Variant Interpretation & ACMG Classification

Every genetic test you order — whether it's an epilepsy panel, an exome, or a genome — will come back with variants classified into one of five tiers. The reason this system matters is not bureaucratic: the tier is a compressed statement of probability, and that probability is what determines whether you are allowed to act on the result clinically.

The ACMG/AMP 2015 guidelines (Richards et al. 2015) created a standardized five-tier system: Pathogenic (P), Likely Pathogenic (LP), Variant of Uncertain Significance (VUS), Likely Benign (LB), and Benign (B). The crucial conceptual move is to read these as points on a continuous probability spectrum rather than discrete categories. Pathogenic corresponds to ≥99% posterior probability that the variant causes disease; VUS is the wide middle band (roughly 10–90%) where the evidence genuinely does not resolve. A variant does not 'become' pathogenic at a magic moment — it accumulates evidence until the probability crosses a threshold we have agreed to call actionable.

Why draw the actionable line at LP rather than only at P? Because a 90% probability of pathogenicity is already high enough that withholding management would do more harm than the residual 10% uncertainty. So both P and LP are actionable — they can inform treatment, guide surveillance, and justify cascade testing of relatives. A VUS is NOT actionable. This is the single most misunderstood point in clinical genetics: a VUS does not mean 'mildly suspicious' or 'probably the answer.' It means the lab weighed everything available and the scale did not tip. Acting on a VUS imports a coin-flip into a clinical decision while feeling like you have a molecular diagnosis — which is exactly the trap.

One important caveat that trips up trainees: for autosomal recessive conditions, a single heterozygous P/LP variant does not diagnose the disease — it identifies a carrier. The actionable content there is recurrence-risk counseling, not treatment. A diagnosis requires a second pathogenic variant on the other allele, confirmed in trans (parental or long-read phasing), because two pathogenic variants sitting on the same allele leave the patient with one fully functional copy and no disease.

Behind each tier is a structured evidence tally. The 2015 framework was originally a qualitative rulebook of combining criteria; Tavtigian et al. 2018 showed it behaves like a naive Bayesian classifier and re-expressed it as an exponentially scaled point system — Supporting/Moderate/Strong/Very Strong map to roughly ±1/±2/±4/±8 points. That insight is practically useful to you: it means evidence adds, so you can look at a VUS, see it sitting at 2 or 5 points, and ask concretely 'what single piece of evidence (usually parental testing) would push it over the line?' You don't need to memorize criterion codes to do that — you need to understand that classification is arithmetic on evidence, not a verdict handed down by the lab.

Every criterion the lab applies is really an attempt to answer one of a few orthogonal questions about a variant. Once you see the underlying question, the criterion codes stop being a list to memorize and become a logic you can reconstruct — and, crucially, anticipate which missing evidence would move the classification.

Population Frequency — the asymmetry that surprises people

The first question is simply: how often does this variant appear in people who do not have the disease? gnomAD (>250,000 exomes and genomes across ancestries) is the reference. The deep point here is an asymmetry. Common is powerful evidence against pathogenicity — a variant present at >5% in any population is benign on that fact alone (BA1), because a rare Mendelian disease cannot be caused by something carried by 1 in 20 healthy people. But the converse is weak: rare is only +1 (supporting), because the human genome is full of rare variants that do nothing. Absence from gnomAD is necessary for most Mendelian pathogenic variants but nowhere near sufficient — which is exactly why so many genuinely rare variants languish as VUS. The right intuition: high frequency closes a case; low frequency merely keeps it open.

Computational Predictions — why no single tool is allowed to win

In silico tools estimate damage from conservation, protein structure, and learned features: REVEL (ensemble, 0–1), CADD (Phred-scaled, >20 ≈ top 1% deleterious), SpliceAI (>0.5 likely splice effect). The reason the framework caps these at supporting strength and counts conflicting predictions as neutral (0) is that they are not independent measurements of biology — they are correlated models trained on overlapping data, so stacking three agreeing tools does not triple your confidence. ClinGen's calibration (Pejaver et al. 2022) put real numbers on this: REVEL ≥0.644 reaches PP3_Supporting and only at the much higher ≥0.932 does it earn Strong. That same paper supplied the benign side (BP4), so a low REVEL is genuine evidence against pathogenicity, not merely absence of evidence for it.

