Classification Methodology
Complete documentation of how Helix Insight processes, annotates, and classifies genetic variants. Every threshold, database version, and classification rule used in production is documented on this page. This documentation is intended for clinical geneticists, laboratory directors, and accreditation auditors.
Variant classification follows the ACMG/AMP 2015 framework (Richards et al., Genetics in Medicine, 2015) implemented through the Bayesian point-based system (Tavtigian et al., Hum Mutat. 2018;39(11):1485-1492. PMID: 30311386) with BayesDel ClinGen SVI calibrated thresholds (Pejaver et al., Am J Hum Genet. 2022;109(12):2163-2177. PMID: 36413997) and SpliceAI integration aligned to ClinGen SVI 2023 recommendations (Walker et al., Am J Hum Genet. 2023;110(7):1046-1067. PMID: 37352859). Optional ClinGen VCEP gene-specific specification overlay is available for approximately 50-60 genes. Classification is strictly evidence-based. No machine learning model determines variant pathogenicity.
Contents
Pipeline Overview
Six-stage processing pipeline transforms a raw VCF file into fully classified variants. Total execution time under 15 minutes for a whole genome (~4M variants) on dedicated hardware.
VCF Parsing
~60sStandard VCF file parsed into columnar in-memory database. Multi-allelic handling, genome build detection (GRCh38 required).
Quality Filtering
~5sConfigurable quality, depth, and genotype quality thresholds. ClinVar-listed pathogenic variants are protected from filtering.
VEP Annotation
~3-4 minEnsembl Variant Effect Predictor for consequence, impact, transcript, and protein annotations. Parallel processing across chromosomes.
Reference DB Annotation
~5-10sPopulation frequencies, clinical significance, functional predictions, gene constraint, phenotype associations, and dosage sensitivity loaded from 7 reference databases.
ACMG Classification
<1sSQL-based ACMG/AMP 2015 classification using Bayesian point framework (Tavtigian et al. 2018). 19 automated criteria evaluated with calibrated evidence strength, point-based classification thresholds applied, continuous confidence scores assigned. Optional VCEP gene-specific overlay for ~50-60 genes.
Export
~5sGene-level summaries exported for streaming. Classified variants persisted to analytical database for downstream services.
Maximum Sensitivity Approach
Helix Insight classifies all variants that pass quality filtering. There is no frequency-based or impact-based pre-filtering at any stage. A common variant (e.g., gnomAD allele frequency 40%) is still classified -- it will receive a Benign classification via BA1, but it is not silently discarded before classification. This design ensures that no variant is excluded from clinical review by an automated filter. The geneticist decides clinical relevance based on the complete classification and annotation data.
Reference Databases
All reference data is stored locally on EU-based infrastructure. No variant data is sent to external APIs during processing. Database versions are fixed per deployment and documented here.
gnomAD
Population allele frequencies (global and population-specific), allele counts, homozygote counts
Used by: BA1 (allele frequency > 5%), BS1 (elevated frequency), PM2 (absent in controls), BS2 (homozygote count)
Source: gnomad.broadinstitute.org
ClinVar
Clinical significance assertions, review star levels, disease associations, submitter information
Used by: PS1 (known pathogenic), PP5 (reputable source pathogenic), BP6 (reputable source benign), ClinVar override logic, quality filter rescue
Source: ncbi.nlm.nih.gov/clinvar
dbNSFP
Functional impact predictions from multiple algorithms, conservation scores
Used by: PP3/BP4 primary tool: BayesDel_noAF with ClinGen SVI calibrated thresholds (Pejaver et al. 2022). Display predictors: SIFT, AlphaMissense, MetaSVM, DANN, PhyloP, GERP (available for clinical review, not used in classification logic)
SpliceAI
Splice impact predictions (4 delta scores: acceptor gain, acceptor loss, donor gain, donor loss)
Used by: PP3_splice (max score >= 0.2), BP4 guard (max score < 0.1), BP7 (synonymous splice check)
Source: Illumina / Ensembl
gnomAD Constraint
Gene-level constraint metrics indicating tolerance to loss-of-function and missense variation
Used by: PVS1 (LoF intolerance), PP2 (missense constraint), BP1 (LoF tolerance)
Source: gnomad.broadinstitute.org
HPO
Gene-to-phenotype associations using Human Phenotype Ontology terms
Used by: PP4 (patient phenotype matching)
Source: hpo.jax.org
ClinGen
Dosage sensitivity scores (haploinsufficiency, triplosensitivity)
Used by: BS1 (inheritance-aware frequency threshold proxy), BP2 (trans with pathogenic in recessive)
Source: clinicalgenome.org
Ensembl VEP
Variant consequence prediction, protein impact, transcript annotation, functional domain mapping
Used by: PVS1 (consequence type), PM1 (Pfam domains), PM4 (in-frame indels), BP3 (non-critical regions), BP7 (synonymous), all impact-based criteria
Source: ensembl.org
ACMG/AMP Classification
Variant classification follows the 2015 ACMG/AMP guidelines with 28 evidence criteria evaluated systematically. 19 criteria are fully automated; 9 require manual curation by the reviewing geneticist.
