Episode Transcript
[00:00:03] Picture a restorer kneeling over an ancient mosaic floor. Most of the tiles are gone. Only a scattering survives. A few here, a few there. She cannot dig up the missing ones, but she has an archive of records from floors laid by the same workshop, and those floors repeat their patterns. So she matches what survives against the archive, and she fills each gap with the tile that the matching floors suggest.
[00:00:30] Now picture that floor as a person's genome, and picture the archive as other people's sequenced genomes.
[00:00:42] This is base by base. Every episode is built with AI tools from a real open Access paper curated by a human, and every claim is checked against the paper before we publish. You'll find this episode's fact check score on our website.
[00:00:56] Now back to the story.
[00:01:03] Here is why that picture matters. A great many genetic studies do not read the whole genome. They use a genotyping chip which measures a chosen set of positions and leaves everything else blank. The blanks are then filled in by a computation called imputation. It compares each person with a reference panel of fully sequenced genomes, and it estimates what the unmeasured positions probably are.
[00:01:27] So when a study reports a gene region tied to a disease, some of the letters behind that result may never have been measured directly. They were inferred.
[00:01:39] That raises an uncomfortable how good is the archive? The largest panel in use until now is called TopMed. In its third release, with 133,597 individuals, it is widely used.
[00:01:55] But what happens if you build a panel more than 3.6 times that size? Would the gaps fill in better for every everyone or only for some populations? And would it change what a real study finds?
[00:02:07] That is what we are going to explore in Italy and across the world.
[00:02:16] Today we celebrate the work of Lars Wienbrandt, David Ellinghouse and colleagues from the Institute of Clinical Molecular Biology at Keele University and the University Medical Center, Schleswig Holstein in Kiel, Germany, who have advanced our understanding of how to fill in the human genome at scale.
[00:02:34] This base by base episode is based on their Open Access article. A haplotype reference panel constructed from 490,319 UK Biobank genomes improves genotype imputation for global populations, published in nature genetics in 2026.
[00:02:55] Let's start with the word in that title that does the most work.
[00:02:59] Haplotype. For most chromosomes, you carry two copies, one from each parent. A haplotype is a run of genetic variants sitting together on one copy. Those runs are inherited in stretches, which is why they work like the repeating patterns of the mosaic. If A person's measured variants match a stretch seen in the panel. The unmeasured ones nearby can be estimated from that same stretch. But a chip does not tell you which variant sits on which which copy.
[00:03:28] So imputation has two steps. The first is phasing, which means sorting a person's variants onto their two chromosome copies. The second is imputation itself, which fills in the blanks. Both lean on the reference panel. How do researchers usually do this? They upload data to an imputation server, a service that runs the computation for them. Servers built on panels such as the Haplotype Reference Consortium and Topmed made this easy because no one needs their own computing resources.
[00:03:58] So what was missing?
[00:04:02] Two things, according to the authors. The first is access.
[00:04:06] Sequencing data from the Topmed third release are only partly available through the database of genotypes and phenotypes. And the panel was assembled from more than 80 individual studies, which prevents its use on standalone servers or on cloud platforms such as the UK Biobank Research Analysis Platform.
[00:04:24] The second is scale. The UK Biobank had released whole genome sequencing data from about half a million participants. Could that become a panel for the research community to impute with?
[00:04:37] There was also a question about whether bigger still helps. A haplotype panel of 78,195 individuals had recently been built from the Genomics England study and used to impute the UK Biobank. That work reported gains for British and South Asian samples, but it also reported a decrease for African American, East Asian and non Finnish European populations at variants with frequencies of 0.01% or higher. And it suggested that Karman variants might already be close to saturated. Is that true? Keep that question in mind.
[00:05:20] So the Keel team set out to build the panel. They began with the UK Biobank whole genome data in the cloud. The scale is hard to imagine. There were 151,561 files totaling 1,509.29 terabytes. The team wrote a custom tool to pull out only what they needed, which shrank the average file from 9.96 gigabytes to 27.86 megabytes.
[00:05:47] Then they filtered for quality, removing variants with too much missing data or low confidence scores.
