AfterQuery/OmicsBench-grader-databases

OmicsBench Grader Databases

Synonym databases for deterministic grading in OmicsBench, a realism-first benchmark for AI coding agents in computational biology.

OmicsBench ships with dozens of prebuilt graders for highly specific biological outputs. This dataset powers the synonym-aware matching behind those graders, so that TP53 and tumor protein p53 are recognized as the same gene, hsa-miR-21-5p and MIMAT0000076 as the same miRNA, and E. coli and Escherichia coli as the same organism.

Format

One read-only SQLite file, grader_databases.sqlite (~22 GB), with five tables -- one per source database. Lookup columns are pre-lowercased at build time (full Unicode str.lower()) and covered by composite indexes, so runtime queries are direct indexed seeks and process memory is O(query result) rather than O(database size).

Table	Columns	Rows	Source	Description
ncbi_gene	organism_lower name_lower gene_id	~219M	NCBI Gene	Gene symbol, synonym, description, Entrez GeneID, and Ensembl ID resolution scoped by organism
ncbi_taxonomy	name_lower canonical_name	~3.2M	NCBI Taxonomy	Scientific name, common name, abbreviation, and synonym resolution for all taxa
hmdb	name_lower accession	~1.5M	HMDB	Metabolite name, synonym, IUPAC name, and HMDB accession resolution
mirbase	name_lower accession	~157K	miRBase	miRNA precursor and mature name/accession resolution, including deprecated entries folded into live replacements
card	name_lower aro_accession	~19K	CARD	Antimicrobial resistance gene name, synonym, and ARO accession resolution

Three of the tables (ncbi_taxonomy, hmdb, mirbase) use WITHOUT ROWID with PRIMARY KEY(name_lower) and are built via INSERT OR IGNORE so the first occurrence of any lowercased name wins. The other two (ncbi_gene, card) are 1:N and allow multiple rows per lookup key -- gene_match and amr_match do set-intersection at query time.

Usage

Install OmicsBench and call the match functions directly; downloading and caching happens on first use.

from omicsbench.grading import (
    gene_match,
    taxonomy_match,
    mirna_match,
    metabolite_match,
    amr_match,
)

gene_match("TP53", "tumor protein p53", organism="Homo sapiens")  # True
taxonomy_match("E. coli", "Escherichia coli")                     # True
mirna_match("hsa-miR-21-5p", "MIMAT0000076")                      # True
metabolite_match("Dextrose", "D-Glucose")                         # True
amr_match("mecA", "PBP2A")                                        # True

Each match function is a thin wrapper over one indexed SQL query plus a case-insensitive string-equality fallback (used when either side is absent from the database, so misses don't crash on out-of-distribution inputs). The public API signatures, return types, and fallback semantics are unchanged from the previous TSV-backed release -- consumers upgrade by bumping their omicsbench package version.

Each lookup module opens a thread-local read-only connection on first use with read-tuning PRAGMAs (query_only=1, cache_size=-20000, temp_store=MEMORY, mmap_size=268435456), so the dataset works correctly under multi-threaded consumers like the OmicsBench async orchestrator. SQLite supports unlimited concurrent readers via shared file locks.

Direct access

If you want to query the tables yourself without installing OmicsBench, open the file read-only:

import sqlite3
from huggingface_hub import hf_hub_download

path = hf_hub_download(
    repo_id="AfterQuery/OmicsBench-grader-databases",
    filename="grader_databases.sqlite",
    repo_type="dataset",
)
conn = sqlite3.connect(f"file:{path}?mode=ro", uri=True)

rows = conn.execute(
    "SELECT gene_id FROM ncbi_gene "
    "WHERE organism_lower = ? AND name_lower = ?",
    ("homo sapiens", "tp53"),
).fetchall()

All lookup columns are pre-lowercased; lowercase your query inputs in Python before binding. The table schemas are stable across releases: any schema change is a breaking revision of this dataset.

Provenance and licensing

The five source databases retain their upstream licenses. OmicsBench packages them into a single SQLite file for efficient runtime access; see the OmicsBench GitHub repository for the full ETL pipeline under scripts/databases/ and its step-by-step runbook in scripts/databases/README.md.

Rebuilds ship as new revisions on this dataset whenever an upstream source publishes a new release.