Overlapping Genes and Multi-Layer Coding

"The skepticism surrounding eukaryotic dual coding is unwarranted: rather than being artifacts, overlapping reading frames are often hallmarks of fascinating biology." - Wen-Yu Chung, Anton Nekrutenko and colleagues, PLoS Computational Biology, 2007¹

Summary

A stretch of DNA can carry more than one message.  Read from a different starting letter, the same letters spell a second protein in a second "reading frame."  Read on the opposite strand, they can spell a third.  Inside a protein-coding gene the same letters can also carry signals for splicing, for proteins that bind DNA, and for the folding of the RNA copy.  Design writers point to this as information compression that only a mind could achieve.  John Sanford writes that "probably all DNA sequences in the genome encrypt multiple codes (up to 12)," and that "this type of data compression can only result from the highest level of information design."²  Evolutionary biologists answer that a second frame is usually born inside an older gene by mutation and selection, a process called overprinting, and that "proteins may arise continuously de novo."³  This article counts how much DNA carries two protein messages, prices what that costs, and asks which account fits the pattern.

The short answer:  two-protein coding is real, common in the smallest genomes and rare among the annotated genes of ours.  Where it occurs it couples two proteins, so most changes that alter one also alter the other.  Design writers are right about that coupling.  Their larger claims, that most DNA carries many codes and that beneficial changes in it must be vanishingly rare, rest on assumptions, not on measurements found here.  Comparisons between related species support overprinting as a route to a second protein frame, with origins inferred rather than observed.

1. The genetic code ties two overlapping frames together.  Under the standard code, about a quarter of single-letter changes in a gene leave its protein sequence unchanged:  24.6% at the base composition of human genes.  Where two frames overlap on the same strand, only 1.2% of changes leave both protein sequences unchanged.  About 46% change both, and 8% put a stop codon into one of them.⁴  About 23% change one protein and leave the other's sequence alone, and another 23% the reverse.  This is arithmetic on the code, not a fitness measurement.  A change that leaves both proteins alone can still matter, a change to both is not always harmful, and how many of the one-protein changes would help is not known.

2. Small viruses are full of overlaps.  Large genomes are not.  Counted for this article from the reference sequences, as a share of each genome's whole length, hepatitis B virus reads 50.0% of its 3,182 letters in two frames, the phage phiX174 15.1% and HIV-1 14.7%.⁵  Of 5,976 reference virus genomes, 53% carry an overlap of 50 letters or more.⁶  The bacterium E. coli reads 0.2% of its genome in two frames, most of it in overlaps of one to four letters where one gene's stop codon shares letters with the next gene's start.  Human mitochondrial DNA reads 0.3% in two frames.⁵

3. Among annotated human genes, two proteins from one stretch are uncommon.  Taking one standard ("canonical") transcript per gene from GENCODE release 50, 21,568 of 33.9 million protein-coding letters, 0.064%, are read in two frames.  This count includes the 13 mitochondrial protein-coding genes, which add two overlapping pairs and a few dozen letters.  Only 28 pairs of genes overlap on the same strand in different frames, and 16 pairs overlap protein-coding sequence on opposite strands.⁷  That is one annotation choice, not a census of every working second frame.  Counting every annotated transcript raises the share to 7.2%, almost all of it from alternative transcripts that read a stretch of one gene in a second frame.  Those transcript models vary in how well they are supported, and the annotation does not say how many make a stable protein.  A 2007 scan found 40 human genes with a second frame of 500 letters or more conserved in mouse, 15 of them high-confidence.¹  A handful of human genes have the second protein shown in cells or tissue:  CDKN2A (p16 and ARF), GNAS (XLalphas and ALEX), POLG (POLGARF), PRNP (AltPrP) and ATXN1 (Alt-ATXN1).

4. RNA from both strands is common.  Proteins from both strands are not.  Counted for this article, 55.3% of human protein-coding genes have an exon of another gene, coding or not, overlapping one of their exons on the opposite strand.  Protein-coding sequence overlaps protein-coding sequence on opposite strands at only 16 gene pairs.⁷  Ribosome profiling, which catches ribosomes sitting on RNA, has catalogued 7,264 translated reading frames that are not annotated proteins;  1,469 of them overlap a known gene's coding sequence in another frame.⁸  When 95,520 proteomics experiments were searched for their products, 24.6% of the 7,264 left a detectable peptide, almost all of them fragments displayed to the immune system, and 15 met the strict standard for a new protein.⁸  Translation, a detected peptide and a working protein are three different findings.

5. The constraint shows up in real genes.  Across 117 overlapping RNA-virus genes, the overlapping parts had shorter evolutionary trees and fewer silent substitutions than the rest of the same genes, while amino-acid-changing substitutions did not differ significantly.⁹⁹  That is the code arithmetic seen in real sequences.  The same paper's ratio of the two rates cannot be relied on:  its Table 1 puts the ratio lower in the overlaps, while its text and supplementary table put it higher, and the usual ratio methods are known to misread overlapping genes.¹⁰¹¹  In hepatitis B virus, drug-resistance mutations in the polymerase gene change the surface protein read from the same letters.¹²  A phiX174 genome redesigned with its overlaps pulled apart grows only when a host plasmid supplies one of its genes, and in one laboratory test it released 16 phages per infected cell against 145 for the natural virus.¹³¹⁴  That gap mixes the effect of separating the genes with the redesign's other changes.

6. Where they can be dated, second frames look like later additions.  In viruses, the newer frame of a pair, identified by comparing relatives, is usually found in one lineage only, codes for floppier protein, and uses codons unlike the rest of the genome.¹⁵¹⁵¹⁶  HIV-1's antisense protein frame is open in most of the pandemic group M and not in the other groups or related ape viruses.¹⁷  Of the 7,264 human translated frames, 89.6% have no detectable protein counterpart in mammals outside the primates, a comparative finding that short proteins make hard to test.¹⁸  POLGARF traces to a mobile element that landed in the gene before placental mammals spread.¹⁹  All these ages are inferred by comparing living species.  None is a birth watched as it happened, and these features are consistent with overprinting without proving it.

7. The design-side claims split.  The claim that overlap constrains change is right, and the data in point 5 support it.  The claim that most DNA carries many overlapping codes is not measured in any source reviewed here.  Sanford's "over half the human genome is read backwards as well as forwards" carries no citation,² and his "up to 12" codes come from a 1989 paper whose body could not be checked here.  The 2013 calculation by George Montañez, Robert Marks, Jorge Fernandez and Sanford is correct arithmetic on assumed inputs:  in its headline case, 12 codes per letter, 99.9% of letters already optimal, preferred letters chosen independently in each code, and every code site defined as one where change matters.  Its own authors write that "the exact degree to which nucleotides participate in two or more codes is still unknown."²⁰  Its result is a conditional probability, not a measured share of beneficial mutations.  No design writer found here engages the overprinting evidence.

Verdict.  Overlapping protein codes are a real feature of compact genomes and an uncommon one among the annotated genes of large genomes.  Where they exist they couple two messages and hold letters in place, which is the design writers' point, and it is correct.  How often a change in an overlap helps has not been measured in any source reviewed here, so the coupling does not show that helpful changes are vanishingly rare.  Two protein frames on one stretch are not a general property of DNA.  Where comparisons between species allow them to be dated, second frames look lineage-limited and built inside older genes, which is what overprinting predicts.  That pattern supports overprinting as an explanation for some second protein frames.  It does not measure the extent of other regulatory layers throughout the genome, settle where the first genes came from, or establish the age of the earth.

Part 1: Reading one stretch of DNA two ways

What "overlapping" means here

A protein-coding gene is read three letters at a time.  Each three-letter word, a codon, names one amino acid or says "stop."  Start one letter later and the same letters fall into different words.  That is a second reading frame, often called +1, and starting two letters later gives +2.  The opposite strand of the double helix offers three more frames, read in the other direction.  Two genes "overlap" when the same letters are read in two of these frames and both readings make a product.

Three other uses of the word need to be kept apart, because the design-side literature mixes them.

1. Two genes whose RNA copies overlap, usually on opposite strands, with no claim that both make protein. A protein gene and a non-coding RNA gene facing each other is the common case.
2. One gene read in two frames in different messages. Alternative splicing can drop or add a few letters, so that downstream letters are read in another frame in one version of the message. Both versions come from the same gene.
3. Other signals laid over coding letters: splice sites, places where regulatory proteins bind, RNA folding, nucleosome positioning. These are often called "codes." Most are preferences rather than codes in the strict sense: a binding site tolerates many sequences, and a nucleosome prefers some letters to others without requiring one.

