Quantifying the Average Human Genome
By Kyle Proffitt
November 3, 2026 | Just over twenty years since the first “complete” human genome sequence was published and about 2 million sequences later, we are still learning how these DNA letters affect human variation and impact health. Researchers from Exeter University, UK have approached the question from a slightly different angle by focusing on how much total variability exists from genome to genome. “What’s normal?” they asked.
Mining two mega-datasets totaling over 900,000 sequenced genomes, they produced the broadest look yet at what a typical human genome contains and found that “normal” can look alarming on paper. “Variation is the norm,” corresponding final author Caroline Wright, professor at Exeter University, told Bio-IT World. “We are all hugely variable; we’ve all got between four and a half and five and a half million positions in our DNA where we differ from the reference genome or we’re different from each other, and that’s normal.” The work was published October in Genome Medicine (DOI: 10.1186/s13073-026-01792-x).
Many Sequenced Genomes
The impetus for the work was twofold. Wright regularly diagnoses disease by interpreting rare genetic variants, and that approach is becoming increasingly common. “We’re doing more and more genome sequencing … and also increasingly doing it in a screening scenario where … we might do genome sequencing first, look to see what variants they’ve got, and then try and predict and prevent disease.” The moment was right, too. Through the All of Us project in the US and the UK Biobank, the team had access to over 900,000 individuals’ genetic sequencing data. “That’s data that’s only been available in the last few years that’s comparable across these different cohorts,” Wright said.
To make the best use of sequencing data for predicting disease, Wright wanted a clearer baseline for normal variation. When asking colleagues, she realized the answer was hard to come by. “Some of these answers are in the literature, some of them are really out of date, they’re all over the place in different publications, and some of them actually no one’s really looked at yet,” she explained.
The Exeter group saw an opportunity. “We should bring this all together into one place, analyze the biggest set of data we can find … and provide a real reference that people can go to,” she said. This would allow researchers to know “what we would expect in terms of variation in the ‘normal’ human genome, if we can call it that—the average—as well as the amount that’s associated with disease.”
Data Analysis
Existing databases of human genomic variation such as ClinVar focus on rare disease-associated variants, while others estimate specific variant frequency within populations. But for a healthy individual, how much variation is likely to exist?
Sequencing data was sorted, largely based on classifications from the 1000 Genomes Project (1KG), into six inferred population groups: African (AFR), Admixed American (AMR), East Asian (EAS), European (EUR), Middle Eastern (MID, All of Us only), and South Asian (SAS). Within each population group, comparisons were made to the GRCh38.p14 reference genome, which Wright noted is “predominantly based on individuals of Western European origin.”
In line with data from other, smaller datasets, they found about 4.4 million changes—the total of both single-nucleotide variants (SNVs) and insertions/deletions (indels) smaller than 50 bp—from the reference genome, per individual for every group except AFR, where the number rises to about 5.2 million.
“The majority of this variation is probably phenotypically neutral,” Wright said. “Some of it will just be neutral mutations that occurred many, many, many generations ago that didn’t have any advantage or disadvantage and were just propagated; some will have some very mild advantage that we haven’t been able to determine yet.” A clue is that most of the changes are fairly common. “Upwards of 98% of variations are shared by more than a percent of people across the planet, so rare variation is actually fairly unusual in an individual genome.” Wright said.
They looked closer at what might not be neutral, though. About 38% of SNV/indels occurred in protein-coding genes, 27% were in non-protein-coding genes (ribosome components, transfer RNA, microRNAs, etc.), and 35% fell in non-gene regions without any known encoded function. Within genes, the vast majority (~97%) are in introns (the parts that get cut out) or untranslated regions.
Within protein-coding regions, the median individual has 9,600 missense (amino-acid-changing) variants, but only 28 are predicted by computer algorithms to have a high likelihood of disrupting the protein and potentially causing pathogenicity. There’s also a median 23 variants predicted to disrupt transcript splicing, and 165 more SNV/indels that could damage protein function through events such as frameshifts or early protein termination. That adds up to at least 216 concerning variations from the reference genome for the median individual. Structural variant analysis added another 25 predicted copy number variants (a floor, since Illumina short-read sequencing isn’t great at detecting these changes; other studies have estimated effects on 160-180 genes). How can we all be this broken?
Variation Is Okay
The raw numbers do not yet say anything about the true importance of the genes involved. For example, a variant in an intron of the HERC2 gene disrupts expression of OCA2, but the result is just blue eyes. A novel rare variant in OCA2 (total loss of which leads to the gene’s namesake disease, oculocutaneous albinism type 2) could plausibly score as pathogenic, and yet, with heterozygous expression, result in that same benign phenotype.
The researchers looked at variants in genes known to cause disease by themselves, filtering for variants with moderate-to-definitive evidence for causing disease present in <1% of the population. “If a variant is rare, it's more plausible that it might be damaging because evolution hasn’t had such a long time to work on it,” Wright said.
With these filters, they identified a median of 1-3 pathogenic/likely pathogenic variants in each person’s genome. Most of these genes are associated with autosomal recessive disease, and most of the individuals are carriers instead of being affected. For autosomal dominant conditions—one bad copy is enough—about 1 in 8 individuals carries at least one ClinVar pathogenic/likely pathogenic variant, and 1 in 2-3 carries at least one high-confidence loss of function prediction.
