55 — Clinical variant interpretation
Before this: Ch 54 — Rare variants and Mendelian disease · Ch 46 — Variant calling · Ch 41 — Data formats · Time: ~55 min
Statistics needed: S1 Probability · S6 Likelihood and Bayes
What you'll be able to do
- Place a variant in the five-tier ACMG/AMP scheme and say what each tier does and does not license
- Name the evidence categories, assign a strength to each, and apply the combining rules to reach a call
- Reconstruct the Bayesian model underneath the checklist, convert criteria to points, and compute the posterior
- Derive a from prevalence, and heterogeneity, and use it to argue both "too common" and "informatively absent"
- Say why three agreeing computational predictors are not three pieces of evidence, and what calibration replaces them with
- Compute an assay's OddsPath from its pathogenic and benign controls, and say why a perfectly discriminating assay with a small control set still cannot reach Very Strong
- Read a record for its evidence rather than its label, and say why classifications move and why penetrance from families overstates risk in an unselected person
This chapter describes how professionals reason. It is not clinical advice, and nothing in it should be used to make a decision about a real person's health.
The core idea
A patient has a disease that looks genetic. Sequencing returns a variant. Someone must decide whether that variant explains the disease — and the decision has consequences: surveillance imaging, prophylactic surgery, testing relatives, reproductive choices.
The naive framing is that the variant either causes disease or it doesn't, and the job is to find out which. That framing makes the output a property of the variant. It is not. The output is a property of what is currently known — how many people have been sequenced, whether anyone ran an assay, whether the family was large enough to observe , whether anyone bothered to curate the gene.
A variant classification is a statement about the evidence, not about the variant. Two labs looking at the same variant on the same day can honestly disagree because they hold different evidence. The same lab can honestly change its mind next year without the variant having changed. "Uncertain significance" is not a claim that the variant is borderline — it is a claim that the evidence has not moved the probability far enough from its starting point in either direction.
The framework everyone uses looks like a checklist: twenty-eight lettered criteria, a set of combining rules, five output tiers. Underneath it is a naive Bayes classifier with hand-set likelihood ratios — likelihood ratios, priors and posteriors are covered in S6, and this chapter assumes them. Each criterion contributes an odds ratio; the strength tiers are a geometric ladder of those odds; the combining rules are a coarse quantisation of "multiply the likelihood ratios into the prior". Recognising that turns an arbitrary-looking rulebook into a model you can criticise — and every important failure mode of the framework is a failure mode of naive Bayes: correlated evidence double-counted, likelihood ratios asserted rather than measured, a prior that is not the same for every gene.
The operative published standard is Richards et al. 2015 (ACMG/AMP), refined by ClinGen Sequence Variant Interpretation working-group recommendations and by gene- and disease-specific Variant Curation Expert Panel specifications. A fourth version is in draft and pilot; §11 says what is changing and why you should not yet use it (reference/verified-facts.md).
1. The five tiers
| Tier | Posterior P(pathogenic) | What it licenses |
|---|---|---|
| Pathogenic | > 0.99 | Diagnosis, management change, predictive testing of relatives |
| Likely pathogenic | 0.90 – 0.99 | In practice the same clinical actions as pathogenic |
| Uncertain significance () | 0.10 – 0.90 | Nothing |
| Likely benign | 0.001 – 0.10 | Removed from consideration; not reported by most labs |
| Benign | < 0.001 | Removed from consideration |
The probability boundaries are not decoration: the 2015 guideline defines "likely" as greater than 90% certainty, and §3 shows the whole rule set is consistent with the boundaries above.
Likely pathogenic is not a hedge. A 90–99% posterior is acted on; P versus LP affects how a report is worded, rarely how a patient is managed.
VUS is the residual, and it is huge. Any variant on which evidence has not accumulated lands there — which, for a novel change in a poorly studied gene, is the default state of the world. A VUS must not drive a clinical decision, must not be used to test relatives, and must not be reported to a patient as "a ". The instruction is asymmetric and deliberate: acting on a VUS as if pathogenic causes harm (unnecessary surgery), and acting on it as if benign causes harm (missed diagnosis), so the correct action is neither — continue the workup by other means.
VUS rates are unequally distributed, and that is a data problem. Patients whose ancestry is under-represented in reference databases and in the published literature receive more VUS results, because both arms of the evidence base are thinner for their variants. The variants are not more ambiguous; the evidence is more absent. Same causal story as polygenic-score failure (Ch 53): biased sampling, not biology.
And a classification attaches not to a variant but to a (variant, gene–disease pair, laboratory, date) tuple. The same variant can be pathogenic for one condition and irrelevant to another, and a classification without its date is nearly uninterpretable.
