GWAS and Noncoding Variants

"Although 88% of TASs were intronic (45%) or intergenic (43%), TASs were not overrepresented in introns and were significantly depleted in intergenic regions." — Hindorff and colleagues, the first survey of the GWAS Catalog, 2009¹

"Nobody is surprised that most of these genetic markers lie outside of protein-coding regions." — Laurence Moran, biochemist, 2022²

Summary

A genome-wide association study, or GWAS, compares the DNA of thousands of people with and without a trait, and flags the spots where a common spelling difference, a single-nucleotide polymorphism or SNP, tracks the trait.  Since 2009 every survey of these studies has found the same thing:  about nine in ten flagged SNPs lie outside the 1 to 2% of the genome that spells proteins.  Design writers and the John Mattick school read this as proof that noncoding DNA does real work.  Junk-DNA defenders reply that nobody ever called regulatory DNA junk, and that a flagged SNP is usually a marker, not a cause.

This article checks the headline numbers against their own controls, recounts them from the current GWAS Catalog, follows the flagged SNPs to their causes, and asks what the results show about how much of the genome matters.

1. The headline is true, and it is mostly a statement about where the genome is. Of 327,456 distinct genome-wide significant variants in the October 2026 GWAS Catalog, 96.4% lie outside protein-coding sequence.³ But 98.7% of the genome lies outside it too. The right question is whether the hits pile up in noncoding DNA more than chance predicts. The study that started the 88% figure, Hindorff 2009, asked that and found they do not: introns held their share and intergenic DNA was depleted.¹
2. Per base, the hits concentrate in coding DNA, then in regulatory DNA, and thin out everywhere else. Against ordinary common SNPs, the 2026 Catalog's hits are 4.5 times enriched in protein-coding sequence, about 2.2 times in UTRs, 1.4 times in ENCODE regulatory elements, flat in introns and 0.7 times in intergenic DNA.³ The early, strongest hits (to 2011) were 10 times enriched in coding sequence. Maurano 2012, the source of "93% noncoding", found 57% of noncoding hits in open chromatin; its own supplement shows 41% of random sites fall there too.⁴
3. A flagged SNP is usually not the cause. Nearby SNPs travel together through the generations, a pattern called linkage disequilibrium, so a GWAS flags a whole block and reports its best marker. One fine-mapping study put the chance that a catalogued SNP is itself causal at about 5%.⁵ When the causes are pinned down, the coding share rises: 14% of predicted causal autoimmune variants were protein-coding against 4% of random SNPs,⁵ and 31% of the best-resolved UK Biobank variants were coding.⁶ The rest, a large majority, are noncoding.
4. Noncoding causes are real, and some are proven to the nucleotide. A single-letter change in an FTO intron breaks a repressor's binding site and doubles two genes' expression in fat-cell precursors.⁷ An enhancer inside BCL11A controls fetal hemoglobin; editing it is the basis of the first approved CRISPR therapy, for sickle-cell disease.⁸⁹ A cholesterol variant near SORT1 creates a binding site that raises the gene's expression in liver.¹⁰ The lactase-persistence enhancer is strongly supported, though its key variant has not been edited in place. None of these sites was ever called junk: each is a switch for a known gene.
5. Heritability points the same way. Regions conserved across mammals are 2.6% of common SNPs but carry about 35% of the trait heritability that common SNPs explain, more per SNP than coding DNA.¹¹ Regulatory marks carry more than their share; inactive chromatin carries less. Enrichment is highest in a small part of the genome.
6. The data cannot fix the share of the genome that matters, and they do not support "most". The site's own Functional DNA article turns the GWAS numbers into "at least 30%" of the genome being sequence-specific. Re-running its arithmetic with other defensible inputs gives answers from about 2% to 30%; the input least distorted by linkage, heritability by annotation, gives about 6 to 10%.¹² That is close to the 10.7% of the genome Zoonomia finds under mammalian constraint.
7. Each side has a habit to break. Design writers cite the noncoding share as if it were an enrichment, and treat markers as causes. Junk-DNA defenders say regulatory DNA was "never" junk, but the regulatory share GWAS reveals is larger than coding, and they rarely put a number on it. Neither side's headline is measured by these studies.

Verdict in one paragraph.  GWAS show that variation in noncoding DNA, mostly in switches that turn genes up and down, explains more of human disease risk than variation in protein-coding DNA.  That settles a point the early junk-DNA literature got wrong in emphasis, and it fits a designed genome rich in control sequence.  It does not show that most noncoding DNA does anything.  Introns and intergenic DNA, 93% of the genome, carry no more than their share of hits, and the best-resolved estimates put the DNA that matters for common traits near a tenth of the genome, not a majority.  A young-earth model loses nothing from this result and gains nothing measurable from it.

