Noncoding RNA and Pervasive Transcription

"Most lncRNAs still remain experimentally untouched or poorly characterized" — John Mattick and Paulo Amaral, RNA, the Epicenter of Genetic Information, 2023¹

"the existence of an RNA does not prove its biological significance, and conversely, the high level of transcriptional noise does not disprove the significance of the observed RNAs" — Kevin Struhl, 2007²

Summary

Only about 1% of human DNA spells out proteins,³ yet most of the genome is copied into RNA at some time in some cell.  Since about 2005 this "pervasive transcription" has been one of the main arguments that most of the genome is functional.  John Mattick and his colleagues argue that the noncoding RNA is a hidden regulatory layer.  Dan Graur, Larry Moran, Alexander Palazzo and others answer that transcription is cheap and sloppy, and that most of the RNA is noise.  This article counts what has been measured.  The short answer is that both slogans overreach, and the tests so far favour a small, real regulatory layer inside a large body of untested transcription.

1. Most of the genome is transcribed, but most of that is introns. ENCODE found 62% to 75% of the genome in RNA across 15 cell lines, and a 2013 reanalysis 85%.^{4 5} Counted for this article from GENCODE release 50, 67% of the genome already lies inside annotated genes, mostly in introns that are cut out and destroyed, while the exons of all genes cover 6.8%.⁶ ENCODE itself said "Most transcribed bases are within or overlapping annotated gene boundaries."⁷
2. Most lncRNAs are rare. ENCODE estimated that "80% of the detected lncRNAs are present in our samples in 1 or fewer copies per cell."⁴ Single-molecule counts found that this is not explained by a few cells holding many copies.⁸
3. The random-DNA control splits by organism. In yeast, 18,000 bases of random sequence made RNA at "levels ... comparable to yeast mRNAs."⁹ In mouse stem cells, 101,000 bases of human DNA written backwards "displayed no activity at all."¹⁰ So "polymerase copies anything" is true for yeast and false for at least one mammalian cell type. A computer model predicts four times less start activity in scrambled human DNA than in the real genome, and its authors list noise-like explanations for the excess as well as function.¹¹
4. The lncRNA catalogue is large and mostly untested. GENCODE lists 34,866 human lncRNA genes, and merged catalogues up to 173,112 lncRNAs. About 2,500 have been studied in any experiment.¹² Mattick and Amaral: "Most lncRNAs still remain experimentally untouched or poorly characterized."¹
5. When lncRNAs are removed, a minority do something measurable. A CRISPR screen of 16,401 lncRNA loci found 499, 3.0%, that changed cell growth in at least one of seven cell lines, against about 10% of protein-coding genes in a single line.¹³ Of 18 mouse lincRNA deletions, 3 were lethal;¹⁴ of 12 more, 11 were normal;¹⁵ of 25 zebrafish lncRNAs chosen for the best odds, none was needed for development, survival or fertility.¹⁶
6. The famous cases are mixed. XIST and FENDRR are clearly needed. MALAT1, one of the most abundant RNAs in the nucleus, can be deleted from mice with "no obvious gross abnormalities."¹⁷ NEAT1 knockouts looked normal, then showed a fertility defect in some females. The HOTAIR knockout phenotype did not replicate.¹⁸ By 2020 about 50 mouse lncRNAs had an effect on removal, and for seven the RNA itself was shown to be responsible by putting it back.¹⁵
7. Conservation picks out a few percent. About 5% of lncRNA sequence is constrained, and about 400 to 2,500 human lncRNAs are deeply conserved by mainstream measures.^{19 20} This article reports conservation as a mark both sides cite, not as a test.
8. Short RNAs are a different case. Transfer, ribosomal, spliceosomal and snoRNAs have known jobs. Of 1,881 human microRNAs in miRBase, 523 meet strict standards.²¹
9. Both sides misstate sources. Graur and colleagues carried Kevin Struhl's 90% noise estimate for yeast into human transcripts without saying so.²² The 2017 volume Theistic Evolution calls 70% to 90% of the genome lncRNA, when lncRNA exons are 2.2%, and cites a review for "Yet when examined, they do have functions" that says the function of "the overwhelming majority" is "currently not known."²³ An ICR article says lncRNAs "constitute about 75% of the genome."²⁴ The published sibling article's estimate that at least 68% of DNA is in functional elements multiplies the transcribed fraction by a localization rate measured on brain-expressed lncRNAs, not on transcribed DNA as a whole, and this article cannot reach that figure.
10. The young-earth reading loses nothing. A created genome since damaged predicts both function and decay, and Robert Carter has written that young-earth creationists "do not require the entire genome to be highly functional."²⁵ What the design side should drop is the use of transcription, conservation patterns and catalogue sizes as if they were tests of function.

Part 1: How much of the genome is copied into RNA

Three surveys, three numbers

A cell copies a stretch of DNA into RNA when it transcribes it.  For a protein-coding gene, the RNA is spliced, sent out of the nucleus and read by ribosomes.  But a cell also makes RNA that codes for no protein.  Some of it, such as the RNA in ribosomes, has been known for sixty years.  The surprise of the 2000s was how much more there is.

