A Mapping-Certified, NK-Enriched Cytotoxic-Context Modifier of the TAP-Loader State in Neuroblastoma

# A Mapping-Certified, NK-Enriched Cytotoxic-Context Modifier of the TAP-Loader State in Neuroblastoma A Preregistered Bulk-Transcriptomic Interaction Protocol with a Mandatory Feature-to-Gene Mapping-Integrity Gate **Version:** V5 (mapping-certified) · **Hypothesis ID:** `GTIM_2026_NB_LOADER_NK_CONTEXT_V5`**Claim class:** `NEUROBLASTOMA_LOADER_NK_ENRICHED_CONTEXT_MODIFIER_CANDIDATE`**Status:** Preregistered hypothesis and analysis plan. No results are reported. Abstract **Background.** Neuroblastoma is an immunologically cold pediatric solid tumour in which the MHC class I antigen-processing machinery — including the peptide-loading complex formed by TAP1, TAP2 and the chaperone tapasin (TAPBP) — is frequently downregulated. Whether a low TAP-loader state is clinically adverse may not be fixed, but conditional on whether cytotoxic lymphocytes capable of recognising MHC-I-deficient cells are present. This interaction has not been formally tested. Separately, reanalyses of the most widely used neuroblastoma expression cohort risk a silent feature-to-gene mapping error that can assign expression rows to the wrong genes. **Hypothesis.** The adverse-outcome association of a low TAP-loader state (APM3\_loader = mean z-score of TAP1, TAP2, TAPBP) is attenuated in tumours with a higher NK-enriched cytotoxic context (mean z-score of NKG7, GZMB, KLRD1). The prespecified interaction coefficient is positive. **Design.** A preregistered, two-sided interaction test (α = 0.01) adjusted for B2M and MYCN, executed only after a mandatory Data-Mapping-Integrity Gate certifies that every expression row resolves to its intended gene by full-platform annotation join, multi-identifier concordance and probe-sequence re-alignment, with a biological positive-control gate. All prior results derived from numeric-identifier coincidence are quarantined and replayed after certified remapping. **Boundaries.** This is an association hypothesis only. It does not prove NK-cell identity or function, does not predict response to any therapy, is not a treatment-selection classifier, and is not clinically actionable. 1. Background and Rationale 1.1 The TAP-loader state in neuroblastoma Neuroblastoma carries a low mutational burden and sparse T-cell infiltration, and repeatedly shows reduced surface MHC class I with downregulation of antigen-processing-machinery components, associated with MYCN amplification and adverse features. The peptide-loading complex — the TAP1/TAP2 transporter and the tapasin (TAPBP) editor — is rate-limiting for assembled, peptide-optimised surface MHC-I. 1.2 Why an interaction, not a main effect The clinical meaning of a low TAP-loader state is plausibly **conditional**. Where cytotoxic lymphocytes capable of MHC-unrestricted killing — NK and γδ T cells — are present, reduced MHC-I may be exploited through missing-self recognition; where they are absent, low MHC-I may simply mark an immune-cold state or facilitate T-cell evasion. This reasoning predicts that the association between the loader state and outcome is **modified by** cytotoxic context — an interaction hypothesis with a predeclared direction, not a main-effect claim. 1.3 Why NK-enriched, not pan-cytotoxic A general cytotoxic signature such as NKG7 plus GZMB conflates effectors that move in opposite directions on a loader-low tumour: MHC-restricted CD8⁺ T cells, which **cannot** recognise MHC-I-deficient cells, and NK and γδ cells, which **can**. A pan-cytotoxic modifier therefore risks self-cancellation or a signal driven by the non-informative arm. Adding KLRD1 (CD94), which is enriched in NK and NK-like lymphocytes, biases the modifier toward the MHC-unrestricted arm whose direction the missing-self logic predicts. **KLRD1 is mandatory and may not be dropped or substituted.** 1.4 Why mapping integrity is a precondition The principal discovery cohort, GSE49710, is hosted on platform **GPL16876, Agilent-020382 Human Custom Microarray 44k (Feature Number version)**. The matrix `ID_REF` values are Agilent feature numbers, **not** Entrez Gene IDs. Assigning rows by numeric coincidence — for example mapping rows "6890", "6891" and "6892" to TAP1, TAP2 and TAPBP because those are the Entrez IDs of those genes — silently attaches expression to the wrong features. Because such an analysis can be statistically valid while measuring the wrong genes, feature-to-gene mapping certification is a precondition for any model or verdict, not a downstream check. The validation cohort GSE85047 is on a different platform (GPL5175, Affymetrix Human Exon 1.0 ST) and requires its own independent mapping certification. 2. Hypothesis **Primary hypothesis.** In neuroblastoma, the association between APM3\_loader and adverse outcome is modified by NK\_enriched\_context such that higher NK\_enriched\_context attenuates the adverse association of low APM3\_loader. With higher expression encoded as a higher score, the prespecified APM3\_loader × NK\_enriched\_context interaction coefficient is positive. **Null hypothesis.