Learn Under the hood
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.
Key takeaways
- Imputation infers unmeasured genotypes from stretches of DNA a person shares with a reference panel of sequenced genomes[2]. Its output file contains a genotype at every imputed position, whether the inference was strong or weak[6].
- Imputation quality (Beagle's DR2, Minimac's Rsq, the INFO score) is the software's own estimate for a variant across a batch of samples, not a check of any one person's call[6].
- With the HRC panel of 32,390 genomes, mean r² against sequencing was 0.98 for common variants and 0.77 for variants rarer than 0.5 percent[2].
- Chips already misread very rare variants: in UK Biobank 16 percent of chip calls below a frequency of 0.001 percent were confirmed[12]. Imputation does not repair that range.
- Our measurement on 18 real chip exports found a coverage defect in our own polygenic-score code, now fixed. Even after the fix the median chip held under half of every score, and the accuracy check that imputation would need first was never run.
A consumer genotyping chip reads a fixed list of positions: between 563,320 and 955,958 rows in the ten exports we measured[1]. Imputation estimates the genotypes in between by finding stretches of DNA that the person shares with a reference panel of sequenced genomes[2]. Large studies depend on it. UK Biobank genotyped its participants on an array and used imputation to raise the number of variants it could test to around 96 million[3]. How well it works depends on how common a variant is and on how many people like the person are in the panel.
How imputation fills a gap
Chromosomes are inherited in long pieces. Two people who share a stretch of DNA from a distant common ancestor carry nearly the same letters along that whole stretch, both at the positions a chip reads and at the positions in between. Imputation exploits this. It compares the measured positions of a person's chromosome with thousands of sequenced chromosomes in a reference panel, finds the panel chromosomes that match each stretch best, and copies the unmeasured letters from them[2].
The software moves along the chromosome with a statistical model that lets the best-matching panel chromosome change wherever two inherited pieces may have been joined. Beagle and Minimac are two such programs. When Beagle 5.0 was published, its authors compared it with Beagle 4.1, IMPUTE4, Minimac3 and Minimac4 and found nearly identical accuracy across all of them; what differed was speed[4]. The quality of the answer is set mostly by the panel: how many chromosomes it holds, and how closely they are related to the person being imputed[2].
Four terms this article relies on
- Reference panel
- A set of sequenced genomes, split into their two chromosome copies, that imputation copies from. The 1000 Genomes Project's 30x panel has 3,202 people[5].
- Allele frequency
- How common a variant is: the share of chromosomes in a population that carry it. The imputation papers cited here call a variant rare below 0.5 percent and low-frequency from 0.5 to 5 percent[2].
- Dosage
- The imputed count of the alternative letter, a number from 0 to 2 that need not be whole. Beagle writes it in a
DSfield[6]. - Imputation quality
- The software's estimate of how well its imputed dosages track the true genotypes at one variant, on a scale from 0 to 1.
What an imputation quality score measures
Each program reports one quality number per variant under its own name. Beagle writes DR2 into the file, Minimac and the Michigan Imputation Server call theirs Rsq or r², and other pipelines report an INFO score. Beagle's manual defines its number in one line.
A “DR2” subfield with the estimated squared correlation between the estimated allele dose and the true allele dose
Two words in that sentence matter. Estimated: the true dose is unknown, so the score is the model's own forecast. Correlation: a correlation is computed across many samples, so the score describes a variant in a batch of people, not the call for any single person. A DR2 of 0.9 says the imputed doses track the truth well across the batch. It does not say that a given person's genotype at that position is 90 percent likely to be right.
There is no agreed cut-off. Researchers filter variants by this score before using them: the TOPMed comparison with UK Biobank exomes kept variants with quality of at least 0.3[7], the HRC panel's own association test used r² above 0.5[8], and FinnGen counted a variant as confidently imputed above an INFO of 0.6[9]. The Michigan server offers the same filter as an upload option, removing imputed variants whose r² is below a chosen value[10].
The filter matters because the output does not mark weak guesses on its own. Beagle's output file holds phased, non-missing genotypes for every sample, whatever the DR2[6]. A program that reads only the genotype column sees a confident-looking call at every one of millions of positions.
Accurate for common variants, weaker for rare ones
The fair test of imputation is to hide genotypes that were measured, impute them, and compare. The Michigan server's authors did this for 100 people of European ancestry: they kept only the positions an Illumina 1M chip reads on chromosome 20, imputed the rest from seven reference panels, and correlated the imputed doses with the people's sequence data[2].
| Reference panel | People in panel | Below 0.5% | 0.5 to 5% | 5 to 50% |
|---|---|---|---|---|
| 1000 Genomes Phase 1 | 1,092 | 0.45 | 0.77 | 0.96 |
| 1000 Genomes Phase 3 | 2,504 | 0.52 | 0.79 | 0.96 |
| Haplotype Reference Consortium v1.1 | 32,390 | 0.77 | 0.90 | 0.98 |
Source: Das et al. 2016, Table 1, Minimac3 column: 100 European-ancestry genomes masked to the Illumina Duo 1M chip on chromosome 20; the lowest band is printed as a minor allele frequency of 0.0001 to 0.5 percent[2].
