Bio Variant Calling Clinical Interpretation

Version Compatibility

Reference examples tested with: bcftools 1.19+, cyvcf2 0.30+, InterVar 2.2+

Before using code patterns, verify installed versions match. If versions differ:

• Python: pip show <package> then help(module.function) to check signatures
• CLI: <tool> --version then <tool> --help to confirm flags

If code throws ImportError, AttributeError, or TypeError, introspect the installed package and adapt the example to match the actual API rather than retrying.

Note: interpretation guidance evolves. Flat 2015 ACMG defaults are OUT OF DATE; verify the current ClinGen SVI recommendations and any gene-specific VCEP specification before classifying. Never apply germline ACMG to a somatic variant.

Clinical Variant Interpretation

"Classify this variant / write the ACMG rationale" -> Assemble independent, calibrated evidence lines and combine them under the correct framework for a pinned (gene, transcript, disease) context.

• Germline Mendelian: ACMG/AMP + ClinGen SVI refinements (below).
• Somatic/tumor: AMP/ASCO/CAP tiers + ClinGen/CGC/VICC oncogenicity (never ACMG).

The governing principle

A variant's clinical significance is NOT a database lookup. It is a Bayesian sum of INDEPENDENT, CALIBRATED evidence relative to a pinned context (genome build, MANE Select transcript, gene disease-mechanism, disease prevalence, framework version). Three traps sink most naive pipelines:

1. Flat 2015 defaults are obsolete. A current classifier applies the ClinGen SVI refinements (graded PVS1, PM2 downgraded, PP5/BP6 retired, calibrated PP3/BP4, Bayesian points). Using the raw 2015 combining rules is a known error.
2. Germline and somatic are different questions with different frameworks. "Does this cause a Mendelian disorder" (ACMG) vs "is this an actionable/oncogenic tumor variant" (Li tiers, Horak oncogenicity). Applying ACMG to a somatic variant is a category error.
3. A ClinVar assertion is a LEAD, not evidence. Concordance between submitters is not independence; 1-star is not usable; re-derive from the underlying data.

Pick the framework FIRST

Before applying generic ACMG, check for a ClinGen Variant Curation Expert Panel (VCEP) specification for the gene (e.g. hearing loss, RASopathy, cardiomyopathy, ENIGMA BRCA1/2). A VCEP spec reweights and constrains criteria and OVERRIDES generic defaults; a 3-star ClinVar assertion often reflects one.

ACMG/AMP germline: the 2015 baseline and its mandatory refinements

The 2015 consensus (Richards 2015 Genet Med 17:405-424) defines five tiers (Pathogenic, Likely Pathogenic, VUS, Likely Benign, Benign) and 28 coded criteria at default strengths: PVS1 (very strong), PS1-4 (strong), PM1-6 (moderate), PP1-5 (supporting); BA1 (stand-alone), BS1-4 (strong), BP1-7 (supporting). A director does NOT interpret with raw 2015 anymore. Apply these ClinGen SVI corrections:

Bayesian points system (Tavtigian 2020 Hum Mutat 41:1734)

Goal: Combine graded evidence into a tier reproducibly instead of matching verbal rule patterns.

Approach: Assign each met criterion a point value by strength (benign subtracts), sum, and threshold. This underlies the emerging points-based ACMG/AMP/CAP/ClinGen overhaul, so prefer it over the 2015 verbal table.

Classification by summed points: Pathogenic >= 10, Likely Pathogenic 6-9, VUS 0-5, Likely Benign -1 to -6, Benign <= -7 (confirm the exact benign cutpoints against Tavtigian 2020 before hard-coding). Benign criteria (BA1/BS/BP) contribute negative points at the same magnitudes.

PVS1 decision tree and the NMD 50-nt rule

Goal: Assign PVS1 the CORRECT strength for a null variant instead of firing it on any "HIGH impact" call.

Approach: Route by gene LOF mechanism, then variant type, then NMD prediction and exon location (Abou Tayoun 2018). Evaluate on the MANE Select transcript, not whichever isoform maximizes severity.

• Gene mechanism gate: PVS1 applies ONLY where loss of function is the established disease mechanism (haploinsufficiency). For gain-of-function / dominant-negative genes a null may be benign -- PVS1 must not fire.
• NMD 50-55 nt rule: a premature termination codon >~50-55 nt upstream of the last exon-exon junction triggers nonsense-mediated decay (true LOF -> full strength). A PTC in the LAST exon, within ~50 nt of the final junction, or in a single-exon gene ESCAPES NMD -- protein is made; downgrade PVS1 (Strong/Moderate/Supporting) by how much functional protein / which domains are lost.
• Transcript relevance: confirm the affected exon is in biologically expressed transcripts; a canonical-splice change in a minor non-expressed isoform is not PVS1.
• "HIGH impact stop_gained" from SnpEff/ANNOVAR is NOT PVS1 -- impact buckets know nothing about NMD or mechanism. Evaluate PVS1 on MANE Select; do not use the worst-consequence transcript. See variant-calling/variant-annotation.

ClinVar: assertions are leads, not evidence

"Look up this variant in ClinVar" -> Read WHO submitted, at what review status, on WHAT evidence -- then re-derive, do not adopt the conclusion.

