You Found Your FoundMyFitness Genetic Report. Here's What It Can and Can't Tell You

Woolf Software
genomics genotyping vcf polygenic-scores wgs
A glowing grid-patterned glass chip inside a dark laboratory instrument, with some grid points lit and others dark, under a hovering robotic lens.

A FoundMyFitness genetic report is best understood as an annotation layer. You upload the raw text file from 23andMe or AncestryDNA. The service looks up a curated set of SNPs, the common single-letter differences between genomes, and it renders what published papers have said about those genotypes. The genotype calls belong to the array vendor, the interpretation is a literature summary, and neither one is a measurement anyone performed on you beyond the original chip run. That framing decides what you should do next. Treat the report as a reading list indexed by your genotype, verify anything you intend to act on, and understand that the file it read has hard limits no interpretation layer can fix.

Audit the input file before you read the output

Before you spend time on the interpretation, it is worth knowing how good the underlying data is, because everything downstream inherits its quality. Open your raw download and look at the structure. A 23andMe file is tab-separated with a comment header and four columns named rsid, chromosome, position and genotype. Coordinates are on GRCh37 and genotypes are reported on the plus strand. AncestryDNA splits alleles into two columns and uses 0 for no-calls. Both are flat text with no quality scores, no read depth and no allele balance. None of the information you would normally use to judge whether a call is trustworthy is present.

Three checks are worth running, in this order:

# no-call rate
awk 'NR>20 && $4 ~ /--/' genome.txt | wc -l
wc -l genome.txt

# probes per chromosome (spot mitochondrial and Y coverage)
awk 'NR>20 {print $2}' genome.txt | sort | uniq -c

# is the variant you care about even on the chip?
grep -P '^rs1815739\t' genome.txt

A healthy consumer array file has on the order of 600,000 to 900,000 lines and a no-call rate under about 2%. If your no-call rate is 5% or higher, the sample was marginal and every downstream call carries that weakness forward. Note also that the array has no notion of phase, meaning it cannot tell which parent contributed which allele, so any “compound heterozygote” claim in an interpretation tool is an inference rather than an observation.

The next step is to convert the file to VCF, the Variant Call Format that standard genomics tools expect, so that real tools can read it:

bcftools convert --tsv2vcf genome.txt -f human_g1k_v37.fasta -s ME -Oz -o me.b37.vcf.gz
bcftools norm --check-ref ws -f human_g1k_v37.fasta me.b37.vcf.gz -Oz -o me.norm.vcf.gz
bcftools index me.norm.vcf.gz

The --check-ref ws option warns on mismatches and swaps ref/alt where the file disagrees with the reference genome, so the warning count is informative. A handful of warnings is normal, while thousands point to a strand or build problem that needs resolving before you go further. If your annotation sources are built on GRCh38, as recent gnomAD releases and recent VEP caches are, lift the coordinates over with CrossMap or the UCSC liftOver chain rather than assuming they transfer unchanged.

The variants the report talks about, and the ones the chip never saw

Understanding what a fitness report can and cannot see comes down to which variants sit on the chip. Most of the SNPs these reports discuss are common and well covered: rs1815739 (ACTN3 R577X), rs9939609 (FTO), rs1801133 (MTHFR C677T), rs762551 (CYP1A2), rs8192678 (PPARGC1A Gly482Ser), and the VDR variants. Those appear on nearly every chip, and array genotyping of common SNPs is accurate.

The gaps are where the interesting problems live, and they fall into four groups:

  • ACE insertion/deletion, one of the most cited “endurance” variants, is a 287 bp Alu insertion in intron 16. Because it is not a SNP, arrays do not type it. Any tool reporting your ACE genotype from array data is relying on a linked proxy such as rs4343 or rs4341, with imperfect linkage disequilibrium between the proxy and the real variant.
  • APOE is defined by two SNPs, rs429358 and rs7412. Chip versions differ in whether both are present in the exported raw data, so grep for both before you believe any haplotype call. If one is missing, the call came from imputation or was never made at all.
  • Rare and de novo variants are invisible. Arrays genotype pre-selected sites, which is a sound design for common-variant association studies and a poor one for anything individually rare. The per-generation human point mutation load is on the order of tens of new variants per genome, none of which a fixed probe set can see 1.
  • Structural variants, repeat expansions, and most of the noncoding genome are absent entirely. Whole-genome sequencing is what moves you from a probe list to the full variant spectrum, including the noncoding regions where most trait-associated signal sits 2.

Given all of this, any rare variant an array does report should be treated as a hypothesis until it is confirmed by an orthogonal method. Clinical laboratories do not act on raw consumer array calls, and the same caution applies to you.

Effect sizes are the part most reports underplay

The biology behind these associations is real, and the effect of any single variant is also small. Physical performance and trainability are polygenic, involving hundreds of contributing loci and heavy environmental modification 3. Family studies of exercise training response established decades ago that the response to a standardized program is itself heritable, roughly half, while individual genotypes explain a few percent at best 4.

So when a report says a genotype is “associated with” lower endurance or higher injury risk, the accurate reading is narrow. In a cohort of thousands, carriers averaged slightly different values, with distributions that overlap almost completely. ACTN3 XX homozygotes are underrepresented among elite sprinters and are also common among people who sprint well. Predictive value from a personal genome is real for a small set of high-penetrance variants and modest for nearly everything else 5.