Part 1: Where the headline numbers come from

What a GWAS can and cannot see

A GWAS reads a few hundred thousand to a few million SNPs in each person, and tests each one for a link to the trait.  It sees only common variants, the ones carried by at least a few percent of people.  It reports the SNP with the smallest p-value in each region, called the lead or index SNP.  That SNP need not be the cause.  It only has to travel with the cause.  Part 3 takes up how often it does.

The genome-wide significance cut-off is p < 5 × 10⁻⁸.  The NHGRI-EBI GWAS Catalog has collected every association at or near that line since 2008.  Every headline figure below is a count of where the Catalog's lead SNPs sit.

Hindorff 2009: 88%, and no excess in noncoding DNA

The first survey of the Catalog, by Lucia Hindorff, Francis Collins and Teri Manolio at the National Human Genome Research Institute, counted 465 distinct trait-associated SNPs from 151 papers.  "43% ( n = 199) of which were located in an intergenic region, 45% ( n = 208) were intronic, 9% ( n = 41) were nonsynonymous, 2% ( n = 10) were in a 5′ or 3′ untranslated region, and 2% ( n = 7) were synonymous."¹  Intronic plus intergenic makes the famous 88%.

Hindorff's team then asked the question the headline skips.  They compared the hits with 100 control sets of random SNPs drawn from the same genotyping arrays, with their linked partners.¹  The answer was in the abstract:  "Although 88% of TASs were intronic (45%) or intergenic (43%), TASs were not overrepresented in introns and were significantly depleted in intergenic regions."¹  The only classes with an excess were amino-acid-changing sites, with an odds ratio of 3.9, and the 5 kb in front of genes, with 2.3.¹  The authors still read the raw shares as "suggesting a greater than anticipated role for noncoding SNPs in common diseases."¹

ENCODE's 2012 summary paper repeated the sum:  "Strikingly, 88% of associated SNPs are either intronic or intergenic," citing Hindorff.¹³  It did not repeat the control.

Maurano 2012: 93%, and open chromatin covers 41% of random sites

John Stamatoyannopoulos's group at the University of Washington mapped DNase I hypersensitive sites, or DHSs, the open patches of chromatin where regulatory proteins bind, in 349 cell and tissue samples.  They laid 5,134 noncoding GWAS SNPs over the maps.  Their introduction gave the figure that later papers cite:  "the majority (~93%) of disease- and trait-associated variants emerging from these studies lie within noncoding sequence."⁴  The supplement puts it the other way:  "only 4.9% of GWAS SNPs lie in coding sequence," where coding means the consensus coding sequence set.⁴

The paper's real finding was an enrichment.  "This revealed a collective 40% enrichment of GWAS SNPs in DHSs."⁴  It also said that "Fully 76.6% of all noncoding GWAS SNPs either lie within a DHS (57.1%, 2,931 SNPs) or are in complete linkage disequilibrium (LD) with SNPs in a nearby DHS (19.5%, 999 SNPs)."⁴

The 76.6% is the number this site's Functional DNA article quotes as evidence of function.  The supplement gives the baseline it needs.  Table S3 lists the share of each kind of site that falls in a DHS.  Random mappable sites:  41.1%.  Ordinary 1000 Genomes SNPs:  42.9%.  All GWAS SNPs:  57.1%, an enrichment of 1.39.  Hits replicated in a second study:  69.8%, 1.70.⁴  So the union of open chromatin across 349 samples already covers about two-fifths of the genome.  A random SNP plus its linked partners would land in or beside one most of the time.  The enrichment is real, and it is larger for the most reliable hits.  It is an enrichment of about 1.4 to 1.7, not a sign that three-quarters of the genome is regulatory.

Schaub and colleagues at Stanford ran the same test the same year with matched random SNPs.  Lead SNPs were enriched 1.12-fold in DNase peaks, 1.22-fold in DNase footprints and 1.25-fold in protein-binding peaks.¹⁴

The 95% and its relatives

John Mattick, the leading scientific advocate of a mostly functional genome, wrote in 2009 that GWAS "shows that the vast majority of variations are located in noncoding regions."¹⁵  In 2013 he put it at "∼95% of all variants associated with complex (as opposed to monogenic) diseases in humans map to non-coding, presumably regulatory, sequences," citing a 2011 cancer-genetics review.¹⁶  That review's Table 1 lists 79 nonsynonymous and 26 synonymous SNPs among about 2,600 Catalog SNPs, about 4% coding.¹⁷  The figure is honest.  Like the 88% and 93%, it is a share with no control.

