Transposable Elements

"Our data suggest that these constitute a minority of all the putative TE-derived enhancers identified through bioinformatic analyses, highlighting the importance of functional tests when assessing the contribution of TEs to gene regulatory networks." — Christopher Todd, Özgen Deniz and Miguel Branco, after deleting and silencing transposon enhancers in mouse stem cells, 2019¹

"Of course, there is much repetitive DNA for which functions have not yet been discovered" — Jonathan Wells, The Myth of Junk DNA, 2011²

Summary

A transposable element, or TE, is a stretch of DNA that can copy itself or move to a new place in the genome.  Some move by cut and paste.  Most of ours copied themselves through RNA, the way a retrovirus does.  Copies of them make up almost half of human DNA.  For forty years they were the main exhibit for "selfish" or "junk" DNA.  Since about 2000, papers have shown TE copies that became genes, promoters, enhancers and immune switches.  Design writers now say that much or most TE DNA probably has a function.  Their critics say that a few copies were put to use and the rest are debris.

This article counts.  The short answer is that the evidence sorts into three layers, and neither slogan describes the biggest one.

1. TEs are 46% of the human genome, and almost none of them can still move. In the complete T2T-CHM13 genome, RepeatMasker marks 3.7 million TE fragments covering 1.43 billion bases: Alu and other SINEs 12.8%, L1 and other LINEs 20.7%, retrovirus-like LTR elements 8.7%, DNA transposons 3.5%.³ Only about 80 to 100 L1 copies per person can still copy themselves,⁴ and the Alu copies they help move number at most a few thousand.⁵ That is about 0.05% of TE copies.⁶ The rest are broken, cut short or scrambled.
2. New insertions do measurable harm. A new Alu insertion turns up in about 1 in 40 births, a new L1 in about 1 in 62.⁷ A 2016 review counted 124 insertions known to cause disease.⁸ In a 1999 count, recombination between Alu copies had caused three times as many cases as Alu insertions.⁹ In 2,954 cancer genomes, 35% carried new somatic insertions.¹⁰
3. Some copies were clearly put to work, and the experiments are good. Knocking out the mouse syncytin-A gene, a captured retrovirus envelope gene, kills embryos at mid-gestation.¹¹ The core of the RAG1 gene that builds antibodies comes from a Transib DNA transposon.¹² The memory gene Arc makes virus-like capsids from a retrotransposon Gag protein,¹³ and mice without it cannot form lasting memories.¹⁴ Deleting one ERV copy, MER41.AIM2, stops a human immune gene from answering interferon.¹⁵ Knocking down L1 RNA stops most mouse embryos at the two-cell stage.¹⁶ These are real functions, most of them shown by removing the copy or its RNA.
4. Most "TE function" counts are of marks, not of function, and the marks fall at or below chance. About 20% of transcription-factor binding sites lie in TEs.¹⁷ But when a 2019 study scored 4.4 million TEs across 127 tissue maps, 47% were in an active state in at least one, and 49% of randomly placed control sequences were too.¹⁸ Counted for this article, 15.7% of TE bases sit inside one of ENCODE's 2.3 million candidate regulatory elements, against 20.3% of the genome as a whole.¹⁹
5. Conservation picks out a few percent. Across 240 mammals, about 11% of the human bases under detectable constraint lie in TEs.²⁰ That is about 36 million bases, or 2.6% of TE sequence.⁶ An earlier 29-mammal map found over 280,000 TE fragments put to use, covering about 7 million bases, 0.5%.²¹ Both methods measure the rest of the genome against old TE copies, assumed neutral, so they cannot score the bulk of TE DNA as constrained.
6. When TEs are tested, a minority work. Of six predicted TE enhancers, chosen for strong evidence and cut out of mouse stem-cell genomes, three had a clear effect.¹ Silencing 34 marked copies of one family changed three genes.¹ Mice missing 1.5 million and 845 thousand bases of gene desert were healthy.²² The copies removed one by one in the studies read for this article total about 71 thousand bases, 0.005% of TE sequence; even ten times that is under 0.1%.⁶
7. The tally. Measured function by removal: a few dozen elements or families. Marked or constrained without a test: about 3% conserved and about 16% marked, overlapping in part. The remaining 80% or more: no evidence either way, and the class-wide tests that touch it look like what neutral decay predicts. "Mostly junk" is a reasonable reading of the class-wide data, not a measurement of each copy. "Mostly functional" is not supported.
8. Both sides have a habit to break. Orgel and Crick predicted in 1980 that the host would "occasionally find some use for particular selfish DNA sequences",²³ so a domesticated TE refutes nothing they said. Some later critics treated "TE-derived" as meaning "nonfunctional" and were wrong about the cases above. On the design side, Wells read a 2007 paper as finding that most conserved human DNA is mobile-element-derived; the paper's figure is about 4% to 5.5%.²⁴ The 2017 volume Theistic Evolution relied on a brain-cell count that a 2016 reanalysis found to be about fifty times too high.²⁵
9. What a young-earth reading must decide. At today's measured rate, 6,000 years would give each person about 5 new Alu copies. Humans carry about 7,000 Alus that chimpanzees lack.²⁶ So on a young-earth timeline those copies were either created where they are, in which case design predicts that they work, though class-wide data so far find young copies the least active, or they arrived in a burst about a thousand times faster than anything measured.⁶ No creationist model found for this article prices either option.

The short form:  transposable elements are not all junk, and the cases where they work are among the best experiments in genome biology.  But the cases are counted in dozens, and the elements in millions.

