What your genome can, and can't, tell you.

Polygenic scores, in plain words: what the number adds up and what it leaves out

A polygenic score is a weighted count of genetic variants, most of them common, ranked against a reference group. The arithmetic is simple. What the rank can tell you depends on how much of the trait the score explains, how close your ancestry is to the people it was built in, and how much of the score your file holds.

Verdict: A polygenic score is a rank on a curve built from other people's genomes. For most traits it explains a few percent of the differences between people, less outside European ancestry, and less again when a file holds only part of the score.

What a percentile means on a genetic report: a rank inside one reference group

A polygenic score card shows one number first, such as 90th percentile. It says where a file's score fell among a reference group of other people. It is silent on how likely the disease is, and it changes if the group changes.

Verdict: A percentile is a position in one reference group. A chance of disease needs the base rate and the score's measured effect, and even at the top of a published breast cancer score most women were not expected to develop it. Change the group and the same file ranks differently.

Confidence tiers on a DNA report: what Tier 1, 2 and 3 mean

A genome report can print a guideline-backed drug interaction and a contested personality gene in the same font. Ours sorts every health and trait section into three confidence tiers first. This is how the tiers are defined, which outside grades sit behind them, and what a tier never says.

Verdict: The tier tells you how much the science behind a section has been tested, before you read the result. It is fixed by section in the report's code, it keeps the outside grade it rests on, and it says nothing about your own risk.

Coverage and callability: what your file let us read, and what it did not

Every genome report has blanks. Some mean the position was read and you carry the common letter there; others mean nobody could look. Which one a blank is depends on the file, and on whether the file kept the evidence to tell the two apart.

Verdict: Depth says how much sequencing landed somewhere; callability says whether anything can be said there. A gVCF, BAM or CRAM carries the evidence for the second and a plain VCF or chip export does not, so the report labels every blank by the file that produced it.

Carrier status vs your own risk: what one copy of a recessive variant means

A carrier result reads like a warning about you. For most recessive conditions it is a fact about a gene you could pass on: one working copy is normally enough. There are known exceptions, and the report labels the results that are not plain carrier status differently.

Verdict: For most recessive conditions a carrier result describes a gene you could pass on, not your own health. A few carrier states have measurable effects, and they are known by name. Two copies, two changes and a single X copy get their own labels, and a chip gets no carrier result at all.

ACMG secondary findings: the 84 genes and why only those

When a clinical lab sequences an exome or a genome to answer one medical question, American guidance asks it to check a short list of other genes as well. Version 3.3 of that list, published in June 2025, has 84 genes. Why it stops at 84, and which genes it turned away, explains what such a result can and cannot tell anyone.

Verdict: ACMG's list is 84 genes long because each gene had to clear the same bar: a serious condition that is often silent until late, a variant likely to cause it, and a medical intervention that changes what happens. It is a minimum for clinical laboratories, and a clear result covers only what was looked for.

Celiac genes in your DNA file: what DQ2 and DQ8 can and can't rule out

Celiac disease almost never develops without one of two immune-gene types, DQ2 or DQ8. About three people in ten carry one, so finding it in a DNA file says little. Not finding it says a great deal, but only when the file could have shown it, and many files cannot.

Verdict: A DQ2 or DQ8 result is common and says little on its own. The informative result is a negative, and it holds only on a file that could have shown a positive: a chip on which both markers were read, or aligned reads typed for both genes. A whole-genome VCF can show a positive but cannot give that negative.

Carrier screening for couples: the one-in-four arithmetic and what a test can find

Most people carry a variant for at least one recessive condition and never know it. It starts to matter only when two partners carry a variant in the same gene. Screening programmes that test couples for over a thousand genes find that in about one couple in fifty, and Finland, with its own founder variants, does not screen couples.

Verdict: One couple in about fifty turns out to share a carrier gene when over a thousand genes are screened, and which genes are on the list decides whom a screen can find. In Finland that list has to include the heritage genes. A comparison of two consumer genome files can point at a shared gene, and it covers fewer genes and kinds of variant than a clinical screen.

