---

Version of record: 10.5281/zenodo.21799856, published 5 August 2026. That identifier is the citable address for this paper and it resolves at https://doi.org/10.5281/zenodo.21799856. It is a version identifier; Zenodo minted a second one that resolves to all versions, and the version identifier is the one to cite.

What this file is. This project's authoritative copy of the manuscript, SUBMIT_THESE/papers/PUBLISH_3_CLINVAR_FUNCTIONAL_EVIDENCE.md, which is the file the deposited PDF was built from. Synced 7 August 2026 by scripts/sync-manuscripts.mjs, which copies the source byte for byte and prepends this note. Nothing in the manuscript below has been rewritten for the website.

What was here before, because nothing on this site is deleted quietly. Until 6 August 2026 this file was a copy taken before the corrections of 4 August 2026 evening and never resynced, so it served the retired 34.1 percent rescaling of the measured current. Earlier on 6 August 2026 it was given a banner reading "superseded revision, do not cite any figure in it". That banner was an accurate description of a stale file and a poor thing to serve on a research site, so the stale file has been replaced with the authoritative text rather than annotated. The retired rescaling divided the measured 68.3 percent by two; O'Neill 2022's own two-allele control reads 218.4 percent of a single allele rather than 200, so the divisor is 2.184, the baseline is 31.3 percent and the comparator for simple loss of one allele is 45.8 percent.

This text matches the version of record, on the evidence available here. The authoritative markdown was last written on 4 August 2026, before the submission PDFs were built that evening and before the deposit the following day, so nothing in it postdates the identifier above.

The limit of that claim, stated rather than hidden. It is an argument from timestamps, and for this paper it is still only that. (This paragraph used to say that no published PDF had ever been downloaded and no published text read back. That was true when it was written and stopped being true on 6 August 2026, so it is corrected here rather than left standing.) The published files of the divergent papers were downloaded from the Zenodo API on 6 August 2026 and diffed against the local text, which is what turned that divergence from an inference into a measurement. This paper was not among them. It is one of the copies believed unchanged on the strength of local modification times, which is the same class of inference the download exercise was run to replace. SUBMIT_THESE/ZENODO_DIVERGENCE_20260806.md records what was checked, and SUBMIT_THESE/V2_STAGING/evidence/DIVERGENCE_VERIFIED_20260806.md records what was not.

A separate limit, and it is not about this text. Six of the ten papers name deposited tables that the archive does not hold, in whole or in part. Paper 10's case was the worst and has been dealt with; the others have not. Do not assume this paper's data availability statement is true of the archive without checking it. The audit is SUBMIT_THESE/PAPER_10_DATA_STATEMENT_FIX.md.

None of this is peer reviewed, and none of it has been through a wet lab. No cell has been edited and no current has been recorded for this variant by this project. Every therapeutic statement in the manuscript below is a prediction.

---

Deposited functional evidence in ClinVar is searchable only by exact term, and no surveyed tool reads it

Ethan Bradley

Independent researcher, no institutional affiliation

ORCID: 0009-0008-8925-7975

Correspondence: the author

Intended venue: Database (Oxford). Preprint first: bioRxiv.

Abstract

ClinVar accepts laboratory functional-evidence submissions in a structured FunctionalData element, and the reason downstream users do not see it is not the access route: no surveyed tool parses it, and six major pipelines already download the release containing it and never read it. In the release of 27 June 2026 (4,531,457 VariationArchive records, MD5 verified against NCBI's checksum), 270,827 records, 5.98 percent, carry functional evidence in 532,132 blocks. One depositor owns every block in 97.17 percent of those; the defensible pool is the 7,661 records carrying a block from another laboratory, and every aggregate is computed on it. Three findings concern retrieval. FunctionalData holds two child elements and only FunctionalConsequence is indexed by the FCNS filter, taking 185 values against 4 and predicting membership with 100.000 percent agreement against 8.611, so a census on the wrong element misstates retrievability. All 185 terms are individually retrievable but the filter supports no wildcard: six constructions meant to return any variant with any functional consequence return zero, so the field is searchable by exact term only, never enumerable. And the one call returning the element fails silently when a flag is omitted, giving a 110-byte empty set under HTTP 200, identical to a nonexistent record's. Of twelve variant-interpretation resources, eleven have a determinable ClinVar route and none parses FunctionalData; for the six already ingesting the release the field is on disk and unread, so the barrier is parser coverage rather than access.

