeddings. NaN when the candidate has no Anc2Vec embedding or there is no centroid.

anc2vec_neighbor_maxcos

Max cosine of the candidate term to any single voting neighbour.

Family:

anc2vec_neighbor

Status:

PRODUCED

Producer:

protea.core._leaf_record_builder._LeafRecordBuilder._anc2vec_features / protea.core._leaf_record_builder._LeafRecordBuilder._anc2vec_fields

Unit:

cosine

Range:

-1.0..1.0; NaN when no embedding

Definition:

Maximum cosine similarity between the candidate term’s Anc2Vec embedding and any individual neighbour embedding in the voting set. NaN when the candidate has no embedding or the neighbour matrix is absent.

anc2vec_has_emb

Whether the candidate GO term has an Anc2Vec embedding.

Family:

anc2vec_neighbor

Status:

PRODUCED

Producer:

protea.core._leaf_record_builder._LeafRecordBuilder._anc2vec_features / protea.core._leaf_record_builder._LeafRecordBuilder._anc2vec_fields

Unit:

flag

Range:

0.0 or 1.0

Definition:

1.0 when the candidate term is present in the Anc2Vec index (so the cosine features are meaningful), 0.0 otherwise. Lets the booster tell a real 0 cosine from an absent embedding.

Family anc2vec_query

anc2vec_query_known_cos

Cosine of the candidate term to the centroid of the query’s known terms.

Family:

anc2vec_query

Status:

PRODUCED

Producer:

protea.core._leaf_record_builder._LeafRecordBuilder._anc2vec_features / protea.core._leaf_record_builder._LeafRecordBuilder._anc2vec_fields

Unit:

cosine

Range:

-1.0..1.0; NaN when no embedding

Definition:

Cosine similarity between the candidate term’s Anc2Vec embedding and the unit centroid of the query protein’s pre-cutoff known annotations. NaN when the candidate has no embedding or the query has no known-term centroid.

anc2vec_query_known_maxcos

Max cosine of the candidate term to any of the query’s known terms.

Family:

anc2vec_query

Status:

PRODUCED

Producer:

protea.core._leaf_record_builder._LeafRecordBuilder._anc2vec_features / protea.core._leaf_record_builder._LeafRecordBuilder._anc2vec_fields

Unit:

cosine

Range:

-1.0..1.0; NaN when no embedding

Definition:

Maximum cosine similarity between the candidate term’s Anc2Vec embedding and any single pre-cutoff known-term embedding of the query. NaN when the candidate has no embedding or the query has no known-term matrix.

anc2vec_query_known_count

How many pre-cutoff known terms the query protein has.

Family:

anc2vec_query

Status:

PRODUCED

Producer:

protea.core._leaf_record_builder._LeafRecordBuilder._anc2vec_fields (q_known_n)

Unit:

count

Range:

bare count, non-negative

Definition:

Count of the query protein’s own pre-cutoff (non-experimental-leakage-free) known annotations, cast to float. Zero for a protein with no prior annotations.

Family emb_pca

emb_pca_query_0

Component 0 of the 16-dim PCA projection of the query PLM embedding.

Family:

emb_pca

Status:

PRODUCED

Producer:

protea.core.operations.export_minijobs._export_features_batch (PCA fit + transform), assembled onto the leaf record in protea.core._leaf_record_builder._LeafRecordBuilder.make_leaf_record

Unit:

PCA component

Range:

unbounded real; NaN when disabled

Definition:

Coordinate 0 of a 16-dimensional PCA projection of the query protein’s PLM embedding. The PCA is fit over the embedding pool at export time and applied to the query vector.

Notes:

Populated when the ``use_embedding_pca`` export flag is enabled; NaN otherwise (LightGBM reads it as missing). A training run measured this family to contribute essentially nothing; that is a property of that run, not of the feature, so no gain number is recorded here.

emb_pca_query_1

Component 1 of the 16-dim PCA projection of the query PLM embedding.

Family:

emb_pca

Status:

PRODUCED

Producer:

protea.core.operations.export_minijobs._export_features_batch (PCA fit + transform), assembled onto the leaf record in protea.core._leaf_record_builder._LeafRecordBuilder.make_leaf_record

Unit:

PCA component

Range:

unbounded real; NaN when disabled

Definition:

Coordinate 1 of a 16-dimensional PCA projection of the query protein’s PLM embedding. The PCA is fit over the embedding pool at export time and applied to the query vector.