Classification Priority Order
Classification logic is applied in strict priority order. Higher-priority rules are evaluated first, and the first matching rule determines the final classification:
BA1 Stand-alone
Allele frequency > 5% is always classified Benign. BA1 is the only stand-alone criterion in the ACMG framework and cannot be overridden by any other evidence, including ClinVar assertions.
Conflicting Evidence
If a variant has pathogenic evidence at moderate strength or above (PVS, PS, or PM criteria triggered) AND strong benign evidence (BS criteria triggered), the variant is classified as VUS regardless of the individual evidence strength. This is a conservative approach that prioritizes clinical safety.
ClinVar Override
ClinVar classification is applied only when no conflicting computational evidence exists. Requires minimum review star level (default: 1 star). ClinVar VUS does not override computational classification.
Bayesian Point System
Each triggered criterion contributes points based on its evidence strength: Very Strong (+8), Strong (+4), Moderate (+2), Supporting (+1) for pathogenic; Strong (-4), Supporting (-1) for benign. Total points determine classification: >= 10 Pathogenic, 6-9 Likely Pathogenic, 0-5 VUS, -1 to -5 Likely Benign, <= -6 Benign. This system is mathematically equivalent to the original 18 ACMG combining rules while filling gaps for evidence combinations not explicitly covered.
Default
Variants that do not meet any of the above criteria are classified as Uncertain Significance (VUS).
Classification Output
Each variant receives one of five standard ACMG classifications, a list of all criteria that were triggered with evidence strength levels (e.g., "PVS1,PM2,PP3_Strong"), a Bayesian point total, and a continuous confidence score derived from the distance between the point total and the nearest classification boundary:
Pathogenic
>= 10 pts
confidence: 0.80-0.99
Likely Pathogenic
6-9 pts
confidence: 0.70-0.90
VUS
0-5 pts
confidence: 0.30-0.60
Likely Benign
-1 to -5 pts
confidence: 0.70-0.90
Benign
<= -6 pts
confidence: 0.80-0.99
Automated Criteria (19 of 28)
These criteria are evaluated automatically for every quality-passing variant. Exact conditions and thresholds are documented below. Each criterion lists the databases it depends on and known limitations.
Pathogenic Evidence
Null variant in gene where loss-of-function is a known disease mechanism
Conditions
Impact = HIGH
Consequence: frameshift, stop_gained, splice_acceptor, or splice_donor variant
Gene constraint: pLI > 0.9 OR LOEUF < 0.35
Exclusions
NMD-rescued transcripts (consequence contains NMD_transcript)
Stop-retained and stop-lost variants
HLA gene family (HLA-A, HLA-B, HLA-C, HLA-DRA, HLA-DRB1, HLA-DRB5, HLA-DQA1, HLA-DQB1, HLA-DPA1, HLA-DPB1, HLA-E, HLA-F, HLA-G, HLA-DMA, HLA-DMB, HLA-DOA, HLA-DOB)
Databases: VEP (consequence, impact), gnomAD Constraint (pLI, LOEUF)
Known limitations
Does not evaluate reading frame rescue via downstream in-frame reinitiation
Does not assess alternative transcript usage or tissue-specific expression
Last-exon truncation logic not implemented (all qualifying exons treated equally)
VCEP gene-specific PVS1 applicability gate available for ~50-60 genes (e.g., PVS1 disabled for gain-of-function genes like MYOC). Generic thresholds used for all other genes.
