Corrections log

This is the versioned record of expert-review corrections to the Atlas. The Atlas is predictive, computational infrastructure (AlphaFold / AlphaMissense / DynaMut2-derived) with curated annotations (ClinVar, gnomAD). Predictions get reviewed, corrected, and validated against wet-lab and published data — and that review happens in the open, here. Review by affected-family domain experts is a credibility asset of this platform, not a footnote.

How this file works

• Entries land under a v-next heading while unreleased, then roll into a numbered version when published to the Atlas site.
• data/corrections/corrections.json is the machine-readable twin of this file and is what the Wolfram Research Intelligence consumes (corrections are appended to the affected variant's card and outrank the original card values). Keep the two in sync — edit both or neither.
• Every entry carries: what was wrong / incomplete, the correction, the reviewer, the date, and references where applicable.

---

v-next — unreleased

Raised by: Dr. Fumihiko Urano (WashU), 2026-08-27, and Dr. Sarah Gladstone (The Snow Foundation), 2026-09-02. Entry 3 was surfaced by the Wolfram Research Intelligence assistant reading its own card. Corrected by Mike Wallace II.

Atlas-wide (data)

1. ER topology at 870–890, and the σ-1 candidacy that keys off it. The classifier placed the lumenal C-terminal domain at 653–890. UniProt O76024 annotates an eleventh transmembrane helix at 870–890. Those residues were therefore labelled Lumenal, and because the σ-1 receptor agonist candidacy badge keys off a Lumenal label, 56 variants (32 missense, 17 synonymous, 4 frameshift, 2 in-frame indel, 1 nonsense) were being surfaced as sigma-1 candidates on a topology call that is wrong. UniProt's curated helix annotations now take precedence over the coarse 3-band model. Atlas-wide σ-1 candidate count falls from 579 to 523.

   How it surfaced: Dr. Urano was asked to review two cavities found by a new docking layer built on the Atlas. He replied that WFS1 sits in a membrane the model does not contain, so the gaps between membrane-spanning segments look like ideal pockets when in life they are full of lipid, and asked that the residues lining the top-scoring cavity be checked against the membrane before any library was docked against it. Running that check against the UniProt annotations confirmed his call (21 of 27 lining residues sit inside annotated helices, drawn from four different ones) and exposed this classifier error as a side effect.

   Scope note, deliberate: only positions inside an annotated helix are reclassified. UniProt's boundaries also imply the loops between helices within 311–652 are not membrane-embedded, but assigning each loop to the cytoplasmic or lumenal side requires an orientation model the Atlas does not have and should not invent. Those positions keep the coarse Transmembrane label and remain a known approximation.

   Refs: UniProt O76024 TRANSMEM annotations — 11 helices: 314–334, 340–360, 402–422, 427–447, 465–485, 496–516, 529–549, 563–583, 589–609, 632–652, 870–890.

Atlas-wide (data) — continued

2. The inheritance scope never reached the machine-readable card. Every card states which disease family its ClinVar label applies to and under which inheritance mode, and for a recessive variant it says plainly that a single copy does not produce the disease and that the classification says nothing about a carrier's own risk. That statement rendered only on the web page. The markdown card — which is the download, the text of the PDF, and what an AI assistant ingests — carried a bare Classification: Pathogenic with no zygosity anywhere in it. Of 54 recessive-only pathogenic variants, 52 shipped a pathogenicity call to a machine reader with no inheritance scope attached; 2 mentioned it by accident of prose. An assistant reading such a card reports the variant as pathogenic full stop. The scope is now injected into every card at build time and renders above the classification it qualifies, never as a footnote after it. All 2,085 cards carry an Inheritance and scope section. (Raised 2026-09-02 · fixed.)

3. A narrative template asserted that conflicting evidence "confirms pathogenicity." Where AlphaMissense scored a variant below threshold, the Category-4 narrative reached for the ClinVar condition list and concluded that clinical evidence confirms pathogenicity. Every variant carrying that phrasing is a ClinVar "Conflicting classifications of pathogenicity" record — 23 of 23, no exceptions. Conflicting means submitters disagree, which is the opposite of confirmation. On several of these the same card simultaneously reported a common allele frequency and homozygotes in gnomAD, so the prose and the structured fields on one card contradicted each other. The claim is neutralised, not inverted: the structured fields are the defensible ones and the prose no longer overrides them. Example (R818C): "AlphaMissense 0.38 below threshold and multi-phenotype clinical does not resolve it (ClinVar: conflicting submissions)." How it surfaced: the Snow Foundation research assistant was grounded in the Atlas on 2026-09-02 and required to read the card before answering. On its first run it read R818C and reported that the card's own narrative contradicted the card's own structured fields. The Atlas caught this through a tool built on top of it. (Surfaced 2026-09-02 · fixed.) Affected: A243V, A422V, A874T, D118A, E737K, F783L, I802M, K876T, L734H, P607L, R138C, R228H, R703C, R732H, R818C, R868H, V412L, V779M, V871G, and 4 others.

