Whole Genome Sequenced and Never Used? What to Do With It

The Short Answer

A question keeps coming up among people who take longevity seriously: my whole genome has been sequenced, sometimes more than once, and I have never used it for anything. What am I missing? Some of them were sequenced a decade ago with short reads and again recently with long reads too, at real expense, and the files have sat untouched both times.

What's missing is a question asked at the right moment. A genome doesn't do anything by being sequenced. It answers questions, and nothing in a typical longevity routine asks it one: not the dashboard, not the clinic, not the company that sold the sequencing.

TL;DR: A genome is a reference book, not a feed: it answers questions, and they have to be asked when a decision is being made. Some are worth asking once: actionable variants (about 1–3.4% of adults in large cohorts carry one, depending on ancestry and the genes checked), drug response and carrier status. Others come up each time blood is drawn, because a genotype can change how a result reads, as it can for HbA1c, bilirubin, neutrophils and PSA. A genome sequenced from blood also records somatic changes of ageing. What it can't tell anyone is how long they'll live.

Educational information, not medical advice. Anything here that could change care belongs with a clinician, a genetic counsellor or a pharmacist. DeepDNA doesn't analyse genome files; this is a guide to what a genome can be asked, and who should answer.

Why Nothing Happened

The 2012 genome was read before most of the dictionary existed

Sequencing turns DNA into data. Interpreting it means looking each variant up in references that say what is known about it. In 2012 there were some: dbSNP, the 1000 Genomes Project, PharmGKB and the first CPIC drug guidelines. Most of what a clinical genomics team uses today came later.

So the 2012 data, if it was interpreted once in 2012, was interpreted against far less than exists today. And interpretation keeps moving. In a JAMA study of 1.45 million people tested for hereditary cancer genes, 7.7% of the distinct variants of uncertain significance were reclassified over the study period; 91.2% of those moved down to benign or likely benign, and 8.7% moved up (Mersch 2018). Reanalysis finds bad news sometimes, but in that study most of what changed was reassuring.

A genome is a reference, not a feed

A longevity routine runs on feedback loops: measure, intervene, measure again. Glucose, ApoB, VO₂max, sleep, an epigenetic clock. Each number moves, so each number gets attention.

The inherited genome doesn't move. No protocol changes it, so it never shows up as a trend, and anything that never moves drops off a dashboard. A dashboard is the wrong place for it anyway. A genome belongs where a reference book belongs, next to decisions: before starting a drug, before anaesthesia, when a biomarker comes back out of range, when a relative is diagnosed with something. Those are the moments it has answers for, and none of them happens on a schedule.

Nobody owns the question

A clinical laboratory reports what was ordered. A consumer service built for genotyping chips reads a few hundred thousand pre-selected positions from a text file (how that works), not a genome. A whole genome arrives as tens of gigabytes of aligned reads per run, plus variant files. Turning that into answers takes bioinformatics pipelines, current databases, and a clinician or genetic counsellor for anything that could change care. When the person who owns the data, the lab that produced it and the doctor who might use it each assume one of the others will do the reading, nobody does it.

For a healthy, well-measured person, "nothing to do" is a common answer

MedSeq, a small randomised pilot trial, gave healthy primary-care adults aged 40 to 65 a genome report or not. Of the 50 who received one, 11 (22%) learned of a monogenic disease risk they didn't know about, and 2 of those 11 (4% of the 50) already had signs of the condition (Vassy 2017). Every sequenced participant also carried at least one variant for a recessive condition, with a median of two.

In that small trial, most people found something to know and little to do. That isn't a failure of the genome. "Nothing actionable in the 84 genes" is a result too, and the only way to get it is to look.

Questions to Ask Once

Is there anything in the actionable genes?

The ACMG secondary-findings list, version 3.3, names 84 genes where a pathogenic variant should be reported because something can be done about it: 28 cancer genes, 41 cardiovascular, 7 metabolic (including HFE, reportable only for C282Y homozygotes) and 8 others, among them RYR1 and CACNA1S for malignant hyperthermia under anaesthesia and TTR for amyloidosis (ACMG SF v3.3, full text).

How often does an unselected adult carry one?

• In UK Biobank, 2% of 49,960 people had a medically actionable variant in the 59 genes listed at the time (Van Hout 2020).
• In Geisinger's MyCode, 3.4% of 175,500 people carried a pathogenic variant in 81 genes, and 87.6% of those told hadn't known about their risk (Savatt 2025).
• In All of Us, 2.26% of participants of European ancestry, 1.62% of African ancestry and 1.32% of Latino ancestry, across 73 genes (Venner 2024).

