rsynk/locale-benchmark-sra50-sampled

LOCALE Benchmark: SRA-50 sampled

A reduced-index copy of SRA-50 v2 for cheap, repeated evaluation, such as tracking a model during training. The 50 runs, 500 queries, and ground truth are identical to SRA-50 v2; only the contigs to index are subsampled, so the whole set embeds in minutes instead of embedding all 8.7M contigs. Scores are not comparable to SRA-50 v2: dropping most contigs raises the ratio of relevant to irrelevant vectors, so absolute numbers come out higher. Use it to compare models against each other, and SRA-50 v2 or SRA-500 v2 to report.

Which contigs are kept

Every contig of every run is kept if it is either of:

1. A relevant contig. The best-aligned contig of every (query, run) pair in the ground truth, so no query loses a positive and queries.parquet is unchanged.
2. A hit contig. Any contig that held a top-100 window for any query in an exact search of SRA-50 v2 with LOCALE (checkpoint 8vqiabk9, 256 bp windows, stride 106) at mutation rates 0.05 and 0.10. Most of these are wrong hits, the near-homologs a model has to rank below the true run, and a uniform subsample would discard nearly all of them. They are hard for that one model; every model evaluated here sees the same set.

Everything else is kept with probability subsample_fraction by a seeded coin flip per contig, so each run keeps about that share of its contigs.

Runs indexed	50 (all of SRA-50 v2)
Queries	500 (all of SRA-50 v2)
Contigs in SRA-50 v2	8,712,391
Relevant contigs kept	1,250
Hit contigs kept	84,308 (1,089 of them also relevant)
Random fraction	0.1
Seed	0
Contigs kept	946,972 (10.9% of SRA-50 v2; 84,469 by rule, the rest by coin flip)

Built by scripts/build_subsample_dataset.py in locale-data; the hit contigs came from benchmark/scripts/hit_contigs.py in locale run on the SRA-50 v2 exact-search hits.

Files

The layout is SRA-50 v2's without its mutations/ provenance directory. Every file except the contigs is a verbatim copy.

File	Contents
accs.txt	the indexed SRA run accessions, one per line
logan_accessions/<run>/<run>.contigs.fa	the kept contigs of each run, in their original order
queries.parquet	one row per query with its ground truth (schema below)
queries_mut0.00.parquet	queries.parquet plus a mutation_rate column (clean)
queries_mut0.05.parquet, queries_mut0.10.parquet	same rows in the same order with query_sequence mutated
queries.fa, queries.fa.fai	the clean queries as fasta (mutation-simulator's input) and its index

Schema of queries.parquet

Column	Type	Meaning
accession	str	run the read was sampled from
query_id	str	<accession>.<spot> from fastq-dump
query_sequence	str	the raw read
contig_accession	list[str]	relevant set: every run with an alignment above the identity threshold; always contains accession
contig_id	list[str]	best-aligned contig in each relevant run
identity	list[f64]	exact matches / read length of that alignment
ratio	list[f64]	contig length / read length
contig_len	list[i64]	contig length
aln_interval_contig	list[struct]	0-based [start, end) of the alignment on the contig
strand	str	+/-: strand of the best hit to the source run

The six list columns are parallel: element i of each describes the same alignment. See the SRA-50 v2 card for how the queries and ground truth were built.

License

The data is licensed under Creative Commons cc-by-4.0; the code that built it is MIT (locale-data). The underlying sequence data comes from the NCBI Sequence Read Archive, which places no restrictions on redistribution; assembled contigs come from Logan.

Citation

For questions, reach out to ryansynk@umd.edu. If you find this work useful, please cite:

@article {Synk2026.05.12.724581,
    author = {Synk, Ryan P. and Pandey, Prashant and Sahinalp, S. Cenk and Duraiswami, Ramani},
    title = {LOCALE: Local-Alignment Embeddings for Noise-Robust DNA Search at SRA Scale},
    elocation-id = {2026.05.12.724581},
    year = {2026},
    doi = {10.64898/2026.05.12.724581},
    publisher = {Cold Spring Harbor Laboratory},
    URL = {https://www.biorxiv.org/content/early/2026/05/14/2026.05.12.724581},
    eprint = {https://www.biorxiv.org/content/early/2026/05/14/2026.05.12.724581.full.pdf},
    journal = {bioRxiv}
}