Only overlap in the strict sense, two protein frames on the same letters, is counted letter by letter in this article.  The others are counted where a source counts them, and named for what they are.

What counts as evidence that a second frame is used

The evidence for a second frame comes in steps, and each step says less than the one after it.

1. An open frame. A stretch with no stop codon in the second frame. Long open second frames are common in ordinary human genes, as Part 5 shows, so an open frame alone proves little.
2. RNA. The stretch is copied into RNA. This shows the letters are copied, not that the second frame is read.
3. Translation. Ribosome profiling catches ribosomes sitting on RNA, and a three-letter rhythm in where they sit shows which frame they read. This shows that a ribosome reads the frame. It does not show that the product survives.
4. A detected peptide. Mass spectrometry finds a fragment of the second protein in cell extracts. Fragments found only among the peptides cells display to the immune system show the protein was made and chopped up, not that it lasts.
5. A stable protein. An antibody against the second protein finds it in normal cells, at the expected size.
6. A function. Removing the second protein, while leaving the first one's amino acids unchanged, harms the cell or the organism, and putting it back repairs the harm.
7. Conservation. The second frame stays open across related species. This counts as evidence of function only if common descent is assumed, and it is sometimes confounded by the first frame's own constraint.

Large catalogues sit on steps 1 to 4.  The handful of proven human cases reach steps 5 and 6.  A number quoted without its step is not a number about function.

What the genetic code charges for a second frame

Every single-letter change in a gene does one of three things to its protein:  nothing, when the new codon names the same amino acid;  swaps one amino acid;  or puts in a stop.  The standard code is built so that about a quarter of changes do nothing.  Most of these silent changes fall on the third letter of a codon.

In a same-strand overlap, the third letter of a codon in one frame is the first or second letter of a codon in the other.  A change that is silent in one frame is therefore usually an amino-acid change in the other.  The table counts every single-letter change in every codon context, weighted by the base composition of human coding DNA, for each way two frames can lie over each other.

The same-strand row is the one that matters most, because it is the commonest kind of long overlap.  There, only 1.2% of changes leave both protein sequences alone, against 24.6% for a gene read one way.  About 23% are silent in one frame and change the other.  Wen-Yu Chung and colleagues got the same shape in 2007 by the same method, averaged over codon positions:  about 1.2% silent in both and about 46% changing both.¹  One opposite-strand arrangement is worse than the same-strand case:  with codons shifted by one letter, no change at all is silent in both frames.  Another is much milder:  when the third letters of both codons fall on the same base pair, 17% of changes are silent in both.⁴

Three limits keep this arithmetic in its place.

1. It counts changes to protein sequence, not changes to fitness. A change that is silent in both frames can still alter splicing, RNA folding or how fast the message is read. The article Synonymous Mutations and Codon Usage measures how often that matters. A change that swaps an amino acid in both proteins is not automatically harmful either; most proteins tolerate many single swaps. How often a real change helps, harms or does nothing is measured in Beneficial and Harmful Mutations, and this article does not repeat it.
2. It treats letters as independent. Real genes have codon preferences and neighbour effects. The shares shift a little with base composition (24.4% silent at even composition, 24.6% at human composition), and not by enough to change the picture.
3. It prices one overlap of two frames. It says nothing about how many other layers sit on the same letters. That is a separate, empirical question, and Parts 2, 3 and 6 take it up.

What the table does show is that a second protein frame removes almost all of the changes that leave both protein sequences alone, and ties most changes in one protein to a change in the other.  A change that alters protein A while leaving protein B's sequence alone is still available, about 23% of changes in each direction, but how many of those would help is a question the code cannot answer.  Any account of overlapping genes has to live with that coupling.

Part 2: How common overlapping genes are

Our counts, from the reference sequences

A simple measure is the number of letters read in two different protein frames.  It needs the annotation of every coding sequence and its frame, and it must count each letter once, however many products use it.  The table gives that count for six reference genomes, counted for this article from the NCBI RefSeq records, and for human protein-coding genes from GENCODE release 50.⁵⁷  The denominators differ.  For the viruses, E. coli and the mitochondrion the share is of the whole genome, most of which codes for protein.  For the human row it is of protein-coding letters only, because only about 1% of the human genome, 33.9 million of 3.1 billion letters, codes for protein;  as a share of the whole 3.1-billion-letter genome the same count is 0.0007%.

These counts depend on what the annotators have accepted, and the error runs both ways.  The SARS-CoV-2 record shows one direction:  its 0.1% leaves out frames that ribosome profiling and patient antibodies later showed to be used, such as ORF3d.²¹  The other direction is that some annotated coding sequences, especially recent and computational models, may not make a working protein.  So a count is a lower bound on real two-frame coding only if every annotated sequence in it is real, and it is not a census of open frames.  Comparisons between rows show the broad pattern, small viruses against large genomes, but because the denominators differ the percentages are not measurements of one identical quantity.  The phiX174 figure shows how definitions change a number.  Bradley Wright and colleagues give the overlapped share of phiX174 as 16.8% in their text and "∼26% of nucleotides involved in encoding more than one gene" in their abstract.¹⁴  Counting letters shared by two annotated genes in any frame, including A* read in the same frame as A, gives 25.8% from the same record;  counting only different frames gives our 15.1%.⁵

Viruses

Three surveys measured how many virus species have overlaps, each on a different sample and threshold.  Nicola Chirico and colleagues, 2010:  "75 per cent of the approximately 2000 known virus species have at least some gene overlap."²²  Robert Belshaw and colleagues, 2007, on 701 RNA virus genomes:  "56% of the viruses having some gene overlap."²³  Timothy Schlub and Edward Holmes, 2020, on 5,976 reference genomes:  "over half the instances of gene overlap are very small, covering <10 nt, and 84 per cent are <50 nt in length.  Despite this, 53 per cent of all viruses still contained a gene overlap of 50 nt or larger."⁶  The share with a 50-letter overlap ran from 19% of double-stranded RNA viruses to 43% of positive-strand RNA viruses and higher in DNA viruses.⁶

How much of a virus genome is overlap depends on its size.  The share of the genome in overlap falls as genomes get longer, in Belshaw's sample, in Chirico's and in Nadav Brandes and Michal Linial's.²³²²²⁴  Brandes and Linial put a ceiling on the absolute amount:  "although viruses span three orders of magnitude in genome length, they almost never have over 1500 overlapping nucleotides, or over four significantly overlapping genes per virus."²⁴  Measured as a share of genome length, Chirico found 4.6% for RNA viruses and 6.6% for DNA viruses of RNA-virus size, and corrected Belshaw's lower figure:  "A lower value for overlap proportion in RNA viruses published previously by one of us ( Belshaw et al . 2007 ) was an error."²²

Why small viruses overlap so much is disputed.  One idea is that a capsid of fixed size limits genome length, so new genes have to fit inside old ones;  Chirico found the size relation only in families with rigid icosahedral capsids, and Brandes and Linial found no such difference.²²²⁴  A second idea is that high mutation rates favour compact genomes;  Chirico found "no obvious relationship between overlap proportion and mutation rate (linear regression, p = 0.50, n = 13)," on a small sample.²²  A third, which Belshaw's title already names, is that overlaps are how small viruses gain new genes at all.  Part 5 takes that up.

Bacteria

In bacteria, overlapping genes are common and almost all tiny.  Zackary Johnson and Sallie Chisholm counted 64,989 overlapping gene pairs in the microbial genomes sequenced by 2003, about one for every six genes, and found that ">85% are <30 bp."²⁵  Most are a stop codon of one gene sharing letters with the start of the next.  Long embedded overlaps are rare.  Before their own find, Lea Fellner and colleagues knew of "only three overlapping gene pairs" in the "extremely well-researched" E. coli.²⁶  Ribosome profiling has since found many candidate starts inside genes:  Sezen Meydan and colleagues found 124 internal start sites shared by two E. coli strains, 74 of them out of frame with the main gene, and tested one of the out-of-frame products in living cells.²⁷  A 2020 review by Zachary Ardern and colleagues says "hundreds of short open reading frames antisense to annotated genes show evidence of both transcription and translation" across bacteria, while its own counts of tested cases are in single and double digits.²⁸

The pattern

Overlap is densest where genomes are smallest.  Hepatitis B virus reads half its letters twice.  E. coli and the human mitochondrion read well under 1% of their genomes twice.  Human coding DNA, on one canonical transcript per gene, reads 0.064% of its coding letters twice, and the whole human genome far less.  Any account of overlapping codes has to explain why annotated two-frame coding is concentrated in the smallest genomes and sparse in the largest.