The ENCODE pilot project of 2007 surveyed 1% of the human genome with microarrays and tag sequencing.  It reported that "93% of bases are represented in a primary transcript identified by at least two independent observations (but potentially using the same technology); this figure is reduced to 74% in the case of primary transcripts detected by at least two different technologies."²⁶  Both figures assume that when two ends of a transcript are seen, everything between them was copied too, and the paper says so.²⁶

The full ENCODE project of 2012 sequenced RNA from 15 cell lines, split into nuclear and cytoplasmic fractions.  Across all of them, "62.1% and 74.7% of the human genome" was "covered by either processed or primary transcripts, respectively, with no cell line showing more than 56.7%" of that total.⁴  A single cell line, on average, covered 39% of the genome with primary transcripts and 22% with processed ones.⁴  Hangauer, Vaughn and McManus pooled public RNA sequencing data in 2013 and found "85.2% of the genome has evidence of transcription, with RNA-seq reads mapping directly to 78.9% of genomic sequence."⁵

So the measured figure is somewhere between about 60% and 85%, depending on how many samples are pooled and how much evidence a base needs.  The published sibling article Functional DNA is Evidence of Design uses the 85% figure, and it is a fair reading of Hangauer.  The question this article asks is the next one:  what is that RNA, and how much of it is there?

Most of it is the inside of known genes

The single most useful fact about pervasive transcription is where it lies.  ENCODE said it plainly:  "Most transcribed bases are within or overlapping annotated gene boundaries (that is, intronic), and only 31% of bases in sequenced transcripts were intergenic."⁷  An intron is the part of a gene that is copied into RNA and then cut out and destroyed before the RNA leaves the nucleus.  Every protein-coding gene with introns produces this RNA every time it is used.

Counted for this article from the current gene catalogue, GENCODE release 50, 67.0% of the human genome lies inside the span of some annotated gene, introns included.  The exons of all those genes, the parts kept in a finished RNA, cover 6.8%.  Protein-coding exons, with their untranslated ends, are 4.4%;  lncRNA exons are 2.2%.⁶  So when 75% to 85% of the genome is seen in RNA, about two-thirds of the genome was always expected to be there, because it sits inside genes.  What is new is roughly 8% to 18% of the genome copied from outside every annotated gene, plus the many new lncRNA exons and isoforms found inside and between them.

This was the point of a 2010 paper by Harm van Bakel, Timothy Hughes and colleagues in Toronto.  In their RNA sequencing of human and mouse tissues, "as much as 92%–93% of all reads can be accounted for by annotated gene regions," and "only 2.2%–2.5% of reads cannot be explained" by genes or known cDNAs.²⁷  They concluded that "the genome is not as pervasively transcribed as previously reported."²⁷  John Mattick's group replied the next year that a rare transcript is not an absent one:  a transcript present at about 10 copies per cell, "a common level of abundance," in 0.1% of cells would have only about an even "chance of being detected by any reads" at van Bakel's depth.²⁸  Van Bakel's team answered that they "strongly disagree with the fundamental argument put forward by Clark et al., which is that the genomic area corresponding to transcripts is more important than their relative abundance."²⁹  Both sides were right about the fact they stressed.  Most bases are copied at some time, and most of the RNA by mass comes from known genes.

How much RNA: copies per cell

The ENCODE transcription paper measured abundance as well as coverage.  "Assuming that 1–4 r.p.k.m. approximates to 1 copy per cell, we find that almost one-quarter of expressed protein-coding genes and 80% of the detected lncRNAs are present in our samples in 1 or fewer copies per cell."⁴  The lncRNAs were also more cell-specific:  29% were "detected in only one of the cell lines," and "only 10% were expressed in all cell lines," against 53% of protein-coding genes expressed in all.⁴  They were concentrated in the nucleus, which the paper called "a centre for the accumulation of ncRNAs."⁴  ENCODE's 2014 summary put it in genome terms:  "For polyadenylated RNA, where it is possible to estimate abundance levels, 70% of the documented coverage is below approximately one transcript per cell."³⁰

Alexander Palazzo and T. Ryan Gregory read the same data as the signature of junk.  Leaving aside about 9,000 transcript species "that are thought to be derived from conserved lncRNA, then most of the annotated noncoding polyadenylated RNAs are present at levels below one copy per cell and are found exclusively in the nucleus."  "In aggregate, these data indicate that the majority of intergenic RNAs are degraded almost immediately after transcription."³¹

An average below one copy per cell can mean two things.  Every cell may carry a stray molecule now and then, or a few cells may carry many while the rest carry none.  The published sibling article argues for the second:  "If an RNA had 500 copies in one cell type and one stage of development, the average could still be less than one"³²  ENCODE's own authors allowed the same:  the low level "may not necessarily be the result of consistent low RNA copy number in all cells within the population interrogated, but may also result from restricted expression in only a subpopulation of cells."⁴

That is a testable claim, and it has been tested once at the level of single molecules.  In 2015, Moran Cabili, John Rinn, Arjun Raj and colleagues counted individual lncRNA molecules by fluorescence in single cells, for 34 lncRNAs "chosen to span diverse characteristics" in three human cell types.⁸  They found "43% of lncRNA-cell pairs having 10 or fewer molecules per cell on average," with a median of 14 molecules against 36 for the mRNAs they examined.⁸  And "the low abundance of lncRNAs observed from cell population measurements cannot be explained by high expression in a small subset of 'jackpot' cells."⁸  Most of these lncRNAs were present at a few copies in most cells.  The test does not settle the question.  The 34 were chosen because they could be seen, they were studied in cultured cells, and a few-copy RNA can still matter.  Mattick and Amaral note that XIST, the RNA that shuts down one X chromosome in every female cell, is itself present at low levels.¹  But the "rare cells with many copies" explanation, which is the one the published article offers, did not hold for the lncRNAs that were measured this way.