Part 1: What they are, and how many

Four classes

Biologists sort human TEs into four groups by how they move.²⁷

1. LINEs (long interspersed elements) are up to about 6,000 letters long and carry the genes for their own copying machinery, including a reverse transcriptase that turns RNA back into DNA. L1 is the only family still active in humans.²⁷
2. SINEs (short interspersed elements) are a few hundred letters long and carry no genes. They borrow the L1 machinery. The main human SINE is Alu, about 300 letters, and Lander's team called it "the only active SINE in the human genome".²⁷ The older MIR family stopped moving long ago.
3. LTR elements are framed by long repeats at both ends. Most are endogenous retroviruses (ERVs), copies of retroviruses that entered the germline, or their remains. Lander's team found that "Most (85%) of the LTR retroposon-derived ‘fossils’ consist only of an isolated LTR", the viral genes between the two end repeats having been cut out by recombination.²⁷
4. DNA transposons move by cut and paste, using a transposase enzyme. In humans "DNA transposons appear to have become completely inactive and long-terminal repeat (LTR) retroposons may also have done so."²⁷

A fifth, small group, SVA, is a composite element that L1 also moves.

The counts

The 2001 draft genome gave the first census.  RepeatMasker, the standard program, finds TE copies by comparing the genome with a library of consensus sequences.  Lander's team recognised "about 45% of the genome as belonging to this class."²⁷  Libraries have grown since, and the complete genome of 2022 fills in the gaps the draft left.  The table compares the draft with a count made for this article from the complete T2T-CHM13 genome.

The 2001 column is Lander's Table 11.²⁷  The CHM13 column is a union of RepeatMasker hits on the 3.12 billion bases of T2T-CHM13 version 2.0, counted with a script in this article's research folder.³  The T2T team's own annotation, with a library built from that genome, gives 46.02%,²⁸ and the RepeatMasker project's run on the older reference with a 2014 library gives 48.49% for all interspersed repeats.²⁹  The differences come from the libraries and the assemblies, not from new copies.  The round figure of "about half" holds whichever count is used.

"Copies" here means RepeatMasker fragments joined into one insertion where the program could join them.  Most are not whole elements.  Of about 549,000 L1 fragments in CHM13, 7,566 are at least 6,000 letters long, and only 320 of those belong to the youngest subfamily, L1HS.³

How many can still move

Almost none.  The best counts come from cloning candidate copies and testing them in cultured cells.

1. L1. Brouha and colleagues searched the 2001 draft and "yielded only 90 L1s with intact ORFs", then tested 82 of them; 40 moved in the assay. "From these data, we predict that there are 80–100 retrotransposition-competent L1s in an average human being. Remarkably, 84% of assayed retrotransposition capability was present in six highly active L1s (hot L1s)."⁴ Beck and colleagues looked at six people's genomes and found that "Each individual contained between three and nine ‘hot’ L1s in their genome".³⁰
2. Alu. Bennett and colleagues tested Alu copies in a cell assay and modelled the genome. All 852 copies matching the consensus exactly were active. Their model allows "up to 1836 “hot” Alu copies", and they call it "a liberal estimate for the number of functional copies".⁵
3. All classes. A 2007 review put it plainly: "only a small proportion (<0.05%) of these elements remain active today."³¹

Counted against CHM13's 3.7 million TE fragments, 100 L1s and 1,836 hot Alus are about 0.05%.⁶

How often a new one lands

Feusier and colleagues sequenced 33 three-generation families and counted new insertions in 437 births:  "an Alu retrotransposition rate of about 1:39.7 births (95% CI 22.4–79.4), a SVA rate of about 1:62.5 births (95% CI 30.6–153.8), and a L1 rate of about 1:62.5 births (95% CI 30.6–153.8)".⁷  Those rates rest on 11, 7 and 7 events, so the ranges are wide.  Older Alu estimates range from twice this rate to a fifth of it, and Feusier's own introduction gives one in 20 births for Alu while the 1999 paper it cites gives one in 200.⁹⁷  One new TE insertion somewhere in about every fifteen to twenty births is a fair summary of the measured rates combined.

Part 2: What new insertions cost

The case that TEs are mostly junk does not rest on their being harmless.  It rests on their being parasites:  copies that spread because they copy themselves, whatever they do to their host.  So the harms are part of the evidence, and they are measured.

Inherited disease

1. Insertions into genes. Hancks and Kazazian's 2016 survey lists 124 insertions known to cause human disease: 76 Alu, 30 L1, 13 SVA and a handful of others. "Most of the 124 disease-causing insertions … reported to date inactivate gene function through insertional mutagenesis or aberrant splicing."⁸ The cases include haemophilia, muscular dystrophies and breast-cancer genes.
2. Recombination between copies. Two Alu copies a few thousand letters apart can pair up by mistake when chromosomes cross over, and the DNA between them is lost or doubled. Deininger and Batzer counted 16 diseases caused by Alu insertion, "about 0.1% of human genetic disorders", and "33 cases of germ-line genetic diseases and 16 cases of cancer caused by unequal homologous recombination between Alu repeats", which "accounts for another 0.3% of human genetic diseases."⁹ That count is from 1999 and is surely low now, but no later survey with a single total was found.
3. Selection against them. Two measurements show the cost from the other side. Most old L1 copies are cut short, but "~ one-third of the human-specific LINE-1 s are full-length, indicating most full-length elements have been selected against throughout primate evolution".⁸ And in a 2002 sample of human exons, 61 alternatively spliced exons contained an Alu while none of the always-used exons did. Sorek, Ast and Graur concluded: "Presumably, evolutionary events that cause a constitutive insertion of an Alu sequence into an mRNA are deleterious and selected against."³²