For variants above 5 percent frequency, every panel gave a mean r² of 0.96 or better. Below 0.5 percent the two 1000 Genomes panels fell to about one half, and a panel with nearly thirteen times as many genomes as Phase 3 lifted it to 0.77. The HRC paper measured the same slope more finely: at a frequency of 0.1 percent, imputing from a 1M chip reached an aggregate r² of 0.64 with the HRC panel and 0.36 with 1000 Genomes Phase 3[8].
0.98
mean r² for variants above 5 percent frequency, HRC panel[2]
0.77
mean r² for variants below 0.5 percent, same panel and chip[2]
52.97%
of UK Biobank exome variants at or below 0.05 percent frequency also found in TOPMed-imputed data at quality 0.3 or more[7]
TOPMed, whose panel was built from deep whole-genome sequencing, ran the largest check: its authors imputed UK Biobank's chip data and compared it with exome sequencing of 49,819 participants. Of 463,182 exome variants with a minor allele frequency above 0.05 percent, 84.86 percent were present in the imputed data at a quality of 0.3 or more; of 3,587,193 rarer variants seen more than once, 52.97 percent were. Where both existed, the average correlation ran from 0.73 for the rarest band to 0.98 for variants above 25 percent[7].
The rare end is where the variants with large effects live. Chips already struggle there: in UK Biobank only 16 percent of 4,757 heterozygous chip calls at variants rarer than 0.001 percent were confirmed by sequencing[12]. We have written about what that means for a pathogenic variant in raw data. Imputation cannot repair it. A variant that only a handful of people carry is present on few panel chromosomes, if on any, and the tables above show what that does to accuracy.
Genes with structural variation are a second weak point. In a UK Biobank analysis of pharmacogenes, phenotypes for CYP2D6 inferred from imputed chip data agreed with the study's integrated call set for 64.86 percent of people, against 99.44 percent for CYP2C19[13]. CYP2D6 also varies by whole-gene deletions, duplications and hybrids, which neither data set called; star alleles explained goes through them.
Ancestry is the third. The panels behind these numbers are mostly European: HRC was built from 20 studies of predominantly European ancestry[8]. For a Finn the relevant question is how many Finns are in the panel. The 1000 Genomes 30x panel holds 99 Finnish samples out of 3,202[14]. FinnGen imputes its participants against SISu v3, a panel of 3,775 Finns sequenced at 25 to 30x[9], and that panel is research data that is not available to us. The variants enriched in Finland, described in our guide to the Finnish disease heritage, are mostly low-frequency variants[9], the band where panel size and ancestry decide accuracy.
Where people impute a 23andMe file
Two kinds of service impute consumer files. The first is the research servers built for study cohorts. The Michigan Imputation Server is free, needs a registered account[10], and imputes with Minimac4[15]. It accepts bgzip-compressed VCF files, one per chromosome, sorted by position, on GRCh37 or GRCh38[16], so a 23andMe or AncestryDNA text export has to be converted before upload. The data travel to the server in Michigan[17], unphased input is phased with Eagle or Beagle, and the results are deleted after seven days[10]. The TOPMed panel is offered the same way, through its own server; the panel itself is never handed to users[7].
The second is consumer services that impute as part of a paid report. SelfDecode's help page says it takes about 650,000 tested variants and predicts up to 200 million. The same page says imputation should not be relied on for high-impact variants such as APOE, BRCA1 and BRCA2, and that for those it uses directly tested data[11].
What we measured before deciding
On 30 August 2026, before writing any imputation code, we measured whether imputation would move our polygenic scores on chip files, and at what cost. Imputed genotypes were only ever going to reach the polygenic scores, and perhaps the ancestry estimate; carrier, clinical and medication results were to stay on measured calls. The licences cleared first: Beagle is free software under the GNU GPL[18], and the 1000 Genomes 30x data were released without access or use restrictions[5].
We then ran 18 real chip exports, public files from the Personal Genome Project, through the production code. At that time the report carried six disease scores. Of the 108 file and score pairs, 97.2 percent came back as not reportable, the other 2.8 percent as an estimate, and none reached a precise percentile.
