The Best Genome Kit Is the One That Gives You a BAM and a gVCF

If you intend to analyze your own data, the best genome kit is a 30x or deeper short-read whole-genome sequence that returns an aligned BAM or CRAM file plus a per-sample gVCF, with no restriction on what you do with them. Everything else is a downgrade. Genotyping arrays (23andMe, AncestryDNA, MyHeritage) do not sequence your genome at all: they measure a fixed panel of roughly 600,000 to 700,000 pre-selected positions and hand you a tab-separated text file. Exomes cover about 1 to 2 percent of the genome and inherit capture-bait bias. Long-read sequencing is the better technology for structural variation but costs more and still has a higher small-variant error rate than short reads in easy regions.

This page is about which kit produces data worth keeping for twenty years. It is not medical advice, and no variant interpretation here substitutes for a clinical geneticist or genetic counselor.

What each tier of product measures

An array is a hybridization assay. Your DNA is fragmented and washed over a chip carrying probes for specific single-nucleotide positions, and fluorescence tells the caller which of two alleles is present. The output is a file of about 600k to 700k rows. Each row holds rsID, chromosome, position, and genotype. That is roughly 0.02 percent of the 3.1 billion base positions in the genome. Array chips also carry no information about anything the chip designer did not choose to include. Novel variants, most insertions and deletions, and nearly all structural variation are therefore invisible by construction.

Short-read whole-genome sequencing is a different category of measurement. DNA is fragmented to a few hundred base pairs, sequenced from both ends, and aligned to a reference genome. At 30x mean coverage, each position is spanned by about thirty independent reads on average, which is enough to distinguish a real heterozygous site from a sequencing error with high confidence across most of the genome. Population-scale sequencing pipelines have converged on this depth for exactly that reason.¹

Long-read sequencing (PacBio HiFi, Oxford Nanopore) reads fragments of 10 to 100+ kilobases. The long fragments span repeats that short reads cannot, which is why long reads resolve structural variants, segmental duplications, and phasing far better than 150 bp paired-end data.² Long reads also make it practical to assemble your genome rather than only align it to a reference, and to read methylation directly from the same molecules.

The comparison that matters: what you can compute

Ask of any kit what analyses the resulting files support. This is where the tiers separate sharply.

From an array text file you can compute polygenic scores over the genotyped sites, check imputed ancestry, and look up common variants with known effects. You cannot call a rare coding variant, because the chip does not measure it. You cannot measure mitochondrial heteroplasmy, copy-number variation with any precision, or short tandem repeat expansions. Imputation using a reference panel will fill in millions of “genotypes” statistically, and those imputed calls are useful for aggregate scores. They are not measurements of your DNA, and using them for single-variant questions is a mistake.

From a 30x WGS BAM plus gVCF you can do most of what a research lab does. That includes single-nucleotide variants and small indels genomewide, pharmacogene star-allele calling with tools like PharmCAT and Stargazer, and HLA typing from reads mapped to chromosome 6. You can also do copy-number calling with read-depth methods and repeat expansion screening with ExpansionHunter. Mitochondrial variants are within reach too, including low-frequency heteroplasmy given the high effective coverage of the mitochondrial genome. You can also re-call everything in five years when the callers improve, because you kept the reads.

From long-read data you add reliable structural variant detection, haplotype phasing across long blocks, and per-CpG methylation without a separate assay. The tradeoff is cost and, for nanopore specifically, a per-base error profile that remains worse than Illumina in homopolymers and other easy contexts.

Why the BAM matters more than the VCF

A VCF is an interpretation of your reads under one pipeline, one reference build, and one set of filters. The reads themselves are the measurement. If a vendor gives you only a VCF, or worse only a PDF report, you lose several options. You cannot re-align to a new reference, inspect a suspicious call, or apply a caller that did not exist when the file was made.

Insist on three things. First, the aligned reads as BAM or CRAM, roughly 45 to 60 GB for a 30x CRAM against GRCh38. Second, a gVCF rather than just a VCF. A gVCF records confidence at reference-matching positions and lets you distinguish “this position is homozygous reference” from “this position was never covered”. Third, the raw FASTQ or at least an unaligned BAM, if you want full freedom to re-align.

Ask which reference build was used. GRCh38 is the sensible default today, with T2T-CHM13 worth having as a secondary alignment for regions GRCh38 handles poorly. Ask whether alt contigs were included, because alt-aware alignment changes calls in HLA and other complex loci.

