Stage 1 results — 26 October 2026

Run on a Windows machine under Application Control, using PowerShell and the Ensembl GRCh37 REST API. No Python, no cloud notebook, no account required. This substitutes for src/stage1.py, which could not be executed locally.

Gate status: PASS. Stage 2 may proceed.

Cohort

531 patients, being every TCGA GBM/LGG case present in both the PanImmune neoantigen file and the pMHC reference summary.

The gap between median and mean, and the five-figure maxima, are the hypermutators. They are a small minority pulling the mean far above the median, which is the expected shape and a sanity check that the cohort is behaving.

Reference burden from the pMHC file (patient-HLA based, the comparator for Stage 3): mean 237.5 binding pMHC in GBM, 79.3 in LGG, range 0–18,807.

Every specimen is a primary tumour

Not one recurrence (sample_type 02) in 531 patients. This settles a question a previous analysis by this author could not answer.

That analysis attributed its hypermutated stratum to temozolomide-induced hypermutation, and conceded only that TMZ attribution "cannot be confirmed." It can be checked, and the attribution is almost certainly wrong: these specimens were resected before adjuvant therapy, so post-treatment alkylator hypermutation is not what this cohort contains.

Two consequences:

1. Findings on post-treatment MMR-deficient hypermutation (Touat et al., 2020) apply only weakly to this cohort.
2. Hypermutation in these samples is more plausibly constitutional or sporadic MMR deficiency — the context in which checkpoint response has been favourable rather than poor (Bouffet et al., 2016).

This should be confirmed with mutational signature analysis before being asserted anywhere. Specimen type establishes timing, not mechanism.

Annotation gate

All 3,600,963 MAF rows carry NCBI_Build = GRCh37. No mixed builds. All 13,539 unique transcripts resolved; no failed REST batches.

Plan threshold was ≤1% mismatch. Observed 0.0000%.

The necessary caveat

This is not independent validation and must not be written up as though it were. MC3's VEP annotation was produced against this same GRCh37 annotation set, so complete agreement largely confirms internal consistency. It is close to tautological.

It is nevertheless the property peptide generation requires: the generator must answer what residue does the callset say is there, not what does the newest annotation think. Resolving MC3 coordinates against Ensembl 108/GRCh38 instead discards about 6.4% of variants and silently mis-generates a further 2.4%, measured on a random 1,500-transcript sample. Both numbers belong in the Methods of any resulting paper.

Stage 2 results — peptide generation

Complete. 8–11mers generated around every variant position, from the GRCh37 protein sequences retrieved in Stage 1.

Internal consistency

For a peptide of length L, exactly L windows contain the mutated residue, so instance counts should divide to roughly the variant count. They do:

All four land within 0.5% of the 39,806 variant total, the shortfall being variants close enough to a protein terminus to lose windows. Nothing is silently dropping out.

Positive control

The pipeline independently reproduces the IDH1 R132H epitope:

TCGA-02-2483  R132H  WVKPIIIGH
TCGA-02-2483  R132H  VKPIIIGHH
TCGA-02-2483  R132H  KPIIIGHHA
TCGA-02-2483  R132H  PIIIGHHAY
TCGA-02-2483  R132H  IIIGHHAYG

The tiling matches the IDH1 reference sequence …GWVKPIIIG[R>H]HAYGDQYRAT… exactly, and the same peptides appear in other R132H-positive cases (TCGA-06-0129 among them), as they must.

Note what this does and does not establish. It confirms the pipeline generates the correct mutant peptide — a software check. It says nothing about immunogenicity, and the clinically validated IDH1 R132H epitope is MHC class II restricted, whereas these are class I candidates.

Artefacts

The attrition table has no unexplained gaps. An unreconciled variant count was the defect that sank the previous attempt; every count here traces end to end.

Stage 3 — pending

Binding prediction is the first step the originating machine cannot run: MHCflurry needs Python, and Application Control forbids it.

Scale: 1,448,687 unique peptides against 6 supertype alleles. The presentation predictor returns the best allele per peptide, so this is 1.45M rows rather than 8.7M — and best-allele is the quantity the analysis uses, since the question is how many peptides bind any allele in the panel.

Restricting to 9mers would cut the input by roughly three quarters. The full 8–11 range is what ANALYSIS_PLAN.md specifies; the subset is a fallback if compute binds. Both are drawn from the same generated set, so the choice can be deferred.

src/stage3.py targets a GitHub Codespace and writes predictions.tsv.gz.