How to download your raw DNA data from 23andMe, Ancestry, MyHeritage and others

Every DNA company keeps the raw data file in a different corner of its settings, behind a different check, and some send it days later by email. These are the routes at eight companies, read on their own help pages on 10 October 2026, and what each file is once you have it.

Verdict: The chip companies all hand over a compressed text file, three of them by emailed link with a deadline. The sequencing providers differ more: Dante offers VCF, BAM and FASTQ from the profile page, Nebula's terms put only the VCF in the portal, and Sequencing.com says its files can be downloaded from the account. Whatever arrives, the archive works here as delivered.

The Finnish Disease Heritage: 39 rare diseases, one bottleneck, and what a DNA file can show

Finland has inherited diseases of its own: 39 rare conditions that are more frequent here than in any other population. Each goes back to small groups of settlers, usually to a single variant that one ancestor carried. That history explains where the diseases cluster, and it decides which DNA tests can find a carrier.

Verdict: The Finnish Disease Heritage is population history written into a few dozen founder variants. Most are short changes that a whole-genome file reads well. None of this is clinical carrier screening, and Finland does not currently offer predictive carrier screening to couples.

Finnish DNA: haplogroup N, Siberian roots and the east–west line

Nearly two in three Finnish men carry a paternal line with deep roots in East Asia. Across the rest of the genome, Finns sit with other Europeans. Both statements are measured and true, and this guide is about the distance between them.

Verdict: Finnish DNA carries a strong eastern signal in the paternal line and a much smaller one in the rest of the genome. Each measure answers a narrower question than it seems to: a Y haplogroup follows one line, a composition compares a genome with today's reference groups, and runs of homozygosity record the size of the population behind it.

GLS2, SLC39A8 and SLC5A11: three genes behind Nightingale's new blood markers

Nightingale Health has added 20 measurements to its blood test, and its first genetic scans of them point at familiar genes. Three of those genes carry common variants that published studies tie to a blood level: GLS2 to glutamine, SLC39A8 to manganese and a long list of other traits, SLC5A11 to myo-inositol. Each moves a group's average by a small, measured amount, and most raw DNA files can read them.

Verdict: All three markers are replicated, point the same way in every study we read, and come with a plausible mechanism. All three are small. A blood test measures the level; the genotype accounts for a sliver of why one person's level differs from another's.

Hemochromatosis in your raw DNA: C282Y, H63D and the penetrance puzzle

C282Y and H63D, the two HFE variants behind most hereditary hemochromatosis, are read by 23andMe's health report and sit in many raw DNA downloads. Reading the letters is the easy part. What two copies of C282Y mean depends on which study you ask, because each counted something different, and the published answers run from under 1 percent to more than half.

Verdict: Two copies of C282Y is the HFE result that carries real weight, more for men than for women and more with age; its published penetrance runs from under 1 to 56 percent because studies counted different things. One copy of C282Y, two copies of H63D, or one of each has shown little or no excess disease in the largest cohort.

Gene tests for antidepressants: what the trials show, and what your DNA file already holds

Gene tests for antidepressants are sold as a faster way to the right drug. What most of them measure is how quickly the liver clears particular drugs. That part has evidence behind it, and if you have a 23andMe export or a sequenced genome, some of the answer may already be in the file.

Verdict: A gene test for antidepressants answers a narrower question than its marketing implies: how quickly a few liver enzymes clear particular drugs. That answer is real, the FDA and CPIC both act on it, and for CYP2C19 a 23andMe export or a genome file may already hold it. No gene test has been shown to say which antidepressant will work for a given person.

A pathogenic variant in your raw DNA data: why most rare chip calls are wrong

A raw-data tool has marked a variant in a DNA file as pathogenic. If the variant is rare and the file comes from a consumer genotyping chip, the measured odds are that the chip misread it. Here are the numbers, the reason, our own check of chip files against the same people's genomes, and what clinical confirmation involves.