Part 3: Overlapping codes in the human genome

The counts, by what they count

Numbers about human overlap range from a few hundredths of a percent to "over half," because they count different things.  The table puts them side by side with their denominators.

Three things stand out.

1. Overlapping RNA is common, overlapping protein is not. Half of protein-coding genes have another gene's exon facing them on the opposite strand. Only 16 gene pairs overlap coding sequence on opposite strands, and only 28 on the same strand in different frames. About half of those 28 involve genes added to GENCODE in the last few years, many from ribosome profiling, and two are in the mitochondrion.⁷
2. Most of the 7.2% comes from alternative transcripts, and the count depends on which transcript models are admitted. Counting every transcript, 12,569 genes have a stretch read in two frames by their own transcripts. Limited to full-length protein-coding transcripts, the figure is 8,226 genes; limited to transcripts whose coding sequence matches the curated CCDS set, 1,346 genes, of which 261 have a two-frame stretch of 150 letters or more.⁷ CCDS is a set of coding sequences that several annotation groups agree on. A transcript outside it is not thereby shown to be unused, and one inside it is not thereby shown to make a stable second protein. The annotation tiers sort models by how they were built and agreed, not by tested function. Han Liang and Laura Landweber's 2006 count on a smaller curated set found two-frame coding in 6.9% of alternatively spliced genes, and gave its share of exons as under 0.5%: "The space of dual or multiple coding regions in the human genome is <0.5% of exons."²⁹
3. The strict counts and the generous counts differ by a factor of about a hundred, and every step up comes from admitting more transcript models. None of them approaches "most DNA." These are counts of protein frames. They do not measure RNA-binding sites, splicing signals or the other proposed layers of Part 6, which are counted separately where a source counts them.

Translated, detected, working: the ribosome profiling catalogue

Ribosome profiling has made the human count of translated frames much larger.  In 2022 Jonathan Mudge and an international group of annotators pooled several studies into a standard set of 7,264 translated reading frames that are not annotated proteins.³¹  The 2026 Nature paper by Eric Deutsch and the same consortium gives the categories:  3,083 upstream frames that end before the main gene starts, 688 that start upstream and run into the main coding sequence in another frame, 720 inside the coding sequence in another frame, 61 that overlap its end, 504 downstream, 1,917 in RNAs annotated as non-coding, and 291 in other transcripts.⁸  So 1,469 of the 7,264 sit at least partly over a known protein in another frame.

Translation is step 3 on the ladder in Part 1.  Deutsch and colleagues then searched 95,520 mass-spectrometry experiments for the products.  The steps fall off fast.

1. In conventional protein digests: "484 peptides passing FDR thresholds that map to 183 out of the 7,264 ncORFs (around 2.5%)."⁸
2. In peptides displayed to the immune system: "3,116 peptides mapping to 1,785 out of 7,264 Ribo-seq ncORFs (24.6%)," almost all on the class I molecules that display fragments of proteins made inside the cell.⁸
3. Meeting the Human Proteome Project's standard for a new protein, after manual review: "we identified 20 candidates for tier 1A status, but further scrutiny reduced this list to 15 ncORFs."⁸

The authors give a reason the strict count is low that has nothing to do with function:  small proteins yield few peptides.  Only "2 out of 36 known proteins under 50 amino acids (5.6%)" pass the same benchmark.⁸  So 15 is a floor for stable small proteins, not a ceiling.  The immune-display count shows the frames are translated and the products made and cut up.  It does not show that the products last or do anything.

On age, the same set looks young.  Clara Sandmann, Jorge Ruiz-Orera and colleagues found that "most of the putative 7,264 sORF-encoded microproteins (n = 6,506; 89.6%) lack significant protein homology across non-primate mammals and could therefore be classified as evolutionarily young."¹⁸  Deutsch's group measured something different, whether the start, the stop and the openness of each frame are kept across species, and found that 2,211 frames (30.4%) scored in the top tenth of untranslated frames of the same type and length, against the 10% expected by chance.⁸  Frames that overlap a known gene score high on this measure, which the authors say is "presumably due to constraint to preserve the CDS."⁸  An internal frame can stay open because the main gene's letters are held in place, which is one reason conservation alone cannot show that a second frame is used.

Function is tested by knocking frames out.  Jin Chen, Jonathan Weissman and colleagues knocked out about 2,350 non-canonical frames by gene editing in stem cells and leukemia cells and found "over 500 ORF knockout 'hits'," including 78 frames that overlap a main gene's start in another frame.³²  The authors note that some upstream-frame effects may act by raising the main protein's level rather than through the small protein itself.³²  John Prensner and colleagues tested 553 candidates, mostly upstream, downstream and non-coding-RNA frames rather than internal ones:  "57 induced viability defects when knocked out in human cancer cell lines."³³  These are the best functional counts available, and few of them are internal second frames.

The cases with proven second proteins

Five human genes have a second protein, read from letters shared with the first in another frame, detected in cells or tissue and studied in its own right.

In 1995 Quelle, Sherr and colleagues called the INK4a arrangement "practically without precedent in mammalian genomes."³⁴  Martin Klemke, Wieland Huttner and colleagues said the same of ALEX in 2001:  "Our results reveal a mechanism of gene usage that is without precedent in mammalian genomes."³⁶  As late as 2005 Anton Nekrutenko and colleagues could list only three other mammalian cases with both products characterized:  INK4a, X-box binding protein 1 and a region shared by 4E-BP3 and MASK.³⁷  The list has grown since, mostly through ribosome profiling, but slowly.  The proven cases show that human genes can carry two protein messages.  They do not show that many do.

Part 4: What overlap costs, measured

Part 1 priced a second frame on paper.  Four kinds of measurement show whether real overlaps pay that price.

Fewer silent changes in overlapping DNA

If a second frame takes away a gene's silent changes, overlapping DNA should collect fewer of them over time than the rest of the gene.  Etienne Simon-Loriere, Edward Holmes and Israel Pagán tested this on 117 overlapping genes from 55 RNA virus species, comparing the overlapping part of each gene with its non-overlapping part.

Three of their results bear directly on constraint, and they agree with each other.

1. Less change overall. The total length of the evolutionary tree, a measure of how much a stretch has changed among related virus strains, "was significantly smaller in OV than in NOV regions" across all genes (P = 0.037), and significantly smaller in 65 of the 117 genes taken one by one.⁹ The effect was clearest where one gene lies wholly inside another (25 of 31 genes), not where genes merely overlap at their ends.⁹
2. Fewer silent substitutions. The rate of silent substitutions was higher outside the overlaps, for all genes together (P = 0.040) and in 84 of 117 genes.⁹
3. No significant difference in amino-acid-changing substitutions. That rate did not differ significantly between the two parts (F = 0.10, P = 0.747).⁹ A non-significant result is not proof of no difference, and the authors report that in terminal overlaps the rate tended to be lower outside the overlap, where conserved protein domains cluster.

That pattern is what the code arithmetic predicts:  a change that would be silent in one frame changes the other protein, so fewer of them survive.  The authors' abstract draws the same conclusion:  "gene overlapping generally results in a decrease in the rate of RNA virus evolution through a reduction in the frequency of synonymous mutations."⁹

The paper's ratio of the two rates, often used as a single measure of selection, does not give a usable answer, because the paper reports it two opposite ways.¹⁰

1. Its Table 1 and one sentence of the text give mean ratios of 0.23 in the overlapping parts and 0.60 in the rest.⁹ Counted row by row for this article, Table 1 has the overlapping part lower in 95 of 117 genes.
2. Two sentences later the text says the ratio "was significantly higher in overlapping than in nonoverlapping fragments" in 81 of 117 genes.⁹ That cannot describe Table 1, where only 22 genes have the overlapping part higher at all.
3. The supplementary table gives each gene's two rates separately. Dividing them, the overlapping part has the higher ratio in 92 of 117 genes, and the difference of the two rates is higher in the overlap in 98. Table 1's direction matches the supplement's in only 35 of 117 genes.

