openvax / topiary / 24434611335

88%

master: 90%

Pull #135

github

Add SelfProteome — nearest-self lookup architecture (part A of #124)

SelfProteome holds a species-tagged, scope-filtered reference protein
corpus indexed by peptide length and answers per-query nearest-
neighbor lookups.  Plugs into TopiaryPredictor via a new
self_proteome= kwarg; result DataFrame gains self_nearest_peptide /
_peptide_length / _edit_distance / _gene_id / _transcript_id /
_reference_offset / _reference_version columns.  Columns are joined
before filter_by / sort_by so they can participate in DSL expressions.

This PR ships the architecture with one axis (sequence-nearest) and
substitutions only.  Follow-ups queued on #124:
  - scope="protected_tissues" + HPA/GTEx tissue filter
  - 1aa insertion / deletion candidates
  - self_nearest_by_binding / self_strongest_nearby second + third axes
  - self_nearest_candidates structured column
  - Seed-and-extend algorithm once the benchmark decides

## Surface

- SelfProteome.from_peptides(dict, peptide_lengths=...) — test helper
  for in-memory reference sets.
- SelfProteome.from_fasta(path, scope=...) — FASTA loader, scope
  limited to "all" or a callable (no gene metadata in FASTA).
- SelfProteome.from_ensembl(species, release, scope=..., cta_source=...)
  — pyensembl-backed loader.  Defaults to scope="non_cta" for human,
  stripping CTA genes via the existing topiary.sources pirlygenes
  integration.  Non-human species must pass cta_source explicitly
  when using "non_cta"; a clear ValueError fires otherwise.
- reference_version property composes an "ensembl-{species}-{release}+
  scope-{scope}+..." string that gets stamped on every output row.
  Custom filters (sets or callables) hash into the string so
  reproducibility holds even without a stable label.
- nearest(peptides) returns a DataFrame with one row per query,
  preserving input order.  Queries whose length isn't represented in
  the reference get None rows rather than raising.

## Algorithm

SIMD-vectorized Hamming distance against... (continued)

169 of 203 new or added lines in 3 files covered. (83.25%)

3480 of 3955 relevant lines covered (87.99%)

2.64 hits per line