Same amino acid change as an established pathogenic variant
Conditions
ClinVar clinical significance: Pathogenic, Pathogenic/Likely_pathogenic, or Likely_pathogenic
ClinVar review stars >= 2
Databases: ClinVar (clinical_significance, review_stars)
Known limitations
Matches on exact variant position and allele, not amino acid position (PM5 amino acid-level matching is disabled)
ClinVar assertions may lag behind current evidence for recently reclassified variants
Located in a mutational hot spot or well-established functional domain
Conditions
Variant overlaps a Pfam protein domain (VEP domains annotation contains "Pfam")
Databases: VEP (domains)
Known limitations
All Pfam domains treated equally regardless of functional importance
Does not distinguish between domain core residues and peripheral positions
Non-Pfam functional domains (e.g., UniProt annotated regions) not considered
Absent from controls or at extremely low frequency in population databases
Conditions
gnomAD global allele frequency < 0.0001 (0.01%)
Frequency data must be present (non-NULL) -- variants without gnomAD data do not qualify
Databases: gnomAD v4.1 (global_af)
Known limitations
Requires gnomAD frequency data to be available for the variant position
Does not apply population-specific frequency adjustments
ClinGen SVI PM2_Supporting downgrade not implemented (full Moderate strength used)
When VCEP gene-specific specifications are enabled, PM2 threshold may differ from the generic 0.01% (e.g., 0% for RASopathy genes where any population frequency argues against pathogenicity)
Detected in trans with a pathogenic variant for recessive disorders
Conditions
Variant flagged as compound heterozygote candidate (compound_het_candidate = true)
Databases: Pipeline-internal (compound heterozygote detection)
Known limitations
Compound heterozygote status inferred from genotype data without formal phasing
Does not confirm that the trans variant is pathogenic
Trio data or long-read phasing would provide definitive confirmation
Protein length change in a non-repetitive region
Conditions
Consequence: in-frame insertion or in-frame deletion
Located within a Pfam functional domain
Not in a repetitive or low-complexity region (domains not containing "tandem", "repeat", "lowcomplexity", or "Seg")
Exclusions
HLA gene family (same exclusion list as PVS1)
Databases: VEP (consequence, domains)
Known limitations
Repetitive region detection based on VEP domain annotations only
Does not evaluate whether the in-frame change disrupts a critical functional residue
Missense variant in a gene with low rate of benign missense variation
Conditions
Consequence: missense variant
Gene constraint: pLI > 0.5
Databases: VEP (consequence), gnomAD Constraint (pLI)
Known limitations
pLI measures loss-of-function constraint, not missense constraint specifically
mis_z (missense Z-score) available but not used for this criterion
Computational evidence supports a deleterious effect (two independent paths)
Conditions
Path A (Missense -- BayesDel): BayesDel_noAF score evaluated with ClinGen SVI calibrated thresholds (Pejaver et al. 2022). Three strength levels: PP3_Strong (>= 0.518, +4 points), PP3_Moderate (0.290-0.517, +2 points), PP3_Supporting (0.130-0.289, +1 point). Scores below 0.130 are indeterminate.
PM1 + PP3 double-counting guard: When PM1 (functional domain) applies alongside PP3_Strong, PP3 is downgraded to PP3_Moderate. Combined PM1 + PP3 capped at Strong equivalent (4 points) per ClinGen SVI.
Path B (Splice evidence): SpliceAI max_score >= 0.2 AND PVS1 does not apply (ClinGen SVI 2023 double-counting guard). Always Supporting strength (+1 point).
Paths A and B are independent. A variant can trigger both if it has a BayesDel score and a splice prediction.
Exclusions
PP3_splice not applied when PVS1 is triggered (prevents double-counting loss-of-function and splice evidence per ClinGen SVI 2023)
BayesDel indeterminate range (-0.180 to 0.129): no PP3 or BP4 applied
Databases: dbNSFP 4.9c (bayesdel_noaf_score), SpliceAI (max_score)
Known limitations
BayesDel_noAF used to avoid circular reasoning with PM2/BA1/BS1 frequency criteria (Pejaver et al. 2022)
PM1 + PP3 cap assumes PM1 = Moderate (2 points). Extensible if future VCEP upgrades PM1 strength.