Cancer

L1 is normally silenced in body cells, and many tumours wake it up.  Lee and colleagues found 194 new insertions in tumours from 43 patients:  "183 L1s, 10 Alus, and 1 ERV".  They appeared in colorectal, prostate and ovarian tumours, "but not in the blood or brain cancers."³³  The pan-cancer study of 2,954 tumour genomes found "19,166 somatically acquired retrotransposition events, which affected 35% of samples".¹⁰  Most such insertions are thought to be passengers.  Some are not:  "Aberrant L1 integrations can delete megabase-scale regions of a chromosome, which sometimes leads to the removal of tumor-suppressor genes".¹⁰

The host keeps them shut

Cells spend real effort holding TEs down, which is itself evidence that unchecked TEs are a threat.  DNA methylation, small RNAs and a large family of KRAB zinc-finger proteins all target them.  Pontis and colleagues found that these zinc-finger proteins "control the timely and pleiotropic activation of TE-derived transcriptional cis regulators during early embryogenesis".³⁴  The same study shows the double edge:  "the same KZFP-controlled TE-based enhancers later serve as developmental and tissue-specific enhancers."³⁴  A suppressed parasite and a regulated switch can be the same piece of DNA.

Alu RNA shows the same double edge.  Pairs of Alu copies in one RNA fold into double strands that look to the cell like a viral genome.  Ahmad and colleagues showed that in Aicardi-Goutières syndrome, a mutant form of the viral-RNA sensor MDA5 attacks "cellular dsRNAs formed by Alu retroelements", and concluded that "the increased efficiency of MDA5 in recognizing dsRNA comes at a cost of self-recognition".³⁵  The cell edits these Alu RNAs heavily, and Part 3 takes up what that editing means.

Pricing the harm

At one new TE insertion in about every 17 births,⁷ and with insertions into genes and Alu recombination together making up well under 1% of known genetic disease,⁹ the burden is real but small per generation.  That fits both readings.  On the parasite reading, most of the genome's 3.7 million copies are old and broken, and the few active ones cost little because the host silences them.  On a design reading, the harm is what a working system does when a few copies escape control.  Neither reading is tested by the harm alone.  What separates them is how much of the remaining 1.4 billion bases does anything.

Part 3: The copies that were put to work

Biologists call a TE copy that the host has taken over "domesticated" or "exapted".  The cases below are the ones usually cited, each with the experiment behind it.  They fall into three kinds:  a host gene that began as a TE gene;  a TE copy that works as a switch for a nearby gene;  and TE RNA or sequence used as raw material.

Genes that began as TE genes

1. Syncytins: placenta-building genes from retrovirus envelopes. A retrovirus uses its envelope protein to fuse with a cell. In 2000, Mi and colleagues found an ERV envelope gene, syncytin, expressed in the placenta's fused outer layer. Expressing it made other cells fuse, and "fusion of a human trophoblastic cell line expressing endogenous syncytin can be inhibited by an anti-syncytin antiserum."³⁶ A screen of all intact human ERV envelope genes found a second fusing one, syncytin-2, expressed specifically in placenta.³⁷ The decisive test came in mice, which have their own, unrelated pair. "Homozygous syncytin-A null mouse embryos die in utero between 11.5 and 13.5 days of gestation."¹¹ Syncytin-B is less critical: "Unexpectedly, syncytin-B null embryos are viable, with only limited late-onset growth retardation and reduced neonate number."³⁸ Lavialle's review counts six syncytin genes captured separately in primates, rodents, rabbits and carnivores, and notes that the primate and mouse genes "are not syntenic in the two species clades", meaning they sit at different places in the genome.³⁹ No human knockout is possible, so the human genes' role rests on cell work and on the mouse analogy.
2. RAG1: the antibody-cutting enzyme from a DNA transposon. Every jawed vertebrate builds its antibody and T-cell receptor genes by cutting and rejoining DNA segments. Kapitonov and Jurka showed in 2005 that the "region of RAG1 required for its catalytic activity is significantly similar to the transposase encoded by DNA transposons that belong to the Transib superfamily", and that the signal sequences RAG1 cuts at came from the ends of a Transib element.¹² That was a sequence comparison. In 2016 Huang and colleagues found a still-working relative, ProtoRAG, in the lancelet, an invertebrate: "11 of the 13 ProtoRAG copies arose during the last 2.7 million years", and its proteins cut and pasted DNA in human cells "using reaction mechanisms similar to those used by vertebrate RAGs."⁴⁰ In 2019 Zhang and colleagues identified two vertebrate changes "that together suppress RAG-mediated transposition more than 1,000-fold."⁴¹ The enzyme that makes antibodies is a transposase that was tamed so that it no longer jumps.
3. Arc: a memory gene that builds virus-like shells. Arc is needed for lasting memory. Mice lacking it "fail to form long-lasting memories for implicit and explicit learning tasks, despite intact short-term memory."¹⁴ In 2018 Pastuzyn and colleagues showed that "Arc self-assembles into virus-like capsids that encapsulate RNA", that neurons release it in vesicles that deliver Arc RNA to other cells, and that "Arc is derived from a vertebrate lineage of Ty3/gypsy retrotransposons, which are also ancestors to retroviruses."¹³ In fruit flies, a Gag-like protein called Arc1 also forms capsid-like structures and carries its own RNA from nerve to muscle, and "Disrupting transfer blocks synaptic plasticity".⁴² The knockout shows the gene matters; whether the capsid-carrying trick is what matters for memory is not yet shown by experiment.

Lander's team listed 47 human genes "with a likely origin in up to 38 different transposon copies" in 2001.²⁷  No current total was found;  later reviews describe cases rather than count them.  Even if the true number were several hundred, these genes are made from a few thousandths of TE sequence.

