Webcast recap | August 12, 2026 Darby Kammeraad, Director of Field Application Sciences · Tommy Carter, PhD, Technical Field Application Scientist
ACMG classification has been quietly rebuilt over the last eight years. The 2015 framework [1] is still the industry baseline, but a decade of ClinGen Sequence Variant Interpretation (SVI) guidance has amended nearly every criterion in it and most labs are still running classifiers that predate those amendments.
VarSeq 3.1.0 closes that gap. Here’s what changed, and what each change is worth on the bench.
Where the guidelines actually stand
| Framework | What it is | VarSeq |
|---|---|---|
| ACMG v3 · 2015 | 28 criteria at fixed strengths, PVS1–BP7, five-tier output<sup>1</sup> | Shipping today (3.0.1) |
| ACMG “v3.5” | Our shorthand for the ClinGen SVI papers amending 2015 | 3.1.0 — next release |
| ACMG v4 | Same evidence, restructured onto a continuous points-based Bayesian scale | Draft today; target early 2027 |
3.1.0 isn’t a stopgap. Calibrated annotations, configurable thresholds and point scoring are the same building blocks v4 needs. So adopting v4 later becomes a change in arithmetic, not a change in evidence.
1. Calibrated computational evidence — one score, not a tool vote
Pejaver et al. calibrated computational predictors directly against the ACMG/AMP strength tiers and showed they can support much stronger evidence than 2015 allowed [2]. VarSeq 3.1.0 acts on that: a single calibrated missense score (BayesDel by default, REVEL via dbNSFP Academic) maps straight to strength, and PP3/BP4 now reach Very Strong instead of capping at Supporting. Pick a recognized predictor and its ClinGen thresholds auto-populate.
On splicing, CI-SpliceAI fires automatically and replaces the four-algorithm consensus (MaxEntScan, NNSplice, GeneSplicer, SpliceSiteFinder). That legacy ensemble was outperformed by deep learning on every benchmark tested [3]. PP3 now carries a description of the predicted event, not just a flag.
Why one calibrated tool beats an ensemble: predictors aren’t equally performant — SIFT, PolyPhen-2 and CADD only reach supporting-to-moderate strength [2]. Conservation (GERP++, PhyloP) is no longer a separate line, because calibrated meta-predictors already incorporate it. And agreement between correlated tools was never independent evidence.
Efficiency win: one calibrated call per line of evidence, no tie-breaks to adjudicate, no aggregate filter to engineer.
2. Gene-aware population frequency
Criteria now compare against FAF95,the 95% confidence lower bound of the highest population MAF, rather than raw AF inflated by low allele counts. BA1/BS1/PM2 cutoffs are supplied per gene, inheritance-aware, with ClinGen defaults [4]. One global cutoff could never serve both common and rare disease genes.
Also: BS1 gains an opt-in Supporting band alongside Strong; PM2 drops to Supporting, rarity is a prerequisite for other evidence, not moderate evidence itself. And a curated Benign Standalone Exception list exempts known-pathogenic-but-common variants from BA1, BS1 and BS2 together [4].
Efficiency win: filter out more benign variation without over-filtering your rare disease genes.
3. Criteria and combining rules, brought into line
- PP5 and BP6 retired — ClinGen recommended discontinuation in 2018. ClinVar and consortium assertions now enter as primary evidence instead of a borrowed conclusion.
- PS1 extends to splicing — same predicted alteration as a known P/LP variant [5]. PS1 and PM5 are now mutually exclusive.
- BP7 requires BP4 first, inside Walker’s conservative boundaries: intronic at or beyond −7/+21, synonymous outside the first and last three exonic bases [5].
- PVS1 evaluated by decision tree, per the SVI loss-of-function recommendations [6] — and PP3/BP4 are withheld wherever PVS1 applies, so one LoF mechanism is never counted twice.
- New ACMG Score fields (8/4/2/1) rank evidence and report conflicts as leaning pathogenic, leaning benign, balanced or absent — not a flat “conflicting.”