If you would rather compute a polygenic score yourself than read a summary of one, you can pull a scoring file from the PGS Catalog and run:

plink2 --vcf me.norm.vcf.gz --score PGS000XXX.txt 1 2 3 header \
       cols=+scoresums no-mean-imputation \
       --out me_prs

Use no-mean-imputation so that missing genotypes are not silently filled with the population mean, a substitution that inflates your apparent score toward average and hides how much of the model your file failed to cover. Then check the ALLELE_CT in the output: if a 1.2-million-variant score matched only 400,000 of your sites, the resulting number is not comparable to the published distribution. Keep in mind too that a single-sample score is meaningless without a reference distribution, so score the 1000 Genomes samples from the same ancestry group and place yourself within that distribution 6. Scores derived in European-ancestry cohorts transfer poorly to other ancestries, which is a property of the training data rather than of you.

The pipeline we would run

If you want to annotate an array file properly, the following three steps cover what the data can support. The first assigns functional consequences, the second adds clinical interpretation where one exists, and the third provides a population-frequency sanity check:

# functional consequence
vep -i me.norm.vcf.gz --cache --assembly GRCh37 --everything \
    --vcf -o me.vep.vcf

# clinical interpretation, where one exists
bcftools annotate -a clinvar.vcf.gz -c INFO/CLNSIG,INFO/CLNREVSTAT \
    me.vep.vcf -Oz -o me.annot.vcf.gz

# population frequency sanity check
bcftools annotate -a gnomad.genomes.sites.vcf.gz -c INFO/AF ...

Two filters keep the output honest. Restrict ClinVar hits to two-star review status or better. That means CLNREVSTAT contains criteria_provided,_multiple_submitters, so you are looking at assertions multiple submitters agree on. And treat anything with a gnomAD allele frequency above about 1% as not a rare pathogenic variant, regardless of what a single old paper claimed, since genuinely damaging variants rarely reach that frequency in the population.

For noncoding sites, where consequence prediction has nothing useful to say, per-position fitness-consequence scores such as fitCons give a genome-wide estimate of whether a position is under selective constraint. That is a far better prior than assuming a variant in an intron cannot matter 7.

Our view is that if your goal is to work with your own molecular data rather than read someone else’s curation of it, an array file is the wrong substrate to build on. Sequence once at 30x, meaning each position of the genome is read about thirty times on average, which is enough depth for confident genotype calls. You get a CRAM file of the aligned reads and a joint-called VCF with genotype quality and depth recorded per site, and every future annotation then runs against that same file. From there, pair it with measurements that change over time. Those include RNA expression, proteins, blood biomarkers and glucose. Genotype tells you about priors. The longitudinal layers tell you what your body is doing this year.

Where a clinician is required

There is a clear line past which this kind of self-directed analysis should stop. If your report flags a variant in a gene with real clinical consequence, take it to a genetic counselor or physician, who will order a validated clinical test rather than interpret a consumer file. We do not interpret variants for medical decisions, and neither should an annotation website. Nothing here is advice about supplements, drugs, or doses.

Questions people also ask

Is Dr. Rhonda Patrick a medical doctor? No. She holds a PhD in biomedical science and did postdoctoral research in aging and metabolism. She is not a physician and does not practice medicine, which is why FoundMyFitness frames its output as educational literature summaries rather than clinical interpretation.

Is FoundMyFitness legitimate? As a literature-annotation product, yes. It cites real papers and tells you which SNPs it could not find in your file, which is more transparency than most competitors offer. Its limits are the limits of the array underneath it and of small per-variant effect sizes.

Are fitness DNA tests accurate? The genotyping is accurate for common SNPs. The predictions are weak, because the traits are highly polygenic and strongly modified by environment 3. In short, accurate genotype, low predictive value.

Which DNA test is best for fitness and health? Whole-genome sequencing at 30x. It covers rare variants, structural variants and noncoding regions that arrays miss entirely, and you keep a file that survives every future reanalysis 2.

Was 23andMe banned in the US? No. The FDA sent a warning letter in 2013 that halted its health reports until specific authorizations were granted starting in 2015. The company later filed for bankruptcy in 2025, and its assets were acquired by a nonprofit founded by its former CEO, which is what prompted the wave of data-deletion questions.

Should I delete my data from 23andMe? Download your raw file first, then decide. Deletion requests and sample-destruction requests are separate actions in the account settings. Once you have the .txt locally, everything in this post runs offline.

Woolf Software 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. Peter D Keightley. Rates and Fitness Consequences of New Mutations in Humans. Genetics, 2012. https://doi.org/10.1534/genetics.111.134668 ↩

  2. 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

  3. Claude Bouchard. The Human Genome, Physical Activity, Fitness, and Health. Kinesiology Review, 2022. https://doi.org/10.1123/kr.2021-0057 ↩ ↩2

  4. Eric J. Devor. Genetics of fitness and physical performance. American Journal of Human Biology, 1998. https://doi.org/10.1002/(sici)1520-6300(1998)10:6<809::aid-ajhb14>3.0.co;2-8 ↩

  5. Brandon L Pierce, Habibul Ahsan. Clinical assessment incorporating a personal genome. The Lancet, 2010. https://doi.org/10.1016/s0140-6736(10)61404-3 ↩

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

  7. Brad Gulko, Melissa J Hubisz, Ilan Gronau, et al. A method for calculating probabilities of fitness consequences for point mutations across the human genome. Nature Genetics, 2015. https://doi.org/10.1038/ng.3196 ↩