The young-earth literature uses the same count.  Jeffrey Tomkins of the Institute for Creation Research wrote in 2013 that "about half of the genetic variation associated with heritable diseases lies in these 'intergenic areas,'" citing Hindorff.¹⁸  Hindorff's figure is 43% of catalogued SNPs, not of disease variation, and Hindorff found intergenic DNA depleted.  Jonathan Wells's The Myth of Junk DNA (2011) and chapter 15 of the 2017 volume Theistic Evolution do not discuss GWAS at all.¹⁹  The fullest design-side use of the data is this site's own Functional DNA article, priced in Part 6.

Part 2: A fresh count from the 2026 Catalog

The method

For this article the whole GWAS Catalog, release of 5 October 2026, was counted against the GENCODE release 50 gene map, with a short Deno script that anyone can rerun.³  The Catalog holds 1,198,394 association rows.  Keeping those at p ≤ 5 × 10⁻⁸ with one clear position leaves 833,007, at 327,456 distinct positions.  Each position was put in one class:  protein-coding sequence first, then UTR, then exon of a noncoding RNA, then intron, then intergenic.  The genome was divided the same way, after removing the 150 million unsequenced bases.

Two controls were used.  The first is the share of genome bases in each class, which is what a hit would show if it landed at random.  The second is the share of ordinary common SNPs in each class, 12.8 million of them from dbSNP build 151.  The second control matters because a GWAS can only flag a site where a common SNP exists, and selection keeps common SNPs out of coding DNA:  coding sequence is 1.32% of the genome but holds 0.80% of common SNPs.

The result

Four things stand out.

1. The headline reproduces. 96.4% of the 2026 hits lie outside protein-coding sequence. Of the hits published to 2011, about 92% do, close to Hindorff's and Maurano's figures.
2. Per site, coding DNA is the richest class by far. A common SNP in coding sequence is 4.5 times as likely as an average common SNP to be a Catalog hit, and was 10 times as likely in the early, strongest studies. UTRs, which control how messenger RNA is used, come second.
3. Regulatory marks carry a modest excess. ENCODE's candidate regulatory elements, version 4, cover 21% of the genome and hold 31% of all hits, 41% of the early ones: 1.4 to 1.9 times their share. That is the same size as Maurano's DHS enrichment.
4. The bulk of the genome carries its share or less. Introns, 63% of the genome, hold 66% of hits, exactly their share. Intergenic DNA, 29%, holds 21%, below its share. These two classes are where any claim about "most of the genome" has to be settled, and the hits are not concentrated there.

The coding excess has shrunk as studies grew.  The early studies of a few thousand people could find only strong effects, and strong effects are often coding.  Biobank studies of half a million people find weak effects everywhere, and many list several linked SNPs per signal.  No linkage pruning was done here, so the 2026 row overweights large signals with many listed partners.  The direction of every enrichment is the same in both eras.

These per-site figures still understate how concentrated the causes are, because a lead SNP is usually a marker for a nearby cause.  That smearing pulls every class toward an enrichment of 1.

Part 3: The flagged SNP is usually a marker, not the cause

Why linkage blurs the map

Chromosomes are shuffled between generations by crossing over, but only a few times per chromosome.  So stretches of tens of thousands of bases are passed down in blocks, and the SNPs in a block rise and fall together in a population.  This is linkage disequilibrium.  If one SNP in a block raises disease risk, every SNP in the block shows the same association, and the one with the best p-value is reported.  Which one wins depends on chance, on which SNPs were on the chip, and on sample size.

So a lead SNP's address says where the cause is to within a block, not what it is.  A block can hold dozens of SNPs spread across introns, intergenic DNA and sometimes an exon.

Fine-mapping: how often the lead SNP is the cause

Statistical fine-mapping weighs every SNP in a block and gives each a posterior probability of being causal.  The SNPs that together reach 95% probability form a "credible set."

1. Autoimmune disease, 2015. Kyle Farh, Bradley Bernstein and colleagues at the Broad Institute fine-mapped 636 signals for 21 autoimmune diseases into 4,950 candidate causal SNPs, about eight per signal. Their model found that "index SNPs reported in the GWAS catalog have on average only a 5% chance of representing a causal SNP."⁵ It picked out one most-likely SNP, above 75% probability, at 12% of loci.⁵
2. Inflammatory bowel disease, 2017. Hailiang Huang and the International IBD Genetics Consortium genotyped 67,852 people densely across 94 regions. "We pinpointed 18 associations to a single causal variant with >95% certainty, and an additional 27 associations to a single variant with >50% certainty."²⁰ Credible sets "ranged in size from one to > 400 variants."²⁰
3. UK Biobank, 49 traits, 2020. Omer Weissbrod, Alkes Price and colleagues fine-mapped 49 traits in an average of 318,000 people and found 2,225 distinct SNPs with posterior probability above 0.95. "Only 39% of the 2,225 PIP>0.95 SNPs were also lead GWAS SNPs."⁶ Across 16 traits chosen to be genetically independent, only 20% of genome-wide significant loci held even one SNP that certain.⁶

The estimates vary because the methods and samples differ.  All agree that the Catalog's SNP is usually a neighbor of the cause.