Verdict: A consumer chip is right more than 99 times in 100 at common variants. At very rare pathogenic variants most of its positive calls are wrong and it misses most carriers, so a rare pathogenic call in chip raw data stays unconfirmed until a clinical test on a new sample repeats it.

SLCO1B1, ABCG2 and statins: what one gene says, and what the pair adds

SLCO1B1 is the best-established gene behind statin muscle side effects, and 23andMe and AncestryDNA files carry the position it is graded on. ABCG2 is the second transporter gene in the 2022 statin guideline, graded for one statin only. For that statin, rosuvastatin, CPIC reads the two genes as a pair, and ABCG2 changes the wording only when both copies carry its variant.

Verdict: SLCO1B1 carries the strongest known genetic signal for statin muscle problems, and its effect differs from statin to statin. ABCG2 changes CPIC's wording only for rosuvastatin, and only when both copies carry its variant: about one Finn in two hundred. Those cells, like every combined cell, rest on extrapolation from single-gene studies rather than on people studied with both variants.

Star alleles explained: how CYP2D6*4 and CYP2C9*3 become a metabolizer group

A pharmacogene result such as `CYP2D6 *1/*4` packs three steps into a few characters: which changes were found on each copy of the gene, what function those copies are given, and the group that function adds up to. Each step has rules, and in CYP2D6 the gene itself can be missing, doubled or fused with its neighbour, which no list of single positions can see.

Verdict: A star allele is a name for a set of changes on one copy of a gene, and `*1` is often what a test reports when it found none of the changes it looked for. The phenotype is CPIC's arithmetic on those names. For CYP2D6, the gene where copies are lost, doubled and fused, only a method that can count copies gives that arithmetic the right inputs.

VUS, ClinVar stars and why a genetic result can change

A variant of uncertain significance, or VUS, is the label a laboratory gives a DNA change when the evidence reaches neither verdict. It is the commonest classification in ClinVar, and most of the ones later resolved have turned out benign. This guide covers what the label means, what ClinVar's review stars measure, and why a genetic result carries a date.

Verdict: A VUS records missing evidence, and the evidence that arrives later has mostly pointed to benign. ClinVar's stars say how many independent reviews stand behind a classification, and for most records that is a single submitter. Both move with time, so a genetic result is only as current as the ClinVar release it names.

VCF, gVCF, BAM and CRAM: what each genome file can tell you

A sequenced genome usually arrives as a folder of files with three- and four-letter names. They are not copies of one another. Each step from the sequencer to the variant list keeps less than the step before, and whatever a file has dropped, no later analysis can get back from it.

Verdict: The most useful folder holds a gVCF rather than a plain VCF, with the BAM or CRAM beside it. Each step down from the reads is smaller and says less, and no later step can put back what an earlier one left out.

APOE4: the 'Alzheimer's gene', and what that phrase gets wrong

Two letters in a raw DNA file decide which version of APOE you carry. The studies behind the 'Alzheimer's gene' label are large and agree on direction. On size they disagree more than the label lets on, most of all for people with two copies of e4.

Verdict: APOE e4 is the strongest common genetic risk factor for late-onset Alzheimer's, and two copies raise the odds a great deal. It is still a risk factor, not a diagnosis: published figures for two copies run from 6 percent within ten years to about 90 percent in families already affected, and many people with Alzheimer's carry no e4 at all.

COMT and the 'warrior vs worrier' gene: a real enzyme, an oversold story

One letter of DNA, the story goes, decides whether you stay calm under fire or think sharply and fret. The enzyme behind it is real and the letter does change it. We followed the story from a 2001 card-sorting study to genome-wide studies of hundreds of thousands of people, and the personality test did not survive the trip.

Verdict: Real gene, real enzyme, real but small effect. It is not a personality test, and there is no good evidence for choosing supplements by COMT genotype.

DNA diet tests: what happened when trials checked the genotype

Genotype-based diet tests promise to say whether low-fat or low-carb suits your body, and that advice built on your genes works better. Two randomised trials tested those promises, one in California and one across seven European countries. Neither found the genotype doing what the tests sell it for.

