Sequences

Page version 1.0.0

Benchmarks

Each section gives recorded runs of Sequences: what they measured, the data they ran on, what they found and the evidence records that hold their numbers. The runs compare search methods, not products.

Exact and pattern counts against GNU grep

Exact text and Pattern count every stored sequence that holds a query. The benchmark of Sequences 0.22.0, run on November 6, 2026, checked each count against GNU grep, which read the same sequence files and shares no code with Sequences.

Data: snapshot 2026-11-05-c456f6bdba29 of the patent sets, with 28,757,943 DNA sequences of 15,612,493,945 bases in all and 10,491,135 proteins of 1,718,205,235 residues.

Matches counted by Sequences and by GNU grep
QueriesSearchesCounted by SequencesCounted by GNU grepEqual
DNA fragments of 15 to 50 bases72145,561145,56172 of 72
Peptides of 8 to 30 residues2031731720 of 20
DNA patterns724,221,00124,221,0017 of 7
Protein patterns66,360,2466,360,2466 of 6

105 of the 105 counts were equal, query by query. A match is a stored sequence that holds the query, counted once for each DNA strand that holds it.

Evidence: biosequence-benchmark-0.22.0.json, version 2.0.0.

The exact scan and BLAST on short DNA fragments

Claims often recite short DNA, such as primers and probes. The same benchmark cut 72 fragments of 15 to 50 bases from stored patent DNA, unchanged, 12 of each length. Each was searched with Exact text, which scans every stored sequence, and with BLAST, and GNU grep listed every stored sequence that holds it.

Share of the sequences GNU grep found that each method listed, by fragment lengthExact text listed 100% at 15 bases, 100% at 18 bases, 100% at 21 bases, 100% at 25 bases, 100% at 30 bases, 100% at 50 bases. BLAST listed 49.5% at 15 bases, 95.7% at 18 bases, 96.9% at 21 bases, 99.3% at 25 bases, 29.0% at 30 bases, 100% at 50 bases.Exact textBLAST0%25%50%75%100%Exact text, 15 bases: 100%BLAST, 15 bases: 49.5%15Exact text, 18 bases: 100%BLAST, 18 bases: 95.7%18Exact text, 21 bases: 100%BLAST, 21 bases: 96.9%21Exact text, 25 bases: 100%BLAST, 25 bases: 99.3%25Exact text, 30 bases: 100%BLAST, 30 bases: 29.0%30Exact text, 50 bases: 100%BLAST, 50 bases: 100%50Fragment length in bases
Stored sequences that hold the fragments: those GNU grep found, and those each method listed
LengthGNU grepExact textBLAST
15 bases38,14238,142 (100%)18,878 (49.5%)
18 bases10,09710,097 (100%)9,658 (95.7%)
21 bases14,01914,019 (100%)13,590 (96.9%)
25 bases9,6939,693 (100%)9,624 (99.3%)
30 bases51,46951,469 (100%)14,931 (29.0%)
50 bases8,8898,889 (100%)8,889 (100%)
All132,309132,309 (100%)75,570 (57.1%)
The chart gives each method's share of the sequences GNU grep found; the table gives the counts.
Fragments whose source sequence came back
LengthExact textBLAST
15 bases12 of 1210 of 12
18 bases12 of 1212 of 12
21 bases12 of 1212 of 12
25 bases12 of 1212 of 12
30 bases12 of 1211 of 12
50 bases12 of 1212 of 12
All72 of 7269 of 72

Exact text listed every sequence GNU grep found, 132,309 of 132,309. BLAST listed 75,570: it listed at most 5,000 sequences for each query, and 41 of the 72 fragments reached that limit. BLAST also did not return the stored sequence that 3 of the fragments were cut from. Automatic, which adds an exact scan for short DNA, listed all 132,309.

Evidence: biosequence-benchmark-0.22.0.json, version 2.0.0.

Finding altered sequences

A claim can reach sequences that differ from the one it recites. The benchmark cut peptides and proteins of 8 to 500 residues from stored patent proteins and changed letters at random, keeping 100% down to 40% of them, 60 queries at each level. A query counts as found when the stored sequence it was cut from comes back. For proteins, Smith-Waterman ran as the automatic choice.