Notes:

Populated when the ``use_embedding_pca`` export flag is enabled; NaN otherwise (LightGBM reads it as missing). A training run measured this family to contribute essentially nothing; that is a property of that run, not of the feature, so no gain number is recorded here.

emb_pca_query_2

Component 2 of the 16-dim PCA projection of the query PLM embedding.

Family:

emb_pca

Status:

PRODUCED

Producer:

protea.core.operations.export_minijobs._export_features_batch (PCA fit + transform), assembled onto the leaf record in protea.core._leaf_record_builder._LeafRecordBuilder.make_leaf_record

Unit:

PCA component

Range:

unbounded real; NaN when disabled

Definition:

Coordinate 2 of a 16-dimensional PCA projection of the query protein’s PLM embedding. The PCA is fit over the embedding pool at export time and applied to the query vector.

Notes:

Populated when the ``use_embedding_pca`` export flag is enabled; NaN otherwise (LightGBM reads it as missing). A training run measured this family to contribute essentially nothing; that is a property of that run, not of the feature, so no gain number is recorded here.

emb_pca_query_3

Component 3 of the 16-dim PCA projection of the query PLM embedding.

Family:

emb_pca

Status:

PRODUCED

Producer:

protea.core.operations.export_minijobs._export_features_batch (PCA fit + transform), assembled onto the leaf record in protea.core._leaf_record_builder._LeafRecordBuilder.make_leaf_record

Unit:

PCA component

Range:

unbounded real; NaN when disabled

Definition:

Coordinate 3 of a 16-dimensional PCA projection of the query protein’s PLM embedding. The PCA is fit over the embedding pool at export time and applied to the query vector.

Notes:

Populated when the ``use_embedding_pca`` export flag is enabled; NaN otherwise (LightGBM reads it as missing). A training run measured this family to contribute essentially nothing; that is a property of that run, not of the feature, so no gain number is recorded here.

emb_pca_query_4

Component 4 of the 16-dim PCA projection of the query PLM embedding.

Family:

emb_pca

Status:

PRODUCED

Producer:

protea.core.operations.export_minijobs._export_features_batch (PCA fit + transform), assembled onto the leaf record in protea.core._leaf_record_builder._LeafRecordBuilder.make_leaf_record

Unit:

PCA component

Range:

unbounded real; NaN when disabled

Definition:

Coordinate 4 of a 16-dimensional PCA projection of the query protein’s PLM embedding. The PCA is fit over the embedding pool at export time and applied to the query vector.

Notes:

Populated when the ``use_embedding_pca`` export flag is enabled; NaN otherwise (LightGBM reads it as missing). A training run measured this family to contribute essentially nothing; that is a property of that run, not of the feature, so no gain number is recorded here.

emb_pca_query_5

Component 5 of the 16-dim PCA projection of the query PLM embedding.

Family:

emb_pca

Status:

PRODUCED

Producer:

protea.core.operations.export_minijobs._export_features_batch (PCA fit + transform), assembled onto the leaf record in protea.core._leaf_record_builder._LeafRecordBuilder.make_leaf_record

Unit:

PCA component

Range:

unbounded real; NaN when disabled

Definition:

Coordinate 5 of a 16-dimensional PCA projection of the query protein’s PLM embedding. The PCA is fit over the embedding pool at export time and applied to the query vector.

Notes:

Populated when the ``use_embedding_pca`` export flag is enabled; NaN otherwise (LightGBM reads it as missing). A training run measured this family to contribute essentially nothing; that is a property of that run, not of the feature, so no gain number is recorded here.

emb_pca_query_6

Component 6 of the 16-dim PCA projection of the query PLM embedding.

Family:

emb_pca

Status:

PRODUCED

Producer:

protea.core.operations.export_minijobs._export_features_batch (PCA fit + transform), assembled onto the leaf record in protea.core._leaf_record_builder._LeafRecordBuilder.make_leaf_record

Unit:

PCA component

Range:

unbounded real; NaN when disabled

Definition:

Coordinate 6 of a 16-dimensional PCA projection of the query protein’s PLM embedding. The PCA is fit over the embedding pool at export time and applied to the query vector.