BayesDel does not reach PP3_Very_Strong per Pejaver calibration data
Patient phenotype is highly specific for a disease with a single genetic etiology
Conditions
Requires patient HPO terms to be provided for the analysis session
Trigger condition A: >= 3 patient HPO terms match the gene HPO profile
Trigger condition B: >= 2 patient HPO terms match AND gene has <= 5 total HPO associations (highly specific gene-phenotype relationship)
Exclusions
Not evaluated when no patient HPO terms are provided
Databases: HPO (gene-phenotype associations)
Known limitations
HPO matching is exact term overlap, not semantic similarity (ontology hierarchy not used at this stage)
Semantic similarity-based matching is performed by the downstream Phenotype Matching Service
Reputable source reports variant as pathogenic
Conditions
ClinVar clinical significance: Pathogenic, Pathogenic/Likely_pathogenic, or Likely_pathogenic
ClinVar review stars >= 1 AND < 2 (lower confidence than PS1)
Exclusions
Does not apply when PS1 already applies (prevents double-counting ClinVar evidence at different strength levels)
Databases: ClinVar (clinical_significance, review_stars)
Known limitations
ClinGen SVI has recommended retiring PP5 as a standalone criterion; retained for maximum sensitivity
Benign Evidence
Allele frequency is above 5% in population databases
Conditions
gnomAD global allele frequency > 0.05 (5%)
Databases: gnomAD v4.1 (global_af)
Known limitations
Uses global allele frequency only; population-specific BA1 thresholds not implemented
BA1 is the only stand-alone ACMG criterion. A variant meeting BA1 is classified as Benign regardless of any other evidence, including ClinVar assertions. When VCEP gene-specific specifications are enabled, the BA1 threshold may be lower than 5% for specific genes (e.g., 0.1% for Cardiomyopathy genes).
Allele frequency is greater than expected for the disorder
Conditions
Autosomal dominant proxy (ClinGen haploinsufficiency score = 3): allele frequency >= 0.001 (0.1%) AND <= 5%
Default / autosomal recessive proxy (haploinsufficiency score < 3 or NULL): allele frequency >= 0.05 (5%) AND <= 5%
Exclusions
Does not apply when BA1 applies (BA1 takes precedence)
Databases: gnomAD v4.1 (global_af), ClinGen (haploinsufficiency_score as inheritance proxy)
Known limitations
ClinGen haploinsufficiency score is a proxy for inheritance mode, not a direct determination
When VCEP gene-specific specifications are enabled, BS1 uses VCEP-defined thresholds that are already mode-of-inheritance-aware, overriding the generic AD/AR proxy logic
Observed in a healthy adult individual for a fully penetrant early-onset disorder
Conditions
gnomAD homozygote count > 15
Databases: gnomAD v4.1 (global_hom)
Known limitations
Does not distinguish between early-onset and late-onset conditions
Fixed threshold; not adjusted for disease penetrance or frequency
Missense variant in a gene for which primarily truncating variants are known to cause disease
Conditions
Consequence: missense variant
Impact: MODERATE
Gene constraint: pLI < 0.1 (gene is tolerant to loss-of-function)
pLI value must be present (non-NULL)
Databases: VEP (consequence, impact), gnomAD Constraint (pLI)
Known limitations
Low pLI used as proxy for "primarily truncating variants cause disease"
Does not directly assess whether missense variants are a known disease mechanism for the gene
Observed in trans with a pathogenic variant for a fully penetrant dominant disorder
Conditions
Compound heterozygote candidate (compound_het_candidate = true)
ClinGen haploinsufficiency score = 30 (dosage sensitivity unlikely)
Databases: Pipeline-internal (compound heterozygote detection), ClinGen (haploinsufficiency_score)
Known limitations
Trans observation inferred without formal phasing
Haploinsufficiency score of 30 is a specific ClinGen code for "dosage sensitivity unlikely"
In-frame insertion or deletion in a repetitive region without a known function
Conditions
Consequence: in-frame insertion or in-frame deletion
Located in a repetitive/low-complexity region (domains contain "tandem", "repeat", "lowcomplexity", or "Seg"), OR not in any Pfam domain, OR no domain annotation available
Databases: VEP (consequence, domains)
Known limitations
Complementary to PM4 (PM4 requires Pfam domain; BP3 requires absence of critical domain)
Computational evidence suggests no impact on gene or gene product
Conditions
BayesDel_noAF score evaluated with ClinGen SVI calibrated thresholds (Pejaver et al. 2022). Two strength levels: BP4_Moderate (<= -0.361, -2 points), BP4_Supporting (-0.360 to -0.181, -1 point).
SpliceAI max_score must be < 0.1 or absent (no predicted splice impact)
Exclusions
BayesDel indeterminate range (-0.180 to 0.129): no BP4 applied
Databases: dbNSFP 4.9c (bayesdel_noaf_score), SpliceAI (max_score)
Known limitations
BayesDel_noAF does not reach BP4_Strong per Pejaver calibration data
SpliceAI guard prevents BP4 for variants with any predicted splice impact
Reputable source reports variant as benign
Conditions
ClinVar clinical significance: Benign, Benign/Likely_benign, or Likely_benign
ClinVar review stars >= 1
Databases: ClinVar (clinical_significance, review_stars)
Known limitations
ClinGen SVI has recommended retiring BP6 as a standalone criterion; retained for maximum sensitivity
Synonymous variant with no predicted impact on splicing
Conditions
Consequence: synonymous variant
Not in a splice region (consequence does not contain "splice_region")
SpliceAI max_score <= 0.1 or absent
Databases: VEP (consequence), SpliceAI (max_score)
Known limitations
Conservation filter intentionally omitted per Walker et al. 2023 Table S13 recommendation ("no improvement in negative predictive value" with conservation filter)
Aligned with ClinGen SVI 2023 (Walker et al.) Figure 4 decision tree for synonymous variant classification.