Keywords: ClinVar, functional evidence, variant interpretation, controlled vocabulary, data retrieval, PS3, BS3

A key to the terms used here

• ClinVar is the public NCBI database where laboratories deposit genetic variants and their interpretations. A variant is a single spelling change in a gene.
• Functional evidence means an actual laboratory measurement of what a variant does to the protein, as opposed to a computational prediction. PS3 and BS3 are the standard codes for functional evidence supporting a pathogenic or a benign interpretation.
• FunctionalData is the structured field in ClinVar's records that holds that evidence. FunctionalEffect and FunctionalConsequence are the two child fields inside it, and which of the two the search index reads is the central finding here.
• XML is the structured text format of the full ClinVar release. A VCF is the compact variant file format most pipelines read instead, and a tab-delimited export is a simple table. The field of interest exists in the first and not in the other two.
• E-utilities are NCBI's public programming interfaces. esearch finds records, efetch retrieves them, esummary returns a short summary, and einfo describes what is indexed. rettype and is_variationid are parameters on those calls.
• A search filter such as FCNS restricts a query to records carrying a particular kind of content. A wildcard is a search term matching anything, and the absence of one is what makes this field searchable but not enumerable: you can ask whether a specific term is present, but you cannot ask for everything.
• Enumerable means you can list every member of a set. Searchable by exact term means you can only confirm the ones you already know to ask about, like a shelf you can check for a named product but cannot inventory.
• An MD5 checksum is a fingerprint of a file used to confirm a download is complete and unaltered.
• HTTP 200 is the code a web server returns for success. A response that returns 200 while containing nothing is a silent failure, the failure mode this paper is mostly about, and the dangerous kind: a loud error gets fixed, a quiet one gets published.
• A VariationID and a VCV accession are the two identifiers ClinVar uses for a record.

1. Introduction

Clinical variant interpretation is limited less by the availability of evidence than by its reachability. Under the ACMG/AMP framework (Richards et al. 2015, PMID 25741868), a well-established functional assay contributes evidence through the PS3 criterion when it shows a damaging effect and BS3 when it shows a neutral one, with the strength of that contribution set by how well the assay has been calibrated (Brnich et al. 2019, PMID 31892348). Laboratories increasingly generate such data at scale, and ClinVar accepts it in a structured form: a FunctionalData element nested within the ObservedIn block of a submitted assertion, carrying a functional effect, a functional consequence, a numeric result and free-text commentary.

The element is not, however, present in every route by which ClinVar is read. It does not appear in esummary responses, whose published document type definition declares 74 elements, none containing the string "functional". It does not appear in the tab-delimited submission summary or in the VCF distributions. NCBI documents this: its own guidance to the FTP files and E-utilities states that submissions including functional data are represented only in the ClinVar XML files under the ObservedIn element, and the schema confirms that FunctionalData is declared within the ObservationSet complex type of which ObservedIn is an instance. The restriction is a published contract, not a hidden behaviour, and this paper does not claim otherwise.

What has not been established is the magnitude. Three questions follow, and none has been answered quantitatively in the literature. How much functional evidence is deposited. How completely it can be retrieved by the search interface ClinVar provides for it. Whether the tools that consume ClinVar read it. This paper answers all three by a complete parse of one release rather than by sampling, and in the course of doing so corrects a claim about the second that I had previously made incorrectly myself.

2. Methods

Every step below is reproducible from public data with no credentials, no institutional access and no licensed software.

2.1 Data

The complete ClinVar variation release was downloaded from the NCBI FTP site at ftp.ncbi.nlm.nih.gov/pub/clinvar/xml/ClinVarVCVRelease_00-latest.xml.gz on 4 August 2026.

Reporting the checksum alongside the release date matters because "the latest release" names a moving target. The checksum names one file.

Two dates appear above and they are not in conflict. ClinVar publishes a new release every week to the same _00-latest path, so the file downloaded on 4 August 2026 is the release dated 27 June 2026, and the checksum is what ties the two together. Every count in this paper is a property of that one file. A reader who repeats this analysis in a later week will get different numbers, and the difference is new deposits rather than a disagreement with what is reported here. Anyone reproducing these figures exactly should confirm the MD5 above before comparing, or retrieve the dated copy from the weekly archive directory; anyone reproducing the method rather than the figures should expect the totals to have grown and the structure of the findings to hold.

2.2 Parsing

The gzipped release was parsed as a stream, without decompressing to disk, using incremental XML parsing (lxml.etree.iterparse) with the element boundary set to VariationArchive. Fields were located by element path rather than by pattern matching over text, and each record was cleared from memory after processing together with its preceding siblings, which holds peak memory below one gigabyte on a file of this size. A complete pass takes 11.6 minutes on a ten-core desktop machine, so the analysis is not compute-limited and requires no cluster.

For each record containing at least one FunctionalData block I extracted the variation identifier and accession, all gene symbols, the aggregate germline classification description, the number of submitters, every FunctionalEffect and FunctionalConsequence value with its Sequence Ontology identifier, numeric results, and ACMG evidence-code strings.

A second pass attributes each FunctionalData block to the depositor that submitted it, by walking each ClinicalAssertion element, reading the SubmitterName attribute of its ClinVarAccession child, and assigning the blocks nested within that assertion to that submitter. This distinction is load-bearing and is discussed in section 3.2.

ACMG evidence codes were identified by regular expression over the text of each functional block, matching the code families and their graded modifiers, case-insensitively. Because depositors write these codes inconsistently, the resulting count is a floor rather than a measurement.

2.3 Retrieval characterisation

Retrieval was tested against the live E-utilities interface rather than inferred from the file. The set of searchable ClinVar fields was enumerated with einfo, which reports 47 fields including FCNS ("Functional consequence") and SBM ("Submitter"). Counts and identifier lists were obtained with esearch, and individual records were inspected with efetch using rettype=vcv. The efetch flag behaviour reported in section 3.4a was tested on 4 August 2026 by issuing the same request with and without is_variationid and recording the response length, the HTTP status and the response body, across three VariationIDs plus one identifier chosen to be absent from the database.

