expression: GTEx per-tissue and TCGA per-tumour-type reference table

reference_expression.tsv.gz, 5,656,063 rows, one flat table with source and context per row so
the two are never averaged together:

gtex gene x tissue 3,955,284 rows 73,321 genes 53 tissues normal-tissue TPM
tcga peptide x cancer 1,700,779 rows 19 tumour types expressionEB++

Keyed differently on purpose. GTEx is gene-keyed (the imputation source and the safety read).
TCGA is peptide-keyed because its source table carries ensp and no ENSP->symbol map ships with
it, so a gene-level join would be a guess; the peptide-level join is exact and additionally
answers whether an exact neoantigen has been seen expressed in that tumour type. SKCM (melanoma)
covers 43,243 peptides.

Verified by lookup rather than by row count: PMEL 44.4 / MLANA 10.7 / TYR 9.3 TPM in sun-exposed
skin against 1.14 / 0.26 / 0.00 in lung and 0.24 / 0.14 / 0.00 in whole blood, while ACTB runs
1467 / 3341 / 4562. Three melanocyte antigens skin-restricted, one housekeeping gene ubiquitous.

Built by bench/neoag/expression_reference.py in 2026-mhcmatch-benchmark. Read by
mhcmatch.expression (0.9.1+).

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

expression/reference_expression.tsv.gz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6274d089b63fa85d81f4c82acf3b63e4d63a1ca56683f76c5e3f2ba56ff20d46
+size 105486870