2. The evidence framework
Criteria are named by side (P or B), strength, and index. On the pathogenic side: PVS (very strong), PS (strong), PM (moderate), PP (supporting). On the benign side: BA (stand-alone), BS (strong), BP (supporting).
| Category | Pathogenic | Benign | What is actually being measured |
|---|---|---|---|
| Population frequency | PM2, PS4 | BA1, BS1, BS2 | Is this variant rarer than a causal for this disease could be — or commoner than one could be? |
| Predicted molecular effect | PVS1, PS1, PM4, PM5, PP3 | BP1, BP3, BP4, BP7 | Does the sequence change plausibly break the protein, by the mechanism this disease uses? |
| Location / | PM1, PP2 | — | Does it fall in a region where variation is not tolerated? |
| Functional assay | PS3 | BS3 | Does a validated experiment show loss (or retention) of function? |
| PP1 | BS4 | Does it track with disease through meioses in families? | |
| De novo | PS2, PM6 | — | Did it arise anew in an affected child of unaffected parents? |
| Allelic / | PM3 | BP2 | For a gene, is it in trans with a known pathogenic allele? |
| specificity | PP4 | BP5 | Is the phenotype so specific it implicates the gene — or already explained by something else? |
| Retired | ~~PP5~~ | ~~BP6~~ | "A reputable source says so." Withdrawn by ClinGen: it launders someone else's unexamined conclusion into your evidence |
Two deserve unpacking.
PVS1 — a null variant (nonsense, frameshift, canonical ±1/±2 splice-site, initiation- loss, whole-exon deletion) in a gene where is an established mechanism of that disease. The final clause is the whole criterion. A stop codon in a gene whose mechanism is is not evidence of pathogenicity, and may be evidence against: a truncated transcript degraded by (Ch 06) produces no poison peptide at all. The ClinGen refinement replaces the flat criterion with a decision tree that sets strength from whether NMD is predicted, how much protein is removed, whether a skipped exon preserves frame, and whether an alternative start codon rescues . A frameshift in the last exon is not the same evidence as one in exon 2, and the tree says so.
PM2 — absent from population databases. The 2015 text made this Moderate; ClinGen recommends Supporting, on a base-rate argument: almost every variant is absent from almost every database, because the space of possible variants vastly exceeds the number sampled. Rarity is what benign and pathogenic rare variants have in common, so observing it barely moves the odds. §4 makes that quantitative.
The combining rules
Pathogenic if any of:
PVS1 + (≥1 Strong | ≥2 Moderate | 1 Moderate + 1 Supporting | ≥2 Supporting)
≥2 Strong
1 Strong + (≥3 Moderate | 2 Moderate + ≥2 Supporting | 1 Moderate + ≥4 Supporting)
Likely pathogenic if any of:
PVS1 + 1 Moderate
1 Strong + 1–2 Moderate
1 Strong + ≥2 Supporting
≥3 Moderate
2 Moderate + ≥2 Supporting
1 Moderate + ≥4 Supporting
Benign if BA1 alone, or ≥2 Strong. Likely benign if 1 Strong + 1 Supporting or ≥2 Supporting. Uncertain otherwise — including when pathogenic and benign criteria are both met and conflict.
Read those as arithmetic and a pattern appears: one Very Strong behaves like two Strong, one Strong like two Moderate, one Moderate like two Supporting. That is not a coincidence.
3. The Bayesian reformulation
Tavtigian and colleagues asked whether the rules above quantise an actual model, and found that they do — to a remarkable degree.
Statistics: the odds form of Bayes — prior odds × likelihood ratio = posterior odds — is covered in S6 §4–5, and S6 §7.3 runs this exact reconstruction in code. Bayes for events is S1 §5.
Posit a naive Bayes classifier. Let π be the prior probability that a variant submitted for classification is pathogenic, and let each criterion i contribute an odds of pathogenicity OddsPathᵢ — the likelihood ratio P(evidence | pathogenic) / P(evidence | benign). Then
Now impose the structure visible in the combining rules: the strength tiers form a geometric ladder, each step squaring the one below. Fix the Very Strong odds at X; then
Very Strong = X Strong = X^(1/2) Moderate = X^(1/4) Supporting = X^(1/8)
Taking logs, that ladder makes the evidence additive in units of the Supporting step — so assign points: Supporting 1, Moderate 2, Strong 4, Very Strong 8, negative on the benign side. Total points N gives OddsPath = X^(N/8), and
With π = 0.10 and X = 350 the boundaries fall out:
| Points N | OddsPath = 350^(N/8) | Posterior | Tier |
|---|---|---|---|
| 10 | 1514 | 0.9941 | Pathogenic (> 0.99) |
| 6 | 80.9 | 0.8999 | Likely pathogenic — the 0.90 boundary, exactly |
| 5 | 38.9 | 0.8122 | |
| 0 | 1 | 0.1000 | VUS — the prior, untouched |
| −6 | 0.0124 | 0.00137 | Likely benign |
| −7 | 0.00594 | 0.00066 | Benign (< 0.001) |
The point thresholds ≥10 / 6–9 / 0–5 / −1 to −6 / ≤−7 reproduce the posterior cut-points 0.99, 0.90, 0.10 and 0.001 essentially exactly. Sixteen of the eighteen 2015 combining rules are internally consistent with this arithmetic. Two are not, and they are worth checking yourself:
≥2 Strong→ Pathogenic. Two Strong is 8 points, OddsPath 350, posterior 35/35.9 = 0.975 — which is Likely pathogenic, not Pathogenic.PVS1 + 1 Moderate→ Likely pathogenic. That is 10 points, posterior 0.994 — which is Pathogenic.