That number was wrong, and it was ours. The scoring code counted a variant as covered when the chip reading produced a variant record, but our chip reader by design produces no record for a two-reference-copy reading. An array that read a scoring variant and found the common genotype was counted as not having read it, so coverage tracked how many risk alleles a person carried instead of what the array measured. Corrected, the median coverage of the six scores came out 37 to 73 percent higher than before. We fixed it the same day, and the fix is pinned by a test that fails on the old code.
Re-measured on the same 18 files, 16 pairs now show an estimate instead of 3, and no score gained a precise percentile. The fix reached real reports. It did not change the decision, because the median chip still held under half of every score. Only imputation could take a score like coronary artery disease from about 16 percent of its variants to the 90 percent a precise percentile needs.
So we priced it. The cheap version was to impute only a window around each scoring variant, the way our Deep Read fetches the regions it reads instead of a whole genome. The six scores used 920 variants spread over 22 chromosomes, and windows around scattered points do not stay small: at a flank of 1 million bases either side they merged into 324 windows covering 29.46 percent of the genome. We had written down 15 percent as the limit before measuring.
Two design rules would bind any later attempt. Once a file is imputed, every scoring variant has a genotype, so coverage would read 100 percent by construction and the rule that withholds thin scores would stop firing without any error; coverage would have to count only variants imputed above a quality threshold. And an outside imputation server would mean sending a customer's genotypes to a third party, which this service does not do, so imputation would have to run on our own machines, as a genome-scale job.
What the report does with each kind of file
| File | Unmeasured positions | What the report shows |
|---|---|---|
| Chip export from an array (23andMe, AncestryDNA, FamilyTreeDNA and others) | Left unread; never imputed | Polygenic scores count the positions the array read, two-reference-copy readings included: a percentile at 90 percent or more, an estimate from 50 percent, no number below that. Rare pathogenic positions are not read. |
| Chip-format export that declares low-pass sequencing in its header | Imputed by the provider before we see it | Read like a chip export, with a caveat sentence saying the genotypes were imputed from a reference panel and are least reliable where a variant is rare |
VCF whose records carry imputation fields (DR2, AR2, IMP or RAF in INFO, DS in FORMAT) | Imputed by whoever made the file | Recognised as imputed: the runs-of-homozygosity section declines it. The other sections read it as a plain VCF and do not yet separate imputed genotypes from called ones. |
| Plain VCF from sequencing | Not imputed; a position with no record is read as the reference genotype and labelled inferred | Polygenic scores are always an estimate; an exome or panel VCF gets no number |
| gVCF | Not imputed; reference blocks with 10 or more reads count as read | The same 90 and 50 percent rules as a chip |
| BAM or CRAM (Deep Read) | Not imputed; read depth is measured directly | Deep Read panels; the scores come from the VCF or gVCF |
Source: Aimosti's chip reader, file detector and scoring engine as of this page's review date. The thresholds and the treatment of each file are set out in full in polygenic scores in plain words; chip reach is on the data page[1].
The third row is a gap we know about. A file that someone imputed elsewhere and then uploaded carries the evidence in its own records, and we detect it, but today only one section acts on it. The free file check names a file's format and build before anything is paid for; it does not yet say that a VCF was imputed.
What Aimosti would (and wouldn't) show you
None of our reports imputes. A chip report reads only the positions the array measured, two-reference-copy readings included, and gives a polygenic score a number only when the chip read at least half of its variants. Files imputed by someone else are read as the table above describes.
What we won't claim
We won't impute a position your file did not read, present an imputation quality score as the accuracy of your own genotype, or count an imputed genotype as coverage. We have not run imputation on a customer file, so we have no accuracy figure of our own to offer.
Bottom line. 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.
Questions people ask
How accurate is DNA imputation?
It depends on how common the variant is. With the HRC panel, mean r² against sequencing was 0.98 above 5 percent frequency and 0.77 below 0.5 percent[2]. A single figure for a whole file mostly describes common variants.
Can imputation find a BRCA1 or BRCA2 variant my chip missed?
Pathogenic variants in those genes are rare, which is where imputation is weakest and where chips already misread most often[12]. SelfDecode, which imputes, says imputation should not be relied on for BRCA1 and BRCA2[11]. Finding them takes sequencing of the genes.
What happens if I upload an imputed file to Aimosti?
A VCF from an imputation server is recognised as imputed from its DR2, AR2, IMP, RAF or DS fields. The runs-of-homozygosity section declines it and the rest of the report reads it as a plain VCF, without yet marking which genotypes were imputed. A chip-format file that declares low-pass sequencing gets a caveat saying its genotypes were imputed.
Why would imputation be less accurate for Finns?