Efficiency win: fewer variants stall at VUS on evidence the guidelines already permitted.
4. Expert interpretations, adopted directly
A matched variant arrives with the curators’ classification, criteria and per-criterion rationale and the classifier adopts them rather than recalculating. ClinGen ECIV is on by default; IARC TP53 and Genomenon Mastermind are available with the appropriate license, and Genomenon-CKB feeds the somatic Cancer CE criterion by exact-variant and same-codon match.
Interpretations recorded against a gene apply only to that gene’s transcripts — so at a multi-gene locus a curated call can’t override its neighbor.
Efficiency win: reuse settled expert curation instead of re-deriving it variant by variant.
5. Configurable, not one-size-fits-all
Predictors, frequency thresholds, exception lists and interpretation sources are all swappable, with ClinGen recommendations as the shipped defaults. Precedence is yours: internal catalog → curated source → Auto Classifier, and a Classification Source field records where every call originated. So a reviewed assessment is never silently overwritten.
Efficiency win: your validated pipeline drives the classifier, not the reverse.
From label to ranked worklist
This is the piece we’d flag for anyone managing throughput. Criteria are weighted 8/4/2/1 (Very Strong → Supporting), benign criteria count negative, and results report as Score Sum, Magnitude, Pathogenic and Benign.
Sum tells you the call. Magnitude tells you how much sits behind it.
Sort by Score Sum and a flat VUS pile becomes a ranked worklist: confident benigns filtered before review, and high-Magnitude/near-zero-Sum variants surfacing exactly where curator judgment is needed. ACMG v4 will turn these same points into the classification itself (≥10 Pathogenic, 6–9 Likely Pathogenic, ≤−4 Benign) — getting fluent in the number now is free preparation.
Guardrail: the score is decision support. The ACMG combining rules, not the number, set the classification.
The demo: one variant, two classifiers
DHCR7 c.964-1G>C, homozygous — a splice acceptor variant in a Smith-Lemli-Opitz case.
VarSeq 3.0.1 (ACMG v3) → BS1 from a single 1000 Genomes threshold. PP5 carrying ClinVar’s assertion as a reputable source. Splice call from the legacy 3-of-4 vote. Conservation reported alongside (GERP++ 15.9, PhyloP 6.5). Result: VUS / Conflicting. Benign and pathogenic criteria both fired with no way to net them — and PP5 was doing work the primary evidence should have done.
VarSeq 3.1.0 (ACMG “v3.5”) → BS1 recalculated on gnomAD 4.1 joint frequencies against the gene threshold. PP5 retired; PS1 applies instead, from the pathogenic match itself. CI-SpliceAI: acceptor loss 1.00, High Impact. PVS1 held at Strong by the decision tree. Result: Pathogenic — Score Sum 4, Magnitude 12. Two Strong pathogenic criteria against one Strong benign: 8 minus 4. Magnitude 12 marks it a well-evidenced call, not a thin one.
Same variant. Same patient. Same evidence sources. The criteria changed — and a case that stalled resolved.
What these updates are worth
| Prioritize, don’t scan | Ranked results put the strongest evidence at the top; review time goes to relevant variants. |
| Less reconciling by hand | One calibrated call per line of evidence, instead of four tools disagreeing and a curator adjudicating. |
| Built on current data | gnomAD 4.1, deep-learning splice models, calibrated thresholds — maintained for you. |
| Answers, not uncertainty | More variants resolve to a real call the first time. Goal: eliminate reanalysis. |
See it on your own data and request a demo or contact [email protected] and find us at ASHG 2026, Booth #525 this fall.