Copies that work as switches

4. MER41 and the interferon response. Chuong, Elde and Feschotte looked for ERV copies bound by the immune signalling proteins STAT1 and IRF1. They found the MER41 family, "with nearly 1,000 copies in humans (N=962) bound by STAT1 and/or IRF1 in at least one cell type."¹⁵ Then they cut single copies out of human cells with CRISPR. "Cells homozygous for the MER41.AIM2 deletion (ΔMER41.AIM2) failed to express AIM2 upon IFNG treatment … thus demonstrating that MER41.AIM2 is necessary for endogenous IFNG-inducible regulation of AIM2."¹⁵ When the edited cells were infected with vaccinia virus, "Secreted levels of activated caspase-1 were markedly reduced", a sign of a much weaker AIM2 inflammasome response.¹⁵ Three more MER41 copies near other immune genes were deleted, and "in contrast to AIM2, deletion of these MER41 elements did not completely abolish IFNG-induced transcript levels of these genes."¹⁵ So four copies were tested by deletion and all four did something; 962 were bound and not tested.
5. MER20 and pregnancy. Lynch and colleagues compared the womb lining of mammals and found that "1,532 genes were recruited into endometrial expression in placental mammals" and that "About 13% of recruited genes are within 200 kb of a Eutherian-specific transposable element (MER20)."⁴³ The MER20 copies carry enhancer marks and "directly bind transcription factors essential for pregnancy". That is a strong pattern, but in the 2011 paper it is a correlation with marks and binding, not a deletion.⁴³
6. HERV-K and SVA enhancers in human embryos. Fuentes, Swigut and Wysocka switched the HERV-K LTR family LTR5Hs on and off across the genome in stem cells: "activation/silencing of LTR5HS is associated with reciprocal up- and down-regulation of hundreds of human genes." They then deleted single copies near six affected genes. "For all six genes, we observed a significant decrease in expression upon deletion of the nearest LTR5HS."⁴⁴ The six were chosen because they had already responded, so this is confirmation of the cases, not a rate. A 2022 study of human and chimpanzee brain progenitor cells found that silencing the young SVA family in human cells "attenuated the expression of ∼25% of the genes that are upregulated in human intermediate progenitors relative to the same cell population in the chimpanzee."⁴⁵

TE RNA and sequence used as raw material

7. L1 RNA in the early embryo. Percharde and colleagues knocked down L1 RNA in mouse embryos with antisense oligonucleotides. "Almost no LINE1 KD embryos reach the blastocyst stage and most arrest at the 2-cell stage."¹⁶ "In embryos, LINE1 RNA is required for Dux silencing, synthesis of rRNA and exit from the 2-cell stage."¹⁶ An independent group found that switching L1 on too early or too late slows mouse development "independently of the coding nature of the LINE-1 transcript".⁴⁶ This is a function of L1 transcription as a class. It does not show that any particular copy is needed, and knockdown hits thousands of copies at once.
8. Alu exons. An Alu inside an intron carries near-splice signals, and a few mutations can turn it into an exon.⁴⁷ The often-quoted figure that about 5% of alternatively spliced exons come from Alu is Sorek, Ast and Graur's count: "Sixty one alternatively spliced internal exons (5.2%) had a significant BLAST hit to an Alu sequence, but none of the constitutively spliced internal exons had such a hit." Of those in coding regions, "The vast majority (84%) … caused a frame-shift or a premature termination codon."³² The authors read it as Alu exons being tolerated only when they are used rarely. It is raw material, not shown function.
9. Alu RNA editing. The enzyme ADAR changes A to I in double-stranded RNA, and Alu pairs are its main target. Bazak and colleagues found that "virtually all adenosines within Alu repeats that form double-stranded RNA undergo A-to-I editing, although most sites exhibit editing at only low levels (<1%)", and estimated "over 100 million human Alu RNA editing sites".⁴⁸ John Mattick and Mark Mehler proposed that this makes Alu "a modular substrate for RNA editing, driven by selection for higher-order cognitive function."⁴⁹ That is a hypothesis. The experiments point the other way: mice whose ADAR1 cannot edit die as embryos, and are rescued by deleting the double-stranded-RNA sensor MDA5.⁵⁰ In humans, the mutant MDA5 of Aicardi-Goutières syndrome is set off by Alu duplexes.³⁵ So the best-tested role of this editing is to keep the cell from mistaking its own repeat RNA for a virus. That is a cost of having Alus, handled by the host, not a job the Alus do.

What the cases show

Every case above is real, and several are among the cleanest experiments in genome biology.  They show that a TE copy can become a gene, a switch, or a regulatory RNA, and that the host depends on some of them.  But they are counted one at a time.  Four MER41 copies were deleted, six LTR5Hs copies, a handful of syncytins.  The families involved have hundreds of thousands of members.  How far the cases generalise is a question for Part 4.

Part 4: Counting at the scale of the whole class

Part 3's cases prove that TE copies can work.  The junk question is about the other 1.4 billion bases.  Four kinds of genome-wide evidence bear on it, and they need to be kept apart, because they measure different things.

Binding and open chromatin

Large fractions of the genome's regulatory-looking sites sit in TEs.

1. "Transposable elements contributed up to 25% of the bound sites" of the stem-cell master regulators OCT4 and NANOG in humans and mice.⁵¹
2. Across 26 pairs of transcription factors in human and mouse blood cells, "Overall, we found that 20% of binding sites were embedded within TEs. This number varied across different TFs, ranging from 2% to 40%."¹⁷
3. In ENCODE's DNase maps of more than 40 cell types, "44% of open chromatin regions were in TEs and that this proportion reached 63% for primate-specific regions." The same authors warn: "Although accessibility does not equate regulatory function".⁵²
4. Zoonomia found that "Most (87%) primate-specific transcription factor binding sites overlap transposable elements, unlike most non–primate-specific sites (30%)", with "the caveat that the binding sites have not been confirmed to have regulatory function".²⁰
5. "6-30% of cap-selected mouse and human RNA transcripts initiate within repetitive elements."⁵³

These numbers answer "what fraction of binding sites are TEs?"  They do not answer the junk question, which is "what fraction of TEs are binding sites?"  With 4.4 million TE fragments in the genome, even a million bound sites would touch a minority of copies.  The two fractions are easy to confuse, and both sides confuse them.

