Genetic Screening: What It Tests, What It Misses, and How to Run It on Your Own Data

Genetic screening is looking for variants in someone who has no symptoms and no particular reason for suspicion. Genetic testing, in the narrow sense, is looking for a cause in someone who already has a phenotype. The distinction matters because it changes the math: in a symptomatic patient the prior probability of a pathogenic variant is high, so a positive result is usually real. In an asymptomatic person the prior is low, so a large fraction of “positive” screening results are false positives, variants of uncertain significance, or real variants with far lower penetrance than the textbook number suggests. If you are technically fluent and want to screen yourself, there is a useful version of this. Get a whole genome at 30x or better, and get the BAM and the gVCF. Run a defined gene list against curated evidence. Treat every hit as a hypothesis that needs a CLIA-grade confirmation test ordered by a clinician before it means anything about you.

The three types, and which one you are doing

The standard taxonomy is newborn screening, carrier screening, and population or predictive screening in adults. Newborn screening is the biochemical heel-stick panel plus, increasingly, sequencing. Carrier screening looks for recessive variants that matter for reproductive planning, not for your own health (the “3-gene” panels people search for are typically cystic fibrosis, spinal muscular atrophy, and fragile X, sometimes with hemoglobinopathies added). Predictive screening in adults is the category most engineers mean: am I carrying a pathogenic variant in a cancer or cardiac gene that will matter in twenty years.

Sequencing an apparently healthy person produces findings at a rate that surprises people. The NC NEXUS trial sequenced healthy newborns and found childhood-onset actionable conditions in a small but non-zero fraction of well infants, alongside a much higher rate of carrier status and uncertain results, which is exactly the yield profile you should expect for yourself.¹ The European Society of Human Genetics laid out the governance problem early. Once you sequence the whole genome, you have generated answers to questions nobody asked. The decision about which of those to return is a matter of policy. That decision is not a technical one.² For your own data, you are the policy.

What the files are

A useful screening dataset is four artifacts:

• sample.bam or sample.cram aligned to GRCh38 (ask for the no-alt analysis set, and ask whether they used alt contigs, because HLA and KIR calls depend on it)
• sample.g.vcf.gz plus index, so you can distinguish “reference” from “no coverage”
• A per-base or per-interval coverage file, mosdepth output is fine
• The FASTQs, if you want to re-align later

Insist on the gVCF. A plain VCF only lists variants. It cannot tell you whether the reason you have no BRCA1 finding is that you are reference there or that the region had two reads. Missing that difference is the single most common misreading of consumer genomic data.

The pipeline we would run

Start with a defined list. We use the ACMG secondary findings list as the core, because it was selected specifically for actionability in asymptomatic people. It holds about 80 genes, mostly for hereditary cancer syndromes, cardiomyopathies, arrhythmias and aortopathies. Malignant hyperthermia and familial hypercholesterolemia are on it as well. Pull the BED from a current gene annotation and check coverage first:

mosdepth --by acmg_sf.bed --thresholds 10,20 sample sample.bam
awk '$7/$4 < 0.95' sample.thresholds.bed.gz   # exons <95% at 20x

If exon 1 of a GC-rich gene drops out, you will see it here. Then annotate:

bcftools view -R acmg_sf.bed sample.vcf.gz -Oz -o acmg.vcf.gz
vep -i acmg.vcf.gz --cache --assembly GRCh38 --everything \
    --plugin SpliceAI,snv=spliceai_snv.vcf.gz,indel=spliceai_indel.vcf.gz \
    --custom clinvar.vcf.gz,ClinVar,vcf,exact,0,CLNSIG,CLNREVSTAT \
    --vcf -o acmg.vep.vcf

Filter on ClinVar CLNSIG of Pathogenic or Likely_pathogenic with CLNREVSTAT of at least two submitters with no conflicts. Filter separately on loss-of-function consequences (stop_gained, frameshift, canonical splice) with gnomAD v4 allele frequency below roughly 1e-4. Keep those two lists apart. The ClinVar list is curated evidence. The LoF list is your own inference and carries a much higher false discovery rate, particularly for genes where loss of function is not the disease mechanism, and for variants in final exons that escape nonsense-mediated decay.

Then the things short-read callers get wrong, which you must handle separately or explicitly declare out of scope:

• Repeat expansions: run ExpansionHunter with the standard catalog for HTT, FMR1, C9orf72, DMPK and ATXN1-3. Expect poor sensitivity for expansions longer than the read insert.
• PMS2 exons 11-15 and SMN1/SMN2 have near-identical paralogs. Standard alignment mismaps them. Use a paralog-aware caller or accept a blind spot.
• CNVs and structural variants: Manta plus GATK gCNV or CNVkit. A single-exon deletion in BRCA1 is a real screening finding that SNV calling never sees.
• Pharmacogenomics: PharmCAT on the VCF gives you CPIC-style diplotypes, but CYP2D6 needs PyPGx or an equivalent copy-number-aware caller.

