Verification Report

This report documents the independent verification of the article Scoring 6,142 CYP2C9 Variants in 30 Seconds: ESM-2 vs Deep Mutational Scanning.

---

Validation Methodology

Deep mutational scanning (DMS) source: Amorosi et al. 2021, American Journal of Human Genetics 108(9):1735–1751 (doi:10.1016/j.ajhg.2021.07.001). The Amorosi dataset reports two independent assays on saturation mutagenesis libraries: a click-seq enzymatic activity assay (6,142 missense variants, TAHA-ABP covalent probe labeling active enzyme in yeast) and a VAMP-seq protein abundance assay (6,370 missense variants, CYP2C9-eGFP fusion in HEK293T human cells). Scores are normalized so synonymous variants ≈ 1.0 and nonsense variants ≈ 0.0.

ESM-2 scoring: 650M-parameter masked-marginal scoring on the NeuroAutomata platform across all 490 positions of CYP2C9 (UniProt P11712). Score polarity aligned natively against the Amorosi activity and abundance scores; no inversion applied. We cross-referenced variants between the ESM-2 landscape and each Amorosi assay; per-domain stratification used UniProt active-site annotations cross-referenced against PDB structure 1R9O.

Global Spearman rho on the activity assay = 0.679 (N = 6,142, p < 10⁻³⁰⁰); global Spearman rho on the abundance assay = 0.634 (N = 6,370, p < 10⁻³⁰⁰). These are ESM-2’s correlations measured against the Amorosi data, not figures reported by Amorosi. The activity correlation exceeds the abundance correlation — ESM-2 captures functional constraint beyond protein stability alone.

Regional stratification carries the structural story. The heme-binding active-site domain (positions 428-437, centered on Cys435 axial heme-iron ligand) achieves Spearman rho = 0.811 (N = 156, p = 1.2 × 10⁻³⁷) — among the highest domain-level correlations we have measured across our benchmarks. This inverts the active-site-weakness pattern observed in PTEN, BRCA1, TPMT, and prior bacterial enzyme benchmarks: heme-thiolate coordination chemistry is universally conserved across ~2 billion years of CYP-family evolution and ESM-2 reads this constraint precisely. By contrast, the SRS5 substrate recognition site (rho = 0.422, N = 29) shows substantially weaker correlation — the enzyme’s substrate specificity has been explored across the CYP superfamily, and ESM-2 reads this exploration as tolerance, not constraint. The distinction (conserved catalytic machinery vs. evolvable substrate specificity) is the methodologically load-bearing regional finding.

Three clinical pharmacogenomic alleles (CYP2C9*2 R144C, *3 I359L, *11 R335W) are checked for whether ESM-2’s directional ranking agrees with published clinical phenotype severity on the DMS activity scale. The article frames these as descriptive ESM-2 directional concordance against published clinical phenotype severity (*11 near-zero > *3 major reduction > *2 mild reduction), not as ACMG/AMP variant classifications. Clinical-allele scores are recorded as quantitative measurements on the DMS activity scale with a descriptive concordance reading, not as classifications. ESM-2 is a ranking tool, not a classifier; categorical thresholds do not generalize across proteins.

A fourth clinical allele, CYP2C9*8 (R150H), is separately disclosed as a documented false negative rather than included as a concordance claim. ESM-2 scores R150H as mildly beneficial (+0.61) against an experimentally decreased-function variant (Amorosi click-seq activity score 0.732). We carry this as a documented, verified limitation rather than a passing mention — the same way we record the Calmodulin named weakness, here at finer per-allele grain. We disclose it openly rather than omit it; the set of concordance claims carries only the three alleles ESM-2 reads concordantly with experiment.

---

Summary

Metric	Value
Claims	15 total
Result	All checked; the page passed verification
Per-claim outcome	7 verified · 4 single-source measurements · 4 reference facts (0 contradicted, 0 uncheckable)
Numbers checked for arithmetic consistency	7 of 15
Categorical classifications made	0 (none authored — see “What Was Checked”)
Source data	CYP2C9 validation dataset (version recorded below at build time)
Verified	2026-05-26
Verified by	Veritas

---

What Was Checked

Source-against-claim review (all 15 claims). Every claim was checked end-to-end against the primary source it cites — whether the source actually says what we claim, and whether the scope matches. A second, stricter check — framing-fidelity, i.e. whether the prose framing is faithful to what the cited authors reported — ran additionally on the verified subset only, not on all 15 (the per-claim outcomes below say which claims received it).

Cited primary sources (independently resolved and checked figure-by-figure for arithmetic consistency, with single-source measurements disclosed as such): Amorosi et al. 2021 AJHG — the click-seq activity DMS dataset (MaveDB urn:mavedb:00000095-a-1) and the VAMP-seq abundance DMS dataset (MaveDB urn:mavedb:00000095-b-1); UniProt P11712 (canonical CYP2C9 sequence reference); and the ProteinGym substitution DMS leaderboard (cross-model baseline, accessed 2026-05-08). Of the seven verified claims, one is taken directly from Amorosi; the other six are ESM-2’s own measurements compared against the Amorosi DMS — including the CYP2C9*8 acknowledged false negative — and each of the seven additionally passed the framing-fidelity check against the Amorosi click-seq activity and VAMP-seq abundance data.

Context sources (referenced for background, outside the formal verification scope): PDB 1R9O (crystal structure for heme-coordination geometry), PharmVar CYP2C9 (clinical allele inventory), and the CPIC warfarin/CYP2C9 guidelines (clinical-context framing).

Arithmetic-consistency check (7 of 15 quantitative claims). Every Spearman rho, sample size, p-value, and percentage-variance figure with arithmetic content was checked for internal consistency — rho values within the valid range (−1 to 1), N values consistent across the activity and abundance cross-reference windows, p-value bounds, percentage-variance derivation (R² = 0.56 from Pearson r = 0.748, the activity↔abundance correlation), and the heme-binding domain N = 156 / SRS5 N = 29 / non-active-site N = 153 / unannotated N = 5,804 partitioning. The CYP2C9*8 false-negative figures were also checked (R150H ESM-2 score +0.61; experimental activity 0.732).

Four single-source measurements. A further 4 claims are single-source DMS measurements with no independent second measurement to cross-check the value against. 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 4 claims (a dataset identifier, the UniProt accession, and structural benchmark-comparison references) are factual or structural references confirmed by source-against-claim review only — there is no arithmetic in them to check.

Categorical-classification check — 0 claims, none authored. An early draft encoded the three concordant clinical pharmacogenomic alleles (CYP2C9*2 R144C, *3 I359L, *11 R335W) as “Likely Pathogenic” categorical predictions. This was caught and recoded before publication — ESM-2 is a ranking tool, not a classifier; we don’t have the calibrated routing needed to make formal ACMG/AMP variant classifications, and the article’s published prose framed these as descriptive directional concordance against published clinical phenotype severity, not as variant classifications. The underlying record now matches the prose: clinical-allele scores are recorded as quantitative DMS-scale measurements with a descriptive concordance reading — not as classifications.

---

How This Page Works

This verification page was generated retroactively against an article first published before our current per-claim verification discipline existed. Working from the published article and our verification records, we wrote a version-pinned claim list to match the existing content, then had it independently verified by Veritas. This lets us bring older posts under the same per-claim verification standard as posts written under the current process, without republishing the original article.

The exact source-data version and verification-result version used at this page’s most recent build are shown in the footer below — 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.