Notes:

Populated when the ``use_embedding_pca`` export flag is enabled; NaN otherwise (LightGBM reads it as missing). A training run measured this family to contribute essentially nothing; that is a property of that run, not of the feature, so no gain number is recorded here.

emb_pca_query_7

Component 7 of the 16-dim PCA projection of the query PLM embedding.

Family:

emb_pca

Status:

PRODUCED

Producer:

protea.core.operations.export_minijobs._export_features_batch (PCA fit + transform), assembled onto the leaf record in protea.core._leaf_record_builder._LeafRecordBuilder.make_leaf_record

Unit:

PCA component

Range:

unbounded real; NaN when disabled

Definition:

Coordinate 7 of a 16-dimensional PCA projection of the query protein’s PLM embedding. The PCA is fit over the embedding pool at export time and applied to the query vector.

Notes:

Populated when the ``use_embedding_pca`` export flag is enabled; NaN otherwise (LightGBM reads it as missing). A training run measured this family to contribute essentially nothing; that is a property of that run, not of the feature, so no gain number is recorded here.

emb_pca_query_8

Component 8 of the 16-dim PCA projection of the query PLM embedding.

Family:

emb_pca

Status:

PRODUCED

Producer:

protea.core.operations.export_minijobs._export_features_batch (PCA fit + transform), assembled onto the leaf record in protea.core._leaf_record_builder._LeafRecordBuilder.make_leaf_record

Unit:

PCA component

Range:

unbounded real; NaN when disabled

Definition:

Coordinate 8 of a 16-dimensional PCA projection of the query protein’s PLM embedding. The PCA is fit over the embedding pool at export time and applied to the query vector.

Notes:

Populated when the ``use_embedding_pca`` export flag is enabled; NaN otherwise (LightGBM reads it as missing). A training run measured this family to contribute essentially nothing; that is a property of that run, not of the feature, so no gain number is recorded here.

emb_pca_query_9

Component 9 of the 16-dim PCA projection of the query PLM embedding.

Family:

emb_pca

Status:

PRODUCED

Producer:

protea.core.operations.export_minijobs._export_features_batch (PCA fit + transform), assembled onto the leaf record in protea.core._leaf_record_builder._LeafRecordBuilder.make_leaf_record

Unit:

PCA component

Range:

unbounded real; NaN when disabled

Definition:

Coordinate 9 of a 16-dimensional PCA projection of the query protein’s PLM embedding. The PCA is fit over the embedding pool at export time and applied to the query vector.

Notes:

Populated when the ``use_embedding_pca`` export flag is enabled; NaN otherwise (LightGBM reads it as missing). A training run measured this family to contribute essentially nothing; that is a property of that run, not of the feature, so no gain number is recorded here.

emb_pca_query_10

Component 10 of the 16-dim PCA projection of the query PLM embedding.

Family:

emb_pca

Status:

PRODUCED

Producer:

protea.core.operations.export_minijobs._export_features_batch (PCA fit + transform), assembled onto the leaf record in protea.core._leaf_record_builder._LeafRecordBuilder.make_leaf_record

Unit:

PCA component

Range:

unbounded real; NaN when disabled

Definition:

Coordinate 10 of a 16-dimensional PCA projection of the query protein’s PLM embedding. The PCA is fit over the embedding pool at export time and applied to the query vector.

Notes:

Populated when the ``use_embedding_pca`` export flag is enabled; NaN otherwise (LightGBM reads it as missing). A training run measured this family to contribute essentially nothing; that is a property of that run, not of the feature, so no gain number is recorded here.

emb_pca_query_11

Component 11 of the 16-dim PCA projection of the query PLM embedding.

Family:

emb_pca

Status:

PRODUCED

Producer:

protea.core.operations.export_minijobs._export_features_batch (PCA fit + transform), assembled onto the leaf record in protea.core._leaf_record_builder._LeafRecordBuilder.make_leaf_record

Unit:

PCA component

Range:

unbounded real; NaN when disabled

Definition:

Coordinate 11 of a 16-dimensional PCA projection of the query protein’s PLM embedding. The PCA is fit over the embedding pool at export time and applied to the query vector.

Notes:

Populated when the ``use_embedding_pca`` export flag is enabled; NaN otherwise (LightGBM reads it as missing). A training run measured this family to contribute essentially nothing; that is a property of that run, not of the feature, so no gain number is recorded here.

emb_pca_query_12

Component 12 of the 16-dim PCA projection of the query PLM embedding.

