rsynk/locale-benchmark-sra62-strain

LOCALE benchmark — SRA62 within-genotype strain discrimination

Query with Hepatitis B virus reads of one genotype; retrieve accessions of the same genotype while a set of different-genotype HBV accessions sits in the same index as hard negatives.

This is the specificity half of a pair. Its sibling, locale-benchmark-sra52-viral, asks whether a model can bridge ~11% divergence. This one asks whether it can refuse to. Same virus, same two genotypes, same measured divergence, opposite ground truth.

Accessions indexed	62 (47 distractors + 5 relevant + 10 hard negatives)
Queries	500 reads, 150 bp
Query genotype	B (10 runs, BioProject PRJNA564199)
Relevant genotype	B (5 runs, PRJNA556730)
Hard-negative genotype	D (10 runs, PRJNA659786)
Relevance density	5 / 62 = 8.1%

Relevance is inverted relative to sra52viral

In sra52viral, a genotype-D query and a genotype-B accession are a match — both are HBV and the genotypes are homologous across ~98% of the genome. Here, that same pair is a non-match. Nothing about the biology changed; the question did.

Read the two together. Either one alone is gameable:

• A model that treats all HBV as identical wins sra52viral and floors this one.
• A model that keys on sequencing-batch artifacts does the reverse.

The informative quantity is the gap between a model's two scores, not either score by itself.

What separates a positive from a hard negative

Consensus sequences assembled from reads, compared position-wise:

comparison	identity	coverage
query project → positives (cross-project, within-genotype B)	0.9838	0.9981
query project → hard negatives (cross-genotype, B vs D)	0.9015	0.8931

About 8 points of nucleotide identity. Both consensuses are read-derived — positions below 10× depth are called N, not filled from the reference — and N-fractions are under 0.3%, so neither inherits its reference.

Ground truth is metadata, not alignment

contig_accession holds the relevant accessions for each query. In sra50 / sra500 that column is alignment-derived. Here it comes from SRA genotype metadata: every genotype-B positive is relevant to every genotype-B query, and no genotype-D accession ever is. No aligner decides what the correct answer is — which matters when the baselines being compared against are themselves k-mer and alignment methods.

Alignment is used only to (a) measure divergence, (b) screen out query reads with no HBV homolog, and (c) label which queries a k-mer or alignment method could discriminate on at all (see stratification below). None of those decide the answer.

Why these accessions

Genotype B exists in only two adequately-screened BioProjects: PRJNA556730 (143 runs in Logan) and PRJNA564199 (29). Because the queries are themselves genotype B, positives spanning both projects would share a project — and therefore primers, adapters and sequencing batch — with the queries, letting a method score well by matching technical artifacts instead of strain identity.

So queries and positives share no BioProject: positives are drawn from PRJNA556730 only, queries from PRJNA564199 only. Positives are the top 5 in-Logan runs ranked on coverage breadth then HBV fraction — method-neutral criteria, the same rule sra52viral used.

The 10 hard negatives are exactly the genotype-D runs that supplied the queries in sra52viral, reused here as index entries.

Every run passed the same three-part QC gate as sra52viral: metadata genotype checked against independent organism-labeled references, HBV fraction ≥ 50%, and coverage breadth ≥ 75% at ≥10×.

Query reads are 150 bp, deliberately

Every genotype-B run in both projects is 300–301 bp; the genotype-D runs that supplied sra52viral's queries are 131–149 bp. Querying at full length would compare 301 bp retrieval against sra52viral's ~140 bp retrieval, and part of any measured strain discrimination would just be query length.

Reads are cut to a random 150 bp window rather than the first 150 bp. The prefix of a 301 bp Illumina read is its highest-quality half, so a prefix would swap one bias for the opposite one rather than remove it. Truncation happens before the HBV screen, so what was screened is what is actually queried.

Query reads are filtered to HBV, deliberately

Logan contigs are assembled from whole runs, and these runs are 50–95% HBV — so every indexed HBV accession also carries human contigs. Unfiltered human query reads would match those human contigs and score as correct retrievals for an entirely non-biological reason.

The filter uses only the query's own genotype-B reference. It does not require reads to align to a positive, which would pre-select the reads that survive cross-project alignment.

Exact k-mer overlap

Fraction of the 500 queries sharing any canonical k-mer (both strands) with each side:

k	vs positives (B)	vs hard negatives (D)
15	100.0%	96.2%
21	100.0%	81.4%
25	99.6%	68.2%
31	94.4%	56.8%

At k=31 an exact k-mer method has real signal to work with — a 37.6-point gap — so this benchmark is not impossible for k-mer search. But 56.8% of queries still share a 31-mer with the hard negatives, so the negatives are genuinely hard for a majority of queries rather than trivially separable.

Stratification: 40% of queries cannot be fooled

Mapping the 500 queries against a genotype-D reference:

	queries
align to genotype D	298 (59.6%)
no genotype-D alignment at all	202 (40.4%)

For those 202, an aligner cannot score the hard negatives — the positives win by default, with no strain discrimination involved. Aggregate recall over all 500 queries therefore overstates discrimination. query_strata.csv ships the per-query flag so results can be reported stratified; the 298-query subset is the honest discrimination measure.

Note the symmetry with the sibling benchmark. sra52viral measured a 37.45% alignment collapse in the D→B direction, where it hurt the baselines: those queries could not reach their positives. Here the same statistic helps them: those queries cannot be fooled by the negatives. Identical quantity, opposite consequence, purely because genotype D flipped from positive to negative.

Files

file	contents
accs.txt	62 SRA run accessions; contigs fetched from the logan-pub S3 bucket
queries.parquet	500 query reads with metadata-derived ground truth
tiers.csv	accession, tier, bioproject, genotype — tier is distractor / positive / hard_negative
query_strata.csv	query_id, maps_to_genotype_D (1/0) — for stratified scoring

queries.parquet schema

Identical to locale-benchmark-sra50, so existing benchmark code reads it unmodified.

column	type	role
accession	str	run the read came from
query_id	str	joins predictions to ground truth
query_sequence	str	the read, 150 bp
contig_accession	list[str]	ground truth — the 5 genotype-B positives
identity	list[float]	oracle score (1.0)
contig_id, ratio, contig_len, aln_interval_contig		schema compatibility only

Distractors and hard negatives are both absent from contig_accession. The distinction between them lives only in tiers.csv; scoring code that ignores that file will compute correct recall but will not be able to say whether errors came from confusing HBV genotypes or from failing to find HBV at all.

Notes for use

Use fresh index and results directories. Neither index_suffix nor experiment_id encodes the dataset, so a 62-accession run lands on the same paths as the 47-accession sra50 and 52-accession sra52viral runs. An existing .done marker will make the benchmark silently load the wrong index and report quietly wrong recall. Check the built index has 62 rows before trusting anything.

Handle empty result sets carefully. A method that returns nothing for a query should score 0 recall for it, not be dropped from the denominator. If scoring code ranks by argsort over a zero-filled score vector, ties resolve to the alphabetically first accessions and can award spurious credit. Set unscored accessions to -inf before ranking.

Do not symlink logan_accessions/ at an existing contig tree. The 10 genotype-D accessions are not in the sra50 tree; a symlink would find the 47 distractors, skip the download, and fail verification on the missing 15.

Reference proxies. The k-mer table, the alignment collapse and query_strata.csv all use the reference genomes LC851451.1 (B) and LC852229.1 (D) as stand-ins for the indexed accessions, not the Logan contigs themselves. They characterise the genotypes, not the specific runs.