Computational Predictors (PP3 / BP4)
PP3 and BP4 use BayesDel_noAF as the primary classification tool with ClinGen SVI calibrated thresholds (Pejaver et al. 2022). BayesDel is a Bayesian framework that integrates deleteriousness scores from multiple underlying predictors into a single calibrated score. The noAF variant (without allele frequency) is used to avoid circular reasoning with PM2, BA1, and BS1 frequency criteria.
| BayesDel_noAF Score | Evidence Strength | ACMG Code | Bayesian Points |
|---|---|---|---|
| >= 0.518 | Strong pathogenic | PP3_Strong | +4 |
| 0.290 to 0.517 | Moderate pathogenic | PP3_Moderate | +2 |
| 0.130 to 0.289 | Supporting pathogenic | PP3_Supporting | +1 |
| -0.180 to 0.129 | Indeterminate | None | 0 |
| -0.360 to -0.181 | Supporting benign | BP4_Supporting | -1 |
| <= -0.361 | Moderate benign | BP4_Moderate | -2 |
PM1 + PP3 Double-Counting Guard
When PM1 (functional domain, Moderate = 2 points) applies alongside PP3_Strong (4 points), the combined evidence would exceed the ClinGen SVI recommended cap of Strong equivalent (4 points). In this case, PP3_Strong is downgraded to PP3_Moderate (2 points), yielding a combined 2 + 2 = 4 points. PP3_Moderate and PP3_Supporting are not affected by this cap.
Why BayesDel
The ClinGen SVI Working Group calibrated four computational tools (BayesDel, MutPred2, REVEL, VEST4) and demonstrated that a single calibrated tool with evidence strength modulation provides more accurate classification than a fixed-threshold multi-predictor consensus. BayesDel_noAF was selected because it explicitly excludes allele frequency from its model (avoiding circular reasoning with PM2/BA1/BS1), is precomputed in dbNSFP 4.9c, and reaches both PP3_Strong and BP4_Moderate in the Pejaver calibration -- providing the widest evidence strength range among available tools.
Display Predictors (Not Used in Classification)
The following predictors are available alongside every variant for the reviewing geneticist to inspect, but they are not used in the PP3/BP4 classification logic. They were used in the weighted consensus approach prior to v3.4 and are retained for clinical reference:
SIFT
dbNSFP 4.9c
AlphaMissense
dbNSFP 4.9c
MetaSVM
dbNSFP 4.9c
DANN
dbNSFP 4.9c
PhyloP (100-way)
dbNSFP 4.9c
GERP++
dbNSFP 4.9c
SpliceAI Integration
Splice impact predictions are integrated following ClinGen Sequence Variant Interpretation (SVI) Working Group recommendations (Walker et al., 2023).
SpliceAI predicts the impact of each variant on mRNA splicing through four delta scores: acceptor gain (DS_AG), acceptor loss (DS_AL), donor gain (DS_DG), and donor loss (DS_DL). The maximum of these four scores is used for classification thresholds.
Scores are sourced from Ensembl precomputed MANE transcript predictions, not computed at runtime. This ensures reproducibility and avoids runtime dependencies on external services.
Thresholds
Supporting evidence for spliceogenicity. Applied as independent PP3 path.
Required for BP4 to apply. Prevents benign classification when splice impact is predicted.
Required for synonymous variant BP7 classification. Confirms no splice impact.
PVS1 Double-Counting Guard
When PVS1 (loss-of-function) is triggered for a variant, PP3_splice is not applied. This prevents double-counting the same biological mechanism (splice disruption leading to loss of function) as both PVS1 and PP3 evidence, per ClinGen SVI recommendation.
Reference: Walker LC et al. Am J Hum Genet. 2023;110(7):1046-1067. PMID: 37352859
Classification Logic
Classification uses the Bayesian point-based framework (Tavtigian et al. 2018, 2020) which is mathematically equivalent to the original 18 ACMG combining rules while providing proper classifications for evidence combinations not explicitly covered by the 2015 guidelines.