Family:

emb_pca

Status:

PRODUCED

Producer:

protea.core.operations.export_minijobs._export_features_batch (PCA fit + transform), assembled onto the leaf record in protea.core._leaf_record_builder._LeafRecordBuilder.make_leaf_record

Unit:

PCA component

Range:

unbounded real; NaN when disabled

Definition:

Coordinate 12 of a 16-dimensional PCA projection of the query protein’s PLM embedding. The PCA is fit over the embedding pool at export time and applied to the query vector.

Notes:

Populated when the ``use_embedding_pca`` export flag is enabled; NaN otherwise (LightGBM reads it as missing). A training run measured this family to contribute essentially nothing; that is a property of that run, not of the feature, so no gain number is recorded here.

emb_pca_query_13

Component 13 of the 16-dim PCA projection of the query PLM embedding.

Family:

emb_pca

Status:

PRODUCED

Producer:

protea.core.operations.export_minijobs._export_features_batch (PCA fit + transform), assembled onto the leaf record in protea.core._leaf_record_builder._LeafRecordBuilder.make_leaf_record

Unit:

PCA component

Range:

unbounded real; NaN when disabled

Definition:

Coordinate 13 of a 16-dimensional PCA projection of the query protein’s PLM embedding. The PCA is fit over the embedding pool at export time and applied to the query vector.

Notes:

Populated when the ``use_embedding_pca`` export flag is enabled; NaN otherwise (LightGBM reads it as missing). A training run measured this family to contribute essentially nothing; that is a property of that run, not of the feature, so no gain number is recorded here.

emb_pca_query_14

Component 14 of the 16-dim PCA projection of the query PLM embedding.

Family:

emb_pca

Status:

PRODUCED

Producer:

protea.core.operations.export_minijobs._export_features_batch (PCA fit + transform), assembled onto the leaf record in protea.core._leaf_record_builder._LeafRecordBuilder.make_leaf_record

Unit:

PCA component

Range:

unbounded real; NaN when disabled

Definition:

Coordinate 14 of a 16-dimensional PCA projection of the query protein’s PLM embedding. The PCA is fit over the embedding pool at export time and applied to the query vector.

Notes:

Populated when the ``use_embedding_pca`` export flag is enabled; NaN otherwise (LightGBM reads it as missing). A training run measured this family to contribute essentially nothing; that is a property of that run, not of the feature, so no gain number is recorded here.

emb_pca_query_15

Component 15 of the 16-dim PCA projection of the query PLM embedding.

Family:

emb_pca

Status:

PRODUCED

Producer:

protea.core.operations.export_minijobs._export_features_batch (PCA fit + transform), assembled onto the leaf record in protea.core._leaf_record_builder._LeafRecordBuilder.make_leaf_record

Unit:

PCA component

Range:

unbounded real; NaN when disabled

Definition:

Coordinate 15 of a 16-dimensional PCA projection of the query protein’s PLM embedding. The PCA is fit over the embedding pool at export time and applied to the query vector.

Notes:

Populated when the ``use_embedding_pca`` export flag is enabled; NaN otherwise (LightGBM reads it as missing). A training run measured this family to contribute essentially nothing; that is a property of that run, not of the feature, so no gain number is recorded here.

Family lineage

lineage_is_ancestor_of_known

Whether the candidate GO term is an ancestor of at least one term the protein is already known to have.

Family:

lineage

Status:

PRODUCED

Producer:

protea_method.lineage.compute_lineage_features

Unit:

indicator

Range:

0.0 or 1.0

Definition:

1.0 when the candidate appears in the is_a / part_of ancestor closure of any pre-cutoff known term of the query protein, else 0.0. A candidate identical to a known term does not count as its own ancestor: the self-overlap is subtracted before the test. A protein with no known terms scores 0.0, a true absence rather than a missing measurement, so this feature is identically zero across the no-knowledge evaluation cell.

Notes:

Opt-in through the compute_lineage_features payload flag. In the sealed v227 to v230 champion it is the second most important feature of the previously-known cell by LightGBM gain, and the third overall, while contributing exactly zero in the no-knowledge cell, where by construction there is nothing to be an ancestor of.

lineage_is_descendant_of_known

Whether the candidate GO term is a descendant of at least one term the protein is already known to have.