Family history is a poor substitute. In the Healthy Nevada Project, 1.33% of participants carried a variant for one of the three CDC Tier 1 conditions (hereditary breast and ovarian cancer, Lynch syndrome, familial hypercholesterolaemia). Ninety per cent had not been identified before, and only 25.2% reported a relevant family history (Grzymski 2020).

For someone who already measures everything, a useful distinction is which of these bloodwork would never show. Familial hypercholesterolaemia shows up in an LDL result. A BRCA2 variant, Lynch syndrome, an inherited arrhythmia or malignant hyperthermia susceptibility does not. Some of the cardiovascular genes matter for training, too: in carriers of desmosomal variants linked to arrhythmogenic right ventricular cardiomyopathy, endurance exercise was associated with earlier and more severe disease (James 2013).

Two rules apply to anything found. A research-grade call from a self-ordered genome has to be confirmed by an accredited clinical laboratory before anyone acts on it. And the result goes to a genetic counsellor or clinician, because penetrance varies widely and many people with a "pathogenic" genotype never develop the condition.

How might I respond to drugs?

This is the part of a genome almost everyone can use. In UK Biobank, 99.5% of participants carried a genotype in 14 well-studied genes that predicts a possibly atypical response to at least one drug, about 10 drugs on average, and nearly a quarter had already been prescribed one of them (McInnes 2021). The Estonian Biobank found 99.8% (Reisberg 2019).

Whether acting on it helps was tested in the PREPARE trial: 6,944 patients in seven European countries, a 12-gene panel, prescribing guided by the Dutch DPWG guidelines. Among patients with an actionable result, clinically relevant adverse drug reactions fell from 27.7% to 21.0%, about 30% lower odds (Swen 2023). The trial was open-label, and critics argue that inflated the effect (Rogers 2023). It remains a large randomised test of pre-emptive panel testing.

Here is how that maps onto drugs that come up in a longevity regimen:

The drugs a longevity stack is built on mostly have no pharmacogenomic guidance. The value sits in the drugs nobody plans for: the painkiller after an injury, the antiemetic and the anaesthetic around a procedure, the antidepressant in a bad year, the antiplatelet after a stent, and, rarely but importantly, chemotherapy. The boxed warnings on the US labels for capecitabine and fluorouracil were revised in late 2025 and early 2026 to advise testing for DPYD variants before starting, unless treatment can't wait (FDA).

This is also where long reads earn some of their price. CYP2D6, which metabolises many antidepressants, opioids and antiemetics, sits next to a near-identical pseudogene and comes in deletions, duplications and CYP2D6–CYP2D7 hybrids (PharmVar). PharmCAT, the reference tool for turning a VCF into CPIC phenotypes, does not call CYP2D6 (except in an experimental research mode) or HLA from a VCF, and takes them as "outside calls" from dedicated tools (PharmCAT docs). Those exist for short-read genomes, such as Cyrius (Chen 2021) and StellarPGx (Twesigomwe 2021), and for PacBio HiFi reads, StarPhase (Holt 2024, preprint), which is for research use only. Our Python walkthrough shows the path from a genotype to a CPIC phenotype for two simpler genes. Phenotypes from any of these tools are research results until a clinical lab confirms them.

What do I carry for my family?

Every participant sequenced in MedSeq carried at least one recessive variant. The ACMG recommends offering carrier screening for 113 genes to anyone pregnant or planning a pregnancy: 97 autosomal recessive genes with a carrier frequency of at least 1 in 200, plus 16 X-linked genes (Gregg 2021). A genome holds most of that answer; a few genes, such as SMN1, HBA1/2 and CYP21A2, need specialised callers or a separate test.

A dominant finding is a family finding as well. In one study of positive exome results, 47% of the relatives who were tested carried the variant too, yet relatives were tested after only 7.6% of the positive results (Stefka 2023). A genome may end up being most useful to someone else.

Questions to Ask Every Time You Draw Blood

This is the part most people miss, and it matters most to people who measure a lot. A genotype doesn't only predict disease. It can change what a routine number means.

These are population findings, often from one ancestry group, and they don't carry over automatically to everyone. They explain why a result can sit somewhere other than a population range would suggest. What any one result means is still a clinician's call.

Some genotypes make a number less informative. In carriers of a G6PD-deficiency allele, HbA1c can read lower than average glucose would explain, because red cells live shorter lives. Clinicians aware of that read HbA1c alongside direct glucose measurements rather than on its own.