Neither discrepancy is scandalous; both are what you get when a committee quantises a continuous model by hand. What the reformulation buys is enormous:
Criteria become measurable rather than asserted. If a criterion is an odds ratio you can estimate it — count how often the evidence appears among known pathogenic versus known benign variants. That single move is what makes §5 and §6 possible, and it is why the field is migrating from "does this criterion apply?" to "at what strength?".
Intermediate strengths become legal. PP3_Moderate, PS3_Supporting, PM2_Supporting are not fudges; they are the model used at its natural resolution.
The prior becomes visible, and contestable. π = 0.10 is the prior for a variant arriving in a classification queue, not for a random position in the genome. Different calibration exercises legitimately use different priors — the ClinGen computational-predictor calibration used π = 0.0441 with X = 1124, giving the same geometric ladder (2.41, 5.79, 33.5, 1124) on a different footing. The framework is parameterised by (π, X), and both are modelling choices.
And the model's weakness becomes explicit. Multiplying likelihood ratios assumes the evidence is conditionally independent given pathogenicity. It very often is not, and every double-counting problem in the rest of this chapter is that assumption failing.
Points-based implementations follow directly, and several VCEP specifications and clinical pipelines now score variants that way. The tiers are unchanged; only the bookkeeping is.
4. Population frequency, and why "rare" is not a number
is the workhorse: 730,947 exomes + 76,215 genomes = 807,162 individuals, aligned to GRCh38 — roughly 1.6 million at a well-covered site (reference/verified-facts.md). It is the single most informative resource in clinical interpretation, and it is used in both directions.
The mistake is to reach for a fixed threshold. Whether a frequency is disqualifying depends entirely on the disease.
Deriving the maximum credible allele frequency
Let D be disease prevalence, g the largest fraction of cases attributable to this gene (genetic heterogeneity), a the largest fraction of those attributable to a single variant (allelic heterogeneity), and ψ the . For a condition, affected carriers of this variant are a fraction D·g·a of the population; since only ψ of carriers are affected, carriers are D·g·a/ψ; and since each carrier holds one of their two alleles:
For a condition affected individuals are homozygous, so under Hardy–Weinberg (Ch 26) the corresponding bound is √(D·g·a/ψ).
Take hypertrophic cardiomyopathy: prevalence ≈ 1/500. The largest single-gene contributor is MYBPC3, at roughly 20% of cases; MYH7 is next, at roughly 14% (Sedaghat-Hamedani et al., Clin Res Cardiol 2018, meta-analysis of 7,675 patients), so for MYH7 take g ≈ 0.14. Extreme allelic heterogeneity gives a ≈ 0.02, and penetrance is age-dependent, ψ ≈ 0.5.
AF_max = (2.0e-3 × 0.14 × 0.02) / (2 × 0.5)
= 5.6e-6 / 1.0
= 5.6 × 10⁻⁶ ≈ 9 alleles in gnomAD's ~1.6 million
A variant seen forty-odd times in gnomAD — a raw frequency of 2.5 × 10⁻⁵, with a lower 95% bound near 1.9 × 10⁻⁵ — is therefore several times commoner than this model says a MYH7 HCM allele could be. Change the assumptions and the answer moves fourfold: at a = 0.005 the bound is 1.4 × 10⁻⁶, about 2 alleles. The threshold is a model output, not a constant — which is why VCEPs publish a disease-specific cut-off rather than leaving it to the curator, and why BA1 at 5%, the 2015 framework's one frequency criterion needing no disease model, is its only stand-alone criterion.
Two cautions before using a number like that in anger, and they are the reason a curator checks for a specification before deriving anything.
Where a VCEP has published thresholds, the thresholds govern — and they are looser than the theoretical bound. The ClinGen Cardiomyopathy Expert Panel's MYH7 specification (v2.0, approved April 2024) sets BA1 at a filtering ≥ 0.1%, BS1 at FAF ≥ 0.01%, and PM2_Supporting at an upper 95% bound ≤ 0.004% in the popmax group — which the panel spells out as no more than about four alleles in an allele number of 230,000. On the panel's numbers our forty-allele variant earns neither: its filtering allele frequency of ~1.9 × 10⁻⁵ is about fivefold below the BS1 line, and forty alleles is far too many for PM2. The gap between 5.6 × 10⁻⁶ and 10⁻⁴ is not an error in either direction — the derivation gives the frequency above which pathogenicity is implausible, and the panel sets the frequency above which a curator may assert benignity and expect to be right nearly always. A benign call is acted on, so the panel buys margin.