Verdict: The two trials built to test this found no benefit from the genotype. In DIETFITS, the three-variant pattern proposed for choosing between low-fat and low-carb did not predict who lost more weight on which diet, and in Food4Me, adding five genotypes to personalised advice did not improve on the same advice without them.

MTHFR: the variant we refuse to report as a clinical finding

Wellness sites sell MTHFR testing as the hidden cause of fatigue, anxiety, miscarriage and heart disease, with methylfolate as the fix. We went through the evidence behind each claim, from half a million Finnish health records to the largest depression and ADHD studies ever run, and this is why our report will not flag it.

Verdict: C677T and A1298C are common, and two copies of C677T are ordinary: about 1 in 17 Finns and 1 in 9 other Europeans have them. The variant raises homocysteine a little, mostly where folate is scarce, but that rise has not shown up as heart disease, clots, miscarriage, depression or ADHD in the large studies built to find it. Mainstream genetics treats it as not clinically actionable, and so does our report.

Sports gene tests: what ACTN3 and ACE can and cannot say

ACTN3 is the gene every sports DNA test reads first, and the science behind it is real. We went through the athlete studies, the training trials and the 2015 expert consensus, and checked how common each genotype is in Finland. The variant is genuine. What it can tell one person is very little.

Verdict: ACTN3 R577X is real biology and the best-replicated finding in sports genetics: people with two X copies make no alpha-actinin-3, and they are scarce among elite sprinters. In ordinary people the variant accounted for about 2 percent of the spread in sprint time or strength gains, and the experts who reviewed the field in 2015 put the predictive value of these tests for training or talent at virtually zero.

BRCA1 and BRCA2: the measured cancer risks, and what a DNA file can show

BRCA1 and BRCA2 are DNA repair genes, and a pathogenic variant in either raises the risk of several cancers by an amount that has now been measured in thousands of carriers followed over time. The numbers are averages over groups, with ranges around them. Which variants a DNA file can show depends on the file: a consumer chip is unreliable here, and a sequenced genome reads the whole gene.

Verdict: A pathogenic variant in BRCA1 or BRCA2 raises cancer risk by a large and well-measured amount, but every estimate is an average over a group of carriers with a range around it. A consumer chip cannot settle whether someone carries one. A sequenced genome reads both genes, and a clinical laboratory confirms what it finds.

Clopidogrel and CYP2C19: what the gene changes, what the trials found

Clopidogrel (Plavix) is switched on by CYP2C19, and two no-function copies of that gene have carried an FDA boxed warning since 2010. The guideline is strict about carriers, the big trials are less clear, and a 23andMe or AncestryDNA file may or may not contain the gene.

Verdict: CYP2C19 makes the active form of clopidogrel, and carriers of no-function alleles make less of it. Whether acting on the genotype improves outcomes depends on the question asked. Switching carriers to ticagrelor was not statistically better in TAILOR-PCI, while POPular Genetics' genotype-guided arm, stepping noncarriers down to clopidogrel, recorded less bleeding. A consumer file can read the gene only when the array carries the right positions.

CYP2D6 and your medicines: star alleles, gene copies and what a DNA file can show

CYP2D6 is the liver enzyme that turns codeine into morphine and tamoxifen into endoxifen, and it clears a long list of antidepressants. CPIC uses it in its prescribing recommendations for nineteen drugs. Reading the gene is harder than reading almost any other on a prescribing list, because a person can carry no copy of it, one, two, three or more, and most DNA files cannot count.

Verdict: CYP2D6 matters to more CPIC guidelines than almost any other gene, and its result rests on a count of gene copies. A file that keeps the reads can settle that count; a chip or variant file cannot, and a CYP2D6 result built from one is a guess about the part that decides the answer.

DPYD and fluoropyrimidine chemotherapy: four variants, one enzyme and what a DNA file can show

DPYD makes the enzyme that clears most of a dose of fluorouracil or capecitabine. Four variants that weaken it are graded by CPIC and sit behind the EMA's 2020 recommendation to test before treatment. What follows is how that score is built, and which of the four a chip, VCF, gVCF or BAM file can show.