Two design choices matter for the validity of what follows. First, every test input was drawn from the computed parse output rather than chosen by hand, so the terms queried are the terms that actually occur in the data. Second, the vocabulary test was run over the complete term list, all 185 distinct FunctionalConsequence values, not a sample, with one esearch per term.

2.4 Tool survey

Twelve variant-interpretation resources were examined by reading their public source: MyVariant.info, cBioPortal / Genome Nexus, clinvar-tsv, Nirvana, ClinVar Miner, the Ensembl ClinVar importer, the VEP plugin set, the Ensembl REST API, Biopython's Entrez module, InterVar, Bioconductor, and OakVar / OpenCRAVAT. For each, the question was whether any code path parses FunctionalData.

The OakVar / OpenCRAVAT annotator modules are distributed through a module store rather than the main source repository. The store endpoint and request form were read from the client source (oakvar/lib/store/db.py, with the default store URL in oakvar/lib/assets/system.yml) and replicated exactly; the store requires authentication and returned HTTP 401 without it. The modules were therefore read from the public module repository instead, where all three ClinVar annotator modules are present.

2.5 Availability

All parser code, the per-variant census, the per-gene summary, the per-term retrieval verification and the tool survey are released as supplementary files (section 8).

3. Results

3.1 How much functional evidence is deposited

Of 4,531,457 VariationArchive records in the release, 270,827 (5.977 percent) contain at least one FunctionalData block, and those records contain 532,132 blocks in total, a mean of 1.96 blocks per record. Counting variants and counting deposited measurements are therefore different questions, and the larger number is the one relevant to how much laboratory work is represented.

3.2 One depositor dominates, and the exclusion rule must be stated precisely

A single depositor, the Dr. Peter K. Rogan Lab at Western University, is associated with the large majority of these records. Any aggregate computed over the full 270,827 describes that submission rather than the field, so a bulk-excluded pool is required. How the exclusion is applied changes the answer materially, because a variant can be submitted by several laboratories at once:

Excluding a variant because the bulk depositor appears anywhere in its submitter list discards 24,594 records in which an independent laboratory also deposited work. But a submitter list is the wrong instrument regardless of how it is thresholded, because it records who touched the record rather than who performed the measurement. A second pass therefore attributes each of the 532,132 functional blocks to the depositor whose submission contains it, by walking each ClinicalAssertion and reading the SubmitterName of its ClinVarAccession. Every block resolved to a depositor: the count of unattributed blocks is zero, so no evidence is lost to this rule.

The denominator used throughout this paper is 7,661: records carrying at least one FunctionalData block deposited by a laboratory other than the single dominant depositor. It is larger than the 7,231 that a submitter-list exclusion yields, because 430 records contain independent bench work alongside a bulk submission and a list-based filter discards all of them. The rule is applied uniformly to every figure below, and readers recomputing these numbers should apply the same one.

One defect in an earlier build of the supplementary census file is worth recording here, because it is a clean example of the failure mode this section is about and because a reader comparing an earlier copy of that file against this table would otherwise find 911 rows unaccounted for. That build wrote only the first four submitter names per record, alphabetically, and then tested for the bulk depositor in that truncated list rather than in the complete one. On records with five or more submitting laboratories where the bulk depositor's name sorted fifth or later, the name was not in the list being searched, and the record was admitted. Exactly 911 records entered the file that way, all of them with five or more submitters, in every case with the bulk depositor at alphabetical rank 5 or worse, and in 892 of the 911 every functional block belonged to the bulk depositor. A separate clause in the same build readmitted 15 records whose first listed gene was SCN5A, the gene in which I carry a variant, which is precisely where a special case should never be. The file distributed with this paper is rebuilt on the attribution rule stated above, carries the complete submitter list in every row, and contains 7,661 rows. No figure in this paper drew on the defective file, and the counts in the table above are unchanged: the exclusion table was computed from the release, and a re-parse of the same checksummed file reproduces all five of its rows. The decomposition 7,231 + 911 + 15 = 8,157 accounts for that earlier file's row count without remainder. The lesson is the one in section 4: the filter ran correctly against the wrong list, and produced a number that passed every internal consistency check I applied to it, including two that summed exactly.

3.3 The retrieval filter indexes one of two adjacent elements

ClinVar provides a search filter for this data, FCNS. Its behaviour is determined by which of two sibling elements it indexes, and the two are easy to conflate because both sit directly inside FunctionalData and both carry Sequence Ontology terms.

The two-term query commonly used to find functional evidence, "functionally abnormal"[FCNS] OR "functionally normal"[FCNS], returns 4,489 variants (1,313 and 3,176 respectively). Testing which element predicts membership in that returned set, across all 270,827 records rather than a sample, that being the full pool including the single bulk depositor:

The FunctionalConsequence contingency table has empty off-diagonal cells in both directions: 4,488 records carry an indexed term and every one is retrieved; 266,339 carry none and none is retrieved. FCNS indexes FunctionalConsequence.

