Skip to content

How to Interpret Your Own Genetic Variants

Oak
A stack of glass laboratory modules through which a glowing DNA strand passes, gaining small beads and sorting into three lit trays.

If you are looking for one genetic variant interpretation tool, the practical answer is that no single tool does the job, and the ones marketed that way are doing one of three separable tasks. The three tasks are annotation (what does this variant touch), evidence aggregation (what do population and clinical databases say about it), and classification (how do the ACMG/AMP rules score that evidence). We run Ensembl VEP for annotation, gnomAD plus ClinVar for evidence, and InterVar or GeneBe for a first-pass classification. Every step before that depends on the variant call being real, which is the part most people skip.

Interpretation starts with whether the variant exists

A variant classification is only as good as the genotype underneath it. A whole-genome sequencing run at 30x mean depth still produces thousands of calls that are artifacts of alignment. Those calls do not reflect biology. Before you ask what a variant means, check that the call is supported.

The fast filters are depth and genotype quality. In a germline VCF from GATK or DeepVariant, we ignore anything with DP below 10 or GQ below 20 unless we are willing to look at the reads. For heterozygous single-nucleotide variants, the allele balance should sit near 0.5, and calls below roughly 0.25 are usually mapping noise or contamination.

Region matters more than most people expect. Segmental duplications and homopolymer runs generate reproducible false positives on short reads. The same is true of the pseudogene-bearing regions around genes like PMS2, SMN1, and CYP21A2. The Genome in a Bottle high-confidence BED files give you a callable mask, and hap.py against HG002 tells you what your pipeline’s error profile looks like before you trust it on yourself. This concern about study design and call quality preceding interpretation is old and well documented in the sequencing literature.1

Normalize and name the variant before anything else

Two VCF lines can describe the same change and fail to match ClinVar because of left-alignment or multi-allelic packing. Normalize first:

bcftools norm -m -any -f GRCh38.fa -Ou in.vcf.gz \
  | bcftools norm -d exact -Oz -o norm.vcf.gz
bcftools index norm.vcf.gz

The first call splits multi-allelic sites and left-aligns indels against the reference. The second removes exact duplicates. Skip this and your indel annotations will silently disagree with every database you query.

Naming is the second half of the problem. HGVS nomenclature is transcript-dependent, so c.1521_1523delCTT means nothing without the transcript accession and version. VariantValidator checks that a description is syntactically legal and that it maps consistently between transcript and genome coordinates.2 We route any variant we intend to discuss with a clinician through it, because a mistyped HGVS string is the most common way a conversation about a variant goes wrong.

Annotation engines: what we run and why

Ensembl VEP is our default. It handles GRCh38 offline with a cache, supports plugins for the predictors you want, and its --pick_allele_gene flag gives one consequence per gene per allele instead of a combinatorial explosion of transcripts.

vep -i norm.vcf.gz --cache --offline --assembly GRCh38 \
  --everything --pick_allele_gene --vcf --compress_output bgzip \
  --plugin SpliceAI,snv=spliceai_snv.vcf.gz,indel=spliceai_indel.vcf.gz \
  --plugin AlphaMissense,file=AlphaMissense_hg38.tsv.gz \
  --custom clinvar.vcf.gz,ClinVar,vcf,exact,0,CLNSIG,CLNREVSTAT \
  -o vep.vcf.gz

ANNOVAR is faster and produces flatter tables, which makes it convenient upstream of InterVar. SnpEff is a reasonable third option and pairs well with SnpSift filtering. The differences between them are mostly transcript-set differences and consequence-ranking differences, and they can change the reported consequence of the same variant. Surveys of the variant analysis toolchain have documented this fragmentation for over a decade.3 Toolkits built around annotate-then-query workflows, such as variant tools, made the same argument early.4

Two database choices carry most of the weight. Use gnomAD v4 for allele frequency, and read the ancestry-specific frequencies rather than the global one. A variant at 3% in one population and absent elsewhere is a different object from one at 0.001% everywhere. Use ClinVar with its review status, because a one-star single-submitter assertion and a three-star expert-panel assertion are not comparable evidence.

Classifiers and what they can and cannot do

The ACMG/AMP framework assigns a variant to one of five tiers using evidence codes such as PVS1 for predicted loss of function and PM2 for absence from population databases. InterVar automates the codes that can be computed from data and leaves the rest for manual review.5 That split is the structure of the problem, and any tool claiming full automation is overreaching.

GeneBe and Franklin are the convenient web front-ends for the same logic, with evergreen annotation sources. ELLA is built for clinical labs that need reproducibility and an audit trail. For a handful of genes, ClinGen Variant Curation Expert Panels have published gene-specific rule specifications that override the generic thresholds, and those specifications should win when they exist.

A recent review of automated interpretation found meaningful disagreement between tools on the same variant, driven by differing evidence thresholds and differing source databases.6 Treat a classifier’s output as a structured hypothesis with its evidence codes attached, and read the codes rather than the label.

In silico predictors and their numbers

Missense and splice predictors are the most useful and most misused part of the stack. SpliceAI outputs four delta scores per variant, and the authors’ suggested cutoffs run from 0.2 for high recall to 0.8 for high precision. AlphaMissense returns a pathogenicity probability with published benign and pathogenic thresholds. Both are evidence. Neither is a verdict.