Some genotypes offer an explanation. A mildly raised unconjugated bilirubin is common in people with the Gilbert genotype, although a genotype can't rule out liver disease. Nor is it a longevity advantage: high bilirubin goes with lower cardiovascular risk in observational data, but a Mendelian randomisation study found no causal effect (Stender 2013). In the Duffy-null study, 23.8% of healthy Black adults with that genotype had counts below 2,000 cells/µL, a level usually flagged as neutropenia; the authors proposed the term "Duffy-null associated neutrophil count" (Merz 2023).

For some markers, the blood test beats the genome. Lp(a) is almost entirely inherited, which is why one measurement is enough. A genome adds little once the number is known, and the KIV-2 repeat that drives it is hard to read from short reads.

And a person's own history is a reference range too. A 2025 study in Nature found that a healthy adult's nine complete-blood-count setpoints are distinguishable from those of 98% of other healthy adults and persist for at least 20 years; in retrospective analyses, personal reference ranges detected several conditions better than population ranges (Foy 2025). For someone with years of bloodwork, the range printed on the report is one reference; their own baseline and their genotype are context a clinician can read it with.

Genetics shapes almost every routine biomarker: a study of 35 blood and urine biomarkers in 363,228 UK Biobank participants found 1,857 associated loci (Sinnott-Armstrong 2021). That is why we think the value lies in reading bloodwork against the genome, not beside it. We've written about how agents misread lab reports and what structured bloodwork needs.

The Part That Changes: What a Blood Sample Recorded

The inherited genome is fixed. The DNA in a blood sample isn't, quite. With age, some blood stem cells acquire mutations that let them out-compete their neighbours, and their descendants come to make up a measurable share of circulating blood. Two genomes sequenced from blood years apart, say one in 2012 and one in 2026, hold a record of that process at two points in time. Very few people have one.

Clonal haematopoiesis (CHIP)

Clonal haematopoiesis of indeterminate potential means a blood-cancer-associated mutation carried by at least 2% of the sequenced DNA in someone without a blood cancer (Steensma 2015). It is rare before 40 and present in 9.5% of people in their seventies. Carriers had about 11 times the rate of blood cancer, 1.4 times the rate of death from any cause and twice the rate of coronary heart disease (Jaiswal 2014). A later study found 1.9 times the risk of coronary heart disease and 4 times the risk of an early heart attack (Jaiswal 2017). Progression to blood cancer remains uncommon, around 0.5–1% a year (Steensma 2015).

Sequencing depth decides what a genome can see. For their study of 97,691 genomes, mostly around 40x, Bick and colleagues tested detection by down-sampling deeply sequenced samples: mutations carried by more than 10% of the DNA were detected well, about half of those at 5–10%, and few below 5%. Reliably catching the 5–10% range takes about 100x (Bick 2020). So a 100x short-read genome can see mid-sized clones, while a genome sequenced at standard depth, as most were in 2012, could only have shown large ones.

CHIP is also one of the few results in a genome that may respond to how someone lives. In UK Biobank and All of Us participants wearing accelerometers, each extra hour a day of moderate-to-vigorous activity went with 14% lower odds of clonal haematopoiesis outside the DNMT3A gene (Gerhardt 2026). Diet quality showed a similar gradient (Bhattacharya 2021). These are associations. No randomised trial has shown that any intervention shrinks CHIP clones or prevents their consequences, though colchicine and IL-1β blockade have exploratory signals in TET2-driven CHIP (LoDoCo2 analysis, CANTOS analysis).

Loss of chromosome Y

In men, some blood cells lose the Y chromosome with age. Detectable loss rose from 2.5% of UK Biobank men at 40 to 43.6% at 70, and it was more common in smokers (Thompson 2019). It has been associated with shorter survival, and in five men sampled 6 to 14 years apart the affected share of cells grew over time (Forsberg 2014). In mice it causes cardiac fibrosis, and in UK Biobank men with loss in more than 40% of white blood cells the rate of cardiovascular death was 1.31 times higher (Sano 2022). Whole-genome data detects it more sensitively than genotyping arrays (Jakubek 2025).

What not to compare

Telomere length can be estimated from sequencing reads, but in a 2,398-genome subset of the TOPMed study, the sequencing plate explained 21.9% of the variation in uncorrected estimates and age only 8.0% (Taub 2022). Our reading is that a change over 10 or 15 years, measured across two sequencing chemistries, would be lost in that noise. Mitochondrial DNA copy number falls with age, by about 2% per decade, but blood cell composition explains about 23% of it (Gupta 2023). Neither is worth comparing between a 2012 run and a 2026 one.