This corrects a claim I published in an earlier draft of this work. That draft reported, on the basis of FunctionalEffect, that FCNS "indexes only two exact Sequence Ontology strings" and therefore missed 45.1 percent of the census, and offered SCN5A VariationID 67951, which carries PS3_strong, as an example of a variant FCNS cannot reach. Both statements are wrong. 67951[ALLUID] AND ("functionally abnormal"[FCNS] OR "functionally normal"[FCNS]) returns 0, which is what that draft observed, but 67951[ALLUID] AND "loss of function"[FCNS] returns 1. The variant was retrievable by the term its depositor actually filed. The 45.1 percent figure is an artefact of measuring the wrong element and should not be cited.

3.4 Every term is retrievable; the vocabulary is not enumerable

Having identified the indexed element, I tested retrievability term by term over the complete vocabulary. For each of the 185 distinct FunctionalConsequence values, "<term>"[FCNS] was queried and compared against my own count.

All 185 of 185 terms return a non-zero result. None is unreachable. Exact agreement between my counts and the live counts is 167 of 185 when both are computed per variant, rising to 179 of 185 once orthographic variants are collapsed.

Two points about that comparison are methodologically relevant. First, esearch counts variants while a naive parse counts blocks, and a variant may carry several blocks bearing the same term; comparing the two produces apparent discrepancies that are purely denominator artefacts. For retained intron the per-block figure is 481,477 against a live count of 244,513, while the per-variant figure is 244,513, an exact match. Second, NCBI's index collapses orthographic variants of the same term: seven terms appear in both space-separated and underscored form, and each form returns the group total, so both spellings of "functionally normal" return 3,176 against separate per-variant counts of 147 and 3,029. Normalisation of this kind makes the filter more capable than a literal reading of the term list suggests, not less.

The limitation lies elsewhere, and it survives the correction. There is no way to ask FCNS for any functional consequence. Six constructions intended to do so all return zero: a bare field name, *[FCNS], a*[FCNS], "*"[FCNS] (rejected as a syntax error), FCNS:FCNS, and a numeric range. The filter answers "which variants carry consequence X" for any X, and never "which variants carry any consequence".

The consequence is a circularity. The vocabulary is open and partly free text, containing entries such as "Severe decrease in peak current due to reduced trafficking" alongside Sequence Ontology identifiers, and 27 terms occur exactly once. To enumerate the field through FCNS, a user must already possess the term list; the only way to obtain the term list is to parse the complete release. Enumeration and exact-term lookup are different capabilities, and only the second is available.

This also settles how the 4,489 figure should be read. It is the size of one two-term slice, not the reach of the filter. The bulk submission, for instance, is thoroughly retrievable through FCNS: Rogan[SBM] AND "retained_intron"[FCNS] returns 244,475, a figure that my independent parse reproduces exactly, while only 13 of those records fall in the two-term union. Setting 4,489 against 270,827, the full pool including the single bulk depositor, as though the difference were unreachable is not a valid comparison.

3.4a One omitted flag turns the working call into a silent empty answer

The single retrieval route that does return the element fails silently when one flag is left out. The failure is worth stating precisely because it does not look like a failure.

Both commands below address the same record, SCN5A p.Asp82Asn, VariationID 4681248, VCV004681248, whose deposited block carries the Sequence Ontology term loss of function, a numeric result of -6.5 and a depositor assessment of PS3_strong from an automated patch clamp assay (PMID 38953211). Both were run on 4 August 2026.

efetch.fcgi?db=clinvar&id=4681248&rettype=vcv&retmode=xml
  110 bytes. The body, with its newlines and whitespace padding preserved exactly:
    <?xml version="1.0" encoding="UTF-8" ?>
    <24 spaces>
    <ClinVarResult-Set><set/></ClinVarResult-Set>
  Stripped of the declaration line, the newlines and the padding, the payload is the
  45-character string <ClinVarResult-Set><set/></ClinVarResult-Set> preceded by the
  39-character declaration. The 110-byte figure is the response as delivered, which is what a caller
  measures.

efetch.fcgi?db=clinvar&id=4681248&rettype=vcv&is_variationid&retmode=xml
  21,614 bytes, including the FunctionalData block in full

The first response arrives under HTTP 200 with no error element, no warning and no message of any kind. It is an empty result set, and it is indistinguishable from a true absence.

Three controls establish that this is a general property of the interface and not a quirk of one record. First, the behaviour is identifier-independent: the flagless form returned the same 110-byte empty set for VariationIDs 67951 and 12345 as well, while the flagged form returned 22,177 and 16,777 bytes for the same two. Second, the identifier is not the problem. esearch for this variant returns exactly this identifier, 4681248, as its only hit, and esummary resolves that same bare identifier to the correct record title without any flag, so the value a caller naturally carries forward from a search is precisely the value that yields nothing here. Third, and most consequential for anyone writing code against this endpoint, the empty response for a valid VariationID without the flag is byte-identical to the empty response for the nonexistent VariationID 999999999 with the flag. The interface returns the same 110 bytes for a missing flag and for a missing variant, so no caller can distinguish a query error from a data absence by inspecting the response.

Omitting rettype entirely does not fail this way. It returns a 194-byte IdList echoing the identifier back, which is visibly not a record. The dangerous case is the one that looks closest to correct: the right endpoint, the right rettype, the right identifier, and one missing flag.

