Dante Labs: What You Get, and What to Do With the Files
Dante Labs sells consumer whole-genome sequencing, typically advertised at 30x short-read coverage. It delivers raw data in FASTQ, BAM and VCF form alongside a set of consumer reports. For a technically fluent buyer the raw files are the only part worth paying for. The reports are thin, and the recurring public complaint is turnaround time and support responsiveness rather than sequencing chemistry. If you buy, plan to verify the data yourself and re-call variants from the BAM or FASTQ before you believe anything downstream.
This page covers what arrives, how to audit it, and what you can and cannot learn from a single short-read genome.
What lands in your download folder
Expect three tiers of files. The FASTQ pair is the unaligned output, usually 150 base paired-end reads, around 100 GB compressed for a 30x human genome. The BAM is the aligned version, 60 to 90 GB depending on compression and whether duplicates were marked. The VCF holds small variants, typically 4 to 5 million single-nucleotide variants and short insertions or deletions per genome.
The first thing to check is the reference build. Older consumer pipelines aligned to hg19 or hs37d5, current ones to GRCh38 with alt contigs. Read the BAM header rather than trusting the product page:
samtools view -H sample.bam | grep -E '^@SQ|^@PG|^@RG' | head -40
The @SQ lines give you contig names and lengths, which pin the build. The @PG lines give the aligner and its command line. If @PG is missing or stripped, treat the alignment as unaudited and plan to realign. Convert to CRAM against the matching reference once you know the build, which takes a 70 GB BAM down to roughly 15 to 20 GB.
Structural variant and copy-number calls are usually absent. If they are present, ask which caller produced them, because consumer pipelines rarely run Manta or a copy-number caller at all.
Audit the data before you trust a single variant
Coverage headline numbers are marketing figures. What matters is the fraction of the genome that reached a depth where calls are reliable. Use mosdepth with quantized output:
mosdepth --by 1000 --fast-mode --quantize 0:1:5:10:20: sample sample.bam
For a genuine 30x genome you want at least 90 percent of the callable autosome at 10x or better, and mean autosomal depth close to 30. Coverage in the mid-twenties is common and workable. Anything under 20x mean depth degrades heterozygous sensitivity noticeably, especially for indels.
Next check library quality. Run samtools stats and look at three numbers: duplicate rate, insert size distribution, and error rate. PCR-free libraries should show duplicate rates in the low single digits. Duplicate rates above 15 percent mean your effective coverage is well below the nominal figure. Insert size modes below 300 base pairs limit what any structural variant caller can do.
Then check contamination and identity. verifybamid2 estimates the fraction of reads from another individual using population allele frequencies. Values above roughly 2 percent should make you distrust low-allele-fraction calls. somalier extract followed by somalier relate gives you inferred sex and an ancestry projection, which catches the one failure mode that wrecks everything downstream: a swapped sample.
Finally, look at the X and Y depth ratio directly in the mosdepth summary. A karyotypically male sample should show roughly half autosomal depth on both X and Y. A mismatch against your known sex means stop and contact the lab.
Re-call variants yourself
We would not build anything on a vendor VCF. Filtering thresholds are undocumented, and you cannot reproduce a result you cannot re-run. Two paths are reasonable.
If the BAM is aligned to a build you want, call directly from it. DeepVariant is the option we prefer for germline short-read data because its default model needs no hand-tuned filter thresholds:
run_deepvariant \
--model_type=WGS \
--ref=GRCh38_full_analysis_set_plus_decoy_hla.fa \
--reads=sample.bam \
--output_vcf=sample.dv.vcf.gz \
--num_shards=32
That runs in a few hours on 32 cores. GATK HaplotypeCaller with VQSR is the alternative, and it is more work for similar sensitivity on a single sample.
If the header is unclear or the build is old, go back to FASTQ and align with bwa-mem2 mem -t 32 -R '@RG\tID:1\tSM:sample\tPL:ILLUMINA', then mark duplicates with samtools markdup and call as above. Full realignment plus calling on a 30x genome is roughly a day of wall-clock time on a 32-core machine.
Annotate with Ensembl VEP. Pull in ClinVar, gnomAD allele frequencies, and a splice predictor such as SpliceAI. Then filter conservatively. Population frequency is the strongest first filter: a variant above 1 percent in gnomAD is very unlikely to explain a rare Mendelian condition. Restrict to ClinVar pathogenic or likely pathogenic entries with at least two-star review status before you look at anything else.