The individual rates point the supplement's way.  With amino-acid changes about the same and silent changes fewer, the ratio in the overlap should rise, not fall.  A mean of ratios across genes need not follow the means of the two rates, so this does not prove Table 1 wrong, but nothing in the paper explains its direction, and the authors' own explanation, fewer silent changes, would push the ratio the other way.  This article therefore does not use the paper's ratios, in either direction.  The authors read their result as "stronger purifying selection in OV than in NOV fragments" linked to constraint on silent change.⁹  The tree lengths and silent rates above support constraint on silent change.  The ratio comparison does not settle anything further.

There is also a reason to distrust the ratio in overlaps at all.  The usual methods for spotting selection assume each gene is free to change on its own.  Niv Sabath, Giddy Landan and Dan Graur showed that applying them to overlapping genes "leads to the false appearance of positive selection even though the gene is in reality subject to negative selection";  for one influenza gene, PB1-F2, the ratio fell from 6.75 to 0.52 when both frames were modelled together.¹¹  A higher ratio in overlapping parts, as the supplement shows, is what that bias would produce.  Silent-substitution counts and tree lengths are more direct evidence of constraint than the ratio.

The same effect is now used to find overlaps.  Andrew Firth searched alignments of about 700 RNA virus species for windows where silent changes are scarcer than expected, and his method picked out 20 of 21 known overlapping genes and 16 of 17 known overlapping non-coding elements.⁴⁰  In hepatitis B virus, which reads half its letters twice, Mizokami and colleagues found the fewest silent changes in the surface gene, which lies wholly inside the polymerase gene, and explained it by the overlap:  "synonymous substitutions in the S gene can cause amino acid changes in its overlapping region in a different frame."⁴¹

The two frames need not share the load evenly or in a fixed way.  In the human GNAS locus, Nekrutenko and colleagues found the two frames taking turns, with amino-acid change shifting from one frame to the other across primate branches;  "none of the ratios is significantly different from 1."³⁷

One mutation, two proteins

Hepatitis B virus shows the cost in a form doctors see.  The polymerase gene lies wholly over the surface-protein gene.  Shuping Tong and Peter Revill:  "Due to the overlap between the P and envelope genes, drug resistant mutations in the P gene can cause amino acid substitutions in or near the 'a' determinant of envelope proteins," and one resistance mutation "can convert codon 172 in the S region into a stop codon, thus truncating the envelope proteins by 55 residues."¹²  Some of these surface changes weaken antibody binding and help the virus escape the immune system.

This is the kind of mutation the design writers have in view:  helpful in one frame and changing the other.  The same review shows both outcomes.  The surface changes that weaken antibody binding leave the virus better off on two counts at once, resisting the drug and dodging antibodies.  The stop codon made by the rtA181T change leaves it worse off on one:  "This mutant was deficient in both virion and HBsAg secretion."¹²  So "changes the other protein" is not the same as "harms the virus."  And the overlap does not stop resistance from arising.  For the older drug lamivudine, "LMV resistant mutations develop in 23% of patients following 12 months of therapy, rising to 80% by five years of treatment," though resistance to the newer tenofovir is uncommon.¹²

Pulling a genome's overlaps apart

The most direct test available is to rebuild a genome without its overlaps and see what happens.  In 2012 Paul Jaschke, Drew Endy and colleagues designed phiX174.1, a 6,302-letter version of the 5,386-letter phage in which every coding sequence and its start signals were separated.  To keep the genome at its natural length they then cut the gene for coat protein F short, making phiX174.1f.  They recovered living phages "containing the øX174.1f genome from E. coli C strains that independently express full-length gene F," so the rebuilt genome works only with gene F supplied from the host.¹³  Their own summary:  "A first fully decompressed phage genome is viable and not irreducibly complex."¹³

In 2020 Bradley Wright, Jaschke and colleagues measured how well the rebuilt phage grows.  On the same host strain, which carries gene F on a plasmid, the natural phage released "145 ± 66 phage produced per cell while decompressed ϕX174 produced only 16 ± 2," and made plaques of 29.2 against 4.1 square millimetres.¹⁴  The rebuilt phage was also less stable when heated and attached poorly to cells at 14 degrees.¹⁴  Lysis timing was the same.

The comparison has three confounds, and they stop the gap from being assigned to overlap removal alone.

1. The plates were not identical. "Wild-type ϕX174 plates contained 2 mM glucose and decompressed ϕX174 plates contained 2 mM rhamnose," the sugar that switches on the plasmid's gene F.¹⁴ The rebuilt phage depends on a gene supplied from a plasmid in the host, a complementation set-up the natural phage does not need, and part of its deficit may come from that or from the different sugar.
2. Separating the genes meant changing many letters to remove the second frames while keeping the first proteins, so the rebuilt genome differs from the natural one in more than its layout. Wright and colleagues point out that the packaged DNA itself differs, which may explain the heat instability.¹⁴
3. These are preprint numbers. The paper was published in ACS Synthetic Biology later in 2020, and that version was not compared here.

With those limits, the redesigned phage released about one-ninth as many phages per infected cell as the natural one.  The experiment shows that phiX174 can replicate in the laboratory with its overlaps separated, provided the host supplies gene F;  it does not show that the redesigned genome is viable on its own.  It shows that the natural, overlapping arrangement does much better than this one redesign, but not how much of the difference comes from the overlaps themselves rather than from the many letter changes, the gene F supply or the plates.

Deleting a young overlapping gene

The bacterial case runs the other way:  remove the new frame and keep the old.  Fellner and colleagues stopped the young overlapping gene nog1 with a single-letter change that leaves the protein of the gene it sits in, citC, unchanged.  Grown alone, the mutant grew like the parent except for slower growth with added magnesium chloride.  In head-to-head competition the mutant lost:  "Δnog1 shows a clear loss of fitness compared to its parental strain when grown in plain LB.  When using MgCl2-supplemented broth, the decrease in fitness of Δnog1 was even more pronounced."²⁶  In their Figure 4b the mutant falls to about a fifth of the mixed culture in plain broth and to about 4% with magnesium chloride, while a control change in citC has no effect.  The authors call the benefit small:  "the gene product provides only a small fitness advantage for its carrier."²⁶  A 2020 review from the same research group reports the result backwards, as a "growth advantage over the wildtype" for the mutant.²⁸  The original figure settles it.

What these measurements do and do not show

They show that overlapping letters collect fewer silent changes than single-frame letters, that one mutation in an overlap really does change two proteins, and that the natural overlapping phiX174 far outgrows one redesign without overlaps.  That is the constraint part of the design-side argument, and it stands.

They do not measure how often a mutation in an overlap is beneficial, and so they do not support the further claim that helpful changes in overlaps are vanishingly rare.  None of these studies counts beneficial mutations, and none supplies the "fraction already optimal" that the Montañez model in Part 7 assumes.  The share of mutations that help, harm or do nothing is measured directly in Beneficial and Harmful Mutations, and the overlap data here do not change those measurements.

Part 5: How second frames arise: the overprinting evidence

What overprinting predicts

In 1992 Paul Keese and Adrian Gibbs gave the idea its modern form.  Many virus genes, they noted, are found only in one family or subfamily, including "a range of overlapping genes that occur in viruses from about 50% of all families."³  Their proposal:  "proteins may arise continuously de novo," and one way is "overprinting," in which an existing stretch of DNA is translated in a different frame.³  If that is how second frames arise, several things should be true, and each can be checked.  None of them is unique to overprinting, so each is a test the idea could fail, not proof that it happened.

1. The newer frame should be found in fewer lineages than the older one. The older gene is shared widely; the overprint only in the group where it arose.
2. Its protein should look unlike ordinary proteins. It begins as whatever the old gene's letters happen to spell in another frame. A new product may lack a compact fold, and disorder can make a second frame easier to establish or retain. The genetic code itself also biases what amino acids that frame can spell.
3. Its codon usage should not match the rest of the genome, because its codons were chosen for the older frame.
4. Long open frames to start from should already be common inside existing genes.
5. New overprints should matter less than old ones, since they have had less time to become built into the cell.

Lineage limits

The newer frame is usually the more narrowly shared.  Corinne Rancurel, David Karlin and colleagues traced the distribution of overlapping virus proteins and "were able to confirm that 17 of these had been created de novo";  "Most proteins created de novo are orphans (i.e., restricted to one species or genus)."¹⁵  Niv Sabath, Andreas Wagner and David Karlin found 12 overprinted genes whose distribution on the family tree was clear enough to place their origin.⁴²  Specific cases repeat the pattern.