Part 2: The control experiment: what does random DNA do?

Why a control is needed

The sibling article on transposable elements found that the cleanest test of a biochemical mark is to ask how often the same mark lands on DNA that cannot have a function.  There, 47% of transposable elements carried an active chromatin state somewhere, and so did 49% of randomly placed control sequences.³³  Transcription needs the same test.  If RNA polymerase copies any DNA it touches at some low rate, then "most of the genome is transcribed" says nothing about function.  If it copies random DNA almost never, then the genome's own transcription needs an explanation.

Kevin Struhl, a Harvard biochemist who studies how yeast genes are switched on, put the noise model in 2007.  "I suggest that ~90% of Pol II initiation events in yeast represent transcriptional noise, and that the specificity of initiation is comparable to that of DNA-binding proteins and other biological processes."²  He was not using "noise" in the sense of cell-to-cell variation.  "The concept of transcriptional noise used here is distinct from noise arising from stochastic variations in the expression level of a given gene in individual cells."²  This matters, because Mattick and Amaral answer the noise model by saying that the concept "was introduced in the 1990s by the observation of the cellular heterogeneity and stochastic fluctuations in the firing of known promoters in bacteria and yeast, not spurious transcription from illegitimate initiation sites."¹  Struhl's paper, the one critics cite, defined it as spurious initiation from the start.  Struhl also named the experiment that would settle it:  "By definition, transcription (and protein binding) from evolutionarily unrelated DNA is noise, so analysis of cells containing large regions of evolutionarily unrelated DNA (for example, Escherichia coli DNA in human cells) will be of interest."²  Sean Eddy made the same proposal in 2013 as a "random genome" project.¹⁰

Yeast: random DNA is transcribed like a gene

Struhl's laboratory ran the experiment in yeast in 2023.  Zlata Gvozdenov, Zeno Barcutean and Struhl built an 18,000-base stretch of computer-generated random sequence and put it into yeast cells.  "Steady-state levels of random-sequence RNAs are comparable to yeast mRNAs, although transcription and decay rates are higher."⁹  Initiation happened at many places, which they read as "very low intrinsic specificity" of the machinery.  They called their result "in excellent accord with calculated estimates that only 10%–20% of elongating RNA Pol II molecules generate mRNAs, such the remaining 80%–90% represent transcriptional noise."⁹  A year later Brendan Camellato and Jef Boeke at New York University put a 101,000-base synthetic locus into yeast:  the human HPRT1 gene and its surroundings written backwards, so that it kept the base composition of human DNA but none of its evolved signals.  They "observed widespread activity of both reversed and native HPRT1 loci in yeast, despite the lack of evolved yeast promoters."¹⁰

So in yeast, pervasive transcription is what random DNA does.  Its observed transcription of most of its own genome is no evidence of function at all.

Mouse stem cells: random-like DNA is silent

The same reversed human locus behaved the opposite way in mouse embryonic stem cells.  "By contrast, the reversed locus displayed no activity at all in mouse embryonic stem cells, and instead exhibited repressive chromatin signatures."¹⁰  A version stripped of CpG pairs lost the repressive marks but stayed silent.  The authors drew two conclusions, one for each side.  "The presence of extremely few RNA-seq reads mapping to the synthetic loci suggests that spurious transcription initiation is not common, and that much low-level transcription observed genome-wide may be a signal artefact or experimental noise, and does not reflect widespread functional transcription.  Conversely, transcription observed at a significantly higher level may well be functional."  And they limited it:  "these results are obtained from embryonic stem cells, and may not translate to all mammalian cells."¹⁰

In 2026 a group at the University of Auckland, with Marcel Dinger among the authors, ran the comparison for the whole human genome by computer.  They used a deep-learning model trained on human transcription-start data, checked that it predicted silence for Camellato's reversed locus, and applied it to reversed and shuffled copies of the human genome.  They found that "transcription is predicted to be sparse across all randomization methods, initiating with at least four-fold lower frequencies than in the native human genome."  They concluded that "most transcription is not a consequence of background noise," and listed three explanations for the excess:  "functional noncoding RNA transcripts, some general requirement for extensive transcription initiation/elongation, and/or mutational biases leading to the frequent appearance of transcription initiation sites by chance."¹¹  Scrambling only the transposable elements made little difference to the prediction.¹¹

What the control shows, priced

1. Yeast. Random DNA is transcribed at about the level of real genes. A model in which most yeast noncoding RNA is noise needs no special pleading.
2. Mammalian stem cells. One 101-kb random-like locus gave essentially zero transcription. So the simplest form of the noise model, "polymerase copies whatever it touches", is false for that mammalian cell type. This is a real point for the design side, and the junk side's slogan "transcription is cheap and random" does not survive it unchanged.
3. Whole human genome, predicted. Native human DNA starts transcription at least four times as often as scrambled human DNA. At most about a quarter of initiation would be expected from sequence that carries no evolved or designed signal.
4. What the excess is. The control does not say the excess is function. Real noncoding DNA is not random: it is full of decayed promoters, enhancers and transposon fragments that once carried start sites. Adey and colleagues' own third explanation, sites created "by chance" through mutational bias, is a noise explanation that their randomization cannot exclude. And Camellato's team, the experimenters, read the low-level genome-wide signal as artefact or noise, and the higher-level signal as possibly functional. That is close to what Struhl said in 2007: an RNA's existence proves nothing either way.