What the causes are, once found

When causes are pinned down, protein-coding variants turn up far more often than among the Catalog's lead SNPs.

Sources:  Farh,⁵ Huang,²⁰ Weissbrod.⁶⁶

Huang's team warned against reading this too far:  "it does not suggest that 30% of all associations are caused by coding variants; rather, it is almost certainly the case that associated coding variants have stronger effect sizes, making them easier to fine-map."²⁰  That warning cuts both ways.  The well-resolved sets overstate the coding share.  The Catalog's lead SNPs understate it.  The true share for common-trait causes lies between about 4% and 30%, and every study puts the noncoding share at a clear majority.  Farh's group put it at "~90%" for autoimmune hits, with "~60%" in immune-cell enhancers.⁵

Testing the candidates one by one

Two kinds of experiment test candidate variants directly.  A massively parallel reporter assay, or MPRA, puts thousands of short DNA pieces, each in both versions, in front of a reporter gene and counts which version drives more output.  CRISPR screens edit or silence candidate sites inside the cell's own chromosomes.

1. Tewhey 2016. 29,173 variants from 3,642 gene-expression association signals were tested in blood cells. "Of the 3432 active sequences, 842 showed allelic skew."²¹ That is 12% of tested pieces active and 2.9% with a difference between versions. The authors estimated that a third to two-thirds of those differences are true causes.²¹
2. Ulirsch 2016. 2,756 variants linked to 75 red-cell trait signals were tested. 32 showed a difference between versions, covering 23 of the 75 signals. "Of the 32 MFVs, only 2 are sentinel variants (9%)"²². The lead SNP was the functional one in two cases.
3. Abell 2022. 30,893 variants at 744 gene-expression signals. "17.7% (132/744) had more than one allelic MPRA hit, indicating that an appreciable number of genetic associations contain multiple regulatory variants in high LD"²³. One signal can have several causes sitting together.
4. Gasperini 2019. CRISPR interference silenced 5,779 candidate enhancers, chosen from open-chromatin maps, in a leukemia cell line. 600 turned down a gene within a million bases. Of candidates sampled without a model's help, "35/948 systematically sampled exploratory DHS (4%)" did.²⁴

These numbers are rates for the sets tested, in one or two cell types.  A piece that does nothing in a leukemia line may work in a neuron.  But they set the scale.  Among variants already flagged by an association, a few percent show a measurable regulatory difference in a given cell.  Among open-chromatin sites picked without bias, about one in 25 controls a gene detectably in the cell where it was open.

Part 4: Four noncoding causes worked out to the mechanism

Each of these began as a GWAS or linkage signal and ended with an experiment that changed the DNA or the protein that reads it.

FTO and obesity

The strongest obesity signal in Europeans sits in an intron of the gene FTO.  For years the assumption was that FTO was the gene involved.  In 2014 Smemo and colleagues showed that the region loops through the chromosome to contact IRX3, a gene a million bases away, and that mice lacking Irx3 weigh 25 to 30% less.²⁵  In 2015 Melina Claussnitzer, Manolis Kellis and colleagues found the letter:  "The rs1421085 T-to-C single-nucleotide variant disrupts a conserved motif for the ARID5B repressor, which leads to derepression of a potent preadipocyte enhancer and a doubling of IRX3 and IRX5 expression during early adipocyte differentiation."⁷  They then edited the letter in fat-cell precursors:  "T-to-C editing doubled endogenous IRX3 and IRX5 expression, to levels seen in risk-allele carriers; starting from the edited preadipocytes, C-to-T re-editing back to the nonrisk allele restored low expression levels"⁷.

The case is as strong as cell work gets.  It is still cell work, in cultured precursors from a cohort of 100 donors, and the paper does not say how many donors' cells were edited.