Gene Constraint (pLI, LOEUF) vs. positional conservation — a distinction trainees blur

Constraint asks whether the human population tolerates losing a copy of this gene: compare observed LoF variants in gnomAD to the number expected by chance. pLI >0.9 flags a gene highly intolerant to heterozygous LoF; LOEUF <0.35 (the newer, continuous metric) says the same. This is a gene-level, within-species selection signal, and it is conceptually different from cross-species conservation (PhyloP, GERP++), which asks whether one specific nucleotide has been frozen across vertebrates for millions of years. They answer different questions and can disagree: a constrained gene can contain individually unconserved positions, and a deeply conserved position can sit in a gene that otherwise tolerates LoF. The practical payoff: constraint tells you whether deleting the gene matters (relevant to PVS1 / truncating variants), while positional conservation tells you whether this particular residue matters (relevant to PP3 / missense variants). Reaching for the wrong metric is a common reasoning error.

Why constrained genes look 'empty' — a survivorship-bias trap. It is worth pausing on why a highly constrained gene carries so few variants in gnomAD, because the naive reading is exactly backwards. gnomAD is a catalog of people who are alive and largely healthy — so, like the WWII bombers that returned hit everywhere except the engine, it shows only the damage that was survivable. Loss-of-function variants are scarce in a constrained gene not because the gene rarely mutates, but because LoF hits there are strongly deleterious and are removed from the population every generation by negative selection. The 'missing' variation is absent from the living, reproducing sample, not from reality. Two consequences follow: (1) absence of LoF variation is a signal of importance, not of safety — it is exactly what makes high pLI / low LOEUF meaningful evidence for haploinsufficiency; and (2) because pathogenic LoF in these genes cannot persist in the population, it keeps re-arising as new mutations — which is precisely why de novo variants dominate severe, sporadic neurodevelopmental disease, and why trio testing (which flags a variant as new in the child) is so powerful.

De Novo Status — strong because it is improbable by chance

A variant present in the child and confirmed absent from both biological parents is among the strongest single lines of evidence (PS2, +4). The logic is purely probabilistic: each person carries only ~1–2 new coding variants, so the chance that the one spontaneous hit happens to land in the one gene that explains the child's phenotype is tiny — and it shrinks further in constrained genes. That is why confirmation matters so much: 'assumed' de novo without testing both parents is downgraded to +2 (Moderate), because non-paternity, sample mix-ups, and low-level parental mosaicism are all real ways an untested 'de novo' can be wrong.

Functional Studies — only as good as their controls

A functional assay can be strong evidence, but the framework deliberately sets a high bar: the readout must measure a disease-relevant function and the assay must be calibrated against known pathogenic and benign controls so the lab can quantify how well 'abnormal in the dish' actually predicts 'pathogenic in the patient.' An elegant experiment with no benign comparators tells you the variant changes something, not that the change causes disease — so it may not qualify at full strength.

Segregation and Phasing

Does the variant travel with disease through the family? Each additional informative affected relative who carries the variant adds evidence, scaling with the number of meioses. Phase is not optional in recessive disease: confirming two variants sit on different alleles (in trans) is what distinguishes a true compound heterozygote from a carrier who happens to have two changes on one chromosome.

Null Variants and PVS1 — the strongest criterion, and the most context-dependent

PVS1 (+8, Very Strong) applies to 'null' variants predicted to abolish the gene product: nonsense, frameshift, canonical ±1,2 splice variants, start-loss, and whole/multi-exon deletions. But its entire force rests on one premise — haploinsufficiency, that losing one working copy is enough to cause disease. That premise is not a property of the variant; it is a property of the gene, established from constraint (pLI >0.9, LOEUF <0.35), known pathogenic LoF variants in ClinVar, and ClinGen curation of the disease mechanism. So PVS1 collapses entirely in gain-of-function genes: if disease requires an overactive protein, a variant that destroys the protein may be benign or even protective. In neurogenetics this is not a footnote — ion channel genes such as SCN8A, SCN1A, and GRIN2A produce phenotypes through GoF or LoF depending on the specific variant, and a missense change can push the channel either way, so functional data may be needed to know which mechanism a novel variant invokes. The discipline is fixed: confirm the gene's established mechanism before you let PVS1 fire.

Now we apply the evidence logic to a concrete report — the kind that lands in your inbox on service. This is a representative teaching example built from real clinical patterns.

The patient: An 8-month-old boy with infantile epileptic spasms (West syndrome), developmental plateau, no family history of epilepsy, normal brain MRI, and hypsarrhythmia on EEG. The neurologist orders a comprehensive epilepsy panel (300+ genes) as a singleton — proband only, parents not included.