Assistant grounding (not an Atlas data change, recorded for the trail)

4. The research assistant was answering variant questions without reading the Atlas. Asked whether R818C is associated with symptoms in the heterozygous form, it answered "Yes." The Atlas card says Conflicting classifications of pathogenicity, AlphaMissense 0.3822 ambiguous, gnomAD global AF 0.507% with 35 homozygotes — a record that argues against a penetrant heterozygous effect. The assistant never consulted it; it searched the web instead, counted a case report with a co-occurring EYA1 deletion as support, and closed by confirming a framework the questioner already held. The assistant now must read the card first, quote the classification verbatim, carry the phenotype scope into the answer, and is barred from turning conflicting into pathogenic or treating a prediction score as clinical evidence. Verified 2026-09-02: it now answers "No" and leads with the homozygote data. The Atlas was right and the tool speaking for it was wrong. (Raised by Dr. Sarah Gladstone 2026-09-02 · fixed.)

---

v2 — released 2026-08-21

Reviewers: Nufar Kotler (domain expert, Jewish Wolfram Network; affected-family advocate) and Dr. Sarah Gladstone (scientific advisor, The Snow Foundation).

v2 ships the three corrections Nufar raised on 2026-08-11 across all 2,085 variants, and rolls in the two Gladstone policy entries that had been sitting in v-next with card re-rendering pending. v1 accepted these corrections and applied two of them to a single card. v2 pays the rest of that debt.

Reviewer-validated 2026-08-17. Sarah re-tested her P885L query against the shipped intelligence rules: "This worked really well... It also eliminates one of the ClinVar issues that folks run into a lot, that WFS1 is listed as having variants causing many POSSIBLE phenotypes, but this often gets interpreted as being DOCUMENTED as causing all of the different phenotypes. Not so. This works much better and more accurately."

Atlas-wide (policy)

1. Population-specific frequency now shown on every variant. Supersedes v1 entry 2 ("variant frequencies", previously open). v1 corrected R558C's Ashkenazi carrier frequency but left every other variant showing only a global gnomAD allele frequency. That is the same error in miniature: a global number standing in for the at-risk population. All 2,085 variants now carry a per-ancestry allele-frequency breakdown across the ten gnomAD v4 ancestry groups, and each card names its highest-frequency population alongside the global figure. Two guardrails ship with it. An ancestry group below the reporting threshold (fewer than 2 observed alleles, or fewer than 500 sampled) is published but never headlined, because a thinly sampled group can post a spuriously high frequency. And allele frequency is not carrier frequency: the derived carrier column is labelled as derived (Hardy-Weinberg, ~2·AF), and a published carrier frequency always outranks it. The underlying data was re-pulled from the gnomAD v4 GraphQL API. The pull is now a committed, re-runnable script (scripts/pull_gnomad.py) with recorded provenance — the previous gnomad_af.json was a hand-run artifact with no reproducible origin. Every allele count in the new pull reproduces the old file exactly; the additions are the population breakdown and the homozygote counts. (Flagged 2026-08-11 · shipped 2026-08-21.)

2. Homozygote counts added — the field did not previously exist. Nufar noted that the presence of homozygotes in general-population databases belongs in the clinical picture, and cited four for R558C. On checking, there was no homozygote field anywhere in the schema: 0 of 2,085 records carried one. gnomAD v4 homozygote counts are now present on every variant, globally and per ancestry group. Where gnomAD has no entry for a variant, the card states that explicitly rather than showing a silent zero — absent and zero are different facts. Where homozygotes do exist, the card says what that means: gnomAD excludes individuals with severe pediatric disease, so homozygotes in this dataset argue against a fully penetrant severe recessive phenotype. That is presented as interpretive signal, never as a reclassification. 122 of the 2,085 Atlas variants have at least one reported homozygote. (Flagged 2026-08-11 · shipped 2026-08-21.)