BCL11A and sickle-cell disease

People who keep making fetal hemoglobin as adults are partly protected from sickle-cell disease and beta-thalassemia.  GWAS tied that trait to SNPs inside an intron of BCL11A, a gene that switches fetal hemoglobin off.  Daniel Bauer, Stuart Orkin and colleagues showed in 2013 that the SNPs sit in an enhancer used only in red-cell precursors:  "Genome engineering reveals the enhancer is required in erythroid but not B-lymphoid cells for BCL11A expression."⁸  Its lead SNP, rs1427407, weakens a binding site for the red-cell factors GATA1 and TAL1.  In 2015 Canver and colleagues cut every possible site across the enhancer with CRISPR and found the core in a stretch of 42 bases at the "+58" site:  "Deletion of h+58 phenocopied deletion of the composite enhancer."²⁶

That finding became a medicine.  Frangoul and colleagues reported in 2021 on patients whose blood stem cells were edited "with CRISPR-Cas9 targeting the BCL11A erythroid-specific enhancer."⁹  The U.S. Food and Drug Administration approved the therapy, Casgevy, in December 2023:  "Casgevy is the first FDA-approved therapy utilizing CRISPR/Cas9."²⁷  Breaking a noncoding enhancer is now a treatment.

Lactase persistence

Most adult mammals stop making lactase, the enzyme that digests milk sugar.  Many Europeans and East Africans do not.  Enattah and colleagues found in 2002 that a variant about 14,000 bases upstream of the lactase gene, C/T-13910, "completely associates with biochemically verified lactase non-persistence" in Finnish families and 236 individuals.²⁸  Tishkoff and colleagues found three different variants in East Africa, a few bases away, "that have derived alleles that significantly enhance transcription from the LCT promoter in vitro."²⁹  Four independent letters in one small enhancer, each favored in a milk-drinking people, make a strong case that the enhancer is the target.  Reporter assays show the European variant raises promoter activity in intestinal cells, and the factor Oct-1 binds it more strongly, though only with a partner factor present.³⁰  No edit of the variant in its native place in intestinal cells was found for this article.  This is the most famous case and the least directly proven.

SORT1 and cholesterol

A region on chromosome 1p13 is tied to LDL cholesterol and heart attack.  Kiran Musunuru, Daniel Rader and colleagues showed in 2010 that "a common noncoding polymorphism at the 1p13 locus, rs12740374, creates a C/EBP transcription factor binding site and alters the hepatic expression of the SORT1 gene."¹⁰  In 62 human livers, people carrying two copies of the minor allele had more than 12 times the SORT1 expression of those with none.¹⁰  Changing Sort1 levels in mouse liver changed blood lipids.  The variant itself cannot be tested in mice, because the binding site does not exist there.

What the four show

Each case is a single letter or a few-hundred-base switch controlling a known gene.  Each sits in the kind of DNA, an enhancer, that molecular biologists had studied since the early 1980s.  They prove that noncoding variants can cause common traits, and that the effect can be large enough to treat a disease.  They do not bear on the 90% of the genome whose status is disputed, because nobody disputed that enhancers work.

Part 5: Heritability by annotation, and the omnigenic picture

Weighing DNA by how much trait variation it explains

Counting hits treats every flagged SNP alike.  A better measure asks how much of a trait's heritable variation each kind of DNA carries.  Stratified LD score regression, published by Hilary Finucane, Alkes Price and colleagues in 2015, does this from GWAS summary data while correcting for linkage.  It uses every SNP, significant or not, so it is not limited to the strongest hits.

1. Conserved DNA carries the most per SNP. "Regions conserved in mammals … showed the largest enrichment of any category, with 2.6% of SNPs explaining an estimated 35% of SNP-heritability on average across traits." The authors add that this is "a significantly higher average enrichment than for coding regions."¹¹
2. Coding DNA carries about a tenth. In the nine-trait meta-analysis of the preprint, coding SNPs were 1.5% of SNPs and carried 10.4% of heritability, a 7-fold enrichment.³¹ Alexander Gusev and colleagues, using a different method on 11 diseases, found that "coding variants, which span 1% of the genome, explained <10% of" the heritability "despite having the highest enrichment."³²
3. Open chromatin carries more than its share. Gusev's DNase sites from 217 cell types held 16% of imputed SNPs and between 38% and 79% of heritability, depending on whether genotyped or imputed SNPs were used.³²
4. Inactive DNA carries less. Boyle, Li and Pritchard summarized: "SNPs in chromatin that is inactive, or active only in irrelevant cell types contribute little or no heritability."³³
5. Constraint at single-base resolution is the strongest marker yet. Zoonomia's comparison of 240 mammals found "3.5% of the human genome as significantly constrained, and likely functional,"³⁴ and common SNPs at those bases carried heritability enriched 7.8-fold, against 7.2-fold for amino-acid-changing SNPs.³⁴

The pattern is consistent across methods.  Trait variation is concentrated in a modest part of the genome:  coding DNA, conserved DNA and active regulatory DNA.  Most of it is noncoding.  The rest of the genome carries less than its share.