How to check a genome you paid for

Do not trust a vendor’s coverage claim. Verify it yourself.

Start with samtools stats and samtools coverage on the CRAM to get mean depth, percentage of the genome above 10x and 20x, duplicate rate, and mapping rate. For a good 30x Illumina library, expect a mapping rate above 99 percent and an optical and PCR duplicate rate under about 10 percent. Expect above 90 percent of the callable genome at 20x or better. A high duplicate rate usually means a low-complexity library, which effectively lowers your real depth.

Then check variant counts. A single human genome aligned to GRCh38 typically yields roughly 4 to 5 million SNVs and several hundred thousand small indels, with the exact figure depending on ancestry and filters. Counts far below that range point to an over-filtered callset. Counts far above it usually mean contamination, which you can quantify directly with VerifyBamID2.

The most useful single check is a benchmark comparison. The Genome in a Bottle project sequenced reference individuals extensively and published high-confidence variant call sets with defined confident regions.³ If you also have access to a NA12878 or HG002 run through the same pipeline, run hap.py against the GIAB truth set restricted to the high-confidence BED file, and read off SNV and indel recall and precision. A modern short-read pipeline should land above 99.5 percent SNV recall and precision inside those regions. Indel numbers will be lower, and that gap is real rather than a pipeline defect.

Finally, be clear-eyed about what “genomewide” excludes. Short reads leave persistent gaps in centromeres, acrocentric short arms, and large segmental duplications. Genes such as SMN1, PMS2, and CYP2D6 sit in regions where short-read calls are unreliable. Long-read data is the correct instrument for those loci.²

What we would buy

If your goal is relative matching and ancestry, an array is the right tool and costs under a hundred dollars. It answers those questions well and cheaply.

If your goal is to have a genome you can analyze for the rest of your life, buy 30x short-read WGS. Choose a provider that returns CRAM, gVCF, and FASTQ on GRCh38. That single purchase supports pharmacogenomics, carrier status, and polygenic scores computed from measured rather than imputed genotypes. It also supports every caller released in the next decade. At current prices this is the best ratio of analyzable information to dollars in personal genomics.

Add long-read sequencing when you have a specific reason. A suspected structural variant, a repeat expansion question, or a gene that short reads cannot resolve would each qualify. So would a wish to phase and assemble your own haplotypes.² Treat it as a second instrument rather than a replacement.

One thing no kit gives you is state. A genome is a fixed text and tells you about disposition. Expression, protein abundance, and metabolic response change month to month. Those require RNA, proteomic, and metabolic measurement alongside the sequence. If you find something in your genome that looks clinically meaningful, take it to a clinician for confirmation in an accredited laboratory before acting on it.

Questions people also ask

Which genome kit is best overall? For a technical reader who plans to do their own analysis, a 30x short-read whole genome delivered as CRAM plus gVCF plus FASTQ. It measures every accessible base rather than a chosen panel, and the files stay useful as tools improve.¹

Is 23andMe a genome sequence? No. It is a genotyping array measuring a few hundred thousand chosen positions, which is a small fraction of one percent of your genome. It cannot detect variants absent from the chip design.

Is 30x coverage enough, or should I pay for 60x? For germline SNV and indel calling, 30x is the established working point and is what population-scale projects use.¹ Higher depth helps mainly for somatic work, mosaicism, and depth-based copy-number calling.

Should I get long reads instead of short reads? Only if you have a specific question short reads cannot answer, such as structural variants, phasing, or genes inside segmental duplications.² For cost-effective small-variant calling, short reads remain the practical choice.

How do I know my sequencing was done well? Compute depth and duplicate statistics yourself with samtools, check contamination with VerifyBamID2, and compare a control sample against Genome in a Bottle truth sets using hap.py inside the published high-confidence regions.³

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 ↩³

2. Glennis A. Logsdon, Mitchell R. Vollger, Evan E. Eichler. Long-read human genome sequencing and its applications. Nature Reviews Genetics, 2020. https://doi.org/10.1038/s41576-020-0236-x ↩ ↩² ↩³ ↩⁴

3. Justin M. Zook, David Catoe, Jennifer McDaniel, et al. Extensive sequencing of seven human genomes to characterize benchmark reference materials. 2015. https://doi.org/10.1101/026468 ↩ ↩²