3. Pathogenicity is now phenotype-scoped, and stratified by review status. Supersedes v1 entry 1 ("pathogenicity presentation", previously applied to R558C only) and folds in Gladstone v-next entry 1 — they are the same edit. Nufar's point: a WFS1 variant may be pathogenic for diabetes or another WFS1-related phenotype, and that tells you nothing about Wolfram syndrome or expected severity. She is right, and the literature backs it. WFS1 covers two clinically distinct disease families — recessive Wolfram syndrome (biallelic, OMIM 222300) and the dominant WFS1-related disorders (DFNA6/14/38 hearing loss OMIM 600965, Wolfram-like syndrome OMIM 614296, Cataract 41 OMIM 116400, dominant optic neuropathy, WFS1-related diabetes). A single "Pathogenic" collapses both. Urano's 2026 severity-scoring preprint (medRxiv 10.64898/2026.03.24.26349216) excludes the dominant disorders from its framework for exactly this reason. Every card now states what the variant is classified for, under which inheritance mode, on what strength of evidence. Zero variants carry a bare pathogenicity label. Where ClinVar states no condition, the card says so outright rather than leaving the label unqualified. Per Gladstone: submissions are stratified by ClinVar review status. 2★+ assertions lead and are marked documented assertion; 0–1★ single-submitter and conflicting entries are marked unverified submission. 700 of 2,085 variants carry a 2★+ assertion; 1,360 are flagged as unverified submissions, and 25 have no ClinVar review status recorded at all. None of the 2,085 presents a submitted condition as a documented phenotype. 4p16.3 contiguous-gene and CNV submissions (4p partial monosomy, 4p16.3 microdeletion/microduplication, and similar) are separated out as not attributable to a single coding change, rather than listed as WFS1 phenotypes. (Flagged 2026-08-11 and 2026-08-14 · shipped 2026-08-21.)

4. Phenotype cannot be generalized across variants. Gladstone v-next entry 2 — card rendering now shipped. The same general region, even one nucleotide up- or downstream, does not predict presentation; variants differ vastly. Phenotype, severity, and onset claims must be precise to the exact variant. Regional or structural similarity supports mechanistic hypotheses only, clearly labelled as hypothesis. Zygosity must be explicit: classic Wolfram syndrome is recessive, and symptoms are never attributed to single-copy carriers absent a variant-specific documented dominant effect. This now renders as a standing caveat on every card, not only as an intelligence rule. (Flagged 2026-08-14 · intelligence rules shipped 2026-08-14 · card rendering shipped 2026-08-21.)

5. Computed values and interpretation are visually separated. A reader must be able to tell which is which in every sentence. AlphaMissense scores, DynaMut2 ΔΔG, and pLDDT are measurements of the model. Phenotype scope is interpretation of submitted records. Every card now states this distinction explicitly alongside the scoped classification. (Shipped 2026-08-21.)

Per-variant

6. R558C (c.1672C>T, p.Arg558Cys) — homozygotes confirmed; population frequency independently cross-checked. The reviewer's cited figure of four homozygotes is confirmed exactly: gnomAD v4 reports 4 R558C homozygotes, all Ashkenazi Jewish. The Ashkenazi allele frequency is 1.327% (393 of 29,608 alleles) against a global allele frequency of 0.035% — a 38-fold enrichment, and an independent confirmation that the global number was never the relevant one for the at-risk population. A carrier frequency derived from that allele frequency (~1 in 38) brackets the published ~1 in 43 (PMID 30014265). The published figure remains the one cited on the card; the derived number is shown only as a cross-check. The adult homozygotes are consistent with the milder, later-onset course documented in the cohort literature recorded in v1 entry 4. (Entered and shipped 2026-08-21.) Refs: gnomAD v4 (retrieved 2026-08-21); PMID 30014265 (Diabetologia 2018).

7. R558H (c.1673G>A, p.Arg558His) — cross-variant phenotype generalization. The card's structural prose concluded that because R558H lacks R558C's free-thiol oxidative-stress liability, "clinically, the consequence is that R558H may present with somewhat milder disease." That is a phenotype claim generalized from a neighbouring variant at the same residue — exactly what entry 4 prohibits. Structural similarity supports a mechanistic hypothesis, never a phenotype claim, and there is no published R558H cohort data on the card to support a severity statement. The mechanistic comparison is retained; the clinical inference is removed and replaced with an explicit statement that presentation does not follow from the structural comparison. Found by an Atlas-wide prose scan while shipping v2 — the only instance of cross-variant phenotype generalization in the 2,085-variant corpus. (Entered and shipped 2026-08-21.)