The omnigenic model

Evan Boyle, Yang Li and Jonathan Pritchard of Stanford looked at height in 2017 and found the signal everywhere.  "We estimate that more than 100,000 SNPs exert independent causal effects on height," spread so widely "that most 100kb windows contribute to variance in height."³³  They proposed an explanation:  "gene regulatory networks are sufficiently interconnected that all genes expressed in disease-relevant cells are liable to affect the functions of core disease-related genes and that most heritability can be explained by effects on genes outside core pathways."³³

This is often read as support for a mostly functional genome.  It is support for something narrower.  The model says that nearly every gene expressed in the relevant cells can nudge a trait, through its regulatory DNA.  It does not say that nearly every base does.  A hundred thousand causal SNPs across three billion bases is one site in 30,000.  And the same paper found the heritability concentrated where chromatin is active in the relevant tissue, not spread evenly over the genome.

Part 6: What this shows about how much of the genome matters

The design-side argument, as this site made it

The most detailed attempt to turn GWAS numbers into a fraction of the genome is on this site.  Functional DNA is Evidence of Design, section "From Disease + Trait Association," reasons as follows.  Coding DNA is 2% of the genome but holds only 4.9% of GWAS hits (Maurano).  So noncoding DNA is "2% / 4.9% = 41% as specific in sequence as coding DNA."  About 71% of coding mutations are harmful, so about 29% of noncoding bases are sequence-specific, and "GWAS studies suggest at least 30% of DNA is sequence-specific."  The article adds that this is an underestimate because noncoding effects are smaller and harder to detect.

The logic is a per-base comparison, which is the right kind of comparison.  Its result depends on its inputs, and three of them do not hold up.

1. The hits are not "function altering mutations." The article calls Maurano's SNPs that. They are the Catalog's lead SNPs, and Part 3 showed these are usually markers. Linkage smears each cause over its block, which makes noncoding DNA look richer, relative to coding DNA, than it is.
2. Coding DNA is not 2% of the genome by Maurano's definition. Maurano counted coding by the consensus coding sequence set. That set covers 32.3 million bases in the hg19 genome, about 1.0%.¹² With 1.0% in place of 2%, the same arithmetic gives 15%.
3. The 76.6% has a baseline. The article quotes it to show the hits map to regulatory DNA. Maurano's own Table S3 shows 41% of random sites in the same open-chromatin set.

Pricing the 30% with other inputs

The same formula was rerun with every defensible input available.¹²

The answers run from 2% to 30%.  The lead-SNP inputs give the high end, because linkage dilutes them toward the genome average.  The fine-mapped inputs give the low end, because coding causes are easier to resolve, as Huang's team warned.  Heritability by annotation corrects for linkage and uses every SNP, so it is the least biased input here, and it gives about 6 to 10%.

That range sits close to independent estimates.  Zoonomia found "At least 332 million bases (~10.7%) in the human genome are unusually conserved across species."³⁵  The site's Functional DNA article argues that conservation undercounts function, and that argument stands as far as it goes.  The GWAS method was offered as a separate line of evidence, though, and on its own best inputs it does not reach 30%.  It lands near 10%.

The formula has deeper limits that no input fixes.  It assumes that a harmful noncoding base shows up in GWAS as readily as a harmful coding base, and that GWAS samples harmful sites in proportion to their number.  Neither is known.  So the method gives an order of magnitude, not a floor.  The honest statement is that GWAS fit a genome in which something like a tenth of the bases matter for common traits, most of them noncoding.  They neither require nor rule out more.

The junk side's replies

Junk-DNA defenders do not dispute the data.

• W. Ford Doolittle, in his 2013 critique of ENCODE, conceded that the new maps "support many current population genetic studies linking human diseases to supposedly nongenic regions."³⁶
• Laurence Moran, a University of Toronto biochemist and the most persistent public defender of junk DNA, put his own budget plainly: "We know that 90% of our genome is junk, about 1% encodes proteins, and another 9% contains lots of functional DNA sequences, including regulatory elements."² On the GWAS figure he wrote: "Nobody is surprised that most of these genetic markers lie outside of protein-coding regions."² His explanation is linkage: "Most of them will be in junk DNA and will have nothing to do with any nearby genes. They just happen to be linked to those genes."²
• On regulatory DNA in particular, Moran wrote in 2013: "Regulatory regions have been part of annotated human genome sequences for thirty years."³⁷
• Moran also turned the figure into a test. Against a Nature paper that called 88% of GWAS loci noncoding, he asked: "If the hyperbole of the ENCODE results was valid wouldnt we expect ~99% of disease-causing variants to be non-coding?"³⁸

The data in Parts 2 and 5 bear out the core of the junk side's reply.  Lead SNPs are mostly markers.  Introns and intergenic DNA hold no more than their share of hits, and inactive chromatin carries little heritability.  The UK Biobank's 2022 whole-genome study found that after removing coding exons, the 1% of the genome with the fewest variants held GWAS hits at 3.2 times the average, and the 1% with the most variants at 0.4 times.³⁹  Signal follows constraint.  If 80% of the genome were sequence-specific, the hits would be spread far more evenly.