The fair summary is that the control experiment has been done once in a mammalian cell, and it removed the cheapest version of the noise model without establishing function.  No one has yet put a megabase of random DNA into a differentiated human cell, which is the experiment both sides should want.

Part 3: Long noncoding RNAs: how many, how conserved, how tested

How many there are depends on who is counting

A long noncoding RNA, or lncRNA, is by convention any RNA longer than 200 bases that does not code for a protein.  The cutoff has no biological meaning.  Mattick and 28 coauthors, in a 2023 consensus statement, called it "arbitrary" and "a convenient size cutoff in biochemical and biophysical RNA purification protocols."³⁴  So "lncRNA" names a size class, the way "small mammal" does, not a kind of molecule with a shared job.

The spread between 28,000 and 173,000 is mostly a matter of how little evidence each catalogue accepts.  Mattick and Amaral cite the high end:  "Well over 100,000 human lncRNAs have been recorded."¹  The useful number for this article is the last row.  About 2,500 human lncRNAs have been studied in any experiment at all.  That is 7% of GENCODE's set and 1.5% of NONCODE's.³⁸

How many are conserved

Conservation, in this literature, means a sequence that has changed less between species than the neutral expectation.  The articles in this section treat it as one lower bound on function under the assumption of common descent, never as the test of function, and the published article Functional DNA is Evidence of Design argues that lack of conservation implies nothing unless common descent is assumed.  The figures are still worth knowing, because both sides cite them.

1. Within lncRNA sequence, about 5% is constrained. Jasmina Ponjavic, Chris Ponting and Gerton Lunter compared 3,122 mouse lncRNAs with human and rat in 2007: "The proportion of conserved sequence (4.1%–5.5%) in these macroRNAs is comparable to the density of exons within protein-coding transcripts (5.2%)."¹⁹ That was enough for them to conclude that the set as a whole carries "the imprint of purifying selection." It also means 95% of the sequence in these transcripts shows no constraint at that resolution.
2. Most lncRNAs are young by the mainstream's dating. Across 11 tetrapods, Anamaria Necsulea and Henrik Kaessmann reported "approximately 11,000 primate-specific lncRNAs and 2,500 highly conserved lncRNAs, including approximately 400 genes that are likely to have originated more than 300 million years ago."²⁰ Across 17 species, Igor Ulitsky's group found that "in each species >70% of lincRNAs cannot be traced to homologs in species that diverged >50 million years ago," while "over a thousand human lincRNAs are likely to have conserved functions in mammals."³⁹ Only 29 were conserved between fish and mammals.³⁹
3. The FANTOM5 count of "potentially functional" lncRNAs is a count of marks. Hon and colleagues combined trait-linked variants, gene-expression links and conservation to "identify 19,175 potentially functional lncRNAs in the human genome."³⁵ That is 69% of their atlas. None of the three criteria is a test of the RNA, and all three would also flag a stretch of DNA whose transcript is a byproduct of an enhancer.

So between about 400 and about 2,500 human lncRNAs are deeply conserved by the usual measures, and over a thousand more are conserved within mammals.  That is between 1% and 10% of GENCODE's 34,866.

How many have been tested, and what happened

Five kinds of test have been run at scale.  Each removes the RNA in a different way, and each has a weakness the other side can point to.

The cell-based numbers and the animal numbers look far apart, and the reason is what each counts.  Guttman's 93% is a molecular effect:  when the RNA was knocked down, some other genes' RNA levels moved.  Such ripple effects are common for any perturbation, and they do not show that the organism needs the RNA.  Liu's 3.0% is a hard phenotype, slower growth, in dividing cells.  The authors compared it with protein-coding genes:  "The fraction of growth modifier lncRNA loci identified in our screens (1–8% per cell line) was less than the fraction of essential protein-coding genes in previous reports (10–11%)."¹³  And 89% of their hits worked "exclusively in one cell type."¹³  So about 3% of lncRNA loci did something measurable in at least one of seven cell lines, against about 10% of coding genes in a typical single line.

The animal tests were run on lncRNAs picked for the best odds.  Goudarzi's team "selected 25 zebrafish lncRNAs based on their conservation, expression profile or proximity to developmental regulators."¹⁶  Sauvageau's 18 were chosen from catalogues with filters for conservation and expression.  Even so, most deletions did nothing visible.  Goudarzi found that "none of the lncRNAs were required for embryogenesis, viability or fertility," and in the one case with a defect, "absence of putative DNA regulatory-elements, but not of the lncRNA transcript itself, resulted in abnormal development."¹⁶