Bayesian Point System (Primary)
Each evidence criterion contributes points based on its strength level. The total determines classification:
Pathogenic Evidence Points
Benign Evidence Points
Classification Thresholds
Pathogenic
>= 10 pts
Likely Path.
6 to 9 pts
VUS
0 to 5 pts
Likely Benign
-1 to -5 pts
Benign
<= -6 pts
High-Confidence Conflict Safety Check
When pathogenic evidence at Strong or Very Strong level conflicts with Strong benign evidence (BS), the variant is flagged for manual review regardless of the point total. This prevents automated resolution of genuinely conflicting high-quality evidence.
ACMG 2015 Combining Rules (Reference)
The original 18 ACMG 2015 combining rules are a special case of the Bayesian point system -- every rule produces the same classification under both approaches. They are retained here as a reference.
Pathogenic (8 rules)
Likely Pathogenic (6 rules)
Benign (2 rules)
Likely Benign (2 rules)
Conflicting Evidence Handling
The Bayesian point system naturally handles most conflicting evidence through point summation. For example, PM2 (+2) and BS1 (-4) yield a net of -2 = Likely Benign. Under the previous v3.3 system, this combination would have been flagged as conflicting and defaulted to VUS. The point-based approach is more nuanced and clinically appropriate. However, when pathogenic evidence at Strong or Very Strong level directly conflicts with Strong benign evidence, the variant is flagged for manual review regardless of point total (see High-Confidence Conflict Safety Check above). BA1 remains a stand-alone override handled at a higher priority level.
VCEP Gene-Specific Specifications
ClinGen Variant Curation Expert Panels (VCEPs) adapt generic ACMG/AMP 2015 criteria to specific genes or diseases. Helix Insight implements approved VCEP specifications as an optional overlay on top of the standard classification.
How It Works
The standard classification pipeline runs first, producing a generic ACMG classification for all variants. For variants in genes with available VCEP specifications, gene-specific thresholds are applied via a lightweight overlay that modifies frequency cutoffs (BA1, BS1, PM2) and criterion applicability (PVS1) based on the published VCEP specification. The Bayesian point total is then recalculated with the modified criteria.
What VCEPs Modify
Gene-specific allele frequency threshold for stand-alone benign (e.g., 0.1% for Cardiomyopathy genes instead of generic 5%)
Gene-specific elevated frequency threshold, already mode-of-inheritance-aware
Gene-specific absent-in-controls threshold (e.g., 0% for RASopathy genes)
Gene-specific applicability gate. Set to FALSE for gain-of-function genes (e.g., MYOC in glaucoma)
Toggle Behavior
VCEP overlay is enabled by default and can be disabled per case in case settings. When enabled, variants in VCEP-covered genes display an audit trail marker (e.g., "[VCEP:Hearing Loss v1.0]") in the criteria string.
Coverage
Approved VCEP specifications are sourced from the ClinGen Criteria Specification Registry (CSpec). Approximately 50-60 genes have published specifications, including panels for Hearing Loss, Cardiomyopathy (MYH7, MYBPC3), RASopathy (PTPN11, BRAF, SOS1), PTEN, CDH1, TP53, PAH, and BRCA1/BRCA2 (ENIGMA). For all other genes, generic ACMG 2015 thresholds are used.
ClinVar Override Logic
ClinVar clinical significance assertions are used as classification evidence, but only under specific conditions that prevent overriding computational evidence when conflicts exist.
When ClinVar override IS applied
ClinVar has a Pathogenic, Likely Pathogenic, Benign, or Likely Benign assertion with at least 1 review star, AND no conflicting computational evidence exists (no BA1, no conflicting pathogenic+benign criteria at moderate+ strength).
When ClinVar override is NOT applied
BA1 applies (frequency > 5% always overrides ClinVar). OR conflicting evidence exists (pathogenic + benign criteria both triggered). OR ClinVar asserts VUS (VUS does not override computational classification). OR ClinVar review stars are below the minimum threshold.
When ClinVar classification is used, the criteria string includes "ClinVar" as the first element (e.g., "ClinVar,PM2,PP3") to make the evidence source explicit. ClinVar review star level is available alongside every variant for the reviewing geneticist to assess assertion quality.
Default minimum review stars for override: 1 (configurable per deployment)
Manual Review Criteria (9 of 28)
These criteria require information that cannot be determined from a single-sample VCF file -- family segregation data, functional study results, confirmed de novo status, or case-level clinical context. They must be evaluated by the reviewing geneticist.
De novo variant (confirmed paternity and maternity)
Requires trio sequencing data and confirmed parental relationships. Cannot be determined from single-sample VCF analysis.