The report: one variant of interest — an SCN2A missense variant returned as VUS. The criteria walkthrough is below.

Trace the arithmetic, not just the verdict. The lab applied two pieces of positive evidence and nothing else fired: absent from gnomAD (PM2_Supporting, +1) and a computational prediction crossing the calibrated bar (REVEL 0.71 ≥ 0.644 → PP3_Supporting, +1). That is 2 points — solidly VUS, well short of the 6 needed for LP. Notice which criteria are blank and why: PS3 is empty because no one has run a function assay on this specific change, and PS2 is empty for a structural reason — a singleton panel physically cannot evaluate de novo status, because the parents were never sequenced.

That blank PS2 is the whole lesson. The variant is not stuck at VUS because the evidence argues against it; it is stuck because the highest-yield evidence was never collected. Order parental testing, and if the variant proves de novo, PS2 adds +4 — total 2 + 4 = 6, which just reaches Likely Pathogenic. The classification did not change because the variant changed; it changed because we gathered the one cheap piece of evidence the test design had omitted. One blood draw from each parent can move a result from 'uncertain, do not act' to 'actionable.' Whenever you see a VUS on a singleton, the reflex should be: what would a trio tell me?

The first walkthrough showed a VUS rescued toward LP by adding evidence. This one shows the more sobering case: a variant that has the strong evidence everyone wants — confirmed de novo — and still does not reach LP. It is the antidote to the reflex of treating 'de novo' as a synonym for 'diagnosis.'

The patient: A 3-year-old girl with global developmental delay, drug-resistant epilepsy (onset 4 months), no distinctive features, neurologically healthy parents, and mild cerebral atrophy on MRI. The team orders trio WES — the right call for a severe early-onset phenotype, precisely because it can detect de novo events a singleton would miss.

The report: a de novo variant in STXBP1 — a thoroughly established developmental and epileptic encephalopathy gene. And yet the lab calls it VUS. Resist the temptation to override the lab; instead, follow the math.

The tally is PS2 (+4, confirmed de novo) plus PM2_Supporting (+1, absent from gnomAD) = 5 points. Six is the LP threshold, so it lands one point short. The natural question — why didn't the computational evidence close the gap? — is the instructive part. REVEL here is 0.48, below the 0.644 calibrated bar, so PP3 does not fire; the ensemble model is simply not convinced this particular amino-acid substitution is damaging. Note the asymmetry from the evidence section in action: a strong gene with a strong inheritance signal is being held at VUS by a missing residue-level signal, because de novo evidence speaks to the gene, not to whether this exact change disrupts the protein.

The lesson is conceptual, not arithmetic. De novo is strong, not decisive. Every healthy person also carries ~1–2 de novo coding variants, most of them inert, so 'arose spontaneously' lowers but does not eliminate the prior that this is an innocent bystander. De novo + rarity buys you 5 points and an honest VUS; crossing into LP needs one more independent line — calibrated computational support, a controlled functional assay, a second unrelated patient with the same de novo change, or a gene-specific VCEP rule that reweights the evidence. The discipline is to let the variant sit at VUS until that evidence actually exists, rather than promoting it because the gene name looks familiar.

The pitfalls below are not random trivia — each one is a place where the default ACMG reasoning quietly assumes something that is false in a corner of neurogenetics. Knowing where the standard logic breaks is what separates reading a report from being fooled by one.

Gain-of-function vs. loss-of-function genes — the assumption hidden in PVS1

The entire null-variant rule assumes that losing the protein is harmful. Some neurogenetic genes break that assumption because disease there requires an abnormally active protein, not a missing one. In those genes a truncating variant — which a naive pipeline would flag as obviously damaging — may be benign or even protective, because nonsense-mediated decay simply removes the offending product. SCN8A (in some contexts), KCNQ3, and certain GRIN2A variants cause disease through GoF. The trap is that the more 'severe-looking' the variant (a clean frameshift), the more confidently it will be mishandled in a GoF gene. Reflex: before you trust a truncating call, confirm the gene's mechanism is LoF.