That last point cuts both ways.  A team of thirteen geneticists led by Chris Ponting warned in 2014 that "Phenotypes arising from deletion of an entire genomic locus cannot be unequivocally attributed either to the loss of the lncRNA per se or to the associated loss of other overlapping DNA regulatory elements."⁴¹  A positive deletion may be a DNA switch, not an RNA.  A negative deletion may hide a function that shows only in an untested tissue, a stressed animal, or a wild population.  The cleanest evidence that an RNA itself acts is a rescue:  delete the gene, then supply the RNA from elsewhere and see the defect go away.  A 2020 review by Fan Gao, Philipp Kapranov and colleagues tabulated every mouse lncRNA reported to have an effect when removed:  about 50 of them.  The table marks an RNA-based rescue for seven:  H19, Malat1, Fendrr, Firre, lncKdm2b, PCFL and Hand2as.¹⁵  Worms and flies, which are cheaper to screen, gave higher rates:  23 of 155 C. elegans lincRNA knockouts affected at least one of six traits, and 33 of 105 testis-specific fly lncRNA knockouts reduced male fertility.¹⁵

The famous cases, claim and retest

The cases cited most often for lncRNA function are not a random sample.  They are the ones with the best evidence.  Their record is mixed.

Kung, Colognori and Lee, writing in the journal Genetics in 2013, summed up the MALAT1 and NEAT1 surprise:  "knockouts of neither NEAT1 nor MALAT1 have robust phenotypes."²³  Abundance, conservation and tidy nuclear location did not predict that the animal needed the RNA.

What the counts add up to

1. Catalogued: about 35,000 human lncRNA genes in GENCODE, and up to 173,000 lncRNA entries in the largest merged catalogue.
2. Studied in any experiment: about 2,500, or 7% of GENCODE's set.
3. Shown to do something measurable when removed, in an unbiased screen: 3% of 16,401 loci, for growth, in at least one of seven cell lines.
4. Shown in a living animal: about 50 mouse lncRNA loci by 2020, with the RNA itself shown by rescue for about 7. Add a few dozen well-studied human and mouse cases such as XIST, TSIX, AIRN and KCNQ1OT1, and the count of lncRNAs with strong evidence that the RNA acts is in the low hundreds at most.
5. Untested: more than 90% of GENCODE's lncRNAs, and nearly all of the larger catalogues.

The regulatory side's own sources say the same thing.  Mattick and Amaral:  "Most lncRNAs still remain experimentally untouched or poorly characterized."¹  The 2023 consensus statement counts "more than 2,000 publications reporting validated lncRNA functions, although most have yet to be followed up in any detail."³⁴  Two thousand publications is not two thousand lncRNAs:  hundreds of papers each have been written on H19, HOTAIR, MALAT1 and XIST.

Part 4: The short RNAs, briefly

Short noncoding RNAs are a different case from lncRNAs.  Several classes have known jobs that nobody disputes:  transfer RNAs, the RNAs of the ribosome and the spliceosome, and the small nucleolar RNAs that guide chemical changes to ribosomal RNA.  Together their genes are a tiny part of the genome.  In GENCODE release 50, the 1,878 microRNA genes cover about 150,000 bases and the 985 snoRNA genes about 115,000, each well under 0.01% of the genome.⁶  The debate over short RNAs is about how many catalogued entries are real, not about how much DNA they occupy.

1. MicroRNAs. The miRBase registry, release 22, lists "1917 annotated hairpin precursors, and 2654 mature sequences" for humans.⁵¹ Its own curators classify "26% of human microRNA annotations" as high confidence.⁵¹ Bastian Fromm and colleagues applied uniform criteria for a real microRNA gene, such as reads from both arms of the hairpin, and found that "of the 1,881 human miRNAs listed in miRBase, only 523 genes belonging to 283 families meet the standards for miRNA annotation."²¹ So about 500 human microRNA genes are well supported, and about 1,400 entries are doubtful. Of the real ones, many are deeply conserved and some are essential: that was one of the first lessons Mattick drew from them.
2. piRNAs. These short RNAs silence transposons in the germ line. piRBase now holds over a hundred million sequences across species, but many entries in such databases are fragments of other RNAs. Juan Pablo Tosar and colleagues found that fragments of transfer and ribosomal RNA, "Although a minority of annotated piRNAs falls in this category, they account for the vast majority of piRNA expression in somatic non-gonadal tissues."⁵²
3. snoRNAs. The snoDB 2.0 database "now contains 2123 non-redundant entries" for human snoRNAs.⁵³ Many have a known target on ribosomal or small nuclear RNA; a large minority are "orphans" with none known.

The short RNAs show the same pattern as the lncRNAs at a smaller scale.  A core of classes has clear, tested functions.  Around it is a catalogue swollen by fragments and low-evidence entries, and the size of the catalogue is not evidence of function.

Part 5: The two positions, in their own words

The regulatory-layer case

John Mattick, an Australian molecular biologist who later ran Genomics England, has made the case for twenty years.  In 2003 he wrote that "ncRNA dominates the genomic output of the higher organisms and has been shown to control chromosome architecture, mRNA turnover and the developmental timing of protein expression, and may also regulate transcription and alternative splicing."⁵⁴  His argument has three parts, and each is a fair argument.