1. HIV-1 ASP. The antisense frame is open in 77% of pandemic group M sequences, with a median length of 182 codons, and not in the other HIV-1 groups or the related ape viruses; Elodie Cassan, Olivier Gascuel and colleagues conclude that it "was created recently and that its emergence in HIV-1 is concomitant with the emergence of the group M."¹⁷
2. SARS-CoV-2 ORF3d. Chase Nelson and colleagues found it in SARS-CoV-2 and "also present in Guangxi pangolin-CoVs but not other closely related pangolin-CoVs or bat-CoVs."²¹
3. Bacterial nog1. "A phylogenetic analysis suggests that the orphan gene nog1 arose by overprinting after Escherichia/Shigella separated from the other γ-proteobacteria."²⁶
4. Human dual-coding regions. Liang and Landweber checked the second frames of human dual-coding regions in other species: "one of two reading frames contains one or more stop codons in 80% of the dual coding regions in chickens, 45% in mice, and 54% in rats," and concluded that "most secondary reading frames in the dual coding regions examined evolved recently in mammals."²⁹
5. ARF. Chickens have no p16INK4a at all, and their ARF stops at the end of its own exon. Yet chicken exon 2 keeps an open frame "capable of encoding 162 amino acids that could have formed the carboxyl terminus of ARF, if splicing had occurred in the register used in mammals."³⁵ The second frame is there in the chicken letters and unused.
6. POLGARF. Gary Loughran, John Atkins and colleagues traced its start to a repeat element that inserted into the gene's leader: "POLGARF evolved ∼160 million y ago due to a mammalian-wide interspersed repeat (MIR) transposition into the 5′ leader sequence of the mammalian POLG gene, which became fixed in placental mammals."¹⁹
7. The ribosome profiling catalogue. 89.6% of the 7,264 human frames have no detectable protein counterpart in mammals outside the primates.¹⁸ This is a comparative finding with a known blind spot: short proteins give homology searches little to work with, so some frames may have relatives that were not detected. It marks the frames as lineage-limited by that test, not as dated by an independent clock.

A frame found in fewer lineages is not by itself proof that it was born late.  The same pattern could arise if relatives lost the frame, if it diverged beyond recognition, or if it was carried in from elsewhere.  Evidence supporting a direction in the cases above includes an outgroup that keeps an open but unused frame, as the chicken does for ARF;  an insertion that supplied the start, as for POLGARF;  or codon usage that marks which frame came first, as in the next sections.  These clues strengthen an inferred direction; they are not observations of a gene's birth.

Unusual proteins

The newer frames make odd proteins.  Rancurel and colleagues found that "almost two-thirds of novel, overlapping protein regions are disordered," against "fewer than one-fourth of nonoverlapping protein regions," meaning they lack a fixed folded shape.¹⁵  Erika Kovacs and colleagues found the same in human genes that code in two frames by alternative splicing.  Of 67 such genes, the older and newer frames could be told apart in 62, and the newer frame was mostly disordered "in 39 out of 62 cases."  Where the second frame is shifted one way, disorder rose from 16.7% in the older frame to 56.3% in the newer;  where it is shifted the other way, both frames were already high, 47.2% and 58.2%.⁴³  Rancurel's group was careful about direction:  overlapping genes may be disordered because new proteins start that way, or because disorder makes an overlap easier to keep, and "our results are compatible with both hypotheses."¹⁵  Sara Willis and Joanna Masel separated the two and estimated that "birth facilitation accounts for 32 ± 13 %" of the extra disorder in overlaps, the rest coming from the genetic code and from which overlaps survive.⁴⁴

Codon usage

Angelo Pavesi, Gkikas Magiorkinis and David Karlin took 25 overlapping virus genes whose older frame was known from family trees.  The older frame's codon usage was closer to the rest of the genome in 21 of the 25, with a median correlation of 0.42 against 0.19 for the newer frame.¹⁶  The newer frame reads letters that were chosen for something else, which is what overprinting predicts.  Codon usage and disorder can also be shaped by selection after a frame arises, so these features are consistent with overprinting, not exclusive proof of it.

Frame bias from the genetic code

Willis and Masel also found that genes born inside other genes favour the +1 frame:  "23 +1 events vs. 1 +2 event, P = 3 × 10 − 6."⁴⁴  They trace this to the code itself:  "+1 frameshifts yield more start codons and fewer stop codons, and hence more and longer ORFs."⁴⁴  The frame most often used is the frame the code leaves most open.

Long open frames are common inside ordinary genes

If overprinting is real, long open frames should already lie hidden in ordinary genes.  The test run for this article took 19,167 human MANE Select coding sequences and looked for ATG-started open frames of 100 codons or more inside each gene, in the five other frames.  It compared the real genes with four kinds of shuffled copies, three replicates each:  bases shuffled;  codons shuffled;  synonymous codons shuffled within the gene, which keeps the protein exactly;  and synonymous codons shuffled only within windows of ten codons, which also keeps the local base composition.⁴⁵

About one human gene in seven hides an ATG-started frame of 100 codons or more in its +1 frame, and one in four or five in the two richer opposite-strand frames.  Shuffles preserving the protein sequence and local base composition reproduce most of that frequency.  Shuffling synonymous codons within short windows, which keeps both, leaves 13.2% to 13.5% in the +1 frame against 14.5% in real genes, and matches or comes close to the real share in the other frames.  For frames of 150 codons or more, about the length Chung's group used, the +1 share is 4.6% in real genes and 3.9% to 4.1% after the window shuffle.  By this test the +1 frame shows no avoidance of stop codons beyond what the protein and the local composition produce:  it carries 4.21 stop codons per 100 codons in real genes and 4.25 to 4.38 in the shuffles that keep the protein or the codons.⁴⁵  The test covers open frames and stop codons under these four nulls.  It does not test other sequence features, and it does not show that any hidden frame is expressed or could become a working gene.

Two readings follow.  First, the raw material for overprinting is common:  thousands of human genes carry long open second frames, at rates the protein-preserving shuffles nearly reproduce without any second protein.  Second, an open frame proves little.  This is why Chung's group asked a harder question.  Their "nearly impossible by chance" referred to frames of 500 letters or more that stay open in both human and mouse alignments, and their simulations gave "Only 0.3% of simulated alignments preserved ARFs with 500 or more nucleotides."¹  An open frame in one genome is common;  a long open frame kept across two species is not.  The two numbers measure different things, and the design-side quotations in Part 7 use the second as though it described the first.

Old overprints matter more than new ones

Sabath and colleagues collected what was known about suppressing each of their 12 overprinted genes:  "Mutations in the youngest de novo genes (taxa 1–3) have little to moderate effect, whereas suppression of older de novo genes (taxa 6 and 10–12) has severe effects."⁴²  nog1 fits the young end:  a measurable but small cost when removed.  This fits a picture in which new frames start with weak or no function and some become built into the virus or cell over time.

What overprinting does not settle

1. No birth has been watched. Every origin in this Part is inferred by comparing living sequences. No source reviewed here observed a second frame arising and becoming useful in real time.
2. The ages are mostly relative. Sabath and colleagues write that "we are unable to estimate the absolute age of de novo genes but only their age relative to the divergence of a viral housekeeping gene."⁴² Cassan's group gives no date in years for ASP. Loughran's "∼160 million y" for POLGARF is a date in years, but it is read off the mainstream calibration of the placental-mammal tree, not measured by an independent clock.¹⁹ "Young" in this section means "found in fewer branches of the tree," and nothing here bears on a number of years.
3. The ages rest on comparing species. Saying that a frame "arose after" a split assumes the species share an ancestor. For strains of one virus or one bacterial genus, few readers on either side would deny a shared ancestor, so the inference is close to common ground there. For human and mouse, or human and chicken, it is the very point in dispute, and the lineage pattern is evidence for overprinting only to a reader who accepts that framework or finds it the best fit to the pattern.
4. The first step is not free, and it is not explained by the open frame alone. A hidden frame becomes a second gene only if it is transcribed, given a start the ribosome uses, and its product does something. Those steps may need mutations, and making an unused product costs the cell something. What overprinting removes is the need to invent two unrelated full sequences at once; it does not show which mutations occurred or how the new frame became established by selection or drift.
5. Overprinting explains the second frame, not the first. Every case here starts with a working gene. Where the first gene came from is the subject of Orphan Genes, How Rare Is a Working Protein? and Gene Duplication and Divergence.
6. A disordered, young, lineage-limited protein can still be important. The pattern says how the frames look, not that they are useless. ARF is a major tumour suppressor in mammals and is absent as an overlap in chickens.