Verdict: A chip file shows almost none of the four DPYD variants, a variant file two, and aligned reads all of them, and even all four leave most early severe toxicity unexplained. A DNA-file result is background to the clinical test.

Factor V Leiden in your DNA file: rs6025, clotting risk and the numbers behind it

Factor V Leiden is the most common inherited clotting tendency in people of European ancestry, and a raw DNA file shows it whenever the file covers its one position. One copy roughly doubles to triples the rate of venous clots in the large Danish cohorts; two copies raise it further, by how much depends on the study. Most carriers never have a clot.

Verdict: Factor V Leiden is a well-measured, common risk factor. One copy carried a hazard ratio of 2.2 to 2.7 in the Danish general-population cohorts, two copies 7 to 18, and the absolute risk in any group depended more on age, weight, smoking, oestrogen and pregnancy than on the genotype alone. A raw file reads it reliably; what it means for one person is a clinical question.

Familial hypercholesterolemia genes: LDLR, APOB, PCSK9 and Finland's founder variants

Familial hypercholesterolemia is one of the commonest inherited conditions, and three genes explain most of it. In Finland a handful of old variants explain most of those, and one of the two commonest is a deletion of 8,535 letters that a chip export or a plain VCF has no way to show.

Verdict: Three genes and a handful of Finnish founder variants explain most familial hypercholesterolemia, but one of the two commonest Finnish variants is a large deletion, and a raw DNA file that lists only small variants cannot show it. A clear FH row covers what the file could carry.

Heart disease polygenic scores: what the studies found, and what one number can carry

Coronary artery disease is a disease of many genes: hundreds of common variants each nudge the odds a little. Scores built from them separate people at the tails, and FinnGen showed the effect in Finns. Added to a doctor's usual risk equation, they have moved its accuracy by about two hundredths, or not at all.

Verdict: Coronary artery disease scores pick out groups at the tails of the distribution whose risk does differ, more so in younger adults, while most people and most cases sit in the middle. Added to a clinical equation that already knows age, cholesterol and blood pressure, they have changed its accuracy by a couple of hundredths or not at all.

HLA-B27 in your DNA file: common in Finland, rarely a diagnosis

About 15 in 100 Finns carry HLA-B27, more than in most of Europe, and most carriers never develop ankylosing spondylitis. A DNA file can point to HLA-B27 through a tag marker. When the tag is missing, that says less than it seems.

Verdict: About one Finn in seven carries HLA-B27, and one to five percent of carriers develop ankylosing spondylitis. A tag marker finds the common European subtypes well and misses others, so a negative tag is weaker evidence than a typed negative. Our report reads neither.

Lipoprotein(a) in your DNA: what two SNPs can say about an inherited level

Lipoprotein(a) is a cholesterol-carrying particle whose level in the blood is set mostly by one gene, LPA. A raw DNA file can carry two LPA variants, rs10455872 and rs3798220, that mark a high level. Neither reads the repeat inside the gene that does most of the work, so what they say runs in one direction only.

Verdict: Lp(a) is the rare blood lipid whose level is mostly written in the genes, and two common variants can flag part of that inheritance from a raw DNA file. The part they cannot see, the number of KIV-2 repeats in LPA, is the larger part, so a flag means something and its absence settles nothing.

Lynch syndrome genes: MLH1, MSH2, MSH6, PMS2, and what a DNA file can show

Lynch syndrome is one name for inherited faults in four DNA repair genes, and the four do not carry the same risk. Prospective studies of thousands of carriers now give the numbers gene by gene. In Finland about half of all cases trace back to one deletion in MLH1, and neither a consumer chip export nor an ordinary variant file shows it.

Verdict: Lynch syndrome is four conditions under one name, and the gene sets most of the measured risk. A chip export cannot settle whether someone carries it, a sequenced genome reads the small variants in all four genes, and the Finnish exon 16 deletion needs a method built to see deletions.