Family:

lineage

Status:

PRODUCED

Producer:

protea_method.lineage.compute_lineage_features

Unit:

indicator

Range:

0.0 or 1.0

Definition:

1.0 when any pre-cutoff known term of the query protein appears in the candidate’s own ancestor closure, else 0.0. Self-overlap is subtracted, so a candidate equal to a known term scores 0.0. A protein with no known terms scores 0.0.

lineage_ancestor_of_count

How many of the protein’s known terms have this candidate among their ancestors.

Family:

lineage

Status:

PRODUCED

Producer:

protea_method.lineage.compute_lineage_features

Unit:

known terms

Range:

bare count, 0 upward, bounded by the known-term count

Definition:

Number of pre-cutoff known terms whose ancestor closure contains the candidate, minus one when the candidate is itself a known term. The graded counterpart of lineage_is_ancestor_of_known.

lineage_descendant_of_count

How many of the protein’s known terms are ancestors of this candidate.

Family:

lineage

Status:

PRODUCED

Producer:

protea_method.lineage.compute_lineage_features

Unit:

known terms

Range:

bare count, 0 upward, bounded by the known-term count

Definition:

Number of pre-cutoff known terms present in the candidate’s ancestor closure, minus one when the candidate is itself a known term. The graded counterpart of lineage_is_descendant_of_known.

Family annotation_meta

qualifier

GO annotation qualifier carried by the reference annotation.

Family:

annotation_meta

Status:

PRODUCED

Producer:

protea.core._leaf_record_builder._LeafRecordBuilder.make_leaf_record (from the reference annotation record)

Range:

GO qualifier vocabulary; empty string when absent

Definition:

The GO qualifier string of the reference annotation that supplied this candidate term (for example ``enables``, ``NOT``), empty when absent. Categorical: encoded to a stable integer code by the lab.

Notes:

Categorical feature.

evidence_code

GO evidence code of the reference annotation.

Family:

annotation_meta

Status:

PRODUCED

Producer:

protea.core._leaf_record_builder._LeafRecordBuilder.make_leaf_record (from the reference annotation record)

Range:

GO evidence-code vocabulary; empty string when absent

Definition:

The GO evidence code of the reference annotation that supplied this candidate term (for example ``EXP``, ``IEA``), empty when absent. Categorical: encoded to a stable integer code by the lab.

Notes:

Categorical feature.

aspect

GO aspect (BP / MF / CC) of the candidate term.

Family:

annotation_meta

Status:

PRODUCED

Producer:

protea.core._leaf_record_builder._LeafRecordBuilder.make_leaf_record (runner.aspect_map)

Range:

one of the three GO aspects

Definition:

The GO aspect of the candidate term, read from the runner’s aspect map (biological process, molecular function or cellular component). Categorical: encoded to a stable integer code by the lab. Also a reserved column.

Notes:

Categorical feature; also present in RESERVED_COLUMNS.

Family plm_context

plm_id

Which protein language model produced the embeddings used for KNN retrieval.

Family:

plm_context

Status:

POOL_INJECTED

Producer:

lab pooled multi-manifest loader (injected at pool stage; absent from raw parquet dumps)

Range:

PLM identifier vocabulary

Definition:

Categorical code identifying the PLM whose embeddings retrieved the candidate. The PROTEA dump does not write this column; the lab’s pooled multi-manifest loader injects it as a per-source constant at stage time so the universal multi-PLM booster sees which PLM a row came from.

Notes:

Categorical feature. See PROTEA FEATURE_LEAKAGE_AUDIT.md for the GO/NO-GO ruling on this column.

Family k_neighborhood

k_context

KNN neighbourhood size (K) used to retrieve this candidate.

Family:

k_neighborhood

Status:

POOL_INJECTED

Producer:

lab pooled multi-manifest loader (injected at pool stage; absent from raw parquet dumps)

Unit:

neighbours

Range:

bare count, positive

Definition:

The K-neighbourhood size for the manifest source that produced this row. The PROTEA dump does not write this column; the lab’s pooled loader injects it as a per-source constant at stage time so a booster trained over several K settings can condition on K.

Family interpro

interpro_hit

Whether an InterPro signature maps the query onto this candidate GO term.

