Claims

each row: the claim, the result that would disprove it, the verdict with its number, the evidence

● Holds (1)

• ●
  Parity with phmmer inside mammals

  Disproved if: Fmax more than 0.05 from phmmer's on any of the 6 mammals

  holds · −0.042 to +0.013 across 6 of 6 · leaderboard

  Scope and related numbers

  Swiss-Prot features, raw calls, setting hp_thomas_dill2_k24_lcFalse, 998 human queries

  • on Pfam domains, the best kmerseek setting for each mammal: −0.193 to +0.130 across 6 of 6; furthest: clupus −0.193
  • settings that pass on Swiss-Prot features (raw or dedup): 2 of 30 setting and dedup choices pass (15 settings): hp_thomas_dill2_k24_lcFalse raw, hp_thomas_dill2_k24_lcFalse dedup

  Source: seanome/2024-kmerseek-analysis@4a4b1171 results.parquet

○ Not yet tested (3)

• ○
  Exact HP seeds find homologs below 40% identity

  Disproved if: Fmax under 0.05 in the under-40% identity bins on Pfam-N, for every hydrophobic-polar setting

  not yet tested · the 0-20%, 20-30%, 30-40% bins hold at most 1 region for any setting and relative; at least 30 are needed to decide · notebook 230 to 232

  Scope and related numbers

  Pfam-N regions, raw calls, kmerseek hp_thomas_dill2 settings, all 15 relatives

  • share of Pfam pairs at 20-30% identity sharing one exact hp_thomas_dill2 23-mer: 0.038
  • recall at a 0.001 false-positive rate for an exact 23-mer seed, Pfam 20-30%: 0.039
• ○
  Short features under 60 aa placed at 50 to 90% identity

  Disproved if: a Kyte-Doolittle hydrophobicity window, with no search, lands on as many of the 87 cases

  not yet tested · the Kyte-Doolittle window was not run on the 87 cases · notebook 244

  Scope and related numbers

  the 87 short Swiss-Prot features of notebook 244 where kmerseek drew the boundary and every other tool's call ran past it

  • short features passing every criterion in notebook 244: 87
  • across all Swiss-Prot features, not only short ones, kmerseek settings above a Kyte-Doolittle scan: 0 of 406
• ○
  The BHF membrane tether is a find other tools miss

  Disproved if: a Kyte-Doolittle window on BHF calls the same region

  not yet tested · no Kyte-Doolittle window has been run on BHF · notebook 241

  Scope and related numbers

  Botryllus histocompatibility factor (BHF)

  • best E-value of any human protein for BHF, any kmerseek setting (notebook 241): 0.609548

⊘ Disproved (2)

• ⊘
  Faster per query than MMseqs2

  Disproved if: kmerseek answers fewer queries per second than MMseqs2 on the same 1,000 Swiss-Prot queries

  disproved 13 Sep 2026 · 2.29 queries per second against MMseqs2's 78.6 · notebook 239

  Scope and related numbers

  1,000 Swiss-Prot queries against Swiss-Prot on one laptop; kmerseek hp k15, scaled 10

  • kmerseek search peak memory, GB: 82.68
  • MMseqs2 -s 5.7 search peak memory, GB: 2.55

  Source: notebooks/239_laptop_scale_swissprot.ipynb @ ba20e8b, cell 1

• ⊘
  Outperforms Foldseek on SCOPe

  Disproved if: inside FoldSeek's 5,713-query SCOPe40 benchmark, kmerseek detects no query that FoldSeek misses

  disproved 30 Aug 2026 · fired on 188 of 5,713 benchmark queries (3.3%); 0 that FoldSeek missed · notebook 076

  Scope and related numbers

  SCOPe40, hp_thomas_dill at k 26, family level

  • per-query family sensitivity, kmerseek better than FoldSeek on 228 shared queries (Wilcoxon): p=0.1171
  • best kmerseek 0.4 AUC over all 5,713 reference queries (family), any setting: 0.295 (mmseqs12.k6.c_max|region_evalue, n = 5,713 queries)
  • best kmerseek 0.4 AUC over only the queries it fires on (family), any setting: 0.999 (uniprot18.k9.exact|f_weighted_target_in_query, n = 753 queries)

  Source: notebooks/076_hp_alphabet_sweep_per_class.ipynb @ 31f53c1, cell 19