Under ACMG/AMP, computational evidence is supporting strength in most rule specifications, and the ClinGen sequence variant interpretation group has recommended treating multiple correlated predictors as one piece of evidence rather than several. Benchmarking work on in silico predictors shows that performance varies substantially across genes and variant classes, and that most predictors were trained on data overlapping the benchmarks used to evaluate them.7 Use one missense predictor and one splice predictor, and do not stack five of them to manufacture confidence.

The pipeline we would run on a personal genome

Start with a normalized, annotated VCF as above. Filter to the callable regions, then to variants with a coding or splice-region consequence. Apply a frequency ceiling appropriate to what you are asking. A 5% allele frequency in any continental population is a standalone benign criterion under the ACMG rules, and for rare dominant conditions we filter far lower.

Then split the output. The first bucket is ClinVar pathogenic or likely pathogenic with two stars or better. That bucket is short and worth careful reading. The second bucket is predicted loss of function in genes with a known loss-of-function mechanism, which is where you check that the transcript is the biologically relevant one and that the variant is not in a terminal exon that escapes nonsense-mediated decay. The third bucket is everything else, which is mostly variants of uncertain significance and should stay that way.

A classification is not a clinical result. A finding you intend to act on needs orthogonal confirmation in a CLIA-certified laboratory and interpretation by a clinical geneticist or genetic counselor, who will weigh your personal and family history alongside the variant. Even in well-studied single-gene conditions, genotype predicts phenotype imperfectly, and the same variant can behave differently across individuals.8 The broader lesson from population-scale sequencing is that most human variation is currently uninterpretable and should be labeled as such.9

Questions people also ask

Is InterVar enough on its own? It is a good first pass and it shows its work by listing which ACMG codes fired. It cannot evaluate segregation, functional assay data, or case-level evidence. Criteria of that kind are exactly what move a variant of uncertain significance off the fence.5

Which reference build should I use? GRCh38, and preferably a build with alt contigs handled explicitly. If a database you need is GRCh37-only, lift over with CrossMap and expect a small percentage of sites to fail or flip strand.

Why do two tools disagree about my variant? Usually because they picked different transcripts, used different ClinVar snapshots, or set different frequency thresholds for the PM2 criterion. Comparative reviews of automated interpretation report this as the normal state rather than the exception.6

Can I interpret variants from a consumer genotyping array? Poorly. Arrays genotype selected sites and impute the rest, so rare variants are either absent or imputed with low confidence. Rare variant interpretation needs sequencing.

What about polygenic scores? They are a different kind of measurement with different assumptions, and they are not classified under ACMG/AMP rules. Do not mix a polygenic score and a monogenic variant classification in one conclusion.

Footnotes

  1. David B. Goldstein, Andrew Allen, Jonathan Keebler, et al. Sequencing studies in human genetics: design and interpretation. Nature Reviews Genetics, 2013. https://doi.org/10.1038/nrg3455 ↩

  2. Peter J. Freeman, Reece K. Hart, Liam J. Gretton, et al. VariantValidator: Accurate validation, mapping, and formatting of sequence variation descriptions. Human Mutation, 2017. https://doi.org/10.1002/humu.23348 ↩

  3. S. Pabinger, A. Dander, M. Fischer, et al. A survey of tools for variant analysis of next-generation genome sequencing data. Briefings in Bioinformatics, 2013. https://doi.org/10.1093/bib/bbs086 ↩

  4. F. Anthony San Lucas, Gao Wang, Paul Scheet, et al. Integrated annotation and analysis of genetic variants from next-generation sequencing studies with variant tools. Bioinformatics, 2011. https://doi.org/10.1093/bioinformatics/btr667 ↩

  5. Quan Li, Kai Wang. InterVar: Clinical Interpretation of Genetic Variants by the 2015 ACMG-AMP Guidelines. The American Journal of Human Genetics, 2017. https://doi.org/10.1016/j.ajhg.2017.01.004 ↩ ↩2

  6. Mireia Costa, Alberto García S, Ana León, et al. The promises and pitfalls of automated variant interpretation: a comprehensive review. Briefings in Bioinformatics, 2025. https://doi.org/10.1093/bib/bbaf545 ↩ ↩2

  7. Panagiotis Katsonis, Kevin Wilhelm, Amanda Williams, et al. Genome interpretation using in silico predictors of variant impact. Human Genetics, 2022. https://doi.org/10.1007/s00439-022-02457-6 ↩

  8. Patrick R Sosnay, Garry R Cutting. Interpretation of genetic variants. Thorax, 2013. https://doi.org/10.1136/thoraxjnl-2013-204903 ↩

  9. Tuuli Lappalainen, Alexandra J. Scott, Margot Brandt, et al. Genomic Analysis in the Age of Human Genome Sequencing. Cell, 2019. https://doi.org/10.1016/j.cell.2019.02.032 ↩

Pothos builds longitudinal molecular profiles of individuals: whole-genome sequencing, single-cell RNA sequencing, and plasma proteomics, delivered as a detailed report, every data file, and stored cells. Build your profile.