And the two bounds are computed on different denominators. The 1.9 × 10⁻⁵ above came from all ~1.6 million alleles; a filtering allele frequency, defined next, is a popmax quantity computed in the single ancestry group where the variant is commonest, where the allele number is much smaller and the interval correspondingly wider. A global lower bound is a stand-in for the real thing, and it is the optimistic stand-in.
Filtering allele frequency, and the other direction
A raw count is a point estimate; you want a bound — and which end of the interval you need depends on which way you argue.
Statistics: confidence intervals on a rare allele count, and the rule of three for zero observations, are covered in S3 §5 and §8.
To argue too common (BS1/BA1) you must be confident the frequency is genuinely high, so take the lower 95% bound. That is the filtering allele frequency: the lower limit of the 95% interval on the allele frequency in the group where the variant is commonest, the group choice guarding against a variant that looks rare globally only because the group carrying it is a small share of the database. If FAF still exceeds AF_max, the variant is too common even under the assumptions most favourable to pathogenicity.
To argue informatively absent (PM2) you need the opposite end. Zero observations in ~1.6 million alleles bounds the frequency above by roughly 3/1.6 × 10⁻⁶ ≈ 1.9 × 10⁻⁶ — the rule of three. That sits below HCM's AF_max of 5.6 × 10⁻⁶, so absence genuinely excludes the "too common" hypothesis. For a disease whose AF_max is 10⁻⁷ it excludes nothing: the database cannot resolve frequencies that low. Absence is informative only when the database resolves the threshold that matters — and even then, most benign rare variants are absent too, which is the base-rate argument for PM2 at Supporting.
Four standing caveats on the database itself:
| Caveat | Consequence |
|---|---|
| gnomAD is not a healthy cohort | Severe paediatric disease is excluded; adults with late-onset conditions are not. For adult-onset disease, carriers are present by design |
| Ancestry composition is uneven | A variant common in an under-sampled population looks rare. Using the highest-frequency group helps only for groups actually sampled |
| varies by site | Allele number varies; a low AF over a low AN is meaningless. Read AN, not just AF |
| Homozygotes are separate evidence | For a fully penetrant early-onset recessive condition, healthy homozygotes are BS2 — a cleaner argument than frequency |
5. Computational predictors
The 2015 text asks for "multiple lines of computational evidence" at Supporting strength. That phrasing invited exactly the wrong behaviour, and the correction is the clearest illustration of the naive-Bayes problem in the whole framework.
They are not independent. Most predictors read the same underlying signal: cross-species at the residue, derived from a multiple sequence alignment. Physical features (buried versus exposed, charge, proline in a helix) are shared too. Worse, the popular "consensus" tools are literal ensembles of each other — REVEL is a random forest over the scores of thirteen other predictors. Recording that SIFT, PolyPhen-2 and REVEL agree is not three observations; it is one observation consulted three times, and multiplying three likelihood ratios for it inflates the evidence by roughly the square of what it should be.
They are contaminated by their evaluation sets. Supervised predictors are trained on labelled variants, and the labels come from and HGMD. Evaluating such a tool on ClinVar measures memorisation as much as generalisation, so published AUCs are systematically optimistic. Honest calibration must exclude training-set overlap — the ClinGen exercise did, which is part of why its thresholds are more conservative than tool developers' defaults.
They are uneven across genes. Accuracy depends on alignment depth, paralogue structure and how well the gene was represented in training, so performance is worst precisely in poorly studied genes — the ones where you needed help. And a predictor trained to recognise "damaging" generalises badly to gain-of-function mechanisms, where the pathogenic variant may be biophysically mild.
The shift to likelihood ratios
Stop asking "is the score above the tool's threshold?" and start asking "what likelihood ratio does this score carry?" ClinGen's calibration took ~11,800 curated ClinVar variants across ~1,900 genes, estimated the local posterior as a function of score, and read off the intervals where each tool reaches Supporting, Moderate, Strong and Very Strong. For REVEL, with π = 0.0441:
BP4 very strong BP4 str BP4 moderate BP4 sup INDETERMINATE PP3 sup PP3 mod PP3 strong
|───────────────|──────────|───────────────|──────────|──────────────|──────────|──────────|───────────|
0 0.003 0.016 0.183 0.290 0.644 0.773 0.932 1
Three things to take from that picture. There is an explicit indeterminate band — middling scores carry no evidence in either direction, and "REVEL = 0.5, mildly suggestive" is reporting noise. The scale is asymmetric: REVEL reaches Very Strong on the benign side but only Strong on the pathogenic side, because a low score is the more reliable signal. And the evidence is counted once: one calibrated predictor at its calibrated strength, not a show of hands.
