Verification Report

This report documents the independent verification of the article TPMT: ESM-2 / VAMP-seq Complementarity.

---

Validation Methodology

DMS source: Matreyek et al. 2018, Nature Genetics 50(6):874–882 (doi:10.1038/s41588-018-0122-z). The Matreyek paper introduced VAMP-seq — Variant Abundance by Massively Parallel Sequencing — and applied the assay to multiple proteins including TPMT, the human thiopurine S-methyltransferase whose deficiency mandates pre-treatment genotyping under CPIC guidelines for 6-mercaptopurine, azathioprine, and thioguanine therapy. EGFP-TPMT fusion was expressed in HEK293T landing-pad cells; cells were FACS-sorted into 4 fluorescence bins and deep-sequenced; Enrich2 scored 8 biological replicates against WT synonymous (normalized to 1.0) and median nonsense at positions 51–219 (normalized to 0.0). Scores are deposited at MaveDB as urn:mavedb:00000013-b-1 (CC0 license, N = 3,689 missense variants scored of 4,655 possible).

ESM-2 scoring: 650M-parameter masked-marginal scoring on the NeuroAutomata platform across all 245 positions of TPMT (UniProt P51580). Score polarity aligned natively against VAMP-seq abundance (positive = stable/abundant, negative = unstable/degraded); both metrics agree on direction, no inversion applied. Cross-referencing variants between the ESM-2 landscape and VAMP-seq scores yielded N = 3,685 matched after WT-identity filtering. Position stratification used SAM-binding active-site annotations to partition active-site (N = 14), non-active-site annotated (N = 150), and unannotated (N = 3,521) positions.

Global Spearman rho = 0.530 (N = 3,685, p = 7.4 × 10⁻²⁶⁶); Pearson r = 0.535. This sits near the ProteinGym ESM-2 650M baseline of 0.414 across 217 assays — a positive but modest gap consistent with VAMP-seq abundance selection on a stable single-domain enzyme. The companion NUDT15 benchmark (rho = 0.526), another pharmacogenomic enzyme in the same thiopurine metabolic pathway scored by VAMP-seq, sits within 0.004 — strong evidence that rho ≈ 0.53 is characteristic for thiopurine-pathway enzymes under VAMP-seq abundance selection.

Regional breakdown carries the structural boundary. The SAM-binding active-site domain achieves Spearman rho = 0.222 at N = 14 (p = 0.45, non-significant) — consistent with the active-site small-N + specificity-engineering boundary observed in PTEN, BRCA1, and prior bacterial enzyme benchmarks. Non-active-site annotated positions (rho = 0.524, N = 150) and unannotated positions (rho = 0.531, N = 3,521) recover the global signal cleanly. The article frames the active-site result as a small-N caveat rather than a discriminative finding — the underlying boundary is methodological, not biological.

Clinical pharmacogenomic alleles (TPMT*2 A80P, *3B A154T, *3C Y240C, *5 L49S, *7 H227Q, *8 R215H) were checked for whether ESM-2’s directional ranking agrees with VAMP-seq abundance. The article frames these as descriptive ESM-2 directional concordance against published clinical phenotype severity (*2 / *3B correctly flagged as low-abundance; *3C borderline; *5 / *7 activity loss not abundance-driven), not as ACMG/AMP variant classifications. The underlying record matches that conservative framing: clinical-allele scores are recorded as quantitative measurements on the VAMP-seq abundance scale with a descriptive concordance reading, not as classification predictions — authored that way from the start, the same approach taken on the earlier CYP2C9 report. The classification check found no mismatch between that record and the published prose.

---

Summary

Metric	Value
Claims	33 total
Result	All checked; the page passed verification (5 numerically verified / 12 single-source measurements / 16 reference facts; 0 contradicted, 0 uncheckable)
Source data	TPMT validation dataset (version recorded below at build time)
Verified	2026-05-18
Verified by	Veritas

---

Claim-by-Claim Verification

What was checked, and how each claim resolves against its primary source, is shown claim by claim below. The exact source-data version and verification-result version used at this page’s most recent build are shown in the page footer.

---

What Was Verified

Framing-fidelity check — all 33 claims. Every claim was checked against its primary source for whether the source actually entails it, whether the scope matches, and whether the prose framing is faithful to what the cited authors reported. Key sources: Matreyek et al. 2018 Nat Genet (TPMT VAMP-seq DMS), UniProt P51580 (TPMT canonical reference), MaveDB urn:mavedb:00000013-b-1 (deposited score set), PharmVar TPMT (clinical allele inventory), the CPIC TPMT / thiopurine guidelines (clinical-context framing), and the ProteinGym substitution DMS leaderboard (cross-model baseline, accessed 2026-05-08).

Numerical check. Of the 33 claims, 5 carried figures that were arithmetic-verified with explicit numeric-range and consistency checks. A further 12 are single-source DMS measurements — single values from one deposited record, with no independent second source available to cross-check them against by construction. We disclose these as exactly that: measured once, not independently corroborated. This is an honest, bounded category, distinct from a verified claim; it is not a coverage gap. The remaining 16 are reference facts (existence, methodological, comparative, literature-sourced, and causal statements) with no arithmetic in them to check — confirmed by the framing-fidelity review above.

Classification check — 0 classification claims. The TPMT clinical-allele claims were written from the start as quantitative measurements on the VAMP-seq abundance scale with a descriptive concordance reading — the same approach taken on the earlier CYP2C9 report — not as classification predictions. No mismatch was found between that record and the published prose. ESM-2 is a ranking tool, not a classifier; categorical thresholds do not generalize across proteins.

One consistency layer not yet active. A further consistency check is planned but was not active for this report.

On retroactive verification. This is an older post, originally written before our current per-claim verification discipline existed; it was brought under verification retroactively — the claim list was written against the already-published article and the existing source-verification records, then independently verified. The source-data version and verification-result version used at this page’s most recent build are both shown in the footer, an evidence chain back to a specific, reproducible state.

---

Verification independence

Verification is performed by Veritas, an independent agent with no involvement in producing the content it checks.