Well-established in vitro or in vivo functional studies show a deleterious effect
Requires curation of published functional assay data. Automated literature extraction of functional evidence is not yet implemented.
Prevalence of the variant in affected individuals is significantly increased compared with controls
Requires case-control study data or odds ratios not available in standard annotation databases.
Novel missense change at an amino acid residue where a different pathogenic missense has been observed
Currently disabled. Requires normalized HGVSp matching against ClinVar at the amino acid position level. Will be enabled when ClinVar preprocessing provides standardized protein-level coordinates.
Disabled (pending implementation)Assumed de novo without confirmation of paternity and maternity
Requires family structure information not available in single-sample analysis.
Cosegregation with disease in multiple affected family members
Requires multi-generational pedigree data and segregation analysis.
Well-established in vitro or in vivo functional studies show no deleterious effect
Requires curation of published functional assay data (benign counterpart of PS3).
Lack of segregation in affected members of a family
Requires family segregation data not available in single-sample analysis.
Variant found in a case with an alternate molecular basis for disease
Requires clinical case-level information about alternative diagnoses.
Quality Filtering
Three configurable quality presets control the stringency of variant filtering. Quality filtering occurs before annotation and classification.
| Preset | Quality (QUAL) | Depth (DP) | Genotype Quality (GQ) | Recommended Use |
|---|---|---|---|---|
| Strict | >= 30 | >= 20 | >= 30 | High-confidence clinical reporting |
| Balanced | >= 20 | >= 15 | >= 20 | Standard clinical analysis (default) |
| Permissive | >= 10 | >= 10 | >= 10 | Maximum sensitivity / research |
ClinVar Rescue Mechanism
Variants with documented clinical significance in ClinVar (Pathogenic or Likely Pathogenic) that fail quality thresholds are not discarded. They are flagged as rescued variants and proceed through the classification pipeline. This prevents clinically significant findings from being silently excluded due to sequencing quality in low-coverage regions -- a deliberate design decision for clinical safety.
Limitations and Disclaimers
Helix Insight is a clinical decision support tool, not a diagnostic device. All classifications require review and confirmation by a qualified clinical geneticist.
9 of 28 ACMG criteria require information not available from single-sample VCF analysis (segregation, functional studies, de novo confirmation). These criteria must be evaluated manually by the reviewing geneticist.
SpliceAI predictions are computational. RNA splicing studies remain the gold standard for confirming splice-altering effects.
Population frequency data from gnomAD may underrepresent certain ethnic groups and geographic populations. Allele frequency thresholds should be interpreted in the context of the patient's ancestry.
ClinVar assertions vary in quality and currency. Review star levels are displayed alongside all ClinVar-derived evidence to enable informed interpretation.
Structural variants (SVs), copy number variants (CNVs), and repeat expansions are not currently classified by this pipeline.
Mitochondrial variants are processed through the same pipeline using nuclear ACMG rules as an approximation. Dedicated mitochondrial classification guidelines (e.g., MitoMap criteria) are not yet implemented.
VCEP gene-specific specifications are implemented as a threshold overlay for BA1, BS1, PM2, and PVS1 applicability. The overlay does not implement VCEP-specific functional assay interpretation (PS3/BS3) or gene-specific segregation logic (PP1/BS4), which require manual curation. Coverage is limited to approximately 50-60 genes with published ClinGen CSpec specifications; all other genes use generic ACMG 2015 thresholds.
PM5 (novel missense at known pathogenic amino acid position) is currently disabled pending standardized protein-level coordinate matching in the ClinVar preprocessing pipeline.
Compound heterozygote detection is inferred from genotype data without long-read phasing or trio analysis. Formal phasing should be performed for clinical confirmation.
Results should always be interpreted in the context of the patient's clinical presentation, family history, and other available clinical information.
Version History
Every methodology change is versioned and documented. The version number corresponds to the classification engine version in production.