Splice-effect prediction is a separate axis with its own calibration, supplying PP3/BP4 for variants a missense predictor cannot see at all — deep-intronic and synonymous changes (Ch 06). Proteome-scale models raise the ceiling on but repeal none of the three problems above: they still need independent calibration, still risk circularity with the label sets, and still cannot know a gene's disease mechanism.
6. Functional evidence and MAVEs
PS3/BS3 ask whether a validated experiment shows the variant damages (or spares) function. The word doing the work is validated. A published assay showing "reduced activity" is not evidence until you know how well it discriminates variants you already know the answer for.
The ClinGen approach converts an assay into an odds ratio using control variants. Run it over known pathogenic and known benign variants; let P₁ be the proportion pathogenic among all controls (the assay-specific prior) and P₂ the proportion pathogenic among controls the assay calls abnormal. Then
Statistics: why a ratio of proportions like this is a likelihood ratio, and what a likelihood ratio buys as evidence, are covered in S6 §4.
and the resulting odds map onto the strength ladder of §3 — roughly 2.1 Supporting, 4.3 Moderate, 18.7 Strong, 350 Very Strong on the (π = 0.10, X = 350) scale. An assay with eleven pathogenic and eleven benign controls that separates them perfectly still cannot exceed Strong, because the control set is too small to estimate an odds ratio of 350. The strength of functional evidence is a property of the assay's validation, not of how impressive the experiment looks.
Multiplexed assays of variant effect
The classical workflow is patient first, variant second, assay third — months of work per variant, and only for variants someone already cared about. Deep mutational scanning inverts it. Synthesise every possible single-amino-acid substitution across a domain or a whole protein, put the library into cells under a selection that depends on the protein working, sequence the population before and after, and read each variant's from its change in frequency (Ch 37). One experiment scores thousands of variants — including ones no patient has yet been found to carry.
That is the point: the functional evidence is already in the database when the patient's result arrives, converting the rate-limiting step of interpretation from months to a lookup. Calibrated against controls with the OddsPath machinery above, well-designed routinely reach Moderate or Strong, and their effect on burden in the genes they cover is large.
The limits are biological. An assay measures one function; a protein with several (catalysis, localisation, partner binding, stability) can be spared in the assayed dimension and broken in another, so a normal score is weaker evidence than an abnormal one unless the assay is known to capture the disease mechanism. A non-native cell type or non-physiological expression can mask effects entirely. And the calibration controls are themselves ClinVar classifications, so the circularity of §5 has not gone away — it has moved.
7. Segregation, de novo, and phase
(PP1/BS4) is a likelihood ratio you can compute exactly, and one of the oldest ideas in the book (Ch 14). If the variant is causal and fully penetrant, every affected relative carries it and every unaffected one does not; under the null each informative is a coin flip. So m meioses all going the right way give LR = 2^m, or LOD ≈ 0.3m. ClinGen guidance, following Jarvik and Browning, sets thresholds for a single family at 3 informative meioses for Supporting (LR 8, LOD ≈ 0.9), 4 for Moderate (LR 16, LOD ≈ 1.2), 5 for Strong (LR 32, LOD ≈ 1.5) — each threshold one meiosis lower when the evidence comes from two or more families. Compare those to §3's ladder — Supporting is 2.1, Moderate 4.3, Strong 18.7 — and the criteria are demanding roughly two- to fourfold more evidence than the nominal strength requires. That conservatism is deliberate: relatives are not independent draws, is rarely complete, and families are reported because they segregate, so the naive 2^m overstates the case.
Two limits bound it. Segregation implicates a locus, not a variant: anything in with the causal change segregates identically (Ch 29). And it saturates — reaching Very Strong requires a larger than most families are, which is why segregation is a supporting player in the modern framework rather than the centrepiece it was in 1990.
De novo evidence (PS2 confirmed, PM6 assumed) is strong because the coincidence is improbable: roughly 60–70 new arise per genome per generation (Ch 16), so one landing in a gene that matches the child's is unlikely by chance. Strength comes from a point system over two axes — how specific the phenotype is for the gene, and whether parentage was genetically confirmed rather than assumed — accumulated across probands, with 0.5 points reaching Supporting, 1 Moderate, 2 Strong, 4 Very Strong. A confirmed de novo in a syndrome with a near-unique genetic cause scores 2 points alone; the same observation in a genetically heterogeneous condition scores 0.5, because dozens of genes could have produced that phenotype. Unconfirmed parentage is not a technicality: a "de novo" variant actually inherited from an unaffected biological father is evidence of the opposite thing.
matters for conditions. Two heterozygous variants in one gene explain disease only in trans — one on each ; in cis they leave an intact copy and explain nothing. PM3 rewards a variant found in trans with a known pathogenic , BP2 penalises the reverse. Phase comes from genotyping the parents or from long reads spanning both sites, and is exactly what short-read calling discards (Ch 46) — a trio is worth far more than a proband alone for this reason as much as for de novo detection.