1. Precision. Mattick and Marcel Dinger argued in 2013 that "the extent of precise dynamic and differential cell- and tissue-specific transcription and splicing observed from the majority of the human genome is a more reliable indicator of genetic function than conservation."⁵⁵ The brain study behind much of this found that of 1,328 lncRNAs probed in the Allen Brain Atlas, 849 were expressed in the adult mouse brain, "and found that the majority were associated with specific neuroanatomical regions, cell types, or subcellular compartments."³
2. Low abundance is expected. "Regulatory RNAs would likewise require relatively low average expression levels, i.e., more localized expression, and more dynamic control."¹ Transcription-factor genes, whose function nobody doubts, are also expressed at low levels.
3. Sampling. "Where tested, these noncoding RNAs usually show evidence of biological function in different developmental and disease contexts, with, by our estimate, hundreds of validated cases already published and many more en route, which is a big enough subset to draw broader conclusions about the likely functionality of the rest."⁵⁵

The case has grown more ambitious.  A 2023 essay by Mattick is titled "A Kuhnian revolution in molecular biology," and its abstract concludes "that most genes in complex organisms specify regulatory RNAs."⁵⁶  It also has honest limits that its makers state.  In the same 2013 paper:  "The fact is we simply do not know."⁵⁵  John Rinn and Howard Chang, whose laboratories found HOTAIR:  "With the mapping of thousands of lncRNA loci, the next challenge is to determine what lncRNAs do."⁵⁷

The noise case

The other side does not say lncRNAs are never functional.  It says that a transcript's existence is not evidence of function, and that the null hypothesis should be the one the evidence must overturn.  Alexander Palazzo and Eliza Lee in 2015:  "in the absence of any such data, the appropriate null hypothesis is that the RNA in question is junk."⁵⁸  W. Ford Doolittle called the inference from transcription to function the weakest of three, "mere existence":  "Because a region is transcribed, its transcript must have some fitness benefit, however remote."⁵⁹  Chris Ponting and Wilfried Haerty, reviewing the genome-wide data in 2022:  "most lncRNA transcript models reflect transcriptional noise or provide minor regulatory roles, leaving relatively few human lncRNAs that contribute centrally to human development, physiology, or behavior."⁶⁰

Two of the noise side's points are claims about origin, not about function.  Palazzo and Eugene Koonin argue that "junk transcripts provide the raw material for the evolution of diverse long non-coding (lnc) RNAs by non-adaptive mechanisms, such as constructive neutral evolution."⁶¹  And the biochemist Larry Moran, the most persistent public critic, gave his own guess for the count:  "It's hard to get a handle on just how many have a proven function but I doubt it's more than 200 in humans."⁶²  In 2018 he endorsed the conclusion of a review by Florian Kopp and Joshua Mendell:  "Clearly, there is no unifying answer – meaningful understanding of lncRNA function (or lack thereof) can only be achieved from detailed study on a case-by-case basis."⁶³  That is also the method this article follows.

The noise side has a citation problem of its own.  Graur and colleagues' much-quoted 2013 attack on ENCODE says that "Some studies even indicate that 90% of the transcripts generated by RNA polymerase II may represent transcriptional noise," citing Struhl.²²  Struhl's figure was for initiation events, and it was for yeast.²  Struhl's own laboratory measured it in yeast sixteen years later and found it right, and the only mammalian test so far went the other way (Part 2).  Graur's sentence carried a yeast estimate into the human genome without saying so.

Where the published article's estimate goes further than the data

The published article Functional DNA is Evidence of Design multiplies two figures:  "If at least 85% of DNA is copied to RNA, and at least 80% of RNA transcripts are taken to specific locations within a cell, then we could estimate that perhaps 85% * 80% = at least 68% of DNA is within functional elements."⁶⁴  This article cannot reach that figure, for two reasons found in the sources.

1. The 80% describes a sample of brain lncRNAs, not 80% of transcribed DNA. It comes from Mattick's account of the brain study above, "In 80% of the cases where we had sufficient resolution to tell," among the 849 lncRNAs seen in mouse brain.⁶⁴ Those lncRNAs' exons are a small part of the genome.
2. Most of the 85% is introns. Part 1 showed that two-thirds of the genome lies inside annotated genes, and ENCODE found that most transcribed bases are intronic. Intron RNA is cut out of the gene's transcript and broken down in the nucleus. It is not trafficked anywhere.

The same article's fourth reason, that "RNAs usually affect development or disease" when tested, rests on Mattick's statement that "almost every time you functionally test a non-coding RNA that looks interesting because it's differentially expressed in one system or another, you get functionally indicative data coming out."  That describes RNAs chosen because they looked interesting.  The systematic tests in Part 3, which deleted lncRNAs chosen to be the likeliest to matter, found effects in a minority.  None of this touches the article's other lines of argument (protein binding, disease association, genetic load), which this article does not test.

How design writers have used the data

Design and creationist writers have mostly repeated the regulatory side's framing.  Where they cite numbers, they can be checked.