Repeat expansions are invisible to standard short-read sequencing

Exome — and standard short-read genome — does NOT reliably detect trinucleotide and other tandem-repeat expansions, because the expanded allele is longer than a short read and the repetitive sequence cannot be uniquely mapped. This is a structural blind spot, not a coverage problem you can fix by sequencing deeper. Friedreich ataxia, most spinocerebellar ataxias, myotonic dystrophy, Huntington disease, Fragile X, and C9orf72 ALS/FTD all live in this blind spot. The dangerous failure mode is a falsely reassuring negative: if the phenotype points to a repeat disorder, a normal exome has not excluded it — order the dedicated repeat-sizing assay (or long-read/optical methods).

Splice variants — certainty falls off a cliff just past the canonical sites

The ±1,2 intronic positions are the invariant splice dinucleotides; variants there almost always disrupt splicing and earn PVS1-level weight. Move just a few bases deeper (±3 to ±8 and beyond) and predictive certainty collapses, because those positions only sometimes matter. SpliceAI can flag likely disruption, but a prediction is not a measurement — RNA studies (RT-PCR from patient tissue, ideally the disease-relevant tissue) show whether the transcript is actually mis-spliced and are far stronger evidence. Beware tissue-specific splicing: a variant silent in blood can be pathogenic in brain.

Mosaicism can fool the variant caller and the pedigree

A post-zygotic variant present in only a fraction of cells carries a variant allele fraction below the ~50% expected for a heterozygote, so it can sink beneath standard bioinformatic filters and be missed entirely — relevant in focal cortical dysplasia and hemimegalencephaly, where the pathogenic variant may exist only in brain. The flip side fools inheritance: low-level gonadal mosaicism in a phenotypically normal parent makes a variant look de novo on a blood-based trio, yet the parent transmits it to more than one child. So when an 'apparently de novo' condition recurs in a sibling, do not conclude the first result was wrong — suspect parental gonadal mosaicism, and counsel recurrence risk accordingly.

A VUS generates anxiety precisely because it sounds like information when it is, by definition, the absence of a conclusion. The communication task is to convey genuine uncertainty without making the family feel either dismissed or doomed — and to avoid the two equal-and-opposite errors clinicians fall into.

The key message to families: 'We found a change in a gene that could be related to your child's condition, but right now the scientific evidence isn't strong enough to say for sure. We call this a variant of uncertain significance. It does not confirm a diagnosis, and it does not rule one out. We will keep watching it as new information comes in.' The framing that lands is uncertain, not bad and not fine — because both of the comforting-sounding shortcuts are clinically dangerous.

Why the two opposite errors both cause harm. Telling a family 'it's probably the cause' manufactures a diagnosis out of a coin flip: it can anchor the entire workup, stop the search for the real etiology, and invite management changes that a VUS does not justify. Telling a family 'it's nothing' is just as wrong, because a meaningful minority of VUS are later upgraded — and a family told to forget it may never return for the reanalysis that would have made the diagnosis. The honest middle is uncomfortable to deliver and is exactly the correct answer.

What to do next

• Order parental testing — usually the single highest-yield move, because (as Walkthrough 1 showed) confirmed de novo status can add the +4 that tips a VUS to LP, and because it can also do the opposite, demonstrating the variant is inherited from an unaffected parent and arguing against it.
• Document the VUS with explicit uncertainty language so a future clinician skimming the chart does not silently convert 'VUS' into 'has a mutation in gene X.' This downstream misreading is a real and recurring source of harm.
• Schedule reanalysis in 1–2 years. The stored data does not change; our interpretation of it does, as gnomAD/ClinVar grow and new gene–disease relationships are validated.
• Refer to genetic counseling for the detailed family conversation and cascade planning.

What NOT to do: never start or change treatment on a VUS alone; never collapse the uncertainty in either direction for the sake of a tidier conversation.

Reanalysis and reclassification — why uncertainty is temporary. Interpretation is a moving target. Reanalyzing previously negative or VUS-only exomes yields new diagnoses in roughly 10–25% of cases, making it one of the highest-yield, lowest-cost interventions in clinical genetics. Among VUS that do get reclassified, the majority drift toward benign (the base rate of harmlessness reasserting itself), but a meaningful fraction are upgraded — which is why the door stays open in both directions, including re-counseling families if a prior LP/P call is ever downgraded. Finally, ClinGen Variant Curation Expert Panels (VCEPs) publish gene-specific rules that replace the generic ACMG criteria for that gene, recalibrating thresholds and defining which assays count; always check for a relevant VCEP before finalizing. See the Genetic Epilepsies and Pharmacogenetics modules for how a confirmed classification then drives real treatment decisions.