The result is a profound statement. “Everybody does have rare, deleterious, pathogenic, disease-causing variation in their genomes,” Wright said. That’s a little bit surprising, she added, because monogenic disease is rare. “Maybe one in 17 people is affected by it; you wouldn’t necessarily expect everyone to have some of those variants in their genome.”
Then there are variants of uncertain significance (VUS), “essentially rare functional variants that a clinical lab has looked at in a diagnostic context and said we’re not sure if this causes disease or not,” according to Wright. “We’ve all got probably over a hundred, maybe two hundred of those in our genomes, in disease-associated genes, which is a really big number, and suggests that the majority of these aren’t causing disease actually.” Since the sampled population is generally healthy, it’s a reasonable hypothesis that most of the discovered variation is harmless.
In time, some VUS “will upgrade to pathogenic with more information, but many of them will downgrade to benign as we learn they’re common or they’re actually not disease associated,” Wright said. “If you start reporting them or saying that you should report these, in large numbers of cases, you will end up generating a huge amount of work, probably for limited benefit.”
Polygenic Scores
The researchers also looked at polygenic scores (PGSs), which combine hundreds to thousands of variants into risk predictors for complex disease. Looking at 28 common diseases, the average person will be in the top 10% risk for about 3 of these. This is mostly basic statistics, but it sounds worse and reveals some inherent danger of overusing these scores. “The more PGSs you calculate in a person, the more they will have … where they’re at the top end of the risk curve,” Wright said. “We need to be cautious of a blanket approach where we calculate them in everyone because that way we’ll definitely identify high-risk scores [for some diseases] in everybody.”
Even when a prediction is informative, the genetic component may be small—relative risk goes up, but absolute risk remains low. “If you’re going to use genetic information, you really need to take into account the context … symptoms or family history, or other clinical factors or environment … to understand disease; genetics by itself is often not deterministic,” Wright cautioned.
De Novo Mutations
The study also measured de novo mutation rates. Within the dataset, they identified 2,334 parent/offspring trios (father, mother, child) and used these relationships to identify the rate of de novo mutations. “Everyone's got maybe 60 de novo mutations in their genome,” Wright reported. About 80% of these come from the father, and that is particularly worsened as the father ages. Again, most of these are probably inconsequential. However, a median one de novo protein-coding variation shows up in each child. “Everyone’s got a de novo coding variant in their genes,” Wright said.
Equity in Sequencing
A major thread in the study is that making sense of any one genome requires an appropriate comparison, and that depends on more extensive and diverse sequencing.
The PGSs illustrate the problem. Non-European samples are significantly more likely to be in the top end of risk curves. One in five EUR samples is top 1% risk for one of the 28 diseases, but for non-EUR, it’s one in three. The paper describes this result as an artifact. Because genome-wide association studies “are derived from European-like populations, they just don’t carry over very well to these other populations … which can often mean you end up predicting people are at much higher risk than they are,” Wright said.
Similarly, an extra 800,000 variants are identified in African-descent individuals partly because GRCh38 is biased toward Western European samples, and many reads do not map neatly. “The more different individuals are from that [GRCh38], the less representative it is, and therefore the more variants they will have relative to it,” Wright explained. There’s a bit more going on here, as the AFR group is an outlier—the other groups have similar variant numbers relative to reference, and the explanation also involves human population history. “The migration of modern humans out of Africa resulted in a population bottleneck and a concomitant loss of genetic diversity,” which means that “non-African populations appear to have a subset of the genetic diversity present in sub-Saharan Africa,” explain Campbell and Tishkoff. In other words, non-Africans are a smaller branch of the diversity Africans retained. “Even in Africa, the most variation is within the African continent,” Wright said.
The AFR gap is the tip of a larger problem. “There are many populations on the earth that are not represented in this data set,” Wright said. “There’s huge swathes of the African continent that are still not included, never mind Indonesian islands and indigenous populations throughout the world … a whole range who similarly might not be well served by the current reference genome.” For these groups, “it may be harder to diagnose them because … particularly recessive diseases have been less well characterized in some cases, they might have more variants of uncertain significance, and their polygenic scores will be less predictive overall.”
The solution, Wright says, is to “get better representation of all types of individuals across the globe in these data sets, because that will allow us to better calibrate the amount of variation.” Beyond more sequencing, she suggested that the pangenome, “which would be trying to align people’s genomes to a more ancestrally appropriate reference,” will solve some of the issue. However, a pangenome re-analysis would require realigning raw reads for all ~900,000 genomes and was not pursued in this project.
Underrepresented groups have the most to lose, but everyone gains from a bigger reference set. “This is what this paper shows: you can't just interpret one genome by itself. You have to do a comparator to know these variants are not part of normal human variation, and there’s no way to do that without comparative populations,” Wright said. “We all benefit from having more diverse data; in particular, you can have a rare variant or a variant that appears to be rare, in one population, which you think, maybe this is causing disease, but if you find that it’s actually quite common in another population because of founder effects and not associated with disease, then you can exclude it in both.”
The Future
This study provides a fresh background for future genetic interpretation. Wright envisions combining this knowledge with additional layers, “maybe proteomic data, methylation data, some transcript data, but matched with individual genome data, because then we can actually start to dig in to this genetic variant is linked with this level of gene expression is linked with this level of protein and then associated with whatever phenotype, and you can actually start to fill in the gaps of the biology … integrating multiomic data at an individual level, perhaps longitudinally, so we can start to actually interpret this variation.”