BRENNEN

Industries / Health / Healthcare / Medical labs / Genetic testing and diagnostics

Deciding what a change means

Judging whether a DNA change found in a patient causes disease, does not, or cannot yet be called, and keeping that judgment current as evidence comes in.

7

Milestones achieved, from 2013 to 2024.

2

To watch for: announced, not yet confirmed outside.

2

Open deliverables. Nobody has shipped them.

The road

AchievedWatch forOpen

Achieved

  1. 2013

    Findings nobody asked for

    An American College of Medical Genetics and Genomics (ACMG) working group led by R. C. Green told labs that run an exome or genome for any reason to also check a minimum list of 56 genes for disease-causing changes in conditions that can be confirmed and treated, and to report them regardless of the reason for the test or the patient's age. It did not favor letting patients decline these findings. After objections, the ACMG Board said on 1 April 2014 that patients should be offered a chance to opt out before testing. The list is now revised yearly; version 3.3 came out in 2025.

    LessonA rule that removed patients' choice did not last a year; consent had to come before the test.

    pmc.ncbi.nlm.nih.govbio-itworld.comeuropepmc.org

    Credit

    R. C. Green, L. G. Biesecker and the ACMG working group; the ACMG Board of Directors

  2. 2014

    A public record of verdicts

    ClinVar, run by the US National Center for Biotechnology Information (NCBI), collects labs' and researchers' classifications of DNA changes with their evidence, keeps every version of each submission, and flags when submitters disagree. M. J. Landrum, D. R. Maglott and colleagues described it in January 2014, with its data free to download.

    europepmc.org

    Credit

    M. J. Landrum, D. R. Maglott and colleagues at NCBI

  3. 2016

    Common, so not the cause

    The Exome Aggregation Consortium (ExAC) pooled the exomes of 60,706 people of several ancestries. Its team reviewed 192 changes that had been reported as disease-causing but were carried by more than 1% of people overall, or of South Asian or Latino people; 163 met ACMG criteria to be reclassified as benign or likely benign, and 126 of those had been reclassified in their source databases by December 2015.

    LessonA change that is too common for how rare its disease is cannot be the cause.

    pmc.ncbi.nlm.nih.gov

    Credit

    M. Lek, K. J. Karczewski and the Exome Aggregation Consortium, in Nature

  4. 2016

    Nine labs, 34% agreement

    Nine US labs in the Clinical Sequencing Exploratory Research (CSER) consortium classified 99 variants with the ACMG-AMP rules published in 2015. Across labs they agreed only 34% of the time, rising to 71% after discussion and a closer reading of the rules.

    LessonShared rules did not by themselves bring agreement; labs had to compare answers and make the rules more specific.

    europepmc.org

    Credit

    L. M. Amendola, G. P. Jarvik and colleagues, in the American Journal of Human Genetics

  5. 2018

    Most uncertain results turn benign

    A variant of uncertain significance (VUS) is a change the lab cannot yet call harmful or harmless. Among 1.45 million people given hereditary cancer tests at a single commercial lab, whose staff at Myriad Genetic Laboratories co-wrote the study, 7.7% of distinct VUS were later reclassified. Of those, 91.2% were downgraded to benign or likely benign, after a median 1.17 years, and 8.7% were upgraded to pathogenic, after a median 1.86 years.

    LessonAn uncertain result is not a positive one; most turn out harmless.

    pmc.ncbi.nlm.nih.gov

    Credit

    J. Mersch, T. Ross and colleagues at UT Southwestern Medical Center and Myriad Genetic Laboratories, in JAMA

  6. 2022

    Scores given evidence weights

    Under the 2015 rules, a computer prediction that a change harms a protein counted only as "supporting" evidence, and labs chose their own score cut-offs. ClinGen's Sequence Variant Interpretation Working Group, led by V. Pejaver and S. E. Brenner, converted the scores of 13 prediction tools into evidence strengths using independent data. Most tools reached the supporting level, several reached "moderate" and some "strong".

    europepmc.org

    Credit

    V. Pejaver, S. E. Brenner and the ClinGen Sequence Variant Interpretation Working Group, in the American Journal of Human Genetics

  7. 2024

    Uncertainty falls unevenly

    Comparing 213,663 people of European-like ancestry with 206,975 of other ancestries in the All of Us and gnomAD databases, M. Dawood and colleagues found more uncertain results among people of non-European-like ancestry in every medical specialty they checked. Lab tests that measure the effect of every possible single change in a gene, applied to BRCA1, TP53 and PTEN, resolved those people's uncertain results at a higher rate and offset the gap.

    LessonEvidence drawn from population frequencies and computer predictions served non-European groups less well.

    europepmc.org

    Credit

    M. Dawood, I. Gallego Romero and colleagues, in Genome Medicine

Watch for

  1. AI scores for every letter

    Google DeepMind released AlphaGenome Atlas on 8 September 2026: precomputed predictions for the effects of about 9 billion possible single-letter changes in the human genome, condensed into one AlphaGenome Variant Impact (AVI) score that combines its AlphaGenome and AlphaMissense models. DeepMind says the AVI score is best in class on many pathogenicity and rare-disease benchmarks, by its own testing. Outside checks have begun: H. Qu and H. Hakonarson derived a cut-off from ClinVar and found that 11,357 of 24,989 uncertain variants (45%) scored above it, and J. P. Miranda and colleagues at Bupa Chile posted a ClinGen-style calibration in November 2026. Both are preprints.