ClinGen VCEP gene-specific specification overlay (optional, enabled by default)
BA1, BS1, PM2: gene-specific frequency thresholds from published VCEP specifications (~50-60 genes)
PVS1: gene-specific applicability gate (disabled for gain-of-function genes)
VCEP audit trail: criteria string includes [VCEP:Panel vX.Y] marker when gene-specific thresholds applied
VCEP toggle: can be enabled/disabled per case in case settings
Source: ClinGen Criteria Specification Registry (CSpec)
Bayesian point-based classification framework (Tavtigian et al. 2018, 2020) replaces sequential 18-rule evaluation
All 18 original ACMG combining rules produce identical results under the point system (backward compatible)
Point system fills gaps: evidence combinations not covered by original 18 rules now classified properly
PP3/BP4: BayesDel_noAF single-tool replaces weighted 6-predictor consensus (ClinGen SVI calibration, Pejaver et al. 2022)
PP3 evidence strength modulation: Strong (>= 0.518), Moderate (0.290-0.517), Supporting (0.130-0.289)
BP4 evidence strength modulation: Moderate (<= -0.361), Supporting (-0.360 to -0.181)
PM1 + PP3 double-counting guard: combined evidence capped at Strong equivalent (4 points)
Continuous confidence scores derived from point distance to classification threshold boundary
Conflicting evidence handled by point summation (more nuanced than binary VUS default)
High-confidence conflict safety check: Strong/Very Strong pathogenic vs. Strong benign flagged for manual review
Legacy predictors (SIFT, AlphaMissense, MetaSVM, DANN, PhyloP, GERP) retained as display data only
SpliceAI PP3 threshold aligned to ClinGen SVI 2023 recommendation (lowered from 0.5 to 0.2)
PP3_splice excluded when PVS1 applies (ClinGen SVI double-counting guard)
BP7 upgraded: synonymous + not splice_region + SpliceAI <= 0.1 (Walker et al. Figure 4)
BP7 conservation filter intentionally omitted per Walker et al. Table S13
Evidence strength modulation (PP3_Moderate for high SpliceAI) deferred pending VCEP specifications
SpliceAI integration: PP3_splice path for splice evidence
BP4 SpliceAI guard: requires max_score < 0.1 for benign consensus
Criteria string distinguishes PP3 (missense) from PP3_splice (splice)
Maximum sensitivity approach: removed all frequency and impact pre-filtering
All quality-passing variants proceed through classification
Clinicians decide clinical relevance using classification + annotations
Conflicting evidence priority: pathogenic + benign evidence produces VUS
BA1 stand-alone override: allele frequency > 5% always classified Benign
ClinVar override restricted to non-conflicting evidence only
SQL-based classification engine (100x performance improvement)
PM2 fix: requires non-NULL frequency data
References
Tavtigian SV, Greenblatt MS, Harrison SM, et al. Modeling the ACMG/AMP variant classification guidelines as a Bayesian classification framework.
Human Mutation. 2018;39(11):1485-1492.
PMID: 30311386Tavtigian SV, Harrison SM, Boucher KM, Biesecker LG. Fitting a naturally scaled point system to the ACMG/AMP variant classification guidelines.
Human Genetics. 2020;139(8):1057-1067.
PMID: 32666219Pejaver V, Byrne AB, Feng BJ, et al. Calibration of computational tools for missense variant pathogenicity classification and ClinGen recommendations for PP3/BP4 criteria.
American Journal of Human Genetics. 2022;109(12):2163-2177.
PMID: 36413997Richards S, Aziz N, Bale S, et al. Standards and guidelines for the interpretation of sequence variants: a joint consensus recommendation of the American College of Medical Genetics and Genomics and the Association for Molecular Pathology.
Genetics in Medicine. 2015;17(5):405-424.
PMID: 25741868Walker LC, Hoya M, Wiggins GAR, et al. Using the ACMG/AMP framework to capture evidence related to predicted and observed impact on splicing: Recommendations from the ClinGen SVI Splicing Subgroup.
American Journal of Human Genetics. 2023;110(7):1046-1067.
PMID: 37352859Jaganathan K, Kyriazopoulou Panagiotopoulou S, McRae JF, et al. Predicting Splicing from Primary Sequence with Deep Learning.
Cell. 2019;176(3):535-548.e24.
PMID: 30661751Karczewski KJ, Francioli LC, Tiao G, et al. The mutational constraint spectrum quantified from variation in 141,456 humans.
Nature. 2020;581(7809):434-443.
PMID: 32461654Cheng J, Novati G, Pan J, et al. Accurate proteome-wide missense variant effect prediction with AlphaMissense.
Science. 2023;381(6664):eadg7492.
PMID: 37733863McLaren W, Gil L, Hunt SE, et al. The Ensembl Variant Effect Predictor.
Genome Biology. 2016;17(1):122.
PMID: 27268795Landrum MJ, Lee JM, Benson M, et al. ClinVar: improving access to variant interpretations and supporting evidence.
Nucleic Acids Research. 2018;46(D1):D1062-D1067.
PMID: 29165669Questions About Our Methodology?
We welcome technical questions from clinical geneticists and laboratory directors. Transparency is foundational to clinical trust.