What one short-read genome can and cannot tell you
Short-read whole-genome sequencing is good at single-nucleotide variants and short indels in unique sequence. It is weak in several specific places, and knowing them saves you from false confidence.
Repeat expansions need a dedicated caller such as ExpansionHunter, and even then long alleles are undercalled. The pharmacogenomic gene CYP2D6 has structural variation that generic callers mishandle, so use Cyrius or an equivalent. Segmental duplications, including much of the region around SMN1 and SMN2, produce mapping artifacts that look like real variants.
Interpretation is the harder limit. Coding variant classification shifts over time as families and functional data accumulate, and reanalysis of an old call can change its meaning entirely 1. A variant that looks clearly pathogenic in one family can present with a narrower phenotype than the textbook description 2. Noncoding variants are worse: enhancer effects are predicted rather than measured, and sequence models for allele-specific enhancer activity are an active research area rather than a settled tool 3. Population frequency filters also carry assumptions, since immune and other gene families show signatures of selection that shape allele frequencies unevenly across populations 4.
DNA also does not tell you which promoter or isoform a tissue uses. Alternative promoter usage is tissue-specific and only visible in transcript data 5. This is the structural reason to pair a genome with RNA sequencing rather than treat the genome as complete.
Where a clinician is required
Consumer sequencing is not a diagnostic test. If your filtered VCF contains a pathogenic variant in a gene with medical implications, the next step is a clinical genetics referral and confirmation in an accredited laboratory, usually by an orthogonal method. Do not change anything about your health based on a consumer VCF, and do not interpret a negative result as reassurance, because coverage gaps and caller blind spots are real. Genetic counseling matters here for the same reason: family implications follow from any germline finding.
Questions people also ask
Is Dante Labs legitimate?
The company delivers real whole-genome data, and many customers receive usable FASTQ and BAM files. The persistent complaints in public reviews concern turnaround time, refunds, and support responsiveness rather than data quality. Treat delivery time as uncertain and verify the data on arrival.
Is 30x coverage enough?
For germline single-nucleotide variants and short indels in unique sequence, yes. Verify the delivered depth rather than the advertised depth, since duplicate reads and uneven coverage often push effective depth into the low twenties. Somatic or mosaic questions need far deeper sequencing.
Should I realign to GRCh38?
If your data is on hg19 or hs37d5 and you plan to use current annotation resources, yes. Liftover of a VCF is lossy at complex sites, and realignment from FASTQ is cleaner. Budget about a day of compute.
Do I need long reads?
Only for specific questions: large structural variants, repeat expansions, phasing across long distances, and difficult regions such as HLA. Short reads cover the majority of interpretable small-variant biology.
What should I do with the reports?
Read them as a starting index, then re-derive every claim from the VCF yourself. The underlying evidence for any given variant is public in ClinVar and the primary literature, and interpretation standards move faster than consumer reports are updated 1.
Footnotes
-
Saeideh Kavousi, Farzad Kamali, Bahareh Rabbani, et al. Unveiling New Insights: Reinterpreting DES Mutation, p. Arg383His , Through a Study of an Iranian Family With Isolated Hypertrophic Cardiomyopathy, Implication for Phenotype–Genotype Correlation Analysis. Clinical Case Reports, 2026. https://doi.org/10.1002/ccr3.73051 ↩ ↩2
-
Bakhtawar Farooq, Zahid Habib Qureshi, Muhammad Tahir Khan, et al. Alport Syndrome–Associated Pathogenic COL4A4 Variant in Sisters With Chronic Kidney Disease: Clinical Findings and Integrative Network Analysis. International Journal of Genomics, 2026. https://doi.org/10.1155/ijog/8868521 ↩
-
R Sathian, P Dutta, F Ay, et al. Genomic language model for predicting enhancers and their allele-specific activity in the human genome. Bioinformatics, 2026. https://doi.org/10.1093/bioinformatics/btag456 ↩
-
Emily C. Walsh, Pardis Sabeti, Holli B. Hutcheson, et al. Searching for signals of evolutionary selection in 168 genes related to immune function. Human Genetics, 2005. https://doi.org/10.1007/s00439-005-0090-0 ↩
-
Lara Berasain, Florencia Foitzick, Tomás Gusmerini, et al. Identification of tissue-specific RBFOX1 alternative promoters in the developing brain, heart and muscle. Scientific Reports, 2026. https://doi.org/10.1038/s41598-026-64263-8 ↩
Pothos builds longitudinal molecular profiles of individuals: whole-genome sequencing, single-cell RNA sequencing, and plasma proteomics, delivered as a detailed report, every data file, and stored cells. Build your profile.