[00:05:55] Next came phasing the panel itself, using a program called ShapeIt5. Think of it as cataloging the archive before anyone can use it. Common variants were phased first to form a scaffold. Rare variants were then placed onto that scaffold.
[00:06:11] Even the X chromosome needed special handling because mail samples had been encoded as diploid across a region where they carried just one copy. The result is the complete panel. 490,319 people 361,441,468 variants to the author's knowledge, the largest imputation panels generated to date.
[00:06:38] They also made a lighter version.
[00:06:40] Most association studies filter out poorly imputed variants, and variants rarer than 0.01% in the panel rarely pass that filter. So the team built a second panel keeping only variants at or above that frequency, 51,939,464 of them. And to use these panels they extended their own software, Eagle Imp. It runs on the project's secure server and in a version called Eagle Impr inside the UK Biobank's cloud platform.
[00:07:11] One detail to hold onto this team built both the panel and the tool being tested now. How do you judge a restoration? The team set up a benchmark. They took genomes from the 1000 Genomes Project, kept only the positions found on a common chip, the Illumina Global Screening Array, and asked each system to rebuild the rest. They did this for 5 super populations and 26 subpopulations. They added 4 real genome wide association datasets from Germany, Italy, Norway and Spain for 35 benchmarks in all. And they compared results against the Topmed imputation server using only variants shared by the panels.
[00:07:57] Scoring used two kinds of measure. The first is estimated R squared, a confidence score the software reports for each variant. Researchers routinely use it to throw out poorly imputed variants, but it is the software grading its own work. The second kind uses a known answer. The 1000 Genomes haplotypes include phasing informed by family trios so the team could check against that truth. Switch error rate counts how often phasing flips between chromosome copies mean absolute error measures how far imputed genotypes land from the real ones. Both directly test the restoration.
[00:08:39] Let's begin with phasing. The complete UK Biobank panel gave lower switch error rates than Topmed for all superpopulations except the American group and for 25 of 26 subpopulations. Then the self reported score, estimated R squared, was higher across all five superpopulations for variants, down to 0.01% frequency and in 22 of 26 subpopulations. The largest gain was in East Asian samples plus 24.39%, nearly a quarter. And the panel achieved that with UK participants as its source.
[00:09:19] It was not a clean sweep for the American super population among variants at 1% or higher quality dipped by 0.7.6% at the lowest frequency threshold. Four subpopulations got worse. The largest decrease was minus 3.33% in the group labeled GWD.
[00:09:40] Other thresholds show showed improvement in 25 and 24 populations of 26.
[00:09:46] So the pattern is broad but it has real exceptions. And remember, this is the software's own confidence. What happens when we check against the known answer?
[00:09:58] Here the picture gets more nuanced. Using mean absolute error, the panel beat TopMed in four of five super populations at least in some frequency range ranges. South Asian and East Asian samples improved in every range. European samples improved at the rarer ranges. But for the African super population topmed performed better in all frequency categories across subpopulations. The lighter UK Biobank panel had lower error in every range for 12 populations. Topmed did for six. Eight were mixed. That is an honest split and the African result deserves to stay in view.
[00:10:38] The team then repeated the benchmark on the real studies. There were 3,351 samples from Germany, 2,090 from Italy, 324 from Norway and 1,725 from Spain. These are Genome Wide association datasets typed on the same chip. Again, the lighter panel did not reduce estimated quality and again the complete panel improved on topmed across the entire allele frequency spectrum. Why does that matter? Because these are the kinds of datasets researchers actually work with, not only reference samples.
[00:11:14] There is also a practical surprise. Remember the lighter panel? Dropping the very rare variants did not lead to a considerable loss in estimated quality. Its overall results made matched the complete panel with slightly improved performance on mean absolute error. And it was fast on average across all 35 benchmarks. The TopMed server took about twice as long as Eagle IMP with the complete panel, even though that panel holds over 3.6 times as many samples. With the lighter panel, runtimes were four to 13 times shorter. Those figures exclude queuing time.