Bulk sequencing also misses most somatic change. Each blood stem cell picks up about 17 new mutations a year, but those only become visible once a clone has expanded (Mitchell 2022).

Methylation in the same files

PacBio HiFi and Oxford Nanopore instruments read DNA methylation natively and store it as tags in the BAM file. If a provider delivered only FASTQ files, those tags are probably gone, so it's worth asking for the BAMs. Clocks built directly on long-read methylation are new: a deCODE clock trained on nanopore data from 7,284 blood samples predicts chronological age with a median error of 2.43 years (Sigurpalsdottir 2026). We found no published validation of GrimAge, PhenoAge or DunedinPACE computed from long-read genomes; those clocks were built on methylation arrays. Any long-read "biological age" is research for now.

What the Long Reads Bought

Short reads are accurate base by base and cheap. Long reads see the parts of the genome short reads struggle with, and a matched pair gets most of both.

The biggest difference is structural variation. Long-read assemblies find about 25,000 structural variants per genome against about 11,000 from short reads, and 91% of the deletions only long reads find sit in repeats and segmental duplications (Zhao 2021). Medically important genes live in those regions. Short reads leave dark stretches in the coding exons of 748 genes (Ebbert 2019), and in 100 samples sequenced at 30x HiFi, callers found 83% of 145 clinically relevant variants that are hard for short reads automatically, with 93% visible in the data (Höps 2025). Genes such as SMN1/SMN2, PMS2, GBA and CYP21A2 have a dedicated long-read tool (Paraphase).

Repeats are the second. More than 60 Mendelian diseases are caused by short tandem repeat expansions (STRchive), and long reads can size them directly (TRGT). The third is phase: whether two variants in the same gene sit on the same copy or on different copies, which for recessive conditions decides whether they matter. Short reads usually need a parent's DNA to settle it; long reads often settle it alone.

Combined, the two are more accurate than either. On a reference sample, DeepVariant's hybrid PacBio-plus-Illumina model scores 0.9992 for single-nucleotide variants, against 0.9985 for PacBio alone and 0.9961 for Illumina alone (DeepVariant metrics).

None of it is used unless the data goes through a pipeline that calls those variant types. A VCF of small variants from the short-read run is a fraction of what was paid for.

What It Won't Tell You: How Long You'll Live

This is probably the expectation that disappoints people most.

Estimates of the heritability of lifespan depend heavily on method. Classic Danish twin data gave 23–26% (Herskind 1996). Genealogies covering millions of people gave 16% (Kaplanis 2018), and once assortative mating was accounted for, "well below 10%" (Ruby 2018). In January 2026 a study in Science argued that the heritability of intrinsic lifespan, with deaths from external causes modelled out, is about 50% (Shenhar 2026). Graham Ruby, first author of the 2018 assortative-mating study, countered that this "changed the trait from overall lifespan to lifespan under a specific set of conditions" (STAT).

Even if half of intrinsic lifespan is inherited, little of it can be read from a genome today. In a study of about a million parental lifespans, each of 12 associated variants was worth between 3 and 13 months per copy; APOE ε4 costs about a year per copy, and the gap between the top and bottom tenth of a polygenic score was about five years, "quite small compared with variability in individual lifespans" (Timmers 2019). People in their late eighties and nineties carry as many common risk alleles as young controls (Beekman 2010), and healthy over-80s as many rare pathogenic variants as anyone else (Erikson 2016).

Polygenic risk scores for disease are real but modest. A score for coronary artery disease identifies 8% of people at three-fold risk or more (Khera 2018). But adding polygenic scores to standard risk factors that already include cholesterol raised the C-index for cardiovascular disease, a 0.5-to-1 measure of how well a model ranks who gets ill, by 0.012 in UK Biobank (Sun 2021), and scores are on average 4.9 times less accurate in people of African ancestry than in Europeans (Martin 2019). For someone who already tracks lipids closely, a score adds only modestly to what they measure. More in our guide to polygenic risk scores.

The one common variant worth knowing is APOE. About 2% of people are ε4 homozygotes, and for them Alzheimer's disease behaves almost like a distinct genetic form, with abnormal amyloid in nearly all of them by 65 (Fortea 2024). That is a population finding, not a forecast for one person: estimates of lifetime risk by 85 are around 50–60% (Genin 2011). It has a direct treatment consequence too: the US labels of lecanemab and donanemab say APOE ε4 status should be tested before starting, because homozygotes have a much higher rate of the brain swelling and bleeding called ARIA (Leqembi label, Kisunla label). This is a result to receive with a genetic counsellor, not from a dashboard. More on our APOE gene page.