Part 6: Other codes laid over coding DNA

Protein frames are only one kind of layer.  Coding letters also carry signals that are not read by ribosomes.  These are real, and some are well measured.  They are also different in kind from a second protein, and the "number of codes" claims in Part 7 depend on counting them as though they were the same.

Trifonov's codes

Edward Trifonov argued from the 1980s that DNA carries several messages at once.  His 1989 paper, by its PubMed abstract, named three:  the protein code with its reading frame, a chromatin code governing how DNA wraps around nucleosomes, and a putative code for DNA loops, and said that the codes overlap "so that some nucleotides belong to several messages simultaneously."⁴⁶  The body of that paper could not be obtained for this article, by three routes, so whether it lists twelve codes is unchecked.  His 2011 review is more sweeping in tone and gives no count:  "one has to admit that the genetic sequences carry many different codes.  If we are to know what the sequences are about, we have to detect and decipher these codes.  The times of surrender to 'junk' and 'selfish DNA' are over."⁴⁷  Trifonov is a mainstream molecular biologist, and his argument is a scientific claim about signals in sequence, not a design argument.

Room in the code for extra signals

Shalev Itzkovitz and Uri Alon asked in 2007 whether the standard genetic code is unusually good at letting coding letters carry other messages.  Compared with alternative codes, it is:  the title of their paper says it is "nearly optimal for allowing additional information within protein-coding sequences."⁴⁸  They offered two evolutionary explanations for how that came about:  "One possibility is that the ability to include parallel codes within protein-coding sequences conferred a selection advantage during the early evolution of the genetic code.  Alternatively, the genetic code might have been fixed in evolution before most parallel codes existed," and they went on to link the code's structure to its resistance to frame-shift errors.⁴⁸

In 2010 Itzkovitz, Eran Hodis and Eran Segal looked for extra signals in the coding regions of about 700 species by asking where codon choice departs from what the protein and codon preferences alone predict.  They found RNA folding signals, avoided start and initiation signals in bacteria, and microRNA target sites in eukaryotes, and wrote that "genomes may have evolved to encode extensive overlapping information within protein-coding regions."⁴⁹  Their measure is a departure from randomised codon choice, so it finds the layers that bias synonymous codons.  It does not count protein frames, and it does not say how many letters are fixed by each layer.

Protein-binding sites inside exons

In 2013 Andrew Stergachis, John Stamatoyannopoulos and colleagues mapped where regulatory proteins sit on DNA inside exons across 81 cell types and reported that "~15% of human codons are dual-use codons."⁵⁰  That is the largest figure for a second layer inside human coding sequence among the sources reviewed here.  It counts codons covered by a protein footprint in at least one cell type, which is not the same as codons whose letters are fixed by that protein.  A 2015 reanalysis disputed the claim that these sites constrain codon choice, and the article Synonymous Mutations and Codon Usage weighs the two papers.  This article does not repeat that argument.

Signals in the untranslated ends

Short frames in the leader of a messenger RNA, before the main start codon, regulate how much protein is made.  Sarah Calvo, Vamsi Mootha and colleagues found that 49% of human transcripts carry at least one, and in reporter tests "uORFs typically reduce protein expression by 30–80%."⁵¹  These are the most common translated "second frames" in the ribosome profiling catalogue, 3,083 of the 7,264.⁸  Most of them lie outside the main coding sequence, so they add a regulatory layer to the message rather than a second protein on the same letters.

What these layers do and do not add up to

1. They are real and measured. RNA structure, splicing signals, binding sites and upstream frames all constrain some letters inside genes.
2. They can constrain synonymous choice. They bias which of several codons for the same amino acid is used. A binding site or a folding signal usually tolerates many sequences. So a letter in a binding site may not fit the Montañez model of Part 7, in which any change at an optimal letter of a code is harmful.
3. No source reviewed here counts how many layers sit on a typical letter. The figures are per layer: 15% of codons in some footprint, 49% of transcripts with an upstream frame, 0.064% of coding letters in two accepted protein frames. None of the sources read here gives a distribution of the number of layers per letter, and none supports "up to 12" as a typical count.

Part 7: The design-side claims, tested

What the design writers say

Seven writers make the overlap argument in print.  Their own words come first.

1. John Sanford, Genetic Entropy, 2005 and 2014. "There is abundant evidence that most DNA sequences are poly-functional, and are therefore poly-constrained. This fact has been extensively demonstrated by Trifonov (1989). For example, most human coding sequences encode for two different RNAs that read in opposite directions (i.e., both DNA strands are transcribed; Yelin et al., 2003)." Then: "Trifonov (1989) has shown that probably all DNA sequences in the genome encrypt multiple codes (up to 12). In computer science, this type of data compression can only result from the highest level of information design." And: "Logically, all poly-constrained DNA had to be designed."² The 2005 edition makes the same three claims.⁵² The 2014 edition adds: "We now know that over half the human genome is read backwards as well as forwards. A given gene produces roughly 7 overlapping transcripts," and concludes that "Mutations that are both unambiguously beneficial and also impactful enough to be selectable, must be vanishingly rare."²
2. George Montañez, Robert Marks, Jorge Fernandez and John Sanford, 2013. A probability model, examined in its own section below. Sanford's book summarises it: "Any change at all will diminish total information with nearly absolute certainty. My colleagues and I have demonstrated this mathematically in a recent paper (Montañez et al., 2013)."²
3. Jonathan Wells, The Myth of Junk DNA, 2011. Wells reports a 2007 study's "overlapping coding regions," a feature "that the researchers concluded 'is nearly impossible by chance.'" and Itzkovitz and Alon's finding that the code is "nearly optimal for allowing additional information within protein-coding sequences."⁵³ He quotes William Dembski: "If these systems are designed ... we can expect the information to be densely packed and multilayered," and concludes that "dense, multi-layered embedding of information is a prediction of intelligent design."⁵³
4. Fazale Rana, Reasons to Believe, 2007. "The apparent widespread occurrence of overlapping genes doesn't make much sense from an evolutionary perspective, because of the cost they represent to the organism." "For me, the only explanation for overlapping genes is the work of a Creator. It's hard to imagine how undirected evolutionary processes could produce overlapping genes."⁵⁴ Reasons to Believe is an old-earth ministry, so this is a design argument, not a young-earth one.
5. Jeffrey Tomkins, Institute for Creation Research, 2014. On the 2013 binding-site paper: "a new, previously hidden genetic code embedded within the same sections of genes that code for proteins—utterly defying all naturalistic explanations for its existence." And: "Some genes overlap parts of other genes in the genome, and now it has been revealed that many genes have areas that contain dual codes within the very same sequence."⁵⁵
6. Ann Gauger, Ola Hössjer and Colin Reeves, 2017. "In addition, this is a strong argument for design: the sophistication required to have codes within codes would not arise by simple mutation and natural selection."⁵⁶
7. Casey Luskin, Science and Culture Today, 2026. "Another discovery in genomics predicted by ID is the presence of overlapping genes, where the same stretch of DNA encodes multiple genes. One strand of DNA might encode one protein while the opposite strand encodes an entirely different functional protein (see, e.g., Adelman et al., 1987 or Mehravar et al., 2021). Such an astonishing degree of information compression is only expected if a mind of genius encoded the information in DNA, not random mutation and blind natural selection."⁵⁷

Claim by claim

Two of these points need saying plainly, because they cut in the design writers' favour.  First, their core intuition is right:  an overlap does tie two messages together, and the code arithmetic, the virus comparisons and the decompressed phage all confirm it.  Second, Rana's observation that overlaps are costly is the same observation Chung's group made, and it is a fair question for any evolutionary account.  The overprinting literature's answer is that new frames begin as passengers on old genes, usually short and disordered, and that mutual constraint grows as both products become needed;  Part 5 lists what that answer still has to show.  The measurements support the coupling claim.  They do not establish either a genome-wide count of overlapping codes or the claim that a mind is the only possible explanation.  Protein-frame annotations do not measure every proposed regulatory layer, and the comparative history of some second frames supports addition within an older gene.

No design writer read for this article cites the overprinting literature.  Searches of Sanford's two editions, the Montañez chapter, Wells's book, Rana, Tomkins, Gauger and Luskin found no mention of Keese and Gibbs, Rancurel, Sabath or Pavesi.