Family:

interpro

Status:

PRODUCED

Producer:

protea.core._interpro_features.apply_interpro_features (default protea.core._leaf_record_builder._LeafRecordBuilder._interpro_default_fields)

Unit:

flag

Range:

False or True

Definition:

True when at least one InterPro member-database signature of the query maps onto this candidate term in the InterPro GO-prediction table. Default False for a (query, term) with no InterPro evidence.

Notes:

The InterPro GO-prediction table is loaded by ``protea.core._interpro_features.load_interpro_go_pred``, which returns an empty table when its source env var is unset, in which case every row keeps the zero/False default and the column carries no signal. Whether the InterPro tables are populated in a given deployment is a database-state question that cannot be settled from the source tree alone, so no ``BROKEN`` status is asserted here. Note that InterPro’s principal contribution to predictions enters through a separate InterPro2GO noisy-OR graft (``protea.core.operations.predict_go_terms._interpro_graft``), not through these reranker columns.

interpro_score

Strength of the InterPro signature->GO mapping for this candidate term.

Family:

interpro

Status:

PRODUCED

Producer:

protea.core._interpro_features.apply_interpro_features (default _interpro_default_fields)

Unit:

score

Range:

non-negative; 0.0 when no InterPro evidence

Definition:

Graded InterPro2GO mapping score for the query onto this candidate term. 0.0 when there is no InterPro evidence.

Notes:

Env-gated table; carries no signal when the InterPro GO-prediction table is unpopulated (see interpro_hit).

interpro_n_signatures

Number of InterPro signatures supporting this candidate term.

Family:

interpro

Status:

PRODUCED

Producer:

protea.core._interpro_features.apply_interpro_features (default _interpro_default_fields)

Unit:

count

Range:

bare count, non-negative

Definition:

Count of distinct InterPro member-database signatures of the query that map onto this candidate term. 0 when there is no InterPro evidence.

Notes:

Env-gated table; 0 for every row when the InterPro table is unpopulated (see interpro_hit).

interpro_db_pfam

Whether a Pfam signature supplied the InterPro mapping.

Family:

interpro

Status:

PRODUCED

Producer:

protea.core._interpro_features.apply_interpro_features (default _interpro_default_fields)

Unit:

flag

Range:

False or True

Definition:

One-hot: True when a Pfam member-database signature contributed the mapping for this candidate term.

Notes:

Env-gated table (see interpro_hit).

interpro_db_panther

Whether a PANTHER signature supplied the InterPro mapping.

Family:

interpro

Status:

PRODUCED

Producer:

protea.core._interpro_features.apply_interpro_features (default _interpro_default_fields)

Unit:

flag

Range:

False or True

Definition:

One-hot: True when a PANTHER member-database signature contributed the mapping for this candidate term.

Notes:

Env-gated table (see interpro_hit).

interpro_db_superfamily

Whether a SUPERFAMILY signature supplied the InterPro mapping.

Family:

interpro

Status:

PRODUCED

Producer:

protea.core._interpro_features.apply_interpro_features (default _interpro_default_fields)

Unit:

flag

Range:

False or True

Definition:

One-hot: True when a SUPERFAMILY member-database signature contributed the mapping for this candidate term.

Notes:

Env-gated table (see interpro_hit).

interpro_db_smart

Whether a SMART signature supplied the InterPro mapping.

Family:

interpro

Status:

PRODUCED

Producer:

protea.core._interpro_features.apply_interpro_features (default _interpro_default_fields)

Unit:

flag

Range:

False or True

Definition:

One-hot: True when a SMART member-database signature contributed the mapping for this candidate term.

Notes:

Env-gated table (see interpro_hit).

interpro_db_cdd

Whether a CDD signature supplied the InterPro mapping.

Family:

interpro

Status:

PRODUCED

Producer:

protea.core._interpro_features.apply_interpro_features (default _interpro_default_fields)

Unit:

flag

Range:

False or True

Definition:

One-hot: True when a CDD member-database signature contributed the mapping for this candidate term.

Notes:

Env-gated table (see interpro_hit).

interpro_db_prosite

Whether a PROSITE signature supplied the InterPro mapping.

Family:

interpro

Status:

PRODUCED

Producer:

protea.core._interpro_features.apply_interpro_features (default _interpro_default_fields)

Unit:

flag

Range:

False or True

Definition:

One-hot: True when a PROSITE member-database signature contributed the mapping for this candidate term.

Notes:

Env-gated table (see interpro_hit).

knn_present

Whether the KNN source proposed this candidate for the (protein, term).