1. The 2017 volume Theistic Evolution. Chapter 15, by Ann Gauger, Ola Hössjer and Colin Reeves, says: "Seventy to 90 percent of the genome is transcribed into RNA but does not produce any protein. Called long noncoding RNAs (lncRNAs), these RNAs can be nested within genes going in either direction, or they can come from DNA with no protein-coding genes."⁶⁵ The figure matches the review by Kung, Colognori and Lee that the chapter cites in its next note, which says "as much as 70–90% of our genome is transcribed at some point during development."²³ But transcribed DNA is not lncRNA. Counted for this article, lncRNA exons cover 2.2% of the genome and lncRNA genes, introns included, 29%.⁶ The chapter goes on: "Yet when examined, they do have functions."⁶⁵ The note for that sentence is the same Kung review, which says "the majority of lncRNAs await further verification" and that "the characteristics and function of the overwhelming majority of these lncRNAs are currently not known."²³ The chapter's sentence is stronger than its source.
2. Jeffrey Tomkins, Institute for Creation Research. A 2025 article says that "numerous lncRNAs, which constitute about 75% of the genome," changed the picture, and that "there are about 20,000 protein-coding genes and over 100,000 lncRNA genes in the human genome."²⁴ The 75% is the transcribed fraction again, not the lncRNA fraction. The "over 100,000" fits only the loosest merged catalogues (Part 3); GENCODE and FANTOM5 count 28,000 to 35,000. A 2017 article's "at least two to one" ratio of lncRNA to coding genes depends on the catalogue: GENCODE gives 1.7 to 1, LncBook 4.7 to 1.⁶⁶
3. Casey Luskin, Discovery Institute. A 2026 newspaper column described the reversed-DNA experiments of Part 2: "If transcription is random and uncontrolled, then scrambling their DNA should not change the amount of RNA that's produced. But if RNA is produced in a controlled, regulated, and functional manner, then transcription in the scrambled sequence should drop dramatically. And that's exactly what they found."⁶⁷ For the mouse cells that is right, and it is a fair point against the cheapest noise model. The column does not mention that the same reversed DNA showed "widespread activity" in yeast, that the experimenters read the low-level genome-wide signal as "artefact or experimental noise," or that the Auckland paper it quotes listed noise-like explanations alongside function.
4. Jonathan Wells, The Myth of Junk DNA (2011). Wells answered van Bakel's 2010 paper by saying that "the Toronto researchers used methods that virtually guaranteed their results," because they looked only at polyadenylated RNA, and he cited a rebuttal by Philipp Kapranov's group.⁶⁸ The criticism of the method was fair. But Kapranov's own data found that the extra RNA in total RNA was mostly from introns: "30%-52% of all informative reads in the human RiboMinus RNA did not overlap exons and fell within intronic regions."⁶⁹ Later total-RNA and single-cell data (Part 1) still found most lncRNAs at a few copies per cell.
5. Robert Carter, Creation Ministries International. Carter made the concession this article's verdict depends on: "I'd like to point out that young-earth creationists do not require the entire genome to be highly functional."²⁵

What a young-earth reading must decide

A young-earth account holds that the human genome was created, and has since been damaged by mutation.  That account predicts function in the created genome and decay since.  It does not by itself predict how much of today's transcription is functional, and nothing in it requires every transcript to matter, as Carter says.  So the young-earth reader is free to accept what the tests show.

What the tests show sorts into three layers.

1. Measured function: a core of RNA classes with known jobs (ribosomal, transfer, spliceosomal, snoRNA, about 500 well-supported microRNA genes) and, among lncRNAs, a set in the hundreds with good evidence, of which a few dozen are shown to act as RNA.
2. Marks without a test: cell-specific expression, nuclear location, conservation in a few percent of sequence, and overlap with trait-linked variants. These describe tens of thousands of lncRNAs. They are the same marks the transposable-element article found to be poor predictors when tested.
3. No evidence either way: most transcribed bases, which are introns of known genes and low-level RNA from between them, and more than 90% of catalogued lncRNAs.

The design argument cannot claim layer 3, and should not claim layer 2 as measured.  It can claim, from Part 2, that mammalian cells do not transcribe random DNA by default, so that most of the transcription in human cells is directed by sequence.  Directed is not the same as functional.  An enhancer that switches on a gene may produce RNA from its own DNA as a side effect, and many lncRNAs come from enhancers.³⁵  Whether that RNA, or only the DNA switch beneath it, matters is the open question, and it can only be settled one locus at a time.

Objections and responses

1. "ENCODE found 75% to 85% of the genome transcribed. That is function." Transcription is an activity, and the question is what the RNA does. About two-thirds of the genome lies inside annotated genes, mostly introns, and intron RNA is cut out and destroyed.⁶ In yeast, random DNA is transcribed about as much as real genes.⁹ Coverage tells where polymerase has been, not what the product is for.
2. "Low abundance does not mean noise. Regulators are rare." True, and Mattick and Amaral make the point well.¹ But rarity is also what noise looks like, so it cannot count for either side. The one single-molecule test found that low lncRNA averages were not caused by a few cells with many copies.⁸ Abundance is neutral evidence; the knockouts are the test.
3. "Cell-specific, regulated expression is the signature of function." It is the signature of regulated transcription. An enhancer that switches on in one cell type will make RNA in that cell type whether or not its RNA matters, and most intergenic lncRNAs start at enhancers.³⁵ The zebrafish case where a defect came from the DNA, not the RNA, shows the difference in practice.¹⁶
4. "Knockouts with no phenotype prove nothing, because of redundancy and untested conditions." This is partly right, and the article does not claim the null-result lncRNAs are proven junk. NEAT1 looked dispensable for three years before a pregnancy defect turned up in some females.⁴⁶ But the argument cuts both ways: a claim of function that cannot be tested by removal is not a measured claim. The honest result is "untested" or "no effect found", not "functional".
5. "Lack of conservation does not mean lack of function." Agreed, and the section's articles never use conservation as the test of function. Mattick and Amaral give real cases of fast-evolving functional lncRNAs.¹ This article reports conservation only as one of the marks the two sides cite, and it is a mark that a young-earth reader need not grant.
6. "Thousands of papers report lncRNA functions." More than 2,000 publications is not 2,000 lncRNAs, and most report knockdown in cancer cell lines, which the 2023 consensus statement itself says are hard to interpret and "most have yet to be followed up in any detail."³⁴ A curated list of lncRNA genes with "prior established functions," used by the GTEx consortium, held 954 genes.⁷⁰ That is the high-end figure for function by a loose standard, and it is under 3% of GENCODE's lncRNA set.
7. "The reversed-DNA experiments show transcription is controlled, so it is functional." They show that mouse stem cells do not transcribe random-like DNA by default.¹⁰ Control of where transcription starts is not the same as a use for what is made, and the experimenters themselves read the low-level genome-wide signal as artefact or noise.¹⁰
8. "Most transcription is junk because evolution made it." That is a claim about origin, and Palazzo and Koonin's model is an origin model.⁶¹ The data in this article do not need any history to be read: the fraction tested is small, and the fraction tested and positive is a minority. The verdict would be the same if the genome were created last week.
9. "The 90% noise figure shows most RNA is junk." The figure was Struhl's estimate for yeast initiation events.² Graur and colleagues applied it to human transcripts without saying so.²² It is confirmed in yeast and has not been measured in a human cell.
10. "The design side predicted function and was right." It predicted function in noncoding DNA, and some of it has been found: XIST, FENDRR, MALAT1's adult effects, NEAT1's pregnancy defect, about 500 real microRNA genes. The prediction that most of the transcribed genome is functional RNA is not supported by the tests run so far, and the design writers checked here stated it more strongly than the sources they cited.