8. ClinVar, and how to read it
is the public archive where laboratories deposit classifications with supporting evidence. As of 8 August 2026 it holds 6.9 million submitted records covering 4.55 million unique variants from 3,485 submitters. It is the reason a lab in one country benefits from a family studied in another, and no serious pipeline runs without it.
It is also an archive of submitted assertions, not a truth set, and the difference is where people get hurt.
Submissions disagree. 165,877 unique variants carry conflicting classifications. Conflicts are not noise to be averaged away; they usually mean the submitters held different evidence, applied different gene-specific criteria, or submitted years apart.
Submitter quality is not uniform, and the star rating exists to say so:
| Stars | Review status | What it means |
|---|---|---|
| 0 | No assertion criteria | No stated method. Treat as an anecdote |
| 1 | Single submitter with criteria, or conflicting classifications | One lab's method — or submitters who disagree. Read every submission |
| 2 | Multiple submitters, no conflicts | Independent agreement |
| 3 | Reviewed by expert panel | A ClinGen VCEP applied a gene-specific specification |
| 4 | Practice guideline | The highest tier, and rare — 663 variants |
About 22,000 variants carry expert-panel review — a small fraction of 4.55 million, and where the framework works as designed: a VCEP publishes a gene-specific specification in advance (which frequency threshold, which domains earn PM1, which assays are calibrated and at what strength) and then applies it. Those records you can lean on; the rest require you to look at the evidence.
Read the evidence, not the label. A one-star "pathogenic" submission from 2016, citing a single case report and a computational prediction, is not evidence that the variant is pathogenic. It is evidence that someone once thought so. The 2015 criteria PP5 and BP6 — which allowed you to count "a reputable source says pathogenic" as evidence — were retired for exactly this reason: they let one lab's unexamined conclusion propagate through everyone else's classification as if it were an independent observation.
That propagation closes a loop worth stating plainly: ClinVar labels train computational predictors; predictor scores supply PP3/BP4 evidence; that evidence feeds new classifications; those classifications are submitted to ClinVar. Any systematic error entering the loop is amplified rather than corrected. Excluding training-set overlap during calibration, and preferring expert-panel records as controls, are attempts to cut the circuit — not solutions.
9. Reclassification, and the obligation it creates
Classifications move, and not symmetrically. In a large hereditary-cancer testing cohort, 7.7% of unique were reclassified and 91.2% of those were downgraded to likely benign or benign, median 1.17 years to an amended report. Because common variants are seen in many patients, roughly a quarter of reported VUS results were affected.
The direction follows from how evidence accumulates. Databases grow, an absent variant acquires a frequency, that frequency usually exceeds what a causal 's could be, BS1 applies, and the variant leaves the uncertain zone downward. Genuinely pathogenic variants are rescued more slowly, by functional data and by in further families. So the base case for a VUS is that it is benign and will eventually be shown to be — which is emphatically not a licence to treat one as benign today.
This creates an obligation with no clean engineering answer. A clinical report is a snapshot of the evidence on its issue date, but it is read years later as a fact. Someone must notice the evidence changed, find the patient, and tell them — and health systems have no reliable callback mechanism. The workable pattern is to make re-evaluation a scheduled batch job over the lab's own case archive rather than an event triggered by a patient's return, and to stamp the report with an as-of date. Who owns the duty to recontact is Ch 58.
10. Secondary findings
Sequence an exome for epilepsy and you also, incidentally, sequenced BRCA1. What do you do with a pathogenic variant nobody asked about?
The terms differ by intent: an incidental finding is stumbled upon, a secondary finding is one you deliberately went looking for outside the indication. The ACMG position is that a defined, minimal list of genes should be actively analysed and reported, with patient opt-out. The current list is ACMG SF v3.3 (2025): 84 genes — hereditary cancer syndromes, inherited cardiovascular conditions, a few metabolic disorders — updated annually.
The list is short on purpose. The inclusion criterion is actionability: there must be an intervention that changes outcome for someone who does not yet know they are at risk. Not "interesting", not "well understood", not "severe". A gene for an untreatable adult-onset neurodegenerative condition fails the test however confident the genetics — a value judgement the working group makes explicitly rather than a scientific conclusion. Reporting is restricted to pathogenic and likely pathogenic variants; never a , which would deliver anxiety with no information.
One statistical point here is easy to miss and matters enormously:
estimated from families ascertained through affected members overstates penetrance in an unselected person. A variant catalogued as pathogenic because it was found in families full of early-onset cancer was, by construction, observed in families where it did something. The same variant found in a population biobank, in someone with no family history, is associated with a substantially lower lifetime risk. The variant did not change; the sampling did.