What a Proper Reanalysis Involves

None of this happens by itself, and most of it isn't a do-it-yourself job. A clinical genomics service, or a bioinformatics team working with one, would start from the aligned reads (BAM or CRAM) rather than a report, keeping the methylation tags of any long-read run, and re-call everything on a current reference with today's pipelines: nf-core/sarek or nf-core/raredisease for short reads, the PacBio WGS Variant Pipeline for HiFi (it bundles DeepVariant, Sawfish, TRGT, HiPhase, Paraphase and StarPhase), and wf-human-variation for nanopore. Where there are two genomes of the same person, their inherited variants should agree, which is a useful first check.

From there the work follows the questions above: actionable genes, pharmacogenes and carrier status, each ending in clinical confirmation and a conversation with a counsellor, clinician or pharmacist; a list of the genotypes that change how the person's tracked biomarkers read; and somatic calls treated as research until targeted sequencing confirms them. Then the interpretation is re-run every year or two, because ClinVar, the ACMG list and the guidelines keep changing, and most reclassified variants of uncertain significance move towards benign (Mersch 2018).

Where an AI Agent Fits

Look at the moments when a genome has something to say: a new prescription, a procedure, an out-of-range result, a relative's diagnosis. More of those conversations now pass through an AI agent, which makes the agent a natural place to put the question to the genome, provided it answers from cited records rather than memory and is never handed the whole genome.

That's the layer DeepDNA builds, so to be precise about today: DeepDNA doesn't accept, process or store genome files, whether BAM, CRAM or VCF. What's live is a free knowledge API and an MCP server, with no key. An agent can screen a medication list for curated drug–gene records and look up genes such as G6PD, UGT1A1, HFE and CYP2D6 with their sources. The server takes gene names, variant IDs (rsIDs) and drug names, never your genotype, and doesn't tell you a phenotype or a metabolizer status. Anyone with Claude, Cursor or another MCP client can connect it and ask, for instance, whether a medication list has pharmacogenomic guidance.

Our /crossover endpoint is designed to return which genetic predispositions have biomarkers outside the lab's own range, inside it, or unmeasured. It's in private beta and not deployed. The API docs list what is live; connecting the MCP server takes one line, and anyone building on genomic and lab data can join the private beta.

That is also how we design for privacy: the files stay with their owner, a tool should receive only the positions a question needs, and the parsing endpoints, in private beta and not yet deployed, are designed to keep nothing once they've answered. A genome is the most identifying data a person has, and under GDPR it is special-category data; DeepDNA is based in Spain, under Spanish and EU law. With any service, ask what it keeps, where it processes, and whether anything is trained on it.

Frequently Asked Questions

What should I do with my whole genome sequencing data?

Ask it specific questions when they matter: once for actionable genes, drug response and carrier status, with clinical confirmation of anything found; whenever bloodwork comes back, whether a genotype changes how a result reads; and every year or two, a fresh interpretation, because the databases change.

Is long-read sequencing worth it over short-read?

For what short reads answer poorly, yes: structural variants, repeat expansions, genes such as SMN1, PMS2 and CYP2D6, HLA, phasing without parents, and methylation from the same run. It only pays off if the data goes through pipelines that call those variant types.

Can my genome tell me my biological age?

Not the inherited sequence. Long-read data can carry DNA methylation, and clocks built directly on it are emerging, but widely used clocks such as GrimAge and DunedinPACE were built on arrays and haven't been validated on long-read genomes. Somatic signals such as clonal haematopoiesis and loss of chromosome Y are linked to ageing outcomes, but they are research measures, not an age.

Can I compare two genomes sequenced years apart?

For clonal haematopoiesis and loss of chromosome Y, with care: depth and chemistry limit what each run detects, and calls need confirming by targeted sequencing. For telomere length and mitochondrial copy number, no: sequencing batch and blood cell composition outweigh years of change. The inherited variants should match, which confirms both files come from the same person.

This article is for educational purposes only and does not constitute medical advice. Genetic findings that could affect care should be confirmed by an accredited clinical laboratory and discussed with a clinician or genetic counsellor; decisions about medicines belong to a prescriber or pharmacist.

This article was created with AI assistance and reviewed by the DeepDNA editorial team.