What It Takes to Discover Your Own Genome

Discovering your genome means ordering a 30x whole-genome sequencing run and taking possession of the raw reads. From there you align them to a reference, call variants, and annotate the roughly four to five million positions where you differ from that reference. The sequencing itself is a solved commodity problem. The interesting work is everything after alignment, and it is the part most consumer products skip. It means deciding which of your variants are well-supported and which are known. It also means deciding which are merely rare, and which are uninterpretable with current evidence. Most of your variants fall in the last category, and a genome that is annotated will say so.

What the data is

A 30x human whole genome produces about 100 gigabases of sequence. It arrives as paired FASTQ files, typically 40 to 60 GB gzipped for the pair, or better as a CRAM file aligned to a specific reference build. CRAM is reference-compressed, so a 30x genome lands around 15 to 20 GB instead of the 80 to 100 GB an equivalent BAM would take. Insist on receiving either FASTQ or CRAM. A vendor that ships you only a VCF has thrown away the evidence layer. You will not be able to re-call, re-align to a newer reference, or inspect a questionable call by eye.

Depth matters more than most buyers realize. Coverage is Poisson-distributed across the genome with GC-dependent bias, so a nominal 30x mean leaves a meaningful fraction of the genome below 10x, where heterozygous sites are routinely missed because one allele fails to be sampled. Thirty-fold is the standard for germline SNV and small-indel calling and is what large population projects converged on for good reason ¹. If you care about structural variants, repeat expansions, or phasing, short reads at 30x will not be enough, and you should plan on long reads instead.

The variant file you end up with should be a gVCF or a VCF with per-site depth, genotype quality, and allele balance retained. Those fields are what let you triage later. A heterozygous call with an allele balance of 0.12 at 8x is noise. The same call at 0.48 and 40x is real.

The pipeline we would run

We would align with bwa-mem2 mem -K 100000000 -Y against GRCh38 including the ALT contigs and the decoy sequence, then mark duplicates and sort with samtools. Using the full analysis set with decoys matters: without them, reads from unplaced sequence pile onto paralogous regions and generate false heterozygous calls in genes like PMS2 and SMN1.

For variant calling we prefer DeepVariant over GATK HaplotypeCaller for a single sample. It is a convolutional model over pileup images, it requires no site-level filtering model of your own, and it has consistently better indel precision on 30x Illumina data. Run it with --model_type=WGS and keep the output gVCF. If you want a second opinion, call with both and look only at the discordant sites, which will be a few tens of thousands and will concentrate in homopolymers and segmental duplications.

Then check the run before believing any of it. mosdepth --by 1000 gives you a coverage histogram and tells you what fraction of the genome sits below 10x. samtools stats gives you insert size distribution and duplicate rate. A duplicate rate above roughly 15 percent means the library was over-amplified from too little input DNA. bcftools stats gives you a transition/transversion ratio, which should sit near 2.0 to 2.1 genome-wide. A Ti/Tv of 1.7 means your call set is contaminated with false positives. Also verify sex chromosome coverage and check for sample contamination with verifyBamID2, because swapped or mixed samples are a real and quiet failure mode.

Annotation comes last. Ensembl VEP with the --everything flag plus gnomAD v4 allele frequencies and ClinVar gives you a workable table. For pharmacogenomics, PharmCAT handles most star-allele assignment, but CYP2D6 requires a copy-number-aware caller such as Cyrius because the gene has a pseudogene neighbor and common hybrid alleles that a generic pipeline silently miscalls.

What you will find, and what it means

The scale is worth internalizing. Population sequencing found that a typical person carries on the order of 250 to 300 loss-of-function variants in annotated genes and 50 to 100 variants previously reported as causing inherited disease ². Almost none of those reported associations will turn out to apply to you. Many were published from small pedigrees, appear at frequencies in gnomAD far too high to cause the disease attributed to them, and have since been reclassified. This is the single most common way a self-directed genome analysis goes wrong: treating a ClinVar “pathogenic” tag as a result rather than as a claim with a date and an evidence level attached.