3.5 Classification status of the non-bulk pool

Within the 7,661 records carrying at least one independently deposited functional block, spanning 622 genes, grouping the aggregate germline classification into coarse buckets:

Two of the 7,661 records carry no gene symbol at all, both intergenic variants on chromosome Y (VariationIDs 3024128 and 3024132, deposited by Human Developmental Genetics, Institut Pasteur), so the per-gene supplementary table has 623 rows: 622 named genes plus one (no gene) bucket holding those two. An earlier draft reported 623 genes by counting that bucket as a gene. The gene count is 622. A record spanning several genes is counted under each of them, which is why the per-gene rows sum to 7,852 rather than to 7,661.

5,359 records, 70.0 percent, carry deposited laboratory functional evidence and still have no confident clinical classification. A further 1,892 records, 24.7 percent of the pool, carry a formal ACMG evidence code in the text of an independently deposited block, most commonly PS3 (1,110), BS3 (1,106), PS3_strong (378) and PS3_moderate (233). In those cases the depositing laboratory has already stated the strength of its own evidence, so this is not raw data awaiting interpretation.

The genes where the pattern is largest are those where a single laboratory has characterised many variants at once:

KCNE1 and JAG1 are the clearest cases: 81 and 87 percent of their variants with functional data have no germline classification, and essentially every one carries an ACMG evidence code.

An important qualification belongs here rather than in the discussion. A ClinVar germline classification requires that a variant has been observed in a person and submitted with clinical context, and functional evidence alone does not and should not trigger one. Many unclassified records are unclassified because no carrier has yet been reported, not because anyone overlooked the data. That reading strengthens rather than weakens the observation: it means functional evidence for a large number of variants has been banked in advance of the first patient, in a field that the ordinary programmatic route does not return.

3.6 No surveyed tool reads the field

Of twelve resources examined, eleven have a determinable ClinVar route and none of the eleven parses FunctionalData. The twelfth, Bioconductor, has no dedicated ClinVar client package and is recorded as not applicable rather than as a negative. Six of the eleven ingest the full XML release in which the element is present, so the omission is not explained by any of them fetching a compact summary.

The final row was undetermined in earlier work and is resolved here. The three ClinVar annotator modules in the public OpenCRAVAT module repository (annotators/clinvar, clinvar_acmg, clinvargene) were read in full, eleven source and configuration files, and the case-insensitive string "functional" occurs zero times in all of them. The annotator issues a single query selecting 38 named columns from a pre-built SQLite table keyed on position, reference and alternate allele; none of the 38 carries functional evidence, so the field cannot reach output regardless of what ClinVar contains.

Nirvana is the clearest illustration of the mechanism. Its parser defines and iterates the ObservedIn element, the very element within which FunctionalData is declared, and extracts only sample and origin values from it. The string "Functional" does not occur in the file. This is a parser standing directly on the data and stepping past it.

4. Discussion

The results compose into a single observation. Functional evidence is deposited in ClinVar at appreciable scale by hundreds of laboratories. It is fully searchable, but only by a term the searcher must already know, and it cannot be enumerated through the search interface at all. And the resources through which most users actually encounter ClinVar do not read it, including several that already hold the data on disk.

The flag behaviour in section 3.4a deserves separate weight, because it is the only one of these findings that actively misleads rather than merely omitting. The other routes fail visibly: a tab-delimited export has no column for the field, and a parser that never mentions the element plainly never reads it. A caller who omits is_variationid gets a well-formed empty answer to a well-formed question, under a success status, for a variant that does in fact carry the evidence. The natural interpretation of that response is that the variant has no functional data. Because the same 110 bytes come back for an identifier that does not exist at all, there is no way to tell the two apart without already knowing the answer.

This has a bearing on the paper's own reliability, and it is worth saying plainly rather than quietly. Four times in this work a wrong parameter or a missing flag produced a zero that looked like a finding about the world. The FCNS element mix-up in section 3.3 is the one that nearly reached print, as a claim that no search filter existed. Unversioned GTEx gene identifiers return HTTP 200 with zero rows. Hand-constructed variant identifiers returned records that did not carry the field. And this flag returns an empty set. Every one of those four was a query defect presenting as an absence of data. That recurrence is not incidental to the argument; it is the argument, restated from the caller's side. An interface that answers a malformed request with a plausible empty result will produce false negatives at a rate nobody measures, because a null result invites no investigation. The practical defence is the one used in section 3.3: predict the query's result from an independent parse of the underlying data, and treat any disagreement as a fault in the query until shown otherwise.

None of these is a defect in isolation, which is why the combination is worth reporting. Restricting a nested element to the XML distributions is a defensible design choice, and NCBI documents it. Indexing a controlled vocabulary term by term is normal information retrieval. Choosing not to parse an optional element is a reasonable decision for any individual tool. The result of all three together is that a laboratory can deposit a well-controlled measurement, with its own ACMG strength assessment attached, into a public database, and a downstream user following ordinary practice will not see it and has no way to discover that it exists.