Marks across the whole class

Two counts turn the question the right way round.

1. Pehrsson 2019. Pehrsson, Choudhary, Sundaram and Wang scored each of 4,430,788 human TE fragments against 127 tissue and cell-type maps from the Roadmap Epigenomics Project. In any one tissue, few are active: "median ≤3% for all active regulatory states". Across all 127, "47% of individual TEs are in an active regulatory chromHMM state and 75% are in a transcribed state at least once."¹⁸ Then they ran the control that matters. They shuffled every TE to a random place in the genome and scored the shuffled set the same way. "The fraction of individual TEs ever in an active regulatory state is slightly lower than expectation, as estimated by shuffling the locations of all TEs ten times. On average, 49% (standard deviation, 0%) of shuffled TEs are in an active regulatory state in at least one epigenome."¹⁸ The authors add that which TEs are active in each tissue is not random: particular subfamilies switch on together.¹⁸ But as a share of the class, the active fraction is what random DNA in the same genome would show.
2. ENCODE's regulatory registry, counted for this article. ENCODE's current registry lists 2,348,854 candidate cis-regulatory elements (cCREs) for the human genome, each a stretch of open chromatin with a promoter, enhancer or CTCF mark in some sample. Together they cover 20.3% of the genome. Overlapping them with RepeatMasker's TEs gives 221 million TE bases inside a cCRE, 15.7% of all TE sequence.¹⁹ If TEs were spread across cCREs like the rest of the genome, they would make up 45.6% of cCRE bases; they make up 35.3%. Old families are enriched: MIR 25.4%, L2 24.0%, ERV1 24.7%, DNA transposons 21.2%. The young, abundant ones are depleted: L1 9.0%, Alu 12.2%.¹⁹

Both counts say the same thing.  TE DNA as a whole carries regulatory marks slightly less often than the average stretch of the genome, and the families old enough to have been sorted by selection carry them more often.  That is the pattern a co-option model predicts:  a large neutral pool, from which a minority of old copies were taken up.  A model in which most TE DNA was built for regulation would predict enrichment across the class, and the data show depletion.

A mark is not a function.  A cCRE is a window of a few hundred letters, and the binding site inside it is smaller, so base counts overstate.  But a mark is the most generous evidence available, and it covers a sixth of TE sequence, not most.

Conservation

Conservation asks whether a stretch of DNA has changed less across species than neutral DNA does.  It assumes common descent, so a young-earth reader may set it aside, but it is the measure the mainstream side relies on, and its numbers are low.

1. Zoonomia, 240 mammals. "We estimate that a minimum of 332 Mb (10.7%) of the human genome is under constraint". Of that, "about 11% of constrained human bases are in transposable elements, with constraint enriched in simple repeats and DNA transposons and depleted in short interspersed nuclear elements, long terminal repeats, and satellite repeats". The authors' reading: "This likely reflects the absence of function within more recently inserted transposable elements."²⁰ Eleven percent of 332 million bases is about 36.5 million bases, 2.6% of TE sequence.⁶
2. Lindblad-Toh 2011, 29 mammals. "Our data revealed >280,000 mobile element exaptations common to mammalian genomes covering ~7Mb", and "Often only a small fraction (median ~11%) of each mobile element is constrained".²¹ Seven million bases is 0.5% of TE sequence.⁶
3. Lowe 2007. Lowe, Bejerano and Haussler found 10,402 constrained noncoding elements of clear TE origin. "Overall, we find that mobile elements have contributed at least 5.5% of all constrained nonexonic elements unique to mammals".²⁴

These methods have a blind spot that cuts in the design side's favour, and it should be stated plainly.  They measure every stretch against old TE copies, which they take as the neutral baseline.  Zoonomia's own words:  "Ancestral repeats are a reasonable proxy for neutrally evolving sequence".²⁰  Rands and colleagues gave the reason:  ancestral repeats "virtually all show the patterns of indel mutation expected under neutral evolution".⁵⁴  So the method can find individual TE bases that are more conserved than the TE average, but it cannot find the TE average itself to be under constraint.  It also cannot see function in a primate-only element, which has no deep alignment.  Conservation is a lower bound on TE function, and a narrow one.

Tests

When TE copies flagged by marks are actually tested, a minority pass.

1. Todd, Deniz and Branco 2019. The team chose six TE copies in mouse stem cells "that display strong evidence of enhancer activity" and cut each out. In trophoblast stem cells both worked, with about fourfold drops in their target genes. "In ESCs, however, only one of the tested TEs had a pronounced effect on the expression of its main target gene"; a second had a small effect and two had none.¹ Then they silenced one whole family, RLTR13D6, with CRISPR interference. Of 76 marked copies, 34 lost their enhancer mark, and "Strikingly, only three genes were significantly differentially expressed upon CRISPRi across both sets of sgRNAs".¹ Their conclusion is the first quotation at the top of this article.
2. Trizzino 2017. The consensus sequences of 69 TE families were cloned in front of a reporter gene in liver cells. "Luciferase expression levels for 66 of the 69 (95.6%) tested TE families were significantly different from the negative control", but "only 17 (25.7%) of these, mostly LTRs and DNA transposons, increased gene expression".⁵⁵ The rest damped it. These were idealised family sequences, not real copies.
3. Du 2022. A massively parallel reporter test of one ERV family, LTR18A, found that "Out of 1506 motif-focused sequences tested, 1004 were classified as active in HepG2, whereas only 52 were classified as active in K562", and concluded that "LTR18A elements have generally lost regulatory activity over time through sequence changes."⁵⁶ The same authors note that "TEs are rarely functionally tested for regulatory activity".⁵⁶
4. Nóbrega 2004. Not a TE test as such, but the largest deletion test of noncoding DNA in a mammal. The team removed two gene deserts from mice, "1,511 kilobases and 845 kilobases in length". The mice "were indistinguishable from wild-type littermates with regard to morphology, reproductive fitness, growth, longevity and a variety of parameters", although the deleted segments held "1,243 non-coding sequences conserved between humans and rodents".²² The paper's abstract does not give the TE share of the deleted stretches. A lab cage is not the wild, and a small fitness cost would not show in a cage, so the test bounds large effects only.