The defensible core of a personal genome is narrow and real. The ACMG secondary findings list covers a small set of genes with actionable, well-established pathogenic variants. Carrier status for recessive conditions is straightforward to read. Pharmacogenomic star alleles in a handful of genes have decent evidence. Beyond that, polygenic scores exist but transfer poorly across ancestries, and their individual-level predictive value is modest. Anything that looks like a diagnosis belongs with a clinical geneticist and a CLIA-certified confirmatory test, because research-grade calls are not the basis for medical decisions.

Why this is cheap now and was not before

The human genome was argued about before it was sequenced, including over whether comprehensive sequencing was a sensible use of money at all ³. Early variation catalogs were built by clever sampling rather than whole genomes, such as the reduced-representation shotgun approach that produced the first large SNP map ⁴. Individual genomes arrived shortly after, including early single-molecule demonstrations that showed the technology path was not going to stay expensive ⁵. Within a decade the reference sequence had reorganized human genetics around variant catalogs and association studies ⁶, and population-scale sequencing turned into routine infrastructure ¹.

The practical consequence is that the sequencing is no longer the bottleneck or the cost center. Interpretation is, and interpretation improves every year against data you already own. That asymmetry is the argument for holding your own FASTQ or CRAM rather than a report.

Where the genome stops

Your genome is fixed and therefore says nothing about state. It does not tell you which genes are being transcribed today or what your circulating protein levels are. The sequence says nothing about how your fasting insulin has drifted or how you respond to a meal. A variant in a metabolic gene is a hypothesis about a mechanism, and the measurement that tests it is RNA, protein, a blood panel, or glucose telemetry. Reading a genome alone tends to produce a long list of plausible stories with no way to adjudicate between them. Pairing it with repeated functional measurements is what makes any of it checkable.

Questions people also ask

What was unique about Ozzy Osbourne’s DNA? His genome was sequenced in 2010, when individual whole genomes were still rare enough to be newsworthy. The reported findings included novel variants in regions near genes involved in alcohol and drug metabolism, plus segments of Neanderthal-derived ancestry, which most people of non-African descent carry. None of it was clinically interpretable, and it is a good illustration of how an interesting genome and an informative one differ.

How much does a genome test typically cost? Research-grade 30x whole-genome sequencing with raw data returned generally runs a few hundred to about a thousand dollars. Clinical whole-genome or exome sequencing with an interpreted report from an accredited lab, ordered through a physician, typically costs several thousand. Genotyping arrays, which read roughly 600,000 to a million preselected positions rather than the whole sequence, cost under a hundred dollars and are a different product with far less coverage of rare variation.

How many generations until you are no longer related? Genealogical relatedness and genetic relatedness diverge quickly. You inherit half your autosomal DNA from each parent and a quarter from each grandparent, but recombination is chunky, so past roughly six to eight generations the expected contribution from any single ancestor is small enough that many ancestors contribute no detectable segments at all. You remain their descendant while sharing no identifiable DNA with them.

What animal is 99% human DNA? Chimpanzees and bonobos. In alignable regions the single-base identity is close to 98.8 percent, though counting insertions and deletions drops the genome-wide figure to roughly 96 percent. The number is sensitive to how you define the comparison, which is why quoted values vary.

Footnotes

1. 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 ↩ ↩²

2. The 1000 Genomes Project Consortium. A map of human genome variation from population-scale sequencing. Nature, 2010. https://doi.org/10.1038/nature09534 ↩

3. P.S Harper. Mapping and sequencing the human genome. Endeavour, 1989. https://doi.org/10.1016/0160-9327(89)90035-5 ↩

4. David Altshuler, Victor J. Pollara, Chris R. Cowles, et al. An SNP map of the human genome generated by reduced representation shotgun sequencing. Nature, 2000. https://doi.org/10.1038/35035083 ↩

5. Dmitry Pushkarev, Norma F Neff, Stephen R Quake. Single-molecule sequencing of an individual human genome. Nature Biotechnology, 2009. https://doi.org/10.1038/nbt.1561 ↩

6. Eric S. Lander. Initial impact of the sequencing of the human genome. Nature, 2011. https://doi.org/10.1038/nature09792 ↩