The gene expression datasets we used include the gene expression levels in 69 non-disease adult tissue/organ samples from humans and 55 non-disease adult tissue/organ samples from mice 15. The weights of these tissue/organs are on a continuum varying by several magnitudes. For reliability, only the largest samples are defined as large tissue/organs and the smallest samples are defined as small tissue/organs (Table 1). The sizes of tissue/organs were determined by searching the literature 161718192021222324 and internet resources (for example, Wikipedia, the free encyclopedia), or estimated by experience. A conservative estimation of the difference in average tissue/organ weight between large tissue/organ samples and small tissue/organ samples is > 50 times.

Tissue/organ-specific genes are those that are expressed only in one particular tissue/organ sample. In total, we found 149 LTS genes and 96 STS genes in humans and 140 LTS genes and 246 STS genes in mice (Table 1, Additional Files 1, 2). As the tissue/organ weights differed by tens