Genomic DNA Analysis: What Your Own Genome Can Tell You, and How to Analyze It

Genomic DNA analysis means reading the DNA sequence of your nuclear and mitochondrial genome, aligning those reads to a reference, and enumerating the places where you differ from that reference. Done at 30x whole-genome coverage, it yields roughly four to five million small variants (single-nucleotide variants and short insertions or deletions), a few thousand structural variants, and copy-number and repeat-length estimates across the genome. A small, well-defined subset of those calls carries clinical weight today: pathogenic variants in a few dozen genes with established penetrance, pharmacogenetic haplotypes with dosing guidelines, and carrier status for recessive conditions. The rest is either association-level signal with modest effect sizes or noise, and the analytical skill is in telling those apart.

Genomic DNA versus “a DNA test”

Genomic DNA is the DNA in your cells: about 3.1 gigabases per haploid copy, plus a 16.6 kb circular mitochondrial genome present in hundreds to thousands of copies per cell. It is the same molecule whether you extract it from saliva or a blood draw. The phrase is used to distinguish it from complementary DNA (cDNA) made from RNA, or from plasmid and synthetic DNA in a lab context.

The distinction that matters commercially is between genotyping arrays and sequencing. A consumer array interrogates 600,000 to 1.5 million pre-chosen positions. It cannot see a variant that is not on the chip, and it cannot see novel or rare variants, which is exactly where most high-impact coding variation sits. Whole-genome sequencing reads the whole molecule with short reads (typically 2x150 bp paired-end) and calls variation without a predetermined list. The move from targeted assays to population-scale sequencing is what changed the character of genomic analysis: the bottleneck stopped being data generation and became interpretation and statistics ¹². If you intend to do your own analysis, an array is a dead end and sequencing is the only sensible input.

The files you should insist on

Ask for these, in this order of importance:

• CRAM (or BAM) aligned to GRCh38 or, better, to a T2T-CHM13-based reference as a secondary alignment. It should come with the reference FASTA and its .fai so the CRAM is decodable. A 30x CRAM runs roughly 25–35 GB, and the equivalent BAM is 80–100 GB.
• gVCF plus a filtered VCF. The gVCF records reference confidence at non-variant sites, which is what lets you distinguish “homozygous reference” from “not covered.”
• FASTQ, if storage permits (~100 GB gzipped). It is the only artifact that lets you re-run alignment against a future reference without loss.
• A QC report: mean and median coverage, fraction of the callable genome at ≥20x, duplicate rate, insert-size distribution, and cross-sample contamination estimate.

Two thresholds we treat as pass/fail: mean autosomal coverage ≥30x with ≥90% of bases at ≥20x, and estimated contamination (VerifyBamID or DRAGEN’s estimate) below 1%. Below 20x, heterozygous sites start dropping out, and a false homozygous-reference call is invisible in a plain VCF.

The pipeline we would run

For alignment, bwa-mem2 mem -t 32 -K 100000000 -Y against GRCh38 with alt contigs, followed by samtools sort and samtools markdup. Illumina’s DRAGEN hardware pipeline is faster and its results are broadly comparable, and if your provider hands you DRAGEN output there is little reason to realign.

For small-variant calling, we prefer DeepVariant over GATK HaplotypeCaller on short reads, mainly for indel precision in low-complexity regions:

run_deepvariant --model_type=WGS --ref=GRCh38.fa \
  --reads=sample.cram --output_vcf=sample.vcf.gz \
  --output_gvcf=sample.g.vcf.gz --num_shards=32

Then normalize before you do anything else, because unnormalized indel representations will silently fail to match ClinVar:

bcftools norm -f GRCh38.fa -m -any -Oz -o sample.norm.vcf.gz sample.vcf.gz
bcftools index -t sample.norm.vcf.gz

Annotate with Ensembl VEP (--cache --offline --everything --pick_allele_gene) or snpEff, joined against ClinVar’s VCF release and gnomAD v4 allele frequencies. Filter aggressively: drop anything with gnomAD popmax allele frequency above 1% before you look at a gene list, since common variants are essentially never highly penetrant.

Small variants are not the whole picture, and skipping the rest is the most common failure in self-analysis. Add Manta or GATK-SV for structural variants, and GATK-gCNV or CNVnator for copy number. Add ExpansionHunter for known repeat expansions. It genotypes loci such as HTT and C9orf72 reasonably well, but short reads cannot reliably size large FMR1 expansions. Use HLA-LA for HLA typing at four-digit resolution, and mutserve plus haplocheck for mitochondrial heteroplasmy and contamination. Short-read WGS also misses much of CYP2D6, a pharmacogene with common hybrid alleles and copy-number variation. Cyrius is the tool built for that specific problem.

What the calls support, and what they do not

Three categories are worth your attention. First, pathogenic and likely-pathogenic variants in genes with established penetrance and defined follow-up. The practical reference set is the ACMG secondary-findings list (v3.2, 81 genes). Reported rates of an actionable finding in unselected adults cluster in the low single-digit percent. The genes are well characterized at the sequence level and have been for decades: the BRCA1 region was fully sequenced and annotated in 1996, well before routine clinical testing ³. Second, pharmacogenetic star alleles in these genes with CPIC guidelines.

• CYP2C19
• CYP2C9
• VKORC1
• DPYD
• TPMT
• SLCO1B1
• CYP2D6

Almost everyone carries at least one. Third, recessive carrier status, which matters for reproductive planning and nothing else.