Rules: 1-star / no-criteria is not evidence. Conflicting interpretations flag genuinely hard variants (penetrance, ancestry, mechanism) -- investigate, do not average. Concordance is not independence (two submitters can copy one original error). PP5/BP6 are retired precisely because an assertion cannot be an evidence input.

Annotate and read ClinVar fields (bcftools / cyvcf2)

Goal: Attach ClinVar assertions as LEADS and surface review status alongside significance.

Approach: Annotate CLNSIG/CLNDN/CLNREVSTAT from the ClinVar VCF, then always carry CLNREVSTAT so a 1-star call is never mistaken for evidence. Download the build-matched ClinVar VCF first (usage-guide.md).

bcftools annotate -a clinvar.vcf.gz \
    -c INFO/CLNSIG,INFO/CLNDN,INFO/CLNREVSTAT input.vcf.gz -Oz -o with_clinvar.vcf.gz

# Surface P/LP leads WITH their review status (never drop CLNREVSTAT)
bcftools view -i 'INFO/CLNSIG~"athogenic"' with_clinvar.vcf.gz \
  | bcftools query -f '%CHROM:%POS %REF>%ALT\t%INFO/CLNSIG\t%INFO/CLNREVSTAT\n'

Population frequency: grpmax filtering-AF, not a global cutoff

Goal: Decide BA1/BS1 (or PM2_Supporting) correctly for THIS disease, not with a universal 1% line.

Approach: Compare the gnomAD grpmax filtering allele frequency to the maximum credible population AF derived from disease prevalence, heterogeneity, inheritance and penetrance (Whiffin 2017 Genet Med 19:1151). A flat cutoff is wrong in both directions.

• Filtering AF (FAF) is the LOWER bound of the 95% CI of the grpmax (genetic-ancestry-group max) AF -- gnomAD v4 exposes it as the fafmax_faf95_max INFO field (fafmax_faf95_max_joint in the joint exome+genome VCF). Using grpmax, not global AF, avoids diluting a variant common in one ancestry across the whole cohort; using the CI lower bound guards against a noisy small-subpopulation estimate.
• Rule: if FAF > the disease's maximum credible population AF, apply BA1/BS1. This is per-disease.
• Presence in gnomAD is NOT benign. Exceptions a director watches for: recessive carriers are healthy (pathogenic alleles sit at carrier frequency, e.g. CFTR); late-onset / reduced-penetrance alleles appear in adult cohorts (BRCA, Lynch); somatic / clonal-hematopoiesis contamination leaks low-AF calls in DNMT3A/TET2; artifacts in homopolymer/segdup regions -- respect gnomAD PASS/quality flags, not raw AF.
• gnomAD ancestry groups are unevenly sampled, so "absent" is much weaker evidence for an under-represented ancestry; PM2/BS1 strength is implicitly ancestry-dependent. gnomAD v2.1.1 is GRCh37 (Karczewski 2020 Nature 581:434); v3/v4 are GRCh38 -- never eyeball "absent" across builds without liftover.

# Illustrative: filter on a grpmax filtering-AF field, keeping absent sites (annotation-dependent)
bcftools view -i 'INFO/fafmax_faf95_max<0.0001 || INFO/fafmax_faf95_max="."' \
    input.vcf.gz -Oz -o faf_filtered.vcf.gz

Pathogenicity predictors: ONE, calibrated

Goal: Convert a computational score into PP3/BP4 at a defensible strength without double-counting.

Approach: Pick ONE predictor that reached >= Strong in the ClinGen calibration and apply it at its calibrated threshold (Pejaver 2022). Stacking correlated tools fakes independence and silently over-calls pathogenic.

• PP3 and BP4 are graded (Supporting/Moderate/Strong) and mutually exclusive. For REVEL (Ioannidis 2016 AJHG 99:877) the well-reproduced SUPPORTING thresholds are PP3 >= 0.644 and BP4 <= 0.290; higher-strength (Moderate/Strong) cutoffs exist -- read them from the Pejaver 2022 supplement or the current ClinGen SVI table rather than hard-coding.
• Use only ONE tool. REVEL is an ensemble of 13 scores (incl. SIFT, PolyPhen), so "REVEL agrees with PolyPhen" is not corroboration -- PolyPhen is INSIDE REVEL.
• SIFT and PolyPhen-2 did not reach even Supporting in the calibration -- a "damaging" call is decorative, not evidence. Raw CADD did not reach Supporting for PP3 (CADD>=20 calibrated to benign-Moderate -- mild evidence AGAINST missense pathogenicity, the opposite of how it is usually invoked); CADD is for genome-wide/non-coding ranking, not missense PP3.
• AlphaMissense (Cheng 2023 Science 381:eadg7492) is proteome-wide and not trained on ClinVar labels, but its developer class cutoffs are NOT ACMG strengths -- check the current ClinGen SVI tool list for its calibrated PP3/BP4 thresholds before assigning a strength.
• Splicing (SpliceAI, Jaganathan 2019 Cell 176:535): delta scores 0-1, developer guidance 0.2 recall / 0.5 recommended / 0.8 precision. A high delta is a PREDICTION; converting it to PS3/PP3 strength needs the ClinGen splicing calibration, and the default scoring window is narrow -- widen it (deep-intronic/pseudoexon variants are otherwise missed). SpliceAI does not report the mis-splicing OUTCOME (exon skip vs intron retention), which determines PVS1 applicability.