Warfarin genetics: CYP2C9, VKORC1, CYP4F2 and what the dosing trials found

The stable warfarin dose differs about twenty-fold between patients, and three genes carry a large share of the spread: CYP2C9 clears the drug, VKORC1 makes the protein it blocks and CYP4F2 removes vitamin K from the cycle. Three randomised trials of genotype-guided starting doses gave three different answers.

Verdict: Three genes explain roughly half of why warfarin doses differ, and VKORC1 is the largest piece. The trials disagree because their comparison arms, outcomes, patients and gene panels differed. The usual two-allele CYP2C9 test misses much of the variation in people of African ancestry, where CPIC recommends against genotype-guided dosing on those results alone. In Finland, direct oral anticoagulants have had more users than warfarin since 2019.

The alcohol flush and ALDH2: what one variant does to acetaldehyde

A red face after half a glass is one of the few things a DNA file explains with a single letter. Most of the work is done by one variant in one enzyme, ALDH2, and it is close to absent in Finland. The same letter turns up in the cancer statistics of the countries where it is common, which is why it deserves a careful reading.

Verdict: ALDH2*2 is one of the best-understood variants in human genetics: one letter, a crippled enzyme, a visible reaction. Its link to oesophageal cancer runs through drinking, and the large cohorts that measured it found no excess risk in carriers who rarely drank. In a Finnish file it is almost always absent, and a chip may not even have looked.

Blood type from DNA: what a raw data file can and cannot say

Most ABO blood groups come down to three letters in one gene, so a genome file can usually name A, B or AB. Group O is harder, because the reference genome is itself an O, and the Rh factor harder still, because Rh-negative usually means a missing gene. This is what the letters say, how often they match a blood test, and what our report reads from each file.

Verdict: A genome file names A, B and AB well, and the best whole-genome method matched a blood test in all 200 people of its blinded test. Group O is the case a variant file cannot prove, because the O deletion is what the reference already says. RhD is mostly a missing gene, which a list of single positions cannot show.

Caffeine and CYP1A2: what the fast or slow label measures

One enzyme, CYP1A2, clears almost all the caffeine people drink, and the variant sold as the fast or slow caffeine gene sits inside it. What it changes is how far smoking or heavy coffee drinking can switch the enzyme up. The 2006 heart attack finding that made it famous did not hold in the largest cohort to test it.

Verdict: rs762551 is a real but conditional marker: its A allele lets CYP1A2 be induced further, which shows mainly in smokers and heavy coffee drinkers. The genome-wide studies found their caffeine signals at other variants, near CYP1A2 and near AHR, and the 2006 genotype-by-coffee heart finding did not replicate in UK Biobank.

Night owl genes: what CRY1, PER3 and 351 chronotype loci can tell you

Whether you are a morning or an evening person is partly inherited, and a 2019 study of 697,828 people found 351 places in the genome that nudge it. Each nudge is tiny. The two famous single-gene night-owl and early-bird variants, found in families, shift sleep by a few minutes in the general population.

Verdict: Chronotype is about half inherited in twin studies, but the inheritance is spread over hundreds of variants with effects of a minute or two each. Even the CRY1 and PER3 family variants shift sleep by only 5 to 7 minutes when found in the general population. A DNA file can rank a tendency; it cannot read anyone's body clock.

Eye colour genes: the HERC2 switch, and why green and hazel are hard to predict

Blue versus brown is one of the few looks a DNA file can call with some confidence, and most of that confidence comes from a single letter near a pigment gene on chromosome 15. Green and hazel are another matter: the largest study so far found 61 regions of the genome involved, and prediction tools still miss most intermediate eyes.

Verdict: One letter near OCA2 settles most of the blue-or-brown question, which is why a single-marker card is worth showing. Everything between blue and brown is spread over dozens of other regions that current prediction tools read poorly. The G allele is commoner in Finland than in any other group gnomAD reports, so for most Finnish files the card's answer will be the blue-associated one.