The Montañez model

The model is the main quantitative design argument examined here, so its assumptions and calculations need to be tested.

What it says, in its own terms.  The model takes a stretch of "functional sequence" in which each letter belongs to between zero and twelve codes.  Membership is defined by consequence:  "A nucleotide that does not contribute to a given code is considered neutral relative to that code."  The authors also "allow for rare nucleotide sites within the functional sequence that are perfectly neutral."²⁰  In each code a letter is already the best letter with probability p(optimal);  the authors use 99.9% "for the analysis and discussion" and call it "a tunable parameter."²⁰  At a code site that is not yet optimal, "only one of the three alternative nucleotides will be an improvement" over the existing letter.²⁰  At an optimal site, "any change is deleterious, assuming no neutral changes."  And:  "There are no absolutely neutral positions in any given code, because by definition such a position is not part of that code."²⁰  So for one code, with r = 0.001 the share not yet optimal, a random change helps with probability "p(B) = (1/3)r = 0.00033."²⁰  The preferred letter of each code "is chosen independently of the other nucleotides that are optimal for codes that may overlap at that position."²⁰  A change that helps one code and harms another counts as neutral, and a change is beneficial overall when it helps a majority of the codes the letter belongs to.  The probability of that is their equation (1), a binomial sum.  For r = 0.001 and L = 12 they print "p(B)L = 4.15 × 10−22 or I[L+] = 71 bits."²⁰

The arithmetic follows from the inputs.  Recomputed for this article, the same inputs give 3.6 x 10^-22, or 71.2 bits.  The printed figure is what a per-code chance of 0.00034 gives, which looks like 1/3 of 0.001 rounded up in the last digit;  the text itself prints 0.00033.⁶³  The difference is about 15% in a number of order 10^-22 and is immaterial to the conclusion.  The two-code values printed on page 146 reproduce exactly.

The two-code column is lower than the three-code column because, with two codes, a change must help both to count as beneficial, while with three it need help only two.  The authors note the same odd-even effect.²⁰

So the result is only as good as the inputs.  The algebra is not in question.  What can be checked is whether each input matches what is known about real overlapping codes.

1. The number of codes, L. The paper's twelve-code sequence is "a hypothetical 100 base pair sequence," and the authors write that "The exact degree to which nucleotides participate in two or more codes is still unknown."²⁰ For protein frames, the annotated share of human coding letters in two frames is 0.064% on one canonical transcript per gene and 7.2% across every annotated transcript; three frames are rarer still. The 17 kinds of code listed on pages 154 and 155 include isochores, nucleosome binding, three-dimensional organisation and "other undiscovered coding structures."²⁰ Several of these are preferences that tolerate many sequences (Part 6), so it is unclear how many of their letters would meet the model's definition of a code site. No source reviewed here gives a count of codes per letter to plug in.
2. The share already optimal, 99.9%. This is assumed, and the paper says so.²⁰ No measurement of the share of letters already at their best possible letter was found for any code in the sources reviewed here. The table shows the answer changing by about nineteen orders of magnitude between r = 0.001 and r = 0.5 at twelve codes.
3. What counts as a code site. Within the model, every change at an optimal site of a code is harmful by construction, because a site where change does not matter is defined as outside that code. That is a modelling choice, and the algebra built on it is consistent. The open question is how many real letters meet the definition. Real outcomes are mixed letter by letter and change by change. At a protein-coding letter, some substitutions are silent and others are not, depending on which letter replaces which. In a same-strand protein overlap, 48% of single-letter changes leave at least one protein sequence unchanged, by the code arithmetic of Part 1, a protein-sequence count rather than a fitness measurement.⁴ The authors themselves say that "each code must be partially 'degenerate' (imperfect) to create the 'flexibility' required to allow other overlapping codes."²⁰ How real mutations in overlaps divide among harmful, neutral and helpful is a measurement, and the measurements are weighed in Beneficial and Harmful Mutations. No site-by-site measurement of an overlap scored in both frames was found among the sources reviewed here.
4. Codes choose their preferred letters independently. This assumption is about which letter each code prefers. It is not the trivial fact that overlapping codes share letters; the model already lets them share. Whether the preferred letters of real overlapping codes are correlated has not been measured in any source reviewed here, and a correlation could push the answer either way. If two codes tend to prefer the same letter, a change that helps one tends to help the other, and beneficial changes become more likely than the model says. If they tend to prefer different letters, the reverse. The code arithmetic of Part 1 shows that effects on protein sequence are correlated between frames, a change silent in one frame being usually an amino-acid change in the other, but that is a statement about effect classes, not about which letter each protein would prefer.
5. Effects are of equal size and cancel. The authors call this "not realistic" and argue that relaxing it would favour their conclusion: "we can logically expect most such ambiguous mutations to have a net deleterious effect. This is because, not only is it more likely for a random change to damage an optimized system than improve it, the nature of that damage will tend to be more pronounced than any potential improvement."²⁰ That is their argument about effect sizes. The model does not measure effect sizes, so the claim stands or falls on the fitness measurements, not on the model.

What the model does and does not show.  It prices the chance that a random change improves an already-optimised multi-code letter, under its stated inputs.  Its output is a conditional probability:  if a letter sits in L codes whose preferred letters are independent, each 99.9% optimised, with code sites defined as places where change matters, then a beneficial change has the stated chance.  It is not a measured fraction of beneficial mutations.  It does not model a new code arising, which is what overprinting is:  a second frame that did not exist before, built on letters chosen for the first.  Sanford's book presents it as a demonstration that "any change at all will diminish total information with nearly absolute certainty."²  The paper shows that for its assumed inputs, not for any measured genome.  Sanford's further claims about selection thresholds and genetic decline are weighed in Genetic Entropy.

What survives.  For two protein frames, the model's qualitative point about coupling holds:  most single-letter changes that alter one protein also alter the other, and only 1.2% leave both protein sequences alone.  The 1.2% is not the chance of a helpful change.  For a change that improves one protein's amino-acid sequence while leaving the other's sequence alone, the candidate category is the changes that alter only the first protein, about 23% of all changes in each direction by the arithmetic of Part 1.⁴  What share of those would help is not known from the code, and a change in protein sequence is not the same thing as a change in the organism's fitness.  Changes that alter both proteins are not all harmful either, as hepatitis B virus shows:  some surface-protein changes made by drug-resistance mutations help the virus escape antibodies (Part 4).  So the protein arithmetic supports coupling and trade-offs between the two frames.  It does not show that beneficial changes in overlaps are vanishingly rare.

Objections and responses

1. "Annotation misses most overlaps.  Ribosome profiling shows thousands more."  Annotation does miss some, and this article counts the ribosome-profiling frames separately:  1,469 translated frames overlap a known coding sequence in another frame, and counting every annotated transcript gives 7.2% of coding letters in two frames.  Even the generous counts leave two-frame coding a small share of human coding DNA and a tiny share of the genome.  Translation is also not function:  of the 7,264 frames, 15 meet the strict protein standard so far, though small proteins are hard to detect.  If future work proves thousands of working second proteins, the amount side of this verdict would change, and the code arithmetic says each would couple its letters to the host gene.

2. "Silent changes in both frames still matter, so the constraint is tighter than 1.2% suggests."  Agreed.  RNA structure, splicing signals and codon preferences can make even both-silent changes costly, as Synonymous Mutations and Codon Usage shows.  That strengthens the point about constraint where overlaps exist.  It does not make overlaps more common.

3. "Overprinting is a story.  A new frame would harm the old gene while it was still useless, so selection would remove it."  The objection is right that the first step is not free.  Making a hidden frame into a gene may need mutations, for a start the ribosome uses or for transcription, and those mutations can change the old protein;  even translating an existing frame costs the cell something.  What overprinting avoids is the need to invent two unrelated full sequences at once, because long open second frames already lie inside about one human gene in seven (Part 5), ribosome profiling finds translation in 1,469 frames that overlap known human genes in another frame (Part 3), and suppressing the youngest overprinted virus genes has "little to moderate effect" (Part 5).⁴²  Whether the steps of a given origin could persist through selection or drift has to be tested case by case.  No source reviewed here watched a second frame arise;  every origin is inferred from comparing living sequences, and the youngest cases are in viruses.