That is ordinary selection bias, and it is why a secondary finding warrants a probabilistic conversation about risk rather than a diagnosis, and why penetrance estimates from unselected biobanks, where they exist, are the right ones to quote.
11. What is coming: ACMG v4
A fourth version is in draft and multi-site pilot, previewed at the ACMG 2025 Clinical Genomics Meeting, with publication anticipated around 2027 and a transition period after that. The direction of travel, as far as the drafts and pilot reports show: evidence codes consolidated around concepts rather than accumulated numbering, so related observations cannot be counted twice — the naive-Bayes problem attacked structurally rather than by errata; decision trees for each evidence type, generalising what the PVS1 tree did for null variants; gene–disease validity entering the classification itself rather than sitting beside it as an unstated precondition; and subdivided by likelihood of pathogenicity, so a variant at posterior 0.85 and one at 0.15 stop sharing a label.
The five tiers and the Bayesian skeleton survive. Until publication, 2015-plus-ClinGen-plus-VCEP is the standard in clinical use, and a report issued today against a draft standard is not interpretable by the people who receive it.
Common misconceptions
| What people think | What's actually true |
|---|---|
| A is a variant that is probably borderline | It is a variant whose evidence has not moved the posterior out of 0.10–0.90. Most VUSs are benign variants nobody has gathered evidence about yet. It licenses no clinical action and no testing of relatives — "cautious" action on a VUS is action on a VUS |
| Classifications are properties of variants | They are properties of (variant, gene–disease pair, lab, date). Two labs can honestly disagree today; one lab can honestly reverse itself next year |
| Three predictors agreeing is strong computational evidence | Predictors share features, share alignments, and are literally built from each other. Use one calibrated predictor, once, at its calibrated strength |
| Absent from means it must be pathogenic; above 1% means it can't be | Almost every variant is absent from almost every database, so absence is informative only if the database resolves the disease's maximum credible frequency — hence PM2 at Supporting. And the "too common" threshold is derived from prevalence, and heterogeneity, varying by orders of magnitude between diseases. Only BA1 at 5% needs no disease model |
| A published functional study establishes PS3 | Not until the assay is calibrated against known pathogenic and benign controls. The strength comes from the validation, not the experiment |
| A truncating variant is pathogenic | Only where is an established mechanism for that disease, and the ClinGen decision tree downgrades it by NMD prediction, position and rescue |
| Pathogenic in means pathogenic | 165,877 variants carry conflicting submissions, and a zero-star submission states no method at all. Read the evidence and the review status |
| A pathogenic secondary finding means the person will get the disease | Penetrance from families ascertained through affected members overstates risk in someone found incidentally. Same variant, different sampling |
Worked example: classifying one variant
This case is constructed. The variant and every observation are invented to exercise the framework; nothing here is a claim about the real residue named.
A 34-year-old proband with echocardiographically confirmed hypertrophic cardiomyopathy. Panel sequencing returns one candidate:
NM_000257.4(MYH7):c.2221G>C p.(Ala741Pro)
The p. term is parenthesised because the protein consequence is predicted, not observed. The gene is MYH7; the protein is MYH7, β-myosin heavy chain. 741 spans c.2221–2223 and falls in the converter domain. A substitution has one representation, so nothing is ambiguous here — but had the candidate been an , the transcript-level HGVS (3′-most on the transcript) and the row (left-aligned on the genome) would shift in opposite genomic directions for a minus-strand gene (Ch 41). Normalise before joining to any database.
Step 0 — is the gene–disease relationship valid, and is there a specification? MYH7 has definitive evidence for HCM. If it did not, no amount of variant-level evidence would rescue the classification. This check comes first — and it comes paired with a second one that curators skip and should not: look up whether a Variant Curation Expert Panel has published a specification for this gene before deriving anything yourself. One has. The ClinGen Cardiomyopathy Expert Panel's MYH7 specification is at v2.0, approved April 2024, and it overrides the general framework in two places below.
Step 1 — mechanism. MYH7 HCM is not a loss-of-function disease; the mechanism is , a mutant myosin incorporated into the sarcomere. So PVS1 is unavailable — and would remain unavailable even for a nonsense variant here, because is not the mechanism. BP1 (" in a gene where truncating variants cause disease") is inapplicable in the opposite direction: missense is the expected class.
Step 2 — population frequency. Absent from v4 (0/~1.6 × 10⁶ ). From §4, AF_max for HCM via MYH7 is 5.6 × 10⁻⁶, and the rule-of-three bound from absence is 1.9 × 10⁻⁶ — below it, so absence is genuinely informative here. The specification agrees by its own route: its PM2_Supporting threshold is an upper 95% bound ≤ 4 × 10⁻⁵ in the popmax group, which zero observations anywhere clears easily.
PM2_Supporting. +1 point.
Step 3 — location. Codon 741 lies in the converter domain, inside the codon 167–931 window the same specification designates as enriched for pathogenic variants and depleted of benign variation.