Issues and unanswered questions

1. No random-DNA test in a differentiated human cell. Camellato's single 101-kb locus in mouse stem cells is the only mammalian measurement, and Adey's genome-wide figure is a computer prediction. A megabase of random DNA in a human cell line, read by total-RNA sequencing, would settle how much of the human background is noise. This is the experiment both sides should ask for.
2. No unbiased census of lncRNA knockouts. The animal tests ran on lncRNAs chosen for the best odds, and the cell screens scored one trait. The International Mouse Phenotyping Consortium has knocked out thousands of genes; a count of its lncRNA-annotated loci with any abnormal phenotype, against a matched set of coding genes, would give the first unbiased rate. It was not attempted here.
3. The "low hundreds" count of RNAs shown to act as RNA is an estimate. It rests on the seven mouse rescues in Gao and colleagues' 2020 table, the classic imprinting and X-inactivation RNAs, and the 954-gene curated list as an upper bound. A scored literature count, rescue by rescue, would firm it up.
4. Full texts not obtained. Mercer, Dinger and Mattick 2009, the chapter's cited source for "Seventy to 90 percent", was read only as an abstract, which says "most of the genome is transcribed in a developmentally regulated manner to produce large numbers of long non-coding RNAs." The figure may be in its body. Also abstract-only: Mattick 2003 and 2023 (BioEssays), Ponting and Haerty 2022, Palazzo and Koonin 2020, Necsulea 2014, Penny 1996, Marahrens 1997, Ripoche 1997, Ulitsky 2016, and Han 2018 (read through Gao 2020).
5. Copy-number calibrations. ENCODE's "1 copy per cell" equals 1 to 4 reads per kilobase per million, a single calibration in cultured lines. How far the 80% figure moves with the calibration has not been computed.
6. Abundance by total RNA. Wells and Kapranov were right that polyadenylated RNA misses much nuclear RNA. The share of total nuclear RNA mass from intergenic, non-intronic transcription, measured in normal human tissue, would answer the van Bakel and Kapranov dispute directly.
7. What the published article should do with its 68% estimate. This article finds that the 80% localization figure applies to a set of lncRNAs, not to transcribed DNA as a whole. Whether the published article's estimate should be revised is Eric's decision; it is recorded in the memory file for the planned update of that article.

Verdict

Most of the human genome is copied into RNA at some time, in some cell.  That is measured, and the critics of 2010 were wrong to minimise it.  But most of what is copied is the introns of known genes, most lncRNAs sit at a few copies per cell, and in the one organism where random DNA has been tested at scale, yeast, random sequence is transcribed as much as genes.  In mammalian stem cells the same test went the other way, so the simplest noise model is wrong for mammals, and transcription there is directed by sequence.  Directed is not the same as useful.

When lncRNAs are removed one at a time, a minority do something measurable:  3% of 16,401 loci in human cells, 3 to 5 of 18 and 1 of 12 in mice, 0 of 25 in zebrafish, and a few dozen famous cases with clear phenotypes, some of which did not replicate.  About 2,500 human lncRNAs have been studied at all, out of 35,000 to 173,000 catalogued.  So the answer to the question is:  some of the RNA is doing something, a measured few hundred to a thousand lncRNAs by the loosest count, and most of it has never been tested.  "Most of the transcribed genome is functional" is not shown by the data, and neither is "it is all junk."  The data favour a small, real regulatory layer inside a large body of untested transcription.

For the young-earth reader nothing is lost.  A created genome damaged by time predicts some function and some noise, and Carter's concession that the genome need not be "highly functional" throughout is the right position.  What the design side should give up is the use of transcription, conservation patterns and catalogue sizes as if they were tests, and the habit, seen in the 2017 volume, at ICR and in Luskin's column, of stating its sources more strongly than they speak.

Sources