Three remedies follow directly from the measurements and none is large. A match-any-value form for FCNS, of the kind other Entrez fields support, would convert exact-term lookup into enumeration and remove the circularity in section 3.4. Adding one element to a parser that already walks ObservedIn is a small change for the six pipelines that already ingest the full release. And an efetch request that supplies an identifier the endpoint cannot interpret in the requested mode should return a warning rather than an empty set, so that a caller can tell a malformed query from an absent record.

The correction reported in section 3.3 has methodological weight beyond this paper. Two sibling elements inside the same container, both carrying Sequence Ontology terms, differ by a factor of 46 in vocabulary size, and only one is indexed. A census that reads the wrong one produces internally consistent numbers and a coverage statistic that is entirely an artefact. My earlier draft made exactly that error and reported a 45.1 percent miss rate that does not exist. The check that caught it was cheap: predicting live query membership from the parse and requiring the contingency table to be clean. Any claim that a database field is unreachable should be tested that way before publication, because a failed query is evidence about the query and not about the world.

5. Limitations

1. Detection is by element presence, not by curation. A record counts if it contains a FunctionalData block. I did not assess whether each block reports a meaningful measurement, and some may be sparse or uninformative.
2. The ACMG code count is a floor. Codes were found by pattern matching over deposit text. Capitalisation is inconsistent in the data and unusual formats will be missed.
3. One release, one snapshot. All counts describe the release of 27 June 2026, identified by checksum. ClinVar publishes weekly and these numbers drift. One concrete illustration arose during this work: VariationID 201627 is retrievable through FCNS, and the assertion carrying its functional data (accession SCV007615693, Avery Lab, Oakland University) has a creation date of 19 July 2026, three weeks after the release date, so the element is absent from the file entirely. Its submission date of 11 June 2026 precedes the release, so the lag is in processing rather than submission. Live queries and file parses disagree by exactly this margin, which is why the file checksum rather than the phrase "latest release" identifies the data here.
4. The vocabulary list ages. 185 terms were found in this release. Depositors may add free-text terms at any time, so any enumeration built on the list is a snapshot.
5. Classification buckets are coarse. Six groupings were imposed on more varied underlying strings and some assignments are judgement calls.
6. The tool survey is a source read, not an execution. I did not run any of the twelve pipelines end to end and did not audit full dependency trees. A tool may have added support after the commit I read. The OakVar module store itself was authentication-gated and was not reached; that finding rests on the public module repository.
7. This measures a database, not clinical practice. It shows that the field is absent from widely used routes. It does not measure how many interpretation decisions were affected, which would require knowing what route each decision used.
8. The documented-versus-hidden distinction matters and is settled against the stronger claim. NCBI documents the XML-only restriction in prose and in the esummary DTD. This paper claims the omission is consequential, not that it is concealed. The same holds for the is_variationid flag in section 3.4a: the flag is documented. What is reported here is the shape of the failure when it is omitted, not a claim that its existence is undisclosed.
9. The flag test covers four identifiers, not the identifier space. Section 3.4a rests on three real VariationIDs and one absent identifier, tested on one day. It establishes that the empty response is not specific to one record and that it collides with the not-found response. It does not establish behaviour for every identifier class the endpoint accepts, and I did not test accession-formatted inputs such as VCV or RCV strings.
10. An earlier build of the supplementary census file was wrong, and internal consistency did not catch it. The defect is described in section 3.2 and accounted for exactly in CENSUS_DISCREPANCY_RESOLVED.md. It matters as a limitation and not only as a fixed bug, because the defective file passed the checks I had: its row count summed, its columns were populated, and the exclusion table in section 3.2 was arithmetically self-consistent in both directions. The error surfaced only when the file's row count was compared against the table it was supposed to support, which is a check I had not run. Any count in this paper that I have not reconciled against an independent recomputation of the same quantity carries the same exposure. The counts in section 3.2 have now been reconciled that way; the retrieval tests, the vocabulary counts and the tool survey rest on separate evidence and have not been recomputed by a second parser.
11. The bulk depositor is identified by submitter-name string. If that laboratory deposits under an additional institutional name I did not recognise, records under that name are counted as independent and inflate the 7,661. Matching the full name and matching the substring "Rogan" both return 263,596, which closes variant spellings of this name but not a different name entirely.

What would falsify the central claims

The element-identification claim would be falsified by a single variant whose FunctionalConsequence carries neither indexed term but which the two-term FCNS union returns, or by one carrying an indexed term that the union does not return. Across 270,827 records, the full pool including the single bulk depositor, the contingency table contains zero of either. The enumeration claim would be falsified by any working match-any-value FCNS construction; six failed. The tool claim would be falsified by any of the twelve reading FunctionalData in a path I did not open. The silent-failure claim in section 3.4a would be falsified by a flagless rettype=vcv request that returns a populated record, or by any warning or error element appearing in that 110-byte response, or by the not-found response differing from it in any byte. The exclusion table and the 7,661 denominator would be falsified by a re-parse of the file with the stated MD5 returning any different value for any of its rows, or by a single FunctionalData block that resolves to no depositor, since the attribution rule assumes every block sits inside a ClinicalAssertion and zero were unattributed across all 532,132. Each of those is a one-command test and the exact commands are printed above.