PM1 (Moderate). +2 points.
Note what is not also claimed. PP2 ("missense in a gene with a low rate of benign missense variation") is true of MYH7, but counting it as well double-counts one observation — precisely the naive-Bayes error §3 warns about. The specification folds it into PM1.
Step 4 — computation, and the first place the specification overrides. REVEL = 0.89. On the general ClinGen calibration of §5 that falls in [0.773, 0.932) and would earn PP3_Moderate. The MYH7 specification does not permit it: it approves PP3 at Supporting only, against a single REVEL threshold of ≥ 0.70, on the reasoning that predictors share inputs and should not be counted as independent criteria. Every higher strength is marked not applicable for this gene. Where a specification exists, it governs, and the general calibration is what you use when one does not. A splice predictor returns a near-zero delta score, so nothing is claimed from either way.
PP3 (Supporting). +1 point. Not PP3_Moderate — that is the general calibration, and this gene has its own.
Step 5 — de novo, and the second override, running the other way. Both parents are clinically unaffected on imaging, neither carries the variant, and parentage is genetically confirmed. HCM is genetically heterogeneous — a dozen sarcomere genes could produce this — so the general SVI point scale of §7 would take the heterogeneity row and score 0.5, reaching Supporting. The specification says otherwise: for most cardiomyopathies it directs the curator to default to "phenotype consistent with gene but not highly specific", which for a confirmed de novo is 1 point — Moderate — and to shift up or down only with stated clinical judgement.
PS2_Moderate. +2 points.
Notice what just happened. Two lookups in the same specification moved this case in opposite directions by the same amount, and a curator who applied the first and not the second would have reported a different classification than a curator who applied neither. Consistency about which rulebook you are in matters more than the individual calls.
Step 6 — what is unavailable, and why. PS4 (case-control excess) needs the variant seen in other cases; this one is private. PM5 needs a different pathogenic missense at codon 741; none is reported — and note that even if one were, the specification forbids counting PM5 alongside PM1, because pathogenic variants at the same codon are part of what defined the cluster PM1 is built on. That is the PP2 problem of Step 3 again, one criterion over: the same observation wearing two names. PP1 needs affected relatives; the case is sporadic and de novo, so there are no informative meioses. PS3 needs a calibrated assay; none has been run.
Step 7 — tally.
PM1 Moderate +2
PS2_Moderate Moderate +2
PM2_Supporting Supporting +1
PP3 Supporting +1
---
6 points
By the 2015 combining rules: two Moderate plus two Supporting → Likely pathogenic. By the points system: 6 points → OddsPath = 350^(6/8) = 80.9 → posterior 8.09/8.99 = 0.900 → Likely pathogenic. The two agree, and the agreement is not luck (§3).
Step 8 — read the margin honestly. This is Likely pathogenic by exactly one point. Drop either supporting criterion and it is 5 points, posterior 0.812 — a . And as Step 5 noted, applying the specification to one of PP3 and PS2 but not the other lands on 5 or 7 points instead: the call rests on two judgements that happen to cancel. A call sitting on a tier boundary should be reported as one, and is the strongest possible argument for generating more evidence.
Step 9 — what would move it. Suppose a calibrated deep mutational scan of the MYH7 motor domain now exists.
| New evidence | Points | Total | Posterior | Call |
|---|---|---|---|---|
| Assay abnormal, OddsPath ≈ 4.3 → PS3_Moderate | +2 | 8 | 0.975 | Likely pathogenic |
| Assay abnormal, OddsPath ≈ 18.7 → PS3 (Strong) | +4 | 10 | 0.994 | Pathogenic |
| Assay normal, calibrated → BS3_Moderate | −2 | 4 | 0.675 | VUS |
One experiment, three different clinical reports. The variant was identical in all three rows — which is the chapter's thesis stated as a table.
Connections
- Back to: Ch 14 — Linkage and mapping and Ch 15 — Pedigrees supply the LOD · Ch 26 — Hardy–Weinberg supplies the frequency bound · Ch 16 — Mutation supplies the de novo rate that makes PS2 strong · Ch 41 — Data formats supplies and the HGVS/ shift conflict · Ch 46 — Variant calling supplies , without which a negative result is uninterpretable · Ch 54 — Rare variants and Mendelian disease supplies the candidate this chapter classifies
- Forward to: Ch 56 — Cancer genomics — somatic variants use a different, tier-based framework built on actionability rather than causation · Ch 57 — Genomics in practice — how a classification becomes a report and a clinical decision · Ch 58 — Ethics, privacy and society — consent for secondary findings, the duty to recontact, and who is disadvantaged by an unrepresentative evidence base · D5 — SCA12 II: population, clinic, therapy — this framework meets a repeat-length variant: D5 §3 runs the counselling calculation for an class the evidence codes were never calibrated on