Lactose intolerance: the one variant most people actually have

Most adults worldwide stop making lactase after childhood. A single regulatory variant near LCT is why a minority keeps making it for life.

What a haplogroup is: one line of descent among thousands

A haplogroup names one branch of a family tree drawn from DNA, following the father's father's line or the mother's mother's back tens of thousands of years. Ten generations back there are 1,024 places in your family tree, and a haplogroup follows one of them.

Verdict: A haplogroup answers a narrow question precisely: which branch of the Y or mitochondrial tree one line of descent belongs to. Its name depends on the tree it was read against and its depth on how much the file can see.

Neanderthal DNA in your genome: how much there is, and how it is found

Most people with ancestry outside Africa carry about 2 percent Neanderthal DNA, and Africans carry some too. That figure is a population estimate that has moved with every new Neanderthal genome. What each person inherited is a scatter of short segments, and a few of them have effects that large studies have measured.

Verdict: Roughly 2 percent of a non-African genome came from Neanderthals, and the exact figure depends more on the method than on the person. The information is in the segments, a few of which have effects measured in large patient studies.

Runs of homozygosity: what identical stretches of DNA record, and why Finns carry more

Every genome has stretches where its two copies match letter for letter, because both came down from one ancestor. How long those stretches are says roughly how long ago that ancestor lived. Finnish genomes carry more of them than most European genomes, and genomes from the north and east of Finland carry the most.

Verdict: Runs of homozygosity are a record of how small and how separate the population behind a genome was, written in lengths. Finnish genomes carry the marks of a few small founding groups, most of all in the north and east. A single total mixes old history with recent history, which is why the report gives it as population history and nothing more.

Why your ancestry percentages differ from one company to the next

Send the same DNA to two companies and the percentages come back different. Neither is lying. Each compares your genome with its own reference panel, using its own method, and a percentage is a statement about that comparison more than about your ancestors.

Verdict: Different percentages from different companies usually mean different reference panels and methods, not a mistake in anyone's DNA. A percentage says which of a company's reference groups your genome resembles, and by how much; it cannot name a group the company never collected.

The Prothrombin row in every 23andMe file, and why we weren't reading it

Every 23andMe export we have tested carries a call for Prothrombin G20210A, one of the two common inherited clotting variants. Until 28 September our reader never found it. This is how a correct row went unread for three and a half months, and what the report says about the variant now.

Verdict: The call was in the file the whole time, under a name our reader did not look for. A rule we added to stop wrong answers also stopped this right one, and the fix had to meet the same standard of proof as the rule.

Lifting a genome from GRCh37 to GRCh38: what CrossMap quietly changes

Many genome files are still delivered on GRCh37 and are translated to GRCh38 before anything is read. On a real gVCF, CrossMap dropped an allele from about one variant call in nine, most of them deletions, and left the genotype pointing at an allele that was no longer there. Its log counts only records it cannot place, so none of this showed up there. Here is the mechanism, and the guard we run around it.

Verdict: A liftover can change a genotype and still report success. Count each record's alleles and its reference length before the lift, compare them after, and treat any record that changed as unread.

A missing line in a VCF is not a clean result

A variant file lists the places where a genome differs from the reference, and most software assumes the person matches it everywhere else. That is usually right. On real files we have watched it fail in three different ways.

Verdict: No record means "matches the reference the file was called against, if this position was read at all". Before a report turns that silence into a genotype, it has to know which reference, and whether the position was read.

Why we don't impute chip data: what imputation fills in, and what it guesses

Imputation turns a chip's few hundred thousand measured positions into tens of millions of genotypes, and large research studies are built on it. It is accurate for common variants and gets worse as variants get rarer. We measured what it would take to add it to our chip reports, found a defect of our own on the way, and did not build it. Here are the numbers.

Verdict: Imputation is a sound research method that is most accurate where a chip already does well and least accurate at the rare variants a report would most want it for. We measured the cost of adding it, ruled out the cheap targeted version on evidence, and never ran the accuracy check that would have to come first, so our chip reports read only what the chip measured.