Add missing pseudosequences for specific training alleles (~6.6k rows dropped)

[iskandr]

Summary

A small set of real, specific alleles present in the curated training data are dropped at training time because they have no pseudosequence in allele_sequences.csv (so they don't match a key and can't be encoded for the pan-allele model). These are genuine alleles (mostly non-human species), not serotypes or junk — adding their pseudosequences would let the dropped measurements train.

Data (from curated_training_data, 994,948 rows)

16 unique alleles / 6,663 rows (~0.67%) are dropped because they resolve to no pseudosequence key. Of these, 12 parse as a specific Allele (recoverable with a pseudosequence); 4 are gene-level labels (RT1-A, Mamu-E, H2-D, H2-K) that are inherently too coarse and would stay dropped.

allele	rows	parses as
Trvu-UB*01:01	5,607	Allele
Mamu-A*07	681	Allele
RT1-A	139	Gene (too coarse)
Ptal-N*01:01	74	Allele
SLA-3*02:02	41	Allele
FLA-E*018:01	36	Allele
Gaga-BF2*13:01	30	Allele
Mamu-E	16	Gene (too coarse)
Eqca-N*006:02	13	Allele
SLA-1*03:02	7	Allele
H2-D	5	Gene (too coarse)
SLA-2*01:06	4	Allele
H2-K*wm7	4	Allele
H2-K*bm8	2	Allele
Mamu-B*04	2	Allele
H2-K	2	Gene (too coarse)

Trvu-UB*01:01 (a trout/fish MHC allele) alone is 5,607 rows — most of the recoverable volume.

Why it matters

These are the actual training rows currently lost to the pan-allele model (not serotypes — serotypes are ≈0% of the curated data; see sibling issue). Recovery requires authoring/adding pseudosequences, not new code:

• Add pseudosequence rows for the 12 specific alleles to allele_sequences.csv (where structural/sequence data exists), OR
• Decide they're out of scope (rare non-human species) and document that the drop is expected.

Proposed work

1. For each of the 12 specific alleles, determine whether a pseudosequence can be derived (the make_allele_sequences.py / filter_sequences.py pipeline under downloads-generation/allele_sequences/).
2. Regenerate allele_sequences.csv including them; retrain picks them up automatically (training ingestion now canonicalizes against whatever keys exist).
3. Leave the 4 gene-level labels dropped (correctly too coarse).

Effort

Small-to-medium — mostly data work in the allele_sequences generation pipeline, no model-code changes. Low volume overall (~6.6k rows, dominated by one fish allele), so prioritize accordingly.

How to reproduce the list

Classify curated_training_data.csv.bz2 alleles against allele_sequences.csv keys via mhcflurry.common.canonicalize_allele_series; the unresolved ones are the table above.

[iskandr]

Related: #293 (serotype expansion for prediction) and #295 (serotypes in training). Unlike those, this is real recoverable training volume (~6.6k rows) and needs pseudosequence data, not serotype logic.