SeqOne is extending its DiagAI Score to include AI-powered prioritization for CNVs and SVs. This update combines a dedicated UP²-SV pathogenicity model with a complete clinical interpretation workflow designed specifically for large variants.

Why Structural Variants Need Their Own AI

Recently, we shared how DiagAI, our proprietary explainable AI-powered variant prioritization tool, reached 98% sensitivity in capturing causal variants within the Top-10 candidates. This validation, conducted on a 565-patient cohort from the Genomics England 100,000 Genomes Project, sets a new bar for SNV prioritization in WGS. However, clinicians know the real story does not stop at small variants.

Copy Number Variants (CNVs) and Structural Variants (SVs) are responsible for a substantial fraction of rare disease diagnoses, yet they remain among the hardest variants to interpret. Each WES, WGS, or long-read case can surface hundreds of structural events. Rule-based ACMG/ClinGen frameworks, while rigorous, were not designed to discriminate the truly pathogenic ones at the speed clinical labs require.

With our Spring 2026 Release, DiagAI extends its clinical-grade prioritization to the world of CNVs and SVs.

‍

Introducing UP²-SV: A Dedicated AI Score for Structural Variants

UP²-SV is the structural-variant counterpart of UP², our XGBoost-based pathogenicity classifier for SNVs. Available in SeqOne’s GermVar suite, it produces a continuous score from −1 (strongly benign) to +1 (strongly pathogenic). This value is normalized into the DiagAI CNV score for sorting and prioritization, matching the exact workflow UP² provides for SNVs.

‍

Under the hood: DiagAI UP2-SV

Extensive training

Trained on over 65,000 SV-labeled ClinVar variants.

Rich data integration

30 engineered features spanning multiple biological dimensions, integrating ClinVar, gnomAD SV, ClinGen (HI/TS), DECIPHER, OMIM, DGV, RepeatMasker, and the Collins Dosage Dataset.

Comprehensive metrics

Captures gene content, dosage sensitivity (pHaplo/pTriplo), exon disruption, repetitive-region overlap, and population frequency.

Advanced prediction

Predicts across three classes (Benign, VUS, and Pathogenic) using gradient boosting, with the same explainability mechanisms as UP² for SNVs.

‍

Validation setting Dataset N Metric Result Interpretation
Held-out benchmark ClinVar SV/CNV 11,319 3-class accuracy 80.1% Measures agreement with ClinVar labels across all classes.
External label benchmark DECIPHER 23,085 3-class agreement 66.7% Lower agreement is expected as DECIPHER labels are often phenotype-dependent.
Clinical diagnostic cohort Internal Cohort1 42 pathogenic SVs Pathogenic recall 97.6% 41/42 known pathogenic SVs were correctly surfaced by UP²-SV.

1 Ravel, Jean-Marie, et al. "Clinical utility of periodic reinterpretation of CNVs of uncertain significance: an 8-year retrospective study." Genome Medicine 15.1 (2023): 39.

‍

Metric note: ClinVar and DECIPHER results measure three-class label agreement across benign, VUS, and pathogenic variants. The clinical cohort result measures pathogenic recall, which is the ability to retain known diagnostic SVs in a clinical interpretation workflow. These metrics answer different questions and should not be directly compared.

DECIPHER agreement is expected to be lower because SV labels are often dependent on phenotype and context. Similar or overlapping variants may be interpreted differently depending on the patient’s clinical presentation, inheritance, and penetrance. UP²-SV evaluates variant-level evidence rather than the full patient-specific context. Therefore, the clinical cohort result is particularly important: UP²-SV retained 41 out of 42 known pathogenic structural variants, supporting its role as a reliable prioritization layer for clinical interpretation.

‍

Why It Matters for Clinical Labs

For exomes, genomes, and long-read analyses, the bottleneck is no longer detection. Instead, the challenge is now variant triage and interpretation.

Combined with our newly retrained UP² for SNVs (now trained on 3M+ variants) and Phenogenius v3.1 (expanded to 6,181 genes and 19,409 HPO terms), the platform now sees more, prioritizes better, and moves faster across both small and large variants.

‍

A Complete Interpretation Experience for CNVs and SVs

UP²-SV does not arrive in isolation; it is part of a fundamentally rebuilt interpretation environment for large variants.

Everything labs already value in the SeqOne SNV workflow has been natively integrated for structural variants:

  • DiagAI-powered sorting and prioritization for CNVs and SVs.
  • Interactive ACMG CNV scoring where you can toggle criteria and watch classifications update in real time.
  • Purpose-built SeqOne Browser for CNV/SV exploration featuring cytoband navigation, event tracks, RefSeq genes, copy number tracks, BAF plots, and DGV.
  • Entity-wide sample frequency for CNV & SV with germline, somatic, and aCGH splits.
  • Unified Compound Heterozygosity views combining small and large variants.
  • The full SNV toolkit applied to large variants, including the column manager, filter presets, bulk actions, and the "Mark as Viewed" feature.


The full Spring 2026 Release is live, with 72+ features and improvements, IVDR-certified from day one. Explore everything at www.seqone.com/releases/spring-2026.

‍

‍

Go Deeper: Read the DiagAI CNV White Paper (2026)

The full methodology, feature engineering, validation cohorts, and a deep-dive CNV case are detailed in our brand-new DiagAI Germline White Paper (2026).

‍

‍

Bringing the same AI-driven sorting and prioritization that accelerates SNV interpretation to deletions, duplications, and beyond. That is DiagAI for structural variants.

‍