Accuracy at low-frequency variants depends on how many closely related genomes the panel holds. The 1000 Genomes 30x panel has 99 Finnish samples[14]; Finland's own SISu panel of 3,775 genomes[9] is not available to us.
Will Aimosti add imputation?
Not before the accuracy test described above has been run and passed, with coverage counted only from well-imputed variants. There is no date for it.
References
- What your DNA file can actually read: measured on 14 real files. Aimosti, 2026. Measured 6 October 2026. Ten chip exports of 563,320 to 955,958 rows read 281 to 495 of 1,457 polygenic-score positions.
- Das S, Forer L, Schönherr S, et al. Next-generation genotype imputation service and methods. Nature Genetics, 2016. doi:10.1038/ng.3656 Table 1, mean r² by minor allele frequency, Minimac3: 1000G Phase 1 (1,092) 0.45, 0.77, 0.96; 1000G Phase 3 (2,504) 0.52, 0.79, 0.96; HRC v1.1 (32,390) 0.77, 0.90, 0.98. Read in the NIH author manuscript, PMC5157836.
- Bycroft C, Freeman C, Petkova D, et al. The UK Biobank resource with deep phenotyping and genomic data. Nature, 2018. doi:10.1038/s41586-018-0579-z
- Browning BL, Zhou Y, Browning SR. A one-penny imputed genome from next-generation reference panels. American Journal of Human Genetics, 2018. doi:10.1016/j.ajhg.2018.07.015
- Byrska-Bishop M, Evani US, Zhao X, et al. High-coverage whole-genome sequencing of the expanded 1000 Genomes Project cohort including 602 trios. Cell, 2022. doi:10.1016/j.cell.2022.08.004
- Browning BL. Beagle 5.5 manual. University of Washington, 2024. Section 5, Output files: the DR2, AF and IMP INFO fields; phased, non-missing genotypes for all samples; the allele dose in the DS format field (section 3.3).
- Taliun D, Harris DN, Kessler MD, et al. Sequencing of 53,831 diverse genomes from the NHLBI TOPMed Program. Nature, 2021. doi:10.1038/s41586-021-03205-y UK Biobank imputed with the TOPMed panel against exome sequencing of 49,819 participants: 84.86% of 463,182 variants with MAF above 0.05% and 52.97% of 3,587,193 non-singleton variants at or below 0.05% present at imputation quality above 0.3; average correlation 0.73 to 0.98.
- McCarthy S, Das S, Kretzschmar W, et al. A reference panel of 64,976 haplotypes for genotype imputation. Nature Genetics, 2016. doi:10.1038/ng.3643 Illumina 1M chip, aggregate r² 0.64 (HRC) against 0.36 (1000 Genomes Phase 3) at a frequency of 0.1%.
- Kurki MI, Karjalainen J, Palta P, et al. FinnGen provides genetic insights from a well-phenotyped isolated population. Nature, 2023. doi:10.1038/s41586-022-05473-8 SISu v3 panel: 3,775 Finnish genomes at 25 to 30x; 16,387,711 of 16,962,023 variants imputed at INFO above 0.6.
- Michigan Imputation Server 2: Getting started. Michigan Imputation Server documentation, 2024.
- Understanding Genetic Imputation: Filling in the Gaps in Your DNA. SelfDecode Help Center. Read 10 October 2026.
- Weedon MN, Jackson L, Harrison JW, et al. Use of SNP chips to detect rare pathogenic variants: retrospective, population based diagnostic evaluation. BMJ, 2021. doi:10.1136/bmj.n214
- McInnes G, Lavertu A, Sangkuhl K, Klein TE, Whirl-Carrillo M, Altman RB. Pharmacogenetics at scale: an analysis of the UK Biobank. Clinical Pharmacology & Therapeutics, 2021. doi:10.1002/cpt.2122 Table 1, phenotype concordance of imputed chip data with the integrated call set: CYP2D6 64.86%, CYP2C19 99.44%.
- 1000 Genomes 30x on GRCh38: the 3,202 samples with their populations. International Genome Sample Resource. 99 of 3,202 samples carry the population code FIN (Finnish in Finland), counted 10 October 2026.
- Michigan Imputation Server 2. Michigan Imputation Server documentation, 2024.
- Michigan Imputation Server 2: Data preparation. Michigan Imputation Server documentation, 2024.
- Michigan Imputation Server 2: Data security. Michigan Imputation Server documentation, 2024.
- Browning BL. Beagle 5.5. University of Washington. Licence: GNU General Public License, version 3 or any later version.
Last reviewed . Every number on this page links to the source it comes from; if one of them has moved, tell us.
Free · in your browser
Find out what your own file can read.
The free check opens your DNA file on your device and shows which format you have and what a report could read from it. The file is never uploaded.