6. Conclusion

270,827 ClinVar records carry deposited laboratory functional evidence in a structured element that is fully searchable by exact term, not enumerable through the search interface, and read by none of the eleven surveyed variant-interpretation resources whose route could be determined. That count must not be quoted on its own: one depositor owns every functional block in 263,166 of those records, 97.17 percent, so 270,827 measures one bulk submission far more than it measures the field. In the 7,661 records, 2.83 percent of the total, that carry at least one block deposited by another laboratory, spanning 622 genes, 70.0 percent still lack a confident clinical classification and 24.7 percent carry a formal ACMG evidence code assigned by the depositing laboratory. The gap is not one of data availability but of retrieval and parser coverage, and closing it requires a match-any-value search form and one additional element in parsers that already read the enclosing block. The single call that does return the element also returns an empty set, silently and under a success status, when one documented flag is omitted, and that empty set is byte-identical to the one returned for a variant that does not exist. A caller who omits the flag therefore records an absence of evidence that is really an absence of a correctly formed request.

7. Data availability

All primary data are public and require no credentials. The release analysed is ClinVarVCVRelease_00-latest.xml.gz, release date 2026-06-27, MD5 2e7e76ebbf668910b8688cc5e4284c1b, from ftp.ncbi.nlm.nih.gov/pub/clinvar/xml/. Because that path always serves the current release, readers reproducing this work should verify the checksum, or retrieve the dated equivalent from the weekly archive directory. Retrieval tests used the public NCBI E-utilities. Tool source was read from public repositories. Venue characteristics were taken from the DOAJ and Crossref APIs.

All derived tables are deposited as a single archive with a permanent identifier. The identifier is recorded in DATA_DOI.txt alongside this manuscript and should be cited as the data source. No ethical approval was required: no human subjects were involved and no individual-level data were accessed.

Corrected, 7 August 2026. The sentence removed here read "Parser code accompanies this manuscript (section 8)." It was false in both places it could have been true. The Zenodo record for this paper, 10.5281/zenodo.21799857, carries one file, the PDF — verified against the Zenodo API on 7 August 2026 — so nothing accompanies the manuscript. And the data archive did not hold the scripts either.

Of the twelve files section 8 lists, eight are deposited and four were not. The eight present are CENSUS_FUNCTIONAL_CENSUS_V2.csv, CENSUS_BY_GENE_V2.csv, CENSUS_FCNS_TERM_VERIFICATION.csv, CENSUS_TOOL_ACCESS_ROUTES_V2.csv, CENSUS_VENUE_VERIFICATION.csv, CLINVAR_FUNCTIONAL_CENSUS.csv, reconcile_gap_rows.csv and FIG1_fcns_element_correction.png.

One of the four was recoverable and is now deposited: CENSUS_DISCREPANCY_RESOLVED.md, the full account of the 911-row defect, added to version 2 of the data archive (10.5281/zenodo.21840036) on 7 August 2026.

Three are not deposited and are not recoverable: census2_parse.py, census2_pass2.py and reconcile_census.py. All three were searched for on 7 August 2026 across the entire project tree and the compute host, and no copy of any of them exists. Section 8 still lists them because they are what produced the census and a reader should know what is missing rather than what was convenient to claim. Every count in section 3.2 remains checkable from CLINVAR_FUNCTIONAL_CENSUS.csv and reconcile_gap_rows.csv, which carry the per-record data both parser passes produced; what cannot be re-derived is the parsing itself.

Archive versioning. The concept DOI 10.5281/zenodo.21799233 always resolves to the current version of the data archive and is the identifier to follow for access. The version current at the time of this revision is version 2, 10.5281/zenodo.21840036. Version DOIs cited elsewhere in this manuscript name the specific version read and are deliberately not rewritten.

8. Supplementary files

The last four files exist because of the defect recorded in section 3.2. They are included rather than quietly replaced so that a reader holding an earlier copy of the census file can reconcile it against this paper row by row instead of taking my word for the difference.

Figure 1

The retrieval filter indexes one of two adjacent elements, and the consequence for coverage. (a) FunctionalData contains two sibling elements carrying Sequence Ontology terms; FunctionalEffect takes 4 distinct values across the release, FunctionalConsequence takes 185. (b) Agreement between each element and live FCNS membership, computed over all 270,827 records carrying functional evidence, the full pool including the single bulk depositor: FunctionalConsequence predicts membership perfectly, FunctionalEffect does not. (c) Each of the 185 distinct FunctionalConsequence terms queried individually against FCNS; every term returns a non-zero result, and the dashed line is equality with the count from the independent parse. Points off the line reflect orthographic normalisation in NCBI's index. Higher agreement in (b) means the element is the one indexed.

Competing interests, and who did the work

Sole author. I designed the analysis, specified and directed the parser code, performed the retrieval tests and the tool survey, and directed and revised the manuscript. The parser code was written by an AI assistant to my specification and the manuscript prose was drafted by one; see the "Use of AI tools" section. I am responsible for every claim made here. I am a carrier of a variant in one of the genes discussed (SCN5A), which is what prompted the initial question; the analysis is a measurement of a database and makes no claim about any variant, patient or clinical decision. No funding was received. No competing financial interests.

