What Your Genome Is, and What You Can Do With the Files

Oak
genomics whole-genome-sequencing vcf bioinformatics variant-interpretation
A sequencing machine in a dark lab feeds three lit glass trays holding drifting fragments, a combed fiber ribbon, and sparse bright points.

Your genome is the roughly 3.1 billion base pairs of DNA in each of your nucleated cells, present in two copies (one from each parent), encoding about 20,000 protein-coding genes plus a much larger amount of regulatory and structural sequence. When someone sells you “your genome,” what you receive is not the molecule but a set of files: read data, an alignment against a reference assembly, and a list of positions where your sequence differs from that reference. Understanding those three artifacts, and their specific failure modes, is most of what separates a useful genome from an expensive PDF.

The three files that matter

Sequencing produces reads. On an Illumina instrument you get paired-end reads of 100 to 150 bases each, delivered as FASTQ: four lines per read, with a Phred quality string encoding per-base error probability. A 30x whole genome is roughly 600 million to 1 billion read pairs, 50 to 120 GB gzip-compressed. Keep these. FASTQ is the only artifact that survives a change of reference assembly or aligner, and both change more often than people expect.

Reads then get aligned to a reference. The output is BAM or, preferably, CRAM, which stores bases by reference difference and cuts a 90 GB BAM to roughly 25 GB. For a human genome in 2026 we would align to GRCh38 with alt contigs plus the HLA and decoy sequences (the “full analysis set”), using bwa-mem2 mem -K 100000000 -Y for reproducible batching, or minimap2 -ax map-ont for nanopore data. T2T-CHM13 is a better assembly in a real sense, complete through centromeres and acrocentric short arms, but most clinical annotation databases are still keyed to GRCh38 coordinates. Our practice is to call on GRCh38 for interpretation and keep CRAM against CHM13 when segmental duplications or the SMN1/SMN2 region matter.

Then come variants, in VCF: one line per variant position, with genotype, depth, and quality fields. A 30x short-read genome yields about 4 to 5 million variants relative to GRCh38, of which roughly 3.5 million are single-nucleotide variants and 500,000 to 600,000 are small insertions and deletions. VCF has been the interchange format for over fifteen years, and the design discussions around what a genome variation file should encode, including reference-relative representation and the need to record no-call regions rather than silently assume reference, are worth reading if you intend to write your own parsers 1.

What sequencing gets wrong, and where

The useful mental model is that a genome is not uniformly observed. Coverage and mappability vary by orders of magnitude across the assembly, and the difficulty is structural rather than a matter of spending more money on depth. This was clear early: the hard parts are repeats, segmental duplications, GC extremes, and the fact that short reads cannot be placed unambiguously where the reference contains near-identical copies 2.

Concretely, short-read 30x whole-genome sequencing reliably calls SNVs in roughly 90 to 92 percent of the genome (the GIAB high-confidence regions), and is materially worse in:

  • Homopolymers and short tandem repeats, where indel calls have high error rates and pathogenic expansions (HTT, FMR1, RFC1) are invisible because the repeat exceeds read length.
  • Pseudogene-shadowed genes: PMS2, SMN1, CYP2D6, GBA, HBA1/HBA2. Reads map ambiguously and callers either drop them or produce false heterozygotes.
  • Structural variants. Short reads detect deletions above roughly 50 bp with moderate sensitivity and inversions and balanced events poorly.
  • HLA and immunoglobulin loci, which are too polymorphic for reference-based calling and need dedicated typers.

Long reads change this picture more than any other single choice. Nanopore sequencing of 1000 Genomes samples at high coverage produced phased assemblies and resolved structural variation and repeat content that short reads had systematically missed, including variants in regions previously inaccessible to reference-based calling 3. If your interest is the hard 8 percent rather than the easy 92, long reads are where to spend. The tradeoff is per-base accuracy in homopolymers and cost per gigabase.

Speed is no longer the constraint either. Targeted analysis on long-read data has been demonstrated from sample to risk assessment in about three hours, which matters for acute settings and tells you something about how much of the classical pipeline latency was convention rather than physics 4.

A pipeline we would run

Assume you have FASTQ and a machine with 64 GB of RAM and a few terabytes of disk.

bwa-mem2 mem -t 32 -K 100000000 -Y -R '@RG\tID:1\tSM:you\tPL:ILLUMINA' \
  GRCh38_full_analysis_set_plus_decoy_hla.fa r1.fq.gz r2.fq.gz \
  | samtools sort -@8 -m2G -o you.bam
samtools markdup -@8 you.bam you.md.bam
samtools view -T ref.fa -C -o you.cram you.md.bam

For germline calling, DeepVariant on a GPU is our default; it outperforms hand-tuned hard filters on indels without your needing to tune anything. GATK HaplotypeCaller with VQSR is defensible if you already have a cohort. Add manta plus smoove for structural variants, ExpansionHunter for known repeat expansions with a supplied catalog, and Stargazer or PyPGx for CYP2D6 star alleles. Check mosdepth output before believing anything: if median autosomal coverage is below 25x or the fraction of the genome at ≥10x is under 0.95, calls in the harder regions are not trustworthy.