None of this is exotic. It is a weekend of work if you have run bioinformatics tooling before, and a week if you have not.

What a positive result means for you

A hit in your VCF is a hypothesis with three failure modes. First, a technical artifact: check the read pileup in IGV before anything else, and look at allele balance, mapping quality, and whether the variant sits in a homopolymer. Second, a classification that has moved: ClinVar entries get reclassified, and a variant called pathogenic in 2016 may be benign now. Always query a current release, and re-run your list annually. Third, and the most important one, penetrance. Estimates for most of these genes come from families ascertained because they were full of disease. In population cohorts the same variants carry substantially lower lifetime risk. A pathogenic BRCA2 variant found by screening a healthy person does not carry the risk quoted from a high-risk clinic cohort.

If you get a hit in a medically actionable gene, take it to a clinical geneticist or genetic counselor and have it confirmed in a CLIA-certified laboratory before acting on it or telling family members. Research-grade sequencing is not diagnostic, and the clinical decision about surveillance belongs to a physician who knows your history.

The contrast with diagnostic testing is worth holding in your head. In symptomatic pediatric patients, whole-genome sequencing reached a diagnostic yield of roughly a third. That was several times what chromosomal microarray delivered on the same patients, and sequencing changed clinical management in a meaningful subset.³ That is the high-prior setting. Screening yourself is the low-prior setting, and your default posterior after a single uncurated hit should be skeptical.

Polygenic scores are a different thing entirely

Monogenic screening asks whether you carry one high-impact variant. Polygenic scores sum thousands of small effects. You can compute them yourself with plink2 --score against weight files from the PGS Catalog, which takes minutes. The problems are calibration and ancestry: most scores were trained on European-ancestry cohorts and lose predictive power outside them, and a raw score is meaningless without a reference distribution to place yourself in. Report yourself as a percentile within an ancestry-matched reference panel, or do not report a number at all. Treat the output as a risk distribution. The number is not a result.

Questions people also ask

What is the difference between genetic testing and genetic screening? Screening is applied to people without symptoms or a specific suspicion, to find a subgroup at elevated risk. Testing is applied to answer a specific question in someone with a phenotype or a known family variant. Same laboratory methods, different prior probability, different interpretation.

Will insurance pay for genetic screening? Usually only when there is a documented indication: family history meeting specific criteria, a known familial variant, or a diagnostic question. Elective screening of a healthy adult is generally self-pay. In the US, GINA prohibits health insurers and employers from using genetic information against you. The law does not cover life, disability, or long-term care insurance, which is a real consideration before you test.

How much does full genetic screening cost? Consumer genotyping arrays run around $100 and read a few hundred thousand pre-chosen sites. Targeted clinical panels are typically a few hundred to about $1,500. Clinical exome or genome sequencing is usually in the low thousands. Research-grade 30x whole-genome sequencing alone has fallen to a few hundred dollars of sequencing cost, with the price of the product reflecting analysis, delivery, and support rather than reagents.

What happens if you test positive? Confirmation first in an accredited laboratory, then a conversation with a genetic counselor about penetrance, surveillance options, and cascade testing for relatives who share the variant. Nothing on this page substitutes for that conversation.

What is the 3-gene carrier screening test used for? Reproductive planning. It checks carrier status for a small set of common recessive or X-linked conditions, typically cystic fibrosis, spinal muscular atrophy, and fragile X. It says little about your own health.

Footnotes

1. Tamara S. Roman, Stephanie B. Crowley, Myra I. Roche, et al. Genomic Sequencing for Newborn Screening: Results of the NC NEXUS Project. The American Journal of Human Genetics, 2020. https://doi.org/10.1016/j.ajhg.2020.08.001 ↩

2. Carla van El, Martina C. Cornel, Pascal Borry, et al. Whole-genome sequencing in health care. European Journal of Human Genetics, 2013. https://doi.org/10.1038/ejhg.2013.46 ↩

3. Dimitri J. Stavropoulos, Daniele Merico, Rebekah Jobling, et al. Whole-genome sequencing expands diagnostic utility and improves clinical management in paediatric medicine. npj Genomic Medicine, 2016. https://doi.org/10.1038/npjgenmed.2015.12 ↩