Moran's 99% test is weaker than it looks.  A genome that was 80% functional, with function evenly spread, would put about 98% of hits outside coding DNA, and the 2026 Catalog has 96.4%.  The headline share cannot tell the two pictures apart.  The per-base enrichments can, and they favor the junk side's picture over an evenly functional genome.

Where the junk side overstates

The junk side's point that regulatory DNA was never called junk is true of regulatory elements as a class.  It is not true of the size of the class.  Through the 1990s and 2000s the standard picture of a disease gene was a broken protein.  The GWAS finding that common-trait risk sits mostly in regulatory DNA, by a ratio of several to one over coding DNA, was a surprise to the people who ran the studies.  Boyle and colleagues called it "a second surprise"³³.  Moran's 9% for everything functional outside coding DNA is a budget, not a measurement, and he does not say how much of it is regulatory.  The GWAS and heritability numbers give the junk side's 9% real content, and they put most of it in regulatory DNA.

What a young-earth model owes

A designed genome is expected to be rich in control sequence, and GWAS show that control sequence is where most common-trait variation lies.  That fits.  But it is equally expected under the junk-DNA view, which always included regulatory DNA in the functional tenth.  The evidence that would separate the two views is the status of the bulk:  the 60% of the genome that is intronic and the 30% that is intergenic, minus their regulatory elements.  GWAS do not show that this bulk matters, and on the per-base figures they lean the other way.  A young-earth model does not need it to matter.  The site's own Functional DNA article says:  "The only junk DNA would be that which has been destroyed by mutations since the original design."  If the bulk of the genome is damaged copies, that sentence covers it.  What design writers should stop doing is citing "93% noncoding" as if it were an enrichment.

Objections and responses

1.  "Noncoding effects are smaller and harder to detect, so the 4.9% coding share overstates coding DNA's role."  This is the Functional DNA article's reason for calling its 30% a floor.  The data run the other way.  Heritability methods use every SNP, including those with tiny effects, and they still put coding DNA at about 7 to 14 times its share.³²³¹  The bias that does exist is linkage, and it hides coding causes behind noncoding markers.  Huang's caution, that coding causes are easier to fine-map, is the one real reason to think fine-mapped sets overstate coding, and the table in Part 6 allows for it by giving both ends.

2.  "Most noncoding hits map to regulatory features, 76.6% in Maurano and 95% in Chen 2015."  Both are shares with no baseline in the sentence quoted.  Maurano's own supplement gives 41.1% of random sites inside the same open-chromatin set.⁴  Chen's 95.09% counts any overlap with very broad databases that include conserved sequence, motifs and expression associations;⁴⁰ random SNPs were not tested, and with sets that broad most of them would match.  The enrichments that survive a control are real and modest, about 1.4 to 1.9 for hits and about 2 to 5 for heritability.

3.  "GWAS see only common variants.  Rare noncoding variants with strong effects are invisible to them."  True, and it limits every number in this article.  Strongly harmful variants are kept rare by selection, so GWAS sample the mild end of each class.  That is why the pricing in Part 6 is an order of magnitude and not a floor.  It does not favor noncoding DNA.  Boyle and colleagues draw the contrast themselves:  Mendelian diseases are "largely caused by protein-coding changes," while complex traits are "mainly driven by noncoding variants"³³.  A rare-variant count would need its own article.

4.  "Many GWAS hits are false positives."  The Functional DNA article answers this objection correctly.  Genome-wide significance is strict, and replication is common.  Maurano's replicated hits were more enriched in open chromatin than unreplicated ones, 1.70 against 1.29.⁴  False positives would dilute the regulatory signal, not create it.

5.  "The omnigenic model shows that the whole genome affects traits."  It shows that most genes expressed in a relevant cell can nudge a trait.  Its own numbers, about 100,000 causal SNPs for height, are one site in 30,000.³³  And its authors found the heritability concentrated in active chromatin, with inactive chromatin contributing "little or no heritability."³³

6.  "Heritability enrichment in conserved DNA assumes common descent, which a creationist need not grant."  The enrichment is an observation, not an inference from ancestry.  Bases that differ little among mammals carry far more trait variation per SNP than other bases.¹¹  A common-design reading can explain why those bases are alike across mammals in its own way.  Either way, the bases that matter for human traits are mostly ones that are also alike across mammals, and those are about a tenth of the genome or less.