How much TE sequence has been tested copy by copy?  The deletions cited in this article remove about 22 copies, at most 71 thousand bases, 0.005% of TE sequence.  Allowing ten times as many from studies not read here gives 0.05%.  Even counting Nóbrega's whole 2.4 million bases as TE would add 0.17%.⁶  Family-wide knockdowns, like the L1 RNA and LTR5Hs experiments, test a family's transcripts all at once and say nothing about any one copy.

The tally

The layers overlap, so the shares do not add.  The table is the article's answer to its question in numbers.  The cases of real function are a small minority, the marks are common but no more common than chance, and most TE DNA has never been tested.

Part 5: The two positions, in their own words

The selfish-DNA case

The modern junk view of TEs began with two papers printed back to back in Nature in April 1980.

1. Doolittle and Sapienza: "Natural selection operating within genomes will inevitably result in the appearance of DNAs with no phenotypic expression whose only ‘function’ is survival within genomes. Prokaryotic transposable elements and eukaryotic middle-repetitive sequences can be seen as such DNAs, and thus no phenotypic or evolutionary function need be assigned to them."⁵⁷ The claim is about explanation: a TE's existence needs no benefit to the host to explain it. It is not a claim that no TE copy does anything.
2. Orgel and Crick: "It seems plausible that most of the latter originated by the spreading of sequences which had little or no effect on the phenotype."²³ They also predicted co-option in so many words: "It would be surprising if the host organism did not occasionally find some use for particular selfish DNA sequences, especially if there were many different sequences widely distributed over the chromosomes."²³

So a domesticated TE was predicted by the founding papers.  It refutes nothing they said.  What would refute them is evidence that most TE DNA serves the host.

The same view, stated by later writers:

3. Cédric Feschotte, whose laboratory found the MER41 switch, wrote in 2008 that "TEs are best viewed as genomic parasites, which essentially owe their survival to their ability to replicate faster than the host that carries them," and that "one needs not bestow TEs with adaptive value to account for their evolutionary persistence." He went on: "In spite of … this selfish and parasitic nature, the movement and accumulation of TEs have exerted a strong influence on the evolutionary trajectory of their hosts"⁵⁸ The leading co-option researcher holds the parasite view.
4. Dan Graur and colleagues, attacking ENCODE in 2013: "Whether transcribed or not, the majority of transposons in the human genome are merely parasites, parasites of parasites, and dead parasites, whose main “function” would appear to be causing frameshifts in reading frames, disabling RNA-specifying sequences, and simply littering the genome."⁵⁹ The first half of that sentence agrees with Part 4's tally. The second half is rhetoric: causing damage is not the "main" thing most TE copies do; most of them, by the measurements, do nothing detectable.
5. Alexander Palazzo and T. Ryan Gregory, 2014: "the vast majority of these elements are inactive in humans, due to a very large fraction being highly degraded by mutation," and "Many examples have been found in which TEs have taken on regulatory or other functional roles in the genome".⁶⁰
6. Laurence Moran, 2014: "the fate of most transposons is death, turning them into pseudogenes." And: "We now know that at least half of our genome consists of defective transposons (pseudogenes) and fragments of transposons. That’s junk by any definition unless it can be shown to have a secondary function."⁶¹

None of these writers claims that every TE copy is junk.  All of them claim that most are, and all of them grant the cases in Part 3.

The design case

1. James Shapiro and Richard von Sternberg, 2005: "There are clear theoretical reasons and many well-documented examples which show that repetitive DNA is essential for genome function." The review discusses "dozens of cases where the functional importance of repetitive elements has been studied in molecular detail."⁶² Its claim is that repeats as a class do essential work; it does not claim to have counted what fraction of copies do.
2. Sternberg, 2010, on the Evolution News site: "Clearly, most DNA really does have function."⁶³ The post cites "the technical literature" for this without naming a paper.
3. Jonathan Wells, 2011, gives a long list of TE functions and says of Francis Collins's use of shared TE copies: "Given the rate at which functions are being discovered, Collins's assumption seems foolhardy, and his argument could eventually collapse in the face of new scientific discoveries." He also concedes, as quoted at the top: "Of course, there is much repetitive DNA for which functions have not yet been discovered".²
4. Jeffrey Tomkins, Institute for Creation Research, 2012: "A recent study in the journal Nature has shown that transposable elements are now absolutely required for the very beginnings of animal life in the earliest stages of embryo development."⁶⁴
5. Robert Carter, Creation Ministries International, 2009, makes the most careful young-earth statement found: "I'd like to point out that young-earth creationists do not require the entire genome to be highly functional."⁶⁵
6. Peter Borger, 2013: "The parasitic view of TEs is wrong", and in the same paragraph: "The major part of ERVs may be degenerate and defunct due to the inherent redundant character of literally hundreds of thousands of copies."⁶⁶
7. Yingguang Liu and Calvin Soper, 2009: "There is a confirmed case of essential ERV function, that is, syncytins"⁶⁷
8. Ann Gauger, Ola Hössjer and Colin Reeves, in the 2017 volume Theistic Evolution, list TE functions as part of the case that humans differ from chimpanzees, under the heading that these kinds of "junk" DNA "have in many cases been shown to have function."⁶⁸

No design writer found for this article gives a number for the fraction of TE DNA that is functional.  The words are "most," "major part," "many cases" and "essential."  Carter and Borger are the most cautious, and their concessions fit the data in Part 4 better than the stronger claims do.