Queries found at each share of letters kept, of 60At 100% kept, Smith-Waterman found 60, BLAST 54 and Exact text 60. At 95% kept, Smith-Waterman found 60, BLAST 54 and Exact text 6. At 90% kept, Smith-Waterman found 60, BLAST 53 and Exact text 0. At 80% kept, Smith-Waterman found 58, BLAST 51 and Exact text 0. At 70% kept, Smith-Waterman found 58, BLAST 49 and Exact text 0. At 60% kept, Smith-Waterman found 55, BLAST 44 and Exact text 0. At 50% kept, Smith-Waterman found 48, BLAST 40 and Exact text 0. At 40% kept, Smith-Waterman found 39, BLAST 33 and Exact text 0.Smith-WatermanBLASTExact text015304560Smith-Waterman, 100% kept: 60Smith-Waterman, 95% kept: 60Smith-Waterman, 90% kept: 60Smith-Waterman, 80% kept: 58Smith-Waterman, 70% kept: 58Smith-Waterman, 60% kept: 55Smith-Waterman, 50% kept: 48Smith-Waterman, 40% kept: 39BLAST, 100% kept: 54BLAST, 95% kept: 54BLAST, 90% kept: 53BLAST, 80% kept: 51BLAST, 70% kept: 49BLAST, 60% kept: 44BLAST, 50% kept: 40BLAST, 40% kept: 33Exact text, 100% kept: 60Exact text, 95% kept: 6Exact text, 90% kept: 0Exact text, 80% kept: 0Exact text, 70% kept: 0Exact text, 60% kept: 0Exact text, 50% kept: 0Exact text, 40% kept: 0100%95%90%80%70%60%50%40%Share of letters kept
Queries found at each share of letters kept, of 60
Letters keptSmith-WatermanBLASTExact text
100%605460
95%60546
90%60530
80%58510
70%58490
60%55440
50%48400
40%39330
The chart gives the queries each method found at each level; the table gives the same numbers.

In all, Smith-Waterman found the most of these queries, then BLAST, then Exact text, and the other altered sets gave the same order:

Queries found, by method
QueriesSearchedSmith-WatermanBLASTExact text
Proteins, letters changed at random, 100% to 40% kept48043837866
Proteins, letters changed as BLOSUM62 weighs them, 95% to 40% kept2101921593
Proteins with 2 to 10 insertions or deletions per 100 residues9083740
DNA fragments of 21 to 50 bases, 95% to 80% kept6053470

Evidence: biosequence-benchmark-0.22.0.json, version 2.0.0.

Within k changes on the fragments Smith-Waterman missed

In the benchmark, Smith-Waterman missed 7 of the 60 altered DNA fragments. Within k changes lists every stored sequence that differs from a query by at most k changes, with no score or E-value cut. On November 15, 2026, in the test runs before its release in Sequences 0.95.0, it searched all 20 fragments of the 4 groups that held those misses: first counting substitutions only, then insertions and deletions too.

Fragments found in the 4 groups that held the misses
FragmentsLetters changedSmith-WatermanWithin k changes
21 bases, 90% kept24 of 55 of 5
21 bases, 80% kept41 of 55 of 5
30 bases, 95% kept24 of 55 of 5
50 bases, 95% kept34 of 55 of 5
The k each group's fragments needed to come back
FragmentsSubstitutions onlyInsertions and deletions too
21 bases, 90% kept22
21 bases, 80% kept43 to 4
30 bases, 95% kept22
50 bases, 95% kept33

No fragment needed a k larger than its changed letters, and k = 4 was the most any needed. Across the benchmark's queries with up to 5 letters changed, Within k changes found the stored sequence each was cut from at k equal to its changes: 186 of 186 protein queries and 117 of 117 DNA queries.

Evidence: 2026-11-15-within-k-changes-release, version 1.0.0; 2026-11-15-pilot-within-k-changes/results/golines.json, version 1.0.0; biosequence-benchmark-0.22.0.json, version 2.0.0.