4. "Lineage limits could mean each kind was created with its own genes, not that the second frame is young."  Presence in one group and absence in its relatives does not by itself prove a late birth.  The frame might have been lost in the relatives, diverged beyond recognition, or arrived by transfer, and separate creation is a further option for anyone who allows it.  The evidence that fixes direction in the best cases is extra:  ASP, ORF3d and nog1 sit inside older genes that their close relatives share, and across 25 virus pairs the frame judged older from family trees usually matches the rest of its genome's codon usage better than the newer frame does (Part 5);  chicken exon 2 keeps an open but unused ARF frame;  POLGARF's start comes from an inserted repeat.  For strains of one virus or one bacterial genus, few readers on either side doubt a shared ancestor, and these features favour a frame arising within the group.  For human second frames, a reader who rejects common descent can read "found only in primates" as "made for primates."  A design account then has to explain why those frames look like overprints:  codon usage taken from the first frame, floppy proteins, frames that stay open in relatives without being used, and in POLGARF a start supplied by a mobile repeat.  Design does not forbid these features.  No design writer reviewed here predicts them.

5. "Small viruses are compact because they were designed to be.  Compression is a mark of design."  Compactness can be designed;  engineers compress data.  The test is whether the overlaps look built in or added on.  Where comparisons between relatives date virus overlaps, the newer frame is usually limited to one lineage and has the codon usage and disorder expected of a later addition, though those features are not exclusive proof (Part 5).  Compression also tracks genome size and capsid type in ways the virus-biology literature still disputes, which is a question about what limits virus genomes rather than about who wrote them.

6. "The decompressed phage shows that overlaps are essential, so phiX174's genome is irreducibly complex."  The engineers who built it say the opposite:  "A first fully decompressed phage genome is viable and not irreducibly complex."¹³  That needs a qualifier both ways.  The redesign replicates only when the host supplies gene F from a plasmid, so it is not shown to be viable on its own.  And it released about one-ninth as many phages per cell as the natural virus, a gap that mixes the effect of separating the genes with the redesign's many letter changes, the supplied gene F and different plates.  The experiment shows the overlaps are not strictly required for replication with help from the host, and that the natural arrangement does far better than this redesign.

7. "Chung's team found dual coding 'nearly impossible by chance.'"  Their phrase refers to long second frames kept open across human and mouse, against a codon-shuffling null they themselves call a possible weakness.  Their conclusion is that such frames are kept by selection and are "hallmarks of fascinating biology," not that they could not arise.  Open second frames in one genome are common (Part 5).

8. "The genetic code is nearly optimal for carrying extra codes.  That alone points to design."  Itzkovitz and Alon's result is real.  It bears on the origin of the genetic code, which is argued in Is the Genetic Code Optimized?, not on how much of a genome is overlapping.  The authors give two evolutionary routes to it, and the article does not settle between them and design.

9. From the other side:  "The Montañez model is a straw man."  It is not.  Its assumptions are written out, its algebra follows from them, apart from a rounding difference in one printed figure, and its point about coupling holds for protein overlaps.  Its weakness is that its decisive inputs, the number of codes per letter, the share already optimal and the independence of each code's preferred letters, are assumed rather than measured, and its definition of a code site makes every change at an optimal site harmful by construction.  It gives a conditional probability, not a measured beneficial fraction.

10. From the other side:  "Overlapping genes are just a quirk of viruses.  Nothing here needs answering."  The design writers have a real point:  overlaps couple letters, and an account of genome change has to price that.  The evolutionary literature does price part of it, in the silent-substitution comparisons, the decompression experiment and the overprinting studies, though the ratio statistics often used for this are unreliable in overlaps (Part 4).  The design writers' errors are about amount, history and the leap from coupling to a vanishing beneficial fraction, not about the existence of coupling.

Issues and unanswered questions

1. All counts here depend on annotation, in both directions. Our GENCODE counts use release 50, count stop codons as coding and include the mitochondrial genes; the canonical count is one transcript per gene by Ensembl's choice. Later releases will add more ribosome-profiling frames, and some recent GENCODE genes behind the 28 same-strand pairs came from exactly that work. Some annotated models may also prove unused. The RefSeq virus records omit frames found later, as SARS-CoV-2 shows. None of these counts measures the non-protein layers of Part 6.
2. Trifonov 1989 was read only in abstract. Springer and ScienceDirect both refused automated access, and no archive copy was found. So Sanford's "up to 12" is unverified, not refuted.
3. The phiX174 numbers come from the 2020 preprint, and the comparison is confounded. The published ACS Synthetic Biology version and the full text of Jaschke 2012 were not read here, and the published numbers may differ. The gap between natural and redesigned phage mixes overlap removal with other letter changes, host-supplied gene F and different plate sugars.
4. No site-by-site fitness scan of an overlap in both frames was found. A deep mutational scan of an overlapping virus region, measuring the effect of every single change on both proteins and on the virus, would test the Montañez inputs directly: how many changes are harmless in each frame, how many help, how many harm. None was found among the sources reviewed here; such a study may exist.
5. Simon-Loriere 2013 reports its ratio result two opposite ways. Its Table 1 and summary means put the ratio of amino-acid to silent substitution rates lower in overlapping parts; its text count of 81 of 117 genes and its supplementary Table S1 put it higher (Part 4). Which is right could not be settled from the paper, and this article uses only the tree-length and silent-rate results, which agree with each other. A corrigendum or the authors' raw output would settle it.
6. The ages of overprints are inferred. Sabath's group gives only relative ages, and Loughran's 160 million years for POLGARF comes from the mainstream calibration of the mammal tree rather than an independent clock. Every origin rests on comparing species, and for human frames the comparisons depend on homology searches that miss short or fast-changing proteins.
7. Several human cases have thin follow-up. A 2012 letter commenting on the AltPrP paper was not obtained. The Alt-ATXN1 paper does not establish how widely the frame is shared. Chung's list of 15 high-confidence genes, in its Table 1, was not read.
8. The human functional screens mostly test other kinds of frames. Prensner's 57 viability hits are mostly upstream, downstream and non-coding-RNA frames, and Chen's 78 start-overlap hits may act partly through the main gene. A screen designed to separate internal second frames from their host genes would sharpen the human count.
9. No young-earth model of overlapping genes was found. The young-earth writers read here cite overlaps as evidence of design but do not say how overlaps came to be distributed as they are, which frames are original, or how fast viral overprints arise. A model would need to say whether second frames were created with the first, arose afterwards, or both, and to predict the codon-usage and lineage patterns of Part 5. Viral cases such as ASP and ORF3d, which young-earth writers could accept as post-creation change, are the natural place to start.

Verdict

Overlapping protein codes are real, and where they occur they couple two messages.  Under the standard code, a same-strand overlap leaves only 1.2% of single-letter changes silent in both proteins, and most changes that alter one protein alter the other.  Real overlaps show the effect:  fewer silent substitutions in overlapping parts of virus genes, mutations that change two proteins at once, and a natural phage that far outgrows a redesign without overlaps.  On the existence of that coupling, Sanford, Montañez and their colleagues are right.

That coupling is not a measured beneficial fraction.  About 23% of changes in a same-strand overlap alter one protein and leave the other's sequence alone, and how many of those, or of the changes that alter both, would help an organism is not measured in any source reviewed here.  The Montañez model is correct algebra on assumed inputs:  the number of codes per letter, the share already optimal, independent preferred letters, and a definition of a code site under which every change at an optimal site is harmful.  Its output is conditional on those inputs.

The design writers' claims about amount are not supported by the counts reviewed here.  Two-frame coding fills half of a hepatitis B genome and 0.064% of human coding letters on one canonical transcript per gene, 7.2% across every annotated transcript.  Other "codes" laid over coding DNA are real and can constrain synonymous choice, but no source reviewed here counts how many sit on a typical letter.  "Up to 12" codes per letter is unmeasured, and "over half the human genome is read backwards as well as forwards" has no cited source.

Where comparisons between species allow a second frame to be dated, it looks added:  limited to one lineage, coding a floppy protein, using codons chosen for the older frame, and in POLGARF started by a mobile repeat.  That is what overprinting predicts, though these features are inferred from living sequences and are not exclusive proof, and the design writers read for this article do not engage it.  The fair conclusion for a skeptical reader is that overlapping genes are a real, constraining feature of compact genomes, uncommon among the annotated genes of large ones, and not evidence that most DNA is a dense multi-layer code.  Whether the first genes, the ones overprinting starts from, could arise without design is a different question, taken up in Orphan Genes and How Rare Is a Working Protein?.

Sources