Where the design citations fail

Four checks against the cited sources turned up claims that the sources do not support.

1. Wells on Lowe 2007. Wells writes: "In 2007, biologists in California found that “the majority of conserved and, by extension, functional sequence in the human genome” seems to be outside of protein-coding exons and to consist of “mobile elements” of “clear repetitive origins.”"² The quoted phrase is real, but it is about noncoding DNA in general: "Thus, the majority of conserved and, by extension, functional sequence in the human genome seems to be nonexonic." On mobile elements Lowe's figure is "at least 5.5% of all constrained nonexonic elements unique to mammals".²⁴ Joining the two phrases turns about one twentieth into a majority.
2. The 2017 volume on brain-cell L1. "LINEs move to new locations in the genome of particular brain cells as they develop, landing near different genes involved in neuronal development. That means many of our brain cells literally rewrite their genetic instructions during the process of becoming mature neurons. The same is true in other tissues of the body as well."⁶⁸ The cited study, Upton 2015, reported "An estimated 13.7 somatic L1 insertions … per hippocampal neuron".⁶⁹ A year later Evrony, Lee, Park and Walsh found "thousands of artifacts being interpreted as somatic mutation events", and their "reanalysis supports a mutation frequency of approximately 0.2 events per cell, which is about fifty-fold lower than reported, confirming that L1 elements mobilize in some human neurons but indicating that L1 mosaicism is not ubiquitous."²⁵ At 0.2 per cell, about four neurons in five carry no new insertion. The reanalysis appeared before the volume did. The claim about "other tissues" has no citation.
3. The 2017 volume on L1 and gene expression. The chapter says LINEs "inhibit gene expression" and play "an important role in the organization of the nucleus."⁶⁸ The cited paper by Han, Szak and Boeke measured that "inserting L1 sequences on a transcript significantly decreases RNA expression", and proposed, as a model, that L1 acts as "a 'molecular rheostat' of target genes."⁷⁰ The measurement is a drag on genes that carry L1 sequence. The regulatory reading is the authors' hypothesis, offered with care, and the volume reports it as a finding.
4. Tomkins on embryos. The Nature study behind Tomkins's "absolutely required" is Macfarlan and colleagues 2012. Its abstract is about a rare state that cultured mouse stem cells pass in and out of, and it says the retroviral sequence involved is "suggesting this foreign sequence has helped to drive cell-fate regulation".⁷¹ The later L1 knockdown in real embryos, Percharde 2018, comes much closer to Tomkins's claim, for L1 RNA as a class.¹⁶

Two smaller points.  Carter wrote that Faulkner's team found "between 6 and 30% of RNAs" start within retrotransposons;⁶⁵ Faulkner's abstract says "repetitive elements", a wider class.⁵³  And Wells's statement that the Shapiro and von Sternberg review "described more than 80 known functions" is not a phrase the review uses;  the review says "dozens of cases", and many of its examples are not TEs.⁶²²

The mainstream side has its own lapses on this ground.  Graur's "main function" line overstates the harm.  Francis Ayala's prediction, quoted by Sternberg, that a function for Alu "seems most unlikely" was too confident:  individual Alu copies have since been shown to become exons and to carry regulatory marks, though no function for Alu as a class has been shown.⁶³  The cases in Part 3 were found by people who took the junk view as a starting hypothesis and tested it.

What a young-earth reading must decide

Carter's concession is the right starting point:  a young-earth model does not need every TE copy to work.  On that reading most TE DNA could be created sequence that has since decayed, or copies made since creation by elements that escaped control, or both.  The data in Part 4 fit that reading as well as they fit the mainstream one, as long as the model does not predict that most TE DNA works.

The harder question is where the copies came from.  Humans carry about 7,000 Alu copies, and about 2,000 L1 copies, that chimpanzees lack, in the parts of the two genomes that align.²⁶  At Feusier's measured rate, a new Alu arises in about one birth in 40.  Over 6,000 years, about 200 generations, each person would inherit about 5 Alu copies that did not exist at creation, and about 2 or 3 would have spread to every human.⁶  So 6,000 years at today's rate falls short of the human-specific Alus by a factor of about 1,400, and of the L1s by about 600.

That leaves two options, and both have costs.

1. The copies were created where they are. Then they are part of the original design, and a design reading predicts that they work. The test is open: human-specific Alus and L1s could be deleted one at a time in human cells, as Todd and Fuentes did for other families. So far the class-wide evidence (Part 4, and Zoonomia's depletion of constraint in young SINEs) points to young copies being the least active.
2. They arrived after creation in a burst. Then the rate in the past was about a thousand times today's, and the model owes a mechanism, and an account of why a burst that large left the human population healthy.

For scale only, and on the mainstream timeline, 6 million years at today's rate would fix about 2,500 to 3,000 Alu copies, against about 7,000 observed.  That is within a factor of about three, and the mainstream reads the gap as a higher Alu rate in the past.  Rates do vary:  the chimpanzee genome paper found Alu threefold more active in the human lineage than in the chimpanzee one.⁶²⁶  No creationist paper found for this article computes either young-earth option.