Repeat runs give the same hits

The benchmark of Sequences 0.22.0 ran every query twice. All 5,182 results, one for each query and method, were the same in both runs: whether the source sequence came back, every Exact text and Pattern count, and how many of the sequences GNU grep found each listed. But 265 of its 1,173 Smith-Waterman and global searches listed a different number of sequences, by up to 1,119, because the sample of scores behind their E-values could change from run to run.

Sequences 0.95.0 ships a build of FASTA36, the program behind those methods, that draws the same sample on every run. On November 15, 2026, that build ran the same 1,173 searches three times: runs A and B on 8 threads each, and run C on 12.

Searches that differed between two runs of the repeatable build: in their count of hits, in the sequences they listed, or in their rows, scores and E-values
RunsSearchesCountSequencesRows
A and B1,173000
A and C1,173000
B and C1,173000

Evidence: biosequence-benchmark-0.22.0.json, version 2.0.0; 2026-11-15-pilot-sw-repeatable, version 1.4.0.

The ST.26 listing check

The listing check, on the Listings tab of Sequence tools, checks a WIPO ST.26 sequence listing before it is filed. From November 7 to 10, 2026, its results were compared with ST.26 validation results on 183 listings: a base listing with no errors, 20 listings with known problems, 161 that each change one thing in the base, and one real listing.

What ST.26 validation and the listing check found, of 183 listings
ResultListings
ST.26 validation found an error115
ST.26 validation found no error68
The listing check agreed: an error where validation found one, none where it found none182
The two differed1

They differed on one listing, whose XML declaration has no encoding: validation reported no error, and the listing check reports one, because paragraph 39(a) of ST.26 gives the declaration with its encoding.

Evidence: the listing check's record of November 8, 2026, version 1.1.1.

A search plan for 1,034 SEQ ID NOs

A search plan turns the SEQ ID NOs a patent's claims recite into searches. On November 14, 2026, the hosted service, then running Sequences 0.85.0, ran a plan for US 12,780,932 B1, whose claims recite 1,034 SEQ ID NOs. Its sequence listing skips 8 of them, so the plan searched the other 1,026, all proteins, with Smith-Waterman on Patent proteins and Swiss-Prot: 2,052 queries of 152,040 letters in all, as 104 jobs, 2 at a time.

The run time expected when the plan was locked, and the time each run took
RunSearchedJobsExpectedTook
Run 1Patent proteins5265 minutes75.9 minutes
Run 2Swiss-Prot5228 minutes6.7 minutes
BothPatent proteins and Swiss-Prot10493 minutes82.5 minutes

Both is the time from the first job's start to the last job's end, and each time is rounded to a tenth of a minute.

All 104 jobs completed, in 82.5 minutes against the 93 the plan expected. The total came in under the estimate; the split between the two runs did not match it.

Evidence: 2026-11-14-query-groups-pilot, version 1.0.0.

Searches an hour with more workers

A Sequences server can hand searches to workers, other servers that take searches from its queue. On November 15, 2026, Sequences 0.92.0 ran batches of 20 and 100 searches, each batch queued at once, on snapshot 2026-11-13-5e447c75c43f: on a server alone, then on a server with 2 and with 3 workers. Half of each batch were proteins searched with Smith-Waterman and half DNA searched with BLAST. Each search ran on 6 threads. Each worker had 6 CPUs of its own; the server alone had 8 and ran one search at a time.

Searches an hour, 20 searches queued at onceThe server alone ran 363.6 searches an hour, 2 workers 598.5 and 3 workers 741.5.20 searches queued at once020040060080020 searches queued at once, the server alone: 363.6363.6Server alone20 searches queued at once, 2 workers: 598.5598.52 workers20 searches queued at once, 3 workers: 741.5741.53 workersServers running searches
Searches an hour, with 20 and with 100 searches queued at once
Servers20 queued100 queued
The server alone363.6Ran beside other work
2 workers598.5552.1
3 workers741.5828.0
The chart gives the batches of 20; the table adds the batches of 100.

With 20 searches queued, 2 and 3 workers ran 1.65 and 2.04 times as many searches an hour as the server alone. With 100 queued, 3 workers ran 1.50 times as many as 2. The batches on one worker, and the server alone's batch of 100, ran beside other work on the same computer, so the record leaves them out of these comparisons.

Evidence: 2026-11-15-scaling-measured, version 1.2.0.