    Achieved whena peer-reviewed calibration on variants AVI was not tuned on sets its evidence strengths, and ClinGen or clinical labs adopt them.

    deepmind.googleeuropepmc.orggithub.com

    Claimed by

    Google DeepMind, September 2026

  2. Points instead of rules

    The 2015 ACMG-AMP rules give each kind of evidence a fixed strength. A committee of ACMG, AMP, the College of American Pathologists and ClinGen has drafted a version 4 that scores evidence in points on a Bayesian scale, with decision trees for applying each kind of evidence. In an April 2026 talk for the Human Genome Organisation (HUGO), Andreas Laner of MGZ, a genetics lab in Munich, presented the draft's scale, on which a change counts as benign below a 1% chance of causing disease, against the 0.1% assumed in the Bayesian reading of the 2015 rules. In November 2026 ClinGen's pages still describe the committee as working on the standards, and one expert panel says it will adapt its rules "as these guidelines are finalized".

    Achieved whenthe standard is published and labs or an outside group report how classifications and agreement between labs change under it.

    hugo-international.orgclinicalgenome.org

    Claimed by

    the ACMG/AMP/CAP/ClinGen Sequence Variant Classification committee, as presented at HUGO, April 2026

Open

  1. 01

    A secondary findings answer key

    A public set of made-up VCF files, the standard variant lists a sequencing pipeline produces, drawn from simulated genomes of no real person, with an answer key giving the correct secondary-findings report for each file under the ACMG list (version 3.3). The files are built to test the list's gene-by-gene rules, and the open tools that automate the search are scored against them.

    The brief
    Credit · unclaimed
  2. 02

    Why labs disagree, tagged

    A monthly public table that tags each conflict in ClinVar, where submitters give the same change different classifications, with its likely cause: an old submission never updated, different conditions named, different population-frequency cut-offs, different use of computer predictions, or case evidence that only one lab cites. It reads only public variant-level fields: dates, conditions, review status, and the evidence text and ACMG criteria codes that many labs write into their submissions.

    The brief
    Credit · unclaimed

Open 01 · the brief

A secondary findings answer key

Credit · unclaimed

The first to publish is credited on this line, and the deliverable moves from Open to Achieved.

What to build

A public set of made-up VCF files, the standard variant lists a sequencing pipeline produces, drawn from simulated genomes of no real person, with an answer key giving the correct secondary-findings report for each file under the ACMG list (version 3.3). The files are built to test the list's gene-by-gene rules, and the open tools that automate the search are scored against them.

Why it is still open

The list carries rules beyond "report every pathogenic change". A public table of version 3.3 has 84 genes in 100 gene-condition rows: 88 rows say report all pathogenic and likely pathogenic changes, and the rest set special rules, such as two changes needed for recessive conditions, HFE reported only for people with two copies of p.C282Y, and one gene reported only for truncating changes. Tools exist and report accuracy on their own data: ACMG_SF_Classifier says 98.78% sensitivity and 99.73% specificity on 7,800 clinical samples at the Research Centre for Medical Genetics in Moscow. NHGRI's own Secondary Genomics Finding Service still bases its list on the 2017 ACMG report and does not accept GRCh38 files. Literature, web and GitHub searches found no shared test set.

github.comgithub.comgenome.gov

Built with what exists

The ACMG SF v3.3 policy statement and the public table of it; ClinVar; gnomAD allele frequencies to draw background genotypes at random, as one open project already does with older ExAC frequencies; bcftools; the open tools SFtool from Babelomics and ACMG_SF_Classifier; Claude Fable 5.2 or another coding agent to write the generator and its tests (see the Note "Three working devices").

europepmc.orggithub.com

  1. First step

    Write out the special rules from the version 3.3 table and generate 20 files with planted changes for those genes.

  2. Done looks like

    At least 500 synthetic files covering every gene and every special rule. Two clinical molecular geneticists check the answer key for a random 50 and agree with it. At least three tools are scored rule by rule. Files, key and code are published with a DOI.

Open 02 · the brief

Why labs disagree, tagged

Credit · unclaimed

The first to publish is credited on this line, and the deliverable moves from Open to Achieved.

What to build

A monthly public table that tags each conflict in ClinVar, where submitters give the same change different classifications, with its likely cause: an old submission never updated, different conditions named, different population-frequency cut-offs, different use of computer predictions, or case evidence that only one lab cites. It reads only public variant-level fields: dates, conditions, review status, and the evidence text and ACMG criteria codes that many labs write into their submissions.

Why it is still open

In 2024 T. E. Lazareva and colleagues found that 5.7% of ClinVar variants carry conflicting interpretations and that 78% of clinically relevant genes hold at least one, but they analyzed patterns by gene, not causes conflict by conflict. Cause-by-cause comparisons have been done by hand on small sets: the nine CSER labs' 99 variants, and 242 disagreements among four large labs working with ClinGen, the US Clinical Genome Resource, in 2017, of which 87.2% were settled by applying current rules or sharing internal data and the rest came down to how each lab applied the rules. Open projects now track how ClinVar classifications change from release to release, one of them counting 479,525 changes between January 2020 and July 2026, without saying why labs disagree.

europepmc.orgeuropepmc.orgeuropepmc.orggithub.com

Built with what exists

ClinVar's submission_summary file, whose evidence text often names the ACMG criteria used (for example "PVS1_vstrong, PM2_mod"); rules for the structured fields; a small extraction model such as Muse Ember for the free text (see the Note "The jobs agents repeat"); Python and DuckDB.

ftp.ncbi.nlm.nih.gov

  1. First step

    Take every conflict in five genes, tag the causes by hand from the submission text, and count how many a set of simple rules gets right.

  2. Done looks like

    Every germline conflict between two or more submitters that state assertion criteria is tagged. Two variant curators blind-check 200 random tags and agree with at least 90%. The table is rerun on three monthly releases, with a list of conflicts that closed and the cause each had been tagged with.