Objections and responses

1. "Absence of evidence is not evidence of absence. Most TEs simply have not been tested, and every year more functions are found." True, and this article says so: 80% or more of TE sequence has no evidence either way. But the untested bulk is not a blank. Two class-wide tests already touch it. Pehrsson's shuffle control and the cCRE count both find TE DNA marked slightly less often than random genome sequence. A model in which most TE DNA was built to regulate genes predicts the opposite. "Functions keep being found" is also true of copies, not of the class: the discoveries in Part 3 are counted in dozens, against millions of copies.
2. "Conservation assumes evolution, so low conservation means nothing to a creationist." Agreed, and the article does not rest on it. The Pseudogenes article and the published Functional DNA article make the same point. Part 4's main evidence, marks against shuffled controls and deletion tests, assumes nothing about common descent.
3. "A biochemical mark in some cell type is evidence of function. ENCODE's critics are just redefining function." Marks are evidence of activity, and activity is where function is found, which is why this article counts them. But a measure that gives the same answer for real TEs (47%) and for randomly placed sequence (49%) cannot tell function from noise for the class. It can still point to particular copies, and the subfamily patterns Pehrsson found are real.
4. "The copies that look dead are scaffolding. DNA bulk matters for nuclear size and cell division." Carter makes this argument, and the published C-Value Paradox article discusses DNA bulk. It is a real effect, but it is not sequence-specific: if TE DNA works as bulk, any DNA of the same length and composition would do. That is a function of the amount of DNA, not of the TEs as TEs, and it does not predict that the copies' sequences matter.
5. "The selfish-DNA view is refuted every time a domesticated TE is found." It is not. Orgel and Crick wrote in 1980 that "It would be surprising if the host organism did not occasionally find some use for particular selfish DNA sequences".²³ The view is refuted only if most copies turn out to serve the host.
6. "The design side was right about junk DNA and the mainstream was wrong, so the design side should get the benefit of the doubt on TEs." On specific cases the design side has been right more often than early critics allowed: syncytins, RAG, Arc and the MER41 switch are now textbook. But on the class as a whole the measurements so far side with the parasite view. The question is decided by counts, and the counts in Part 4 are the best available.
7. "Lab knockouts miss subtle effects. A healthy mouse in a cage can still lack something it needs in the wild." True. The Nóbrega deletions bound large effects only, and the tissue-culture tests in Part 4 test one cell type each. That is a reason not to call the untested bulk proven junk. It is not a reason to call it functional.
8. "The 2001 figure was 45%; you have quietly raised it to 46% or 48%." The table in Part 1 gives all three counts and their sources. They differ because of the repeat libraries and the assemblies. Nothing in the argument depends on which is used.
9. "Shared TEs between humans and chimpanzees are proof of common ancestry, and this article dodges it." That argument belongs to Shared Genetic Mistakes and Common Ancestry and, for humans and chimpanzees, Shared Retroviruses and Pseudogenes. This article asks only whether TEs work. Part 5's pricing touches origin only to show what a young-earth model must explain.

Issues and unanswered questions

1. No census of tested TEs exists. No database lists the TE copies whose function has been tested by deletion, knockdown or reporter, with results good and bad. The "under 0.1% tested" figure here is an order-of-magnitude bound from the studies read.⁶ A scored literature count would firm it up, as it would for pseudogenes.
2. Copy number of human TE-derived genes. Lander listed 47 in 2001.²⁷ No current count was found.
3. The cCRE overlap is generous and coarse. The count here is by bases, with cCRE windows wider than the sites inside them. An overlap by binding-site summits, and a shuffled control run on the same cCRE set, would give a sharper figure. The shuffled control was not run for this article; Pehrsson's was, on a different mark set.
4. Disease totals are old. The Alu recombination counts are from 1999. Payer and Burns's 2019 review was not read in full.
5. Human-specific Alu deletions. No study found has deleted human-specific Alu copies one by one and measured the result. It is the cleanest test of the "created in place" option in Part 5.
6. Young-earth origin of TE copies. No creationist model found for this article gives a rate or a mechanism for the human-specific Alus and L1s. Borger's "VIGE" idea proposes that TEs were designed to generate variation, but gives no rate that can be compared with the 7,000 human-specific Alus.
7. Brain mosaicism. Evrony's reanalysis cut the somatic L1 rate fiftyfold, and later work by Upton's group or others on the rate was not checked for this article. Whether neurons use L1 insertions for anything is untested.
8. Some locators are section-level. Several PMC and HTML sources have no page numbers; their footnotes give sections. The Lavialle review's page numbers were not checked against the printed folios.

Verdict

Transposable elements are not all junk.  Some of the best experiments in genome biology show TE copies that became placenta genes, the antibody-cutting enzyme, a memory gene, immune switches and an RNA the early embryo needs.  Design writers who pointed to these cases were right, and early critics who treated "TE-derived" as "useless" were wrong about them.

But the cases are dozens, and the copies are millions.  Across the class, about 2.6% of TE sequence is conserved, about 16% carries a regulatory mark in some cell type, and the share that is marked is no higher than for randomly placed DNA.  When marked copies are tested, about half or fewer do anything measurable.  Under 0.1% of TE sequence has been tested copy by copy.  On that evidence, "most TE DNA has no sequence-specific function" is the better-supported reading, though it is a reading of class-wide data, not a measurement of each copy.  "Most TE DNA is functional" is not supported by any count found for this article, and the design writers who say so give no number.

For a young-earth reading, nothing here forces the conclusion that TEs are evidence against design;  a model in which created DNA has decayed fits the class-wide data.  What the model cannot do is claim TEs as evidence for design in bulk, and it owes an account of the thousands of human-specific copies:  created in place, in which case design predicts a function the tests have not yet found, or arrived in a burst a thousand times faster than today.

Sources