Annotation is where most self-directed projects go wrong. Run VEP with --cache --offline --everything --assembly GRCh38, then join against ClinVar with the submission-level file rather than the aggregated one, because star ratings and conflicting interpretations are the signal you need. Expect roughly 2 to 5 variants per genome with a plausible pathogenic assertion in a medically actionable gene, and expect most apparent hits to be classification artifacts: a variant asserted pathogenic in 2009 by one lab, now reclassified, or a variant whose allele frequency in gnomAD is 0.3 percent and therefore cannot cause a rare dominant disease.

Interpretation in a healthy person

If you are asymptomatic, the base rates dominate. Whole-genome sequencing in healthy people produces a great deal of variation of uncertain significance, and the careful reviews are clear that population screening in the absence of phenotype or family history has a modest positive predictive value and a substantial burden of follow-up 5. That is a statement about arithmetic, not about the technology. Penetrance estimates for most variants come from ascertained families and shrink, often dramatically, when measured in unselected cohorts.

Where genomes earn their keep is when there is a phenotype to anchor them. Sequencing critically ill newborns first, rather than after a diagnostic odyssey, raises the rate of precise genetic diagnosis and does so more equitably than referral-gated testing 6. In early-onset advanced heart failure, whole-genome sequencing identifies causal variants in a meaningful fraction of patients who would otherwise carry an idiopathic label 7. The pattern is consistent: a genome is a hypothesis-resolving instrument, strongest when pointed at a specific question.

This is also why we treat the genome as one layer rather than the answer. Your DNA is fixed and tells you about capacity. RNA expression, protein abundance, and metabolic response tell you what that capacity is doing this year. A variant of uncertain significance in a splice region becomes far less uncertain when you can look at the transcript.

Anything you find that looks medically consequential should be confirmed in a clinical laboratory on a fresh sample and interpreted with a genetic counselor or physician. Research-grade calls carry sample-swap and contamination risk that no amount of filtering removes, and classification requires clinical context you cannot supply from a file.

Questions people also ask

What is a genome in simple words? It is the complete DNA instruction set of an organism, written in four letters, about 3.1 billion letters long in humans, copied in nearly every cell. The project to read the first human reference took more than a decade of international effort and was contentious from the start about cost and value 8.

What is your genome, specifically? Your genome is your two inherited copies of that sequence, differing from the reference at roughly 4 to 5 million positions, plus structural differences that reference-relative formats represent poorly. No two people share it, and your somatic cells accumulate additional mutations over your lifetime.

What is MyOme, and what are the benefits? MyOme is a company offering clinical whole-genome sequencing with curated interpretation reports, typically ordered through a clinician. The benefit of that model is clinical-grade calling and a report you can act on with a physician. The limitation is that you generally receive an interpretation rather than the full read-level data, so you cannot re-analyze against a new assembly or run your own callers.

Should I choose short reads or long reads? Short-read 30x is the cost-efficient choice for SNVs and small indels across the accessible genome. Choose long reads if you care about structural variation, phasing, repeat expansions, or the pseudogene-shadowed genes, and accept higher cost and weaker homopolymer accuracy 3.

Oak builds longitudinal molecular profiles of individuals: whole-genome sequencing, RNA sequencing, proteomics, blood biomarkers, and continuous glucose data, integrated into one model of you. Build your profile.

Footnotes

  1. Martin G Reese, Barry Moore, Colin Batchelor, et al. A standard variation file format for human genome sequences. Genome Biology, 2010. https://doi.org/10.1186/gb-2010-11-8-r88 ↩

  2. D. C. Koboldt, L. Ding, E. R. Mardis, et al. Challenges of sequencing human genomes. Briefings in Bioinformatics, 2010. https://doi.org/10.1093/bib/bbq016 ↩

  3. Jonas A. Gustafson, Sophia B. Gibson, Nikhita Damaraju, et al. High-coverage nanopore sequencing of samples from the 1000 Genomes Project to build a comprehensive catalog of human genetic variation. Genome Research, 2024. https://doi.org/10.1101/gr.279273.124 ↩ ↩2

  4. Miranda Galey, Paxton Reed, Tara Wenger, et al. 3-hour genome sequencing and targeted analysis to rapidly assess genetic risk. 2022. https://doi.org/10.1101/2022.09.09.22279746 ↩

  5. Noralane M. Lindor, Stephen N. Thibodeau, Wylie Burke. Whole-Genome Sequencing in Healthy People. Mayo Clinic Proceedings, 2017. https://doi.org/10.1016/j.mayocp.2016.10.019 ↩

  6. Tara L. Wenger, Abbey Scott, Lukas Kruidenier, et al. SeqFirst: Building equity access to a precise genetic diagnosis in critically ill newborns. The American Journal of Human Genetics, 2025. https://doi.org/10.1016/j.ajhg.2025.02.003 ↩

  7. Erik Linnér, Tomasz Czuba, Olof Gidlöf, et al. Whole genome sequencing in early onset advanced heart failure. Scientific Reports, 2025. https://doi.org/10.1038/s41598-025-88465-8 ↩

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