Everything else is quantitative and weak. Polygenic scores built from genome-wide association summary statistics are computable from your VCF with plink2 --score, and they are real in the sense that they predict variance in a population. They are also poorly portable across ancestries, because the discovery cohorts skewed European and linkage patterns differ. Careful handling of population structure and shared data infrastructure were central concerns of the consortia that produced these datasets in the first place ⁴. Treat a polygenic score as a percentile within a reference population, and do not treat a percentile as a prognosis. For ancestry, project your genotypes onto principal components computed from the 1000 Genomes phase 3 panel with plink2 --pca and --score on the loadings, rather than trusting a vendor’s proprietary label.

A candid account of reading one’s own retail genome captures the failure mode: the numbers arrive, they change depending on which provider computed them, and they rarely resolve into a decision ⁵. If you want your genome to change a decision, define the decision first.

Risks, mostly informational

The physical risk is a blood draw or a saliva tube. The real exposures are informational and permanent. Your genome identifies you, identifies your relatives, and cannot be revoked once disclosed. Re-identification does not require the raw file: even aggregate statistics such as allele frequencies from a study cohort can leak whether a specific individual was a participant, which is why controlled-access repositories exist ⁶. On the order of a few dozen independent common SNPs are enough to distinguish an individual.

Practical posture: keep the CRAM and VCF encrypted at rest on hardware you control, treat any upload to a third-party interpretation service as public disclosure, and read the consent terms on research sharing carefully. Sequencing a minor raises the additional problem that the person whose data it is cannot consent and will inherit findings, including adult-onset risk, that they did not choose to learn ⁷.

Where a clinician is not optional

We can tell you how to call variants and how to judge the evidence. We cannot tell you what a result means for your body, and neither can a pipeline. A pathogenic call in a cancer-predisposition or cardiac gene needs orthogonal confirmation in an accredited clinical laboratory, because research-grade pipelines produce false positives, especially for indels in repetitive sequence. Interpretation, family-history context, surveillance planning, and any medication decision belong with a genetic counselor or physician. Bring them the annotated variant with coordinates, the transcript, the read depth and allele balance, and the ClinVar review status. That is a productive conversation. A percentile from a polygenic score is not.

Questions people also ask

What can DNA tests really tell you? Reliably: carrier status for recessive conditions, pharmacogenetic metabolizer phenotypes, and pathogenic variants in a few dozen high-penetrance genes. Probabilistically: population-relative risk percentiles for common traits, with wide uncertainty and limited transfer across ancestry groups. Not at all: most of what people hope for, including anything resembling a prediction of when or whether you will get a common disease.

What is the difference between genomic DNA and normal DNA? Genomic DNA is the chromosomal and mitochondrial DNA in your cells, distinguished from cDNA copied from RNA or from plasmid and synthetic constructs used in the lab. In a consumer context the phrase is often used loosely to mean “your whole genome” as opposed to a targeted panel.

What’s the difference between genetic testing and genomic testing? Genetic testing usually targets one gene or a defined panel to answer a specific question. Genomic testing reads the whole genome or exome and answers questions you did not pre-specify, which is more informative and also more prone to findings of uncertain significance ¹.

What diseases can be diagnosed with DNA testing? Monogenic conditions with well-established variant-disease relationships. Examples are cystic fibrosis, Huntington disease, hereditary hemochromatosis, familial hypercholesterolemia, and the hereditary cancer syndromes. Diagnosis is a clinical act that requires a clinician and a confirmatory accredited test. A VCF line is not a diagnosis.

What cancers can be detected through genetic testing? Germline sequencing identifies inherited predisposition. It is most useful in BRCA1, BRCA2, PALB2 and the Lynch syndrome mismatch-repair genes, and in TP53 and CDH1. Germline sequencing does not detect cancer. Detecting an existing tumor is a different assay on tumor tissue or cell-free DNA.

Is genomic testing worth it? If you will do the analysis yourself, keep the raw files, and treat the output as a baseline to be re-interrogated as knowledge improves, yes. If you want a single answer from a report, an array-based product will disappoint you and a genome will overwhelm you.

Footnotes

1. 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 ↩ ↩²

2. Rachel L Goldfeder, Dennis P Wall, Muin J Khoury, et al. Human Genome Sequencing at the Population Scale: A Primer on High-Throughput DNA Sequencing and Analysis. American Journal of Epidemiology, 2017. https://doi.org/10.1093/aje/kww224 ↩

3. T M Smith, M K Lee, C I Szabo, et al. Complete genomic sequence and analysis of 117 kb of human DNA containing the gene BRCA1. Genome Research, 1996. https://doi.org/10.1101/gr.6.11.1029 ↩

4. The GAIN Collaborative Research Group. New models of collaboration in genome-wide association studies: the Genetic Association Information Network. Nature Genetics, 2007. https://doi.org/10.1038/ng2127 ↩

5. Martin Richards. Reading the runes of my genome: a personal exploration of retail genetics. New Genetics and Society, 2010. https://doi.org/10.1080/14636778.2010.507486 ↩

6. Xiaoyong Zhou, Bo Peng, Yong Fuga Li, et al. To Release or Not to Release: Evaluating Information Leaks in Aggregate Human-Genome Data. Lecture Notes in Computer Science, 2011. https://doi.org/10.1007/978-3-642-23822-2_33 ↩

7. David P. Dimmock, David P. Bick. Ethical Issues in DNA Sequencing in the Neonate. Clinics in Perinatology, 2014. https://doi.org/10.1016/j.clp.2014.08.016 ↩