Family:

interpro

Status:

PRODUCED

Producer:

protea.core._leaf_record_builder._LeafRecordBuilder._interpro_default_fields (and the InterPro union path)

Unit:

flag

Range:

False or True

Definition:

Presence flag: True when the KNN retrieval source contributed this candidate. Used when pooling KNN and InterPro candidates so a true zero is distinguishable from an absent source. True on every KNN leaf record.

interpro_present

Whether the InterPro source proposed this candidate for the (protein, term).

Family:

interpro

Status:

PRODUCED

Producer:

protea.core._interpro_features.apply_interpro_features (default _interpro_default_fields)

Unit:

flag

Range:

False or True

Definition:

Presence flag: True when the InterPro source contributed this candidate. Default False on a KNN leaf record with no InterPro evidence; set True by the InterPro post-pass / union path.

Notes:

Env-gated table (see interpro_hit).

Family classifier

classifier_score

Full-catalogue direct classifier score for this candidate GO term.

Family:

classifier

Status:

DECLARED_ABSENT

Producer:

protea.core.operations.predict_go_terms._classifier.apply_classifier (gated by compute_classifier, default False; export default emits NaN via _lafa_default_fields)

Unit:

score

Range:

producer-dependent; NaN in the default export

Definition:

Per-candidate score from the direct full-catalogue predictor (the first-place LAFA classifier). The producer ``protea.core.operations.predict_go_terms._classifier.apply_classifier`` (and the export post-pass ``protea.core.training_dump._classifier_postpass.apply_classifier_to_frame``) computes it, but only when the ``compute_classifier`` flag is set. It defaults to ``False`` on the export payload, so the default research-dataset export never fills this column.

Notes:

ADR-D45: the default export emits NaN (was a well-defined 0.0 before PROTEA #710). The sealed 0.4063 champion was trained without the classifier family precisely because it was absent. The schema fingerprint pins names, not values, so a booster selecting this family still passes the schema-sha guard.

classifier_present

Whether the direct classifier proposed this candidate GO term.

Family:

classifier

Status:

DECLARED_ABSENT

Producer:

protea.core.operations.predict_go_terms._classifier.apply_classifier (gated by compute_classifier, default False; export default emits NaN via _lafa_default_fields)

Unit:

flag

Range:

0.0 or 1.0 when produced; NaN in the default export

Definition:

Presence flag for the direct full-catalogue classifier source. Produced by the same classifier producer, gated by ``compute_classifier`` (default False), so the default export never fills it.

Notes:

ADR-D45: DECLARED_ABSENT in the default export. See classifier_score.

Family self_prior

self_prior_score

Score from the query protein’s own pre-cutoff non-experimental annotations.

Family:

self_prior

Status:

DECLARED_ABSENT

Producer:

protea.core.operations.predict_go_terms._post_knn_pipeline.apply_self_prior (gated by compute_self_prior, default False; export default emits NaN via _lafa_default_fields)

Unit:

score

Range:

producer-dependent; NaN in the default export

Definition:

Self-prior signal: how strongly the query protein’s own pre-cutoff non-experimental annotations support this candidate term. Produced by ``protea.core.operations.predict_go_terms._post_knn_pipeline.apply_self_prior``, gated by ``compute_self_prior`` (default False), so the default research-dataset export leaves it at the NaN default.

Notes:

ADR-D45: DECLARED_ABSENT in the default export. See classifier_score.

Family association

association_total

Conditional probability of the candidate term given the query’s known terms.

Family:

association

Status:

DECLARED_ABSENT

Producer:

protea.core.operations.predict_go_terms._post_knn_pipeline.apply_association (gated by compute_association, default False; export default emits NaN via _lafa_default_fields)

Unit:

probability

Range:

0.0..1.0 when produced; NaN in the default export

Definition:

Total cross-aspect association: the conditional probability of the candidate term given the query protein’s pre-cutoff known terms, over all branches. Produced by ``protea.core.operations.predict_go_terms._post_knn_pipeline.apply_association``, gated by ``compute_association`` (default False), so the default export leaves it at the NaN default.

Notes:

ADR-D45: DECLARED_ABSENT in the default export. See classifier_score.

association_cross

Cross-branch-only conditional probability of the candidate given known terms.