Use of AI tools

This work was carried out with AI coding and research assistants (Anthropic Claude, via Claude Code). That use is disclosed here rather than left to inference.

Analysis code. The great majority of the analysis code in this project -- parsers, genome scans, regeneration scripts and verification scripts -- was written by an AI assistant working to my specification. I set what each script had to compute, chose the thresholds and the decision rules, and checked the output against the claims it is used to support.

Manuscript text. The prose of this manuscript was drafted by an AI assistant. I directed the drafting and revised the result, and I am responsible for every claim it makes.

Scientific decisions. The questions asked, the thresholds set, what was allowed to count as a refutation, and what was published were mine.

Verification, which does not depend on any of the above. Where a claim in this manuscript is regenerable from deposited inputs, the script that regenerates it and that script's own output are in the data deposit. Reproduction does not require trusting any account of who wrote what.

What no AI system did. No AI system generated, altered or selected any experimental measurement; this project contains no wet-lab data of any kind. All primary literature cited was retrieved from PubMed, PMC and publisher sources. Every reference in this manuscript has been machine-resolved against its own record, including a check that each PMID's first author and year match the author and year printed beside it in the text.

References

This paper rests mainly on database releases, schemas and source repositories rather than on a literature corpus, so the list below is short by nature. It is divided accordingly. Each of the three journal articles was verified against Europe PMC on 4 August 2026 by querying its PubMed identifier and reading back the author list, journal, year, volume, issue and pages recorded here; none is reproduced from memory. Where an article is cited in the text it is cited by first author and PMID so the two lists can be checked against each other.

Journal articles

1. Richards S, Aziz N, Bale S, Bick D, Das S, Gastier-Foster J, Grody WW, Hegde M, Lyon E, Spector E, Voelkerding K, Rehm HL; ACMG Laboratory Quality Assurance Committee. Standards and guidelines for the interpretation of sequence variants: a joint consensus recommendation of the American College of Medical Genetics and Genomics and the Association for Molecular Pathology. Genet Med 2015;17(5):405-424. doi:10.1038/gim.2015.30. PMID 25741868. Cited in section 1 for the framework in which PS3 and BS3 are defined.

2. Brnich SE, Abou Tayoun AN, Couch FJ, Cutting GR, Greenblatt MS, Heinen CD, Kanavy DM, Luo X, McNulty SM, Starita LM, Tavtigian SV, Wright MW, Harrison SM, Biesecker LG, Berg JS; Clinical Genome Resource Sequence Variant Interpretation Working Group. Recommendations for application of the functional evidence PS3/BS3 criterion using the ACMG/AMP sequence variant interpretation framework. Genome Med 2019;12(1):3. doi:10.1186/s13073-019-0690-2. PMID 31892348. Cited in section 1 for how assay calibration sets the strength of a functional evidence code.

3. Ma JG, O'Neill MJ, Richardson E, Thomson KL, Ingles J, Muhammad A, Solus JF, Davogustto G, Anderson KC, Shoemaker MB, Stergachis AB, Floyd BJ, Dunn K, Parikh VN, Chubb H, Perrin MJ, Roden DM, Vandenberg JI, Ng CA, Glazer AM. Multisite validation of a functional assay to adjudicate SCN5A Brugada syndrome-associated variants. Circ Genom Precis Med 2024;17(4):e004569. doi:10.1161/circgen.124.004569. PMID 38953211. Cited in section 3.4a as the assay underlying the worked example, VariationID 4681248.

Data, schemas and interfaces

4. ClinVar VCV XML release, release date 2026-06-27, downloaded 4 August 2026 from ftp.ncbi.nlm.nih.gov/pub/clinvar/xml/ClinVarVCVRelease_00-latest.xml.gz. 5,824,540,370 bytes, MD5 2e7e76ebbf668910b8688cc5e4284c1b, matching the .md5 sidecar published alongside it. This file is the primary source for every count in the paper.

5. ClinVar VCV XML schema, ClinVar_VCV_2.6.xsd, declared in the header of the release above and published at ftp.ncbi.nlm.nih.gov/pub/clinvar/xsd_public/. Source for the declaration of FunctionalData within the ObservationSet complex type of which ObservedIn is an instance.

6. NCBI E-utilities, einfo, esearch, efetch and esummary against db=clinvar, queried 4 August 2026. Source for the 47 indexed fields including FCNS, for all live retrieval counts, and for the rettype=vcv and is_variationid behaviour reported in section 3.4a.

7. The esummary document type definition for db=clinvar, published by NCBI. Source for the statement that its 74 declared elements include none containing the string "functional".

8. Public source repositories of the twelve variant-interpretation resources surveyed in section 3.6, read 4 August 2026. Each resource, the specific route examined and the evidence per row are recorded in CENSUS_TOOL_ACCESS_ROUTES_V2.csv rather than duplicated here.

9. DOAJ and Crossref public APIs, queried for the venue characteristics recorded in CENSUS_VENUE_VERIFICATION.csv.

I did not perform a systematic literature search for prior quantitative censuses of this element. The three articles above are cited for the interpretive framework and for one assay, not as a survey of prior work, and I make no claim that no earlier census exists.

Back to all ten papers