Still open

• Wet-lab variant scoring as validation reference. Published wet-lab variant-scoring data will be incorporated so computational predictions can be checked against experimental scores. Article citation still pending from Nufar — this entry stays open until it arrives. Related: double-blind wet-lab validation of Atlas predictions planned at WashU (Dr. Urano) and Belgium (Dr. Loncke). (Flagged 2026-08-11 · open.)

• P885L card prose. The Gladstone per-variant correction (v1 log, v-next entry 3) is live as a card overlay and in the intelligence, and P885L now carries the Atlas-wide phenotype scoping shipped above. The specific prose claim that its ClinVar list represents "the broadest clinical spectrum documented for any single position" is superseded by the scoped rendering but has not been rewritten in the source card narrative. (Open — prose re-render.)

---

v1 — released 2026-08-11

Reviewer: Nufar Kotler — domain expert, Jewish Wolfram Network (~38-person international community group); cybersecurity professional; affected-family advocate. Feedback delivered on the 2026-08-11 RareResearch.AI × Jewish Wolfram Network call. Her written comments and the wet-lab variant-scoring article she is sending will be added verbatim as entries when received.

Atlas-wide (policy)

1. Pathogenicity must be presented as a range/spectrum, not a single label. Single labels ("Pathogenic/Likely pathogenic") flatten real clinical variability — the same homozygous variant can produce widely different ages of onset and severity. Where ClinVar submissions conflict or published cohort data shows a spectrum, cards must show the range and the evidence for each end of it. (Flagged 2026-08-11 · applied to R558C in v1 — see entry 4 · superseded by v2 entry 3, shipped Atlas-wide 2026-08-21.)

2. Variant frequency corrections. Nufar flagged specific variant-frequency errors in Atlas content. (Flagged 2026-08-11 · R558C corrected in v1 — see entry 4 · superseded by v2 entry 1, shipped Atlas-wide 2026-08-21.)

3. Wet-lab variant scoring as validation reference. Published wet-lab variant-scoring data exists and will be incorporated as a validation reference so computational predictions can be checked against experimental scores. Article citation pending from Nufar. (Flagged 2026-08-11 · open. Related: double-blind wet-lab validation of Atlas predictions planned at WashU (Dr. Urano) and Belgium (Dr. Loncke).)

Per-variant

4. R558C (c.1672C>T, p.Arg558Cys) — frequency error on the card, and pathogenicity context. (a) Frequency. The card stated R558C is carried by "approximately 1 in 3" Ashkenazi Jewish individuals, with a claimed supporting range of "27–34%." This is wrong by an order of magnitude and has no published basis. The verified published carrier frequency is ≈2.32% (~1 in 43) in Ashkenazi Jews (1.32% Ashkenazi/Sephardi mixed; 0.04% Sephardi; gnomAD global AF 0.035% is not the relevant number for the at-risk population). R558C is still the Ashkenazi Jewish founder variant and the largest identifiable carrier group — the corrected number carries that argument on its own. (b) Pathogenicity. The card's single "Pathogenic/Likely pathogenic" label and AlphaMissense 0.849 (LPath) are correct as annotations but incomplete as clinical interpretation: published cohort data shows R558C homozygotes present a substantially milder, later-onset form (mean diabetes onset ~17–19 yrs, optic atrophy ~29 yrs, Wolfram diagnosis ~30 yrs; R558C protein more stable than other recessive pathogenic variants). Verified published data outranks the single predictive label. (Entered 2026-08-11 · applied and released in v1, 2026-08-11: card text corrected in data/variants/R558C.json, the source card Molecular Atlas/R558C/R558C_variant_card.md, the public markdown copy, and the printable HTML/PDF card; an "Expert Review & Corrections" section now appears on the card itself.) Refs: PMID 30014265 (Diabetologia 2018); PMID 36134655 (JCI Insight 2022); Neurol Genet 2021, doi:10.1212/NXG.0000000000000578.

Related fixes outside the Atlas

• Wolfram Research Intelligence system prompt claimed R558C carrier frequency of "1 in 3" in the Ashkenazi Jewish population — off by more than an order of magnitude. Corrected to ≈2.3% (~1 in 43) on 2026-08-11, same refs as entry 4.

---

Log created 2026-08-11 by Mike Wallace (RareResearch.AI) in response to the first external expert review of the Atlas. v1 released 2026-08-11. v2 released 2026-08-21, shipping the Kotler corrections across all 2,085 variants and the Gladstone card rendering.

The superseded figures stay visible in this log on purpose. A corrections log that hides what it got wrong is not a corrections log.