Family:

association

Status:

DECLARED_ABSENT

Producer:

protea.core.operations.predict_go_terms._post_knn_pipeline.apply_association (gated by compute_association, default False; export default emits NaN via _lafa_default_fields)

Unit:

probability

Range:

0.0..1.0 when produced; NaN in the default export

Definition:

Cross-branch component of the association signal: the conditional probability of the candidate term given the query’s known terms, restricted to terms in a different GO aspect. Same producer and gating as ``association_total``.

Notes:

ADR-D45: DECLARED_ABSENT in the default export. See classifier_score.

association_present

Whether the association source proposed this candidate GO term.

Family:

association

Status:

DECLARED_ABSENT

Producer:

protea.core.operations.predict_go_terms._post_knn_pipeline.apply_association (gated by compute_association, default False; export default emits NaN via _lafa_default_fields)

Unit:

flag

Range:

0.0 or 1.0 when produced; NaN in the default export

Definition:

Presence flag for the cross-aspect association source. Same producer and gating as ``association_total`` (``compute_association``, default False), so the default export leaves it at the NaN default.

Notes:

ADR-D45: DECLARED_ABSENT in the default export. See classifier_score.

Family protst_text

protst_text_score

ProtST text-to-GO transfer score for this candidate GO term.

Family:

protst_text

Status:

DECLARED_ABSENT

Producer:

protea.core.operations.predict_go_terms._protst_text.apply_protst_text (gated by compute_protst, default False; export default emits NaN via _lafa_default_fields)

Unit:

normalised vote

Range:

0.0..1.0 when produced; NaN in the default export

Definition:

Normalised cosine-weighted vote for the candidate term from the query protein’s ProtST kNN neighbourhood. The producer takes the query’s ProtST protein embedding, retrieves its top-30 cosine-nearest reference proteins (both banks L2-normalised), casts a vote per reference GO term weighted by the neighbour cosine, and normalises by the per-query maximum vote. Reference GO terms are restricted to the pre-cutoff annotation set (leakage-free). Produced by ``protea.core.operations.predict_go_terms._protst_text.apply_protst_text``, gated by ``compute_protst`` (default False), so the default research-dataset export leaves it at the NaN default.

Notes:

ADR-D45: DECLARED_ABSENT in the default export. The ProtST reference bank is a separate EmbeddingConfig; whether it is populated in a given deployment is a database-state question, so no BROKEN status is asserted here. See classifier_score.

protst_vote_fraction

Fraction of the query’s ProtST neighbours that carry this candidate term.

Family:

protst_text

Status:

DECLARED_ABSENT

Producer:

protea.core.operations.predict_go_terms._protst_text.apply_protst_text (gated by compute_protst, default False; export default emits NaN via _lafa_default_fields)

Unit:

fraction

Range:

0.0..1.0 when produced; NaN in the default export

Definition:

Coverage feature for the ProtST transfer: the fraction of the query’s 30 ProtST cosine-nearest neighbours whose pre-cutoff annotations include the candidate term. Decouples a single very close neighbour from a broad consensus, mirroring ``neighbor_vote_fraction``. Same producer and gating as ``protst_text_score`` (``compute_protst``, default False), so the default export leaves it at the NaN default.

Notes:

ADR-D45: DECLARED_ABSENT in the default export. See protst_text_score.

protst_present

Whether the ProtST source contributed to this candidate at all.

Family:

protst_text

Status:

DECLARED_ABSENT

Producer:

protea.core.operations.predict_go_terms._protst_text.apply_protst_text (gated by compute_protst, default False; export default emits NaN via _lafa_default_fields)

Unit:

flag

Range:

0.0 or 1.0 when produced; NaN in the default export

Definition:

Presence flag for the ProtST transfer source: 1.0 when at least one ProtST neighbour voted the candidate term (so the score is a measured value), 0.0 when ProtST had coverage but no support. Distinguishes a measured zero from an absent source, letting the booster gate on ProtST coverage, mirroring ``interpro_present`` / ``knn_present``. Same producer and gating as ``protst_text_score`` (``compute_protst``, default False).

Notes:

ADR-D45: DECLARED_ABSENT in the default export. See protst_text_score.

Programmatic access

The same registry is importable at runtime for tooling that needs the machine-readable form:

from protea_contracts.feature_docs import FEATURE_DOCS, FeatureStatus

doc = FEATURE_DOCS["interpro_hit"]
print(doc.status, doc.producer)