References
- Richards S, Aziz N, Bale S, et al. Standards and guidelines for the interpretation of sequence variants: a joint consensus recommendation of the ACMG and the AMP. Genet Med. 2015;17(5):405–424. doi:10.1038/gim.2015.30
- Pejaver V, Byrne AB, Feng B-J, et al. Calibration of computational tools for missense variant pathogenicity classification and ClinGen recommendations for PP3/BP4 criteria. Am J Hum Genet. 2022;109(12):2163–2177. doi:10.1016/j.ajhg.2022.10.013
- Fortier N, Rudy G, Scherer A. Analyzing the performance of deep learning splice prediction algorithms. PLoS One. 2026;21(5):e0348885. doi:10.1371/journal.pone.0348885
- Ghosh R, Harrison SM, Rehm HL, Plon SE, Biesecker LG; ClinGen SVI Working Group. Updated recommendation for the benign stand-alone ACMG/AMP criterion. Hum Mutat. 2018;39(11):1525–1530. doi:10.1002/humu.23642
- Walker LC, de la Hoya M, Wiggins GAR, et al.; ClinGen SVI Splicing Subgroup. Using the ACMG/AMP framework to capture evidence related to predicted and observed impact on splicing. Am J Hum Genet. 2023;110(7):1046–1067. doi:10.1016/j.ajhg.2023.06.002
- Abou Tayoun AN, Pesaran T, DiStefano MT, et al.; ClinGen SVI Working Group. Recommendations for interpreting the loss of function PVS1 ACMG/AMP variant criterion. Hum Mutat. 2018;39(11):1517–1524. doi:10.1002/humu.23626
- Brnich SE, Abou Tayoun AN, Couch FJ, et al.; ClinGen SVI Working Group. Recommendations for application of the functional evidence PS3/BS3 criterion using the ACMG/AMP sequence variant interpretation framework. Genome Med. 2020;12:3. doi:10.1186/s13073-019-0690-2
- Biesecker LG, Byrne AB, Harrison SM, et al.; ClinGen SVI Working Group. ClinGen guidance for use of the PP1/BS4 co-segregation and PP4 phenotype specificity criteria for sequence variant pathogenicity classification. Am J Hum Genet. 2024. doi:10.1016/j.ajhg.2023.11.009
- ClinGen Sequence Variant Interpretation Working Group. SVI Recommendation for De Novo Criteria (PS2 & PM6), Version 1.1. Approved March 18, 2018; updated May 5, 2021.
Questions during Webcast
- What’s your take on Google’s alpha missense? Answer: Strong signal, and we’re watching it closely but it isn’t a drop-in substitute for REVEL or BayesDel in an ACMG workflow currently, and the reason is calibration rather than accuracy.
Under ACMG “v3.5,” PP3/BP4 strength comes from calibrated score intervals published by ClinGen’s SVI (Pejaver et al. 2022). AlphaMissense isn’t among the calibrated predictors, so there are no ClinGen-endorsed thresholds mapping its scores to Supporting through Very Strong. Assigning an evidence strength without those intervals means setting your own cutoff. The practice the calibration effort was designed to replace.
There’s a technical dimension too. AlphaMissense predicts protein-level consequence, while REVEL and BayesDel are meta-predictors integrating multiple orthogonal signals, which is part of why they span a wider calibrated range across the missense spectrum.
Where it’s useful today: as a triage filter (keep it if AlphaMissense predicts damaging), as orthogonal context during manual VUS review, and as a candidate input signal for future predictor development. - Some labs submit functional evidence (PS3) in ClinVar to confirm LoF from splicing variants. Do you use this data? E.g. https://www.ncbi.nlm.nih.gov/clinvar/variation/432231/#new-submission-functional-data Answer: Honestly, we are exploring this added capability and this is a great example. RNA data confirming a splicing consequence is exactly the kind of evidence that should move a variant off the VUS pile, so we see the value clearly. One guideline note worth mentioning: the ClinGen SVI Splicing Subgroup (Walker et al. 2023) recommends capturing RNA splicing assay data through the repurposed PVS1_Strength code, reserving PS3/BS3 for assays measuring functional impact that RNA-splicing assays don’t directly capture. Today, ClinVar submissions enter VarSeq 3.1.0 as primary evidence rather than through the retired PP5 shortcut, and curated evidence comes in through configurable interpretation sources like ClinGen ECIV. Making fuller use of submitted functional data is something we want to explore further and the swappable evidence-source design in 3.1.0 gives us a natural place to put it.