** The interaction coefficient is zero after adjustment for B2M and MYCN. 3. Frozen Score Definitions | Score | Genes | Formula | Status ||---|---|---|---|| **APM3\_loader** | TAP1, TAP2, TAPBP | mean z-score(TAP1, TAP2, TAPBP) | FROZEN || **NK\_enriched\_context** | NKG7, GZMB, KLRD1 | mean z-score(NKG7, GZMB, KLRD1) | FROZEN || **B2M** | B2M | z-score(B2M) | Separate integrity/adjustment variable; never scored with the loader | B2M is excluded from APM3\_loader because B2M and the loader genes are co-regulated, which would create partial circularity when B2M also informs the MHC-I surface phenotype. Definitions are frozen and may not be modified after any outcome field is examined. 4. Data-Mapping-Integrity Gate Execution is blocked unless `identifier_namespace_verified = PASS`. The lock `NUMERIC_ID_COINCIDENCE_IS_NOT_GENE_MAPPING` applies. The gate proceeds, strictly blind to all outcome fields, as follows: 1. **Platform identity.** Confirm the exact GPL accession and matrix ID namespace before any join. 2. **Full annotation join.** Resolve every matrix feature through the complete official platform annotation: `matrix_ID_REF → platform_feature_ID → gene`. 3. **Multi-identifier concordance.** For each of the eight frozen genes, require concordant GeneSymbol and EntrezGeneID, plus Ensembl, RefSeq or UniProt where available. Reject control, ambiguous, multi-gene or conflicting features. 4. **Sequence-level confirmation (mandatory for the eight frozen genes).** Re-align each candidate feature's probe sequence to the current genome assembly and confirm it maps uniquely to the intended gene locus. Annotation-identifier concordance alone is insufficient, because all identifiers can inherit a common upstream annotation error. A gene whose feature fails sequence confirmation is `PRIMARY_GENE_NOT_SOURCE_CERTIFIED`. 5. **Probe-collapse rule with a predeclared concordance threshold.** When a gene maps to multiple validated features, compute pairwise across-sample Spearman correlations among those features, outcome-blind. If the minimum pairwise ρ ≥ 0.50, collapse by the per-sample median of standardised features. If the minimum pairwise ρ < 0.50, return `MULTI_PROBE_DISCORDANCE_DATA_TILE_REQUIRED`. Probe selection by outcome association, p-value, effect size or survival performance is prohibited. 6. **Biological positive-control gate (blocking, not advisory).** After mapping, the join must pass predeclared biological checks or execution halts: - *MYCN concordance.* MYCN expression must be bimodal and concordant with the known MYCN-amplification labels in the cohort. - *Sex concordance.* XIST and Y-linked genes consistent with recorded sex where available. - *Housekeeping sanity check.* Standard housekeeping genes within expected variability bounds. Failure returns `BIOLOGICAL_QA_FAILED_MAPPING_SUSPECT`. 7. **Outcome-blind mapping freeze and receipt.** Freeze the verified feature set and issue a mapping receipt recording, per gene, the selected platform feature IDs, the annotation source, version and date, the sequence-confirmation result, the concordance metric and collapse decision, the biological-QA status, and a SHA-256 hash of the frozen mapping. 5. Prior-Result Quarantine All previous GSE49710 results derived from direct numeric `ID_REF`-to-Entrez interpretation are quarantined: treated as neither supported, null, nor falsified, and labelled `PREVIOUS_RESULT_MAPPING_DEPENDENT_NOT_INTERPRETABLE`. This explicitly includes the founding discovery anchor — the APM-low/adverse-event association — and every dependent model. The entire lineage is provisionally un-anchored until the discovery anchor is replayed on the certified mapping. 6. Primary Model After mapping certification and outcome-blind score construction: - **Event endpoint (GSE49710):** `adverse_event ~ APM3_loader + NK_enriched_context + APM3_loader × NK_enriched_context + B2M + MYCN` - **Time-to-event endpoint (GSE85047, TARGET-NBL):** `Surv(time, event) ~ APM3_loader + NK_enriched_context + APM3_loader × NK_enriched_context + B2M + MYCN` **Primary test:** the `APM3_loader × NK_enriched_context` interaction coefficient.**Significance threshold:** α = 0.01, two-sided, with a predeclared positive direction.**Score form:** continuous z-scores. No median split enters the primary model. A predeclared median-split view may be used for visualisation only. 7. Power, Estimand and Replication Strategy Interaction tests require substantially larger samples than main effects, and α = 0.01 is conservative. An uninformative result is a realistic outcome and is reported honestly as `INTERACTION_NOT_ESTIMABLE_UNDERPOWERED`. - **Predeclared power screen.** For each cohort, report the minimum detectable interaction effect at α = 0.01 and 80% power given its sample size and event count, computed before interpreting the estimate.- **Primary promotion