[00:11:52] Now to the real test. Does any of this change a study's result?
[00:11:57] The real world dataset with the greatest gain was Italy. So the team repeated a published Genome Wide association study of severe COVID 19 with respiratory failure with 835 cases and 1255 healthy controls. They imputed the same people twice, once through the TopMed server and once once with the lighter UK Biobank panel. Across most of the genome the results were largely similar except in one place, a region on chromosome 3 called 3P 2131 that is a well established region and the most strongly associated with COVID19 susceptibility in this Italian dataset only the UK Biobank imputation reached genome wide significance. There the lead variant changed too. And here is what makes that interesting. The new lead variant is the same one reported as the lead in three independent studies of critically ill or severely ill COVID 19 patients.
[00:12:58] One of those followed 2,244 critically ill patients from 208 UK intensive care units.
[00:13:06] So the sharper tile matched what larger studies had already already seen.
[00:13:12] The team then asked which variant was most likely causal using a Bayesian method called fine mapping with susi. With the UK Biobank imputation, that same variant ranked first in a credible set meant to contain the causal variant with at least 95% probability.
[00:13:30] With TopMed it ranked fourth and the credible set was larger and less precise.
[00:13:35] The two lead variants were only moderately linked with an r squared of 0.6625.
[00:13:41] Remember, this is a statistical ranking at 1 locus in 1 data set. It does not prove biology.
[00:13:53] So what about that question of saturation? The authors argue that imputation of common variants at frequencies of 1 per percent or higher is not saturated for European study populations. That contrasts with the Genomics England benchmarks. The association gain was shown at one locus alongside the wider benchmark evidence. Still, for European populations not well represented in current panels, the authors recommend imputing with their UK Biobank panel.
[00:14:23] They go further because of gains of up to 2024.39% across 35 study populations. They believe this is the most suitable panel for imputing worldwide populations when no population specific panel is available.
[00:14:39] For a practical reader that opens the use of a very large panel inside the UK Biobank's own cloud. The software and the applet code are freely available under an open license and users studying common variants can take the lighter panel for faster runs without noticeable loss of accuracy.
[00:14:58] But we should note the limits. First, the benchmark comes from the people who built the tool, and the Genomics England panel could not be included because it is not publicly available.
[00:15:09] Second, topmed keeps real strengths. It holds more panel specific variants than the UK Biobank panel does, which can help with variants specific to non European populations. And third, the African results on mean absolute error still favored TopMed.
[00:15:26] Fourth, many of these figures rest on estimated R squared, the software's own confidence score.
[00:15:35] There is a deeper limit too, and the authors state it for many populations worldwide there is still room for improvement, especially for Rare variants below 0.01% frequency. Estimated quality is still low for most subpopulations and mean absolute error and correlation can themselves be inaccurate at low frequencies.
[00:15:56] So what would help. The authors suggest a global super panel with millions of haplotypes or several population specific panels offered through servers and cloud platforms.
[00:16:12] So here is where this leaves us a panel of 490,319 genomes improved phasing and estimated imputation quality for most populations tested, with the largest gain in East Asian samples tested against known answers, it won in many groups and lost in others, including the African superpopulation. And in one Italian COVID 19 study, better imputation brought another known signal to genome wide significance and sharpened it at a single locus. The panel also ran faster than Topmet, not slower.
[00:16:49] Remember the restorer? A bigger archive showed her more of the patterns, and some tiles she had placed slightly wrong could be set right.
[00:16:57] Yet some tiles remain uncertain because rare variants still have low imputation quality in many populations.
[00:17:05] Every variant filled in this way is an estimate, and the quality of that estimate depends on who is in the archive.
[00:17:12] What does this mean for the future of genetic studies if the gaps in our data depend on whose genomes we chose to sequence?
[00:17:25] This episode was based on an Open Access article under the CCBY 4.4.0 license. You can find a direct link to the paper and the license in our episode description. If you enjoyed this, follow or subscribe in your podcast app and leave a five star rating. If you'd like to support our work, use the donation link in the description. Thanks for listening and join us next time as we explore more science base by base.