7.  "A noncoding hit that is only a linked marker still sits near a real cause, and most real causes are noncoding.  So the junk side is wrong to dismiss the data."  Correct.  Moran's sentence that most hits "will be in junk DNA and will have nothing to do with any nearby genes"² is true of the markers and says nothing about the causes.  Fine-mapping puts the causes mostly in noncoding DNA, and Part 4's cases are noncoding.  The junk side's real claim is about how much DNA these causes occupy, and on that claim the data support it.

8.  "Introns hold their share of hits, so introns are as functional as the genome average."  The genome average is low.  An enrichment of 1.0 says introns are neither richer nor poorer than a random stretch of DNA.  With linkage smearing every class toward 1, the per-base figures cannot say whether intron function is higher or lower than average.  They do not support a claim that introns are mostly functional.

9.  "Shi Huang's group showed that common SNPs are functional as a group."  A 2012 preprint and a 2014 paper from Shi Huang.s group link the overall count of minor alleles in an individual to traits in mice, flies, yeast and 21 human disease datasets.⁴¹  The preprint does not say that all common SNPs are functional;  that wording, and "about 0.1% of the human genome," appear in a comment Huang posted on Moran's blog.⁴²  Even taken at face value, 0.1% of the genome is a small fraction.  The method has not been taken up by others, and the 2014 journal paper was not obtained for this article.

10.  "GWAS prove the genome is designed."  They prove that much of what matters in the genome for common traits is control sequence.  A designer would plausibly write that.  So would the evolutionary account, which always held that the functional fraction includes regulatory DNA.  The finding does not separate the two.

Issues and unanswered questions

1. No linkage pruning in the Catalog count. The 2026 row counts every listed variant. A count of one lead per independent signal would be the cleaner number, and needs reference-panel linkage data not used here.
2. The common-SNP control is mixed-ancestry. dbSNP's common set counts a SNP that is common in any population, while most GWAS are of Europeans. A European-only control would change the enrichments somewhat, not their direction.
3. The pricing formula's assumptions are untested. It assumes harmful noncoding and coding bases are equally visible to GWAS. No study measures that. A simulation from a population-genetic model with a set functional fraction would show how the method's output tracks the true fraction.
4. Rare variants are outside this article. A per-base count of rare disease-causing variants by class, from ClinVar or exome-plus-genome studies, would extend the picture to strong effects.
5. Two famous cases are thinner than told. No edit of the lactase-persistence variant in its native place in intestinal cells was found. A claimed CRISPR confirmation of the SORT1 variant was found only on a secondary page, without a primary. The FTO paper does not say how many donors' cells were edited.
6. Sources not obtained. Finucane 2015's published supplementary tables (the preprint's Table S3 is used); Hindorff 2009's supplement, which would settle a contradiction in its text about whether introns were enriched; Eddy 2012 and 2013 in full; Moran's 2023 book; Huang's 2014 journal paper; Frangoul 2021 beyond its abstract.
7. No design-side or creationist author has written a GWAS-specific argument beyond the headline share. Searches of Evolution News, ICR, CMI and Answers in Genesis, and of Wells 2011 and the 2017 Theistic Evolution volume, found none. This site's own Functional DNA article is the fullest version, and it needs the corrections in Part 6.

Verdict

The headline is right and misread.  About 96% of GWAS hits lie outside protein-coding DNA, because about 99% of the genome does.  Measured per base and against controls, the hits concentrate in coding DNA first and regulatory DNA second, while introns and intergenic DNA carry their share or less.  Most flagged SNPs are markers.  The causes behind them, once found, are still mostly noncoding, and the best of them are switches proven to the single letter, one of which is now the target of an approved CRISPR therapy.

So GWAS show that regulatory DNA explains more common disease risk than protein-coding DNA does.  That was a real discovery, and it corrected a protein-centered picture of disease.  It does not show that most of the genome is functional.  The arithmetic that turned the GWAS numbers into "at least 30%" on this site uses markers as causes and a coding share twice too large;  with its least biased inputs it gives about 6 to 10%.  The junk side's budget of roughly a tenth functional fits the GWAS data, and the design side's "most of the genome" does not get support from them.  A young-earth model is not harmed:  it expects a designed genome dense in control sequence, and it allows for decayed DNA since creation.  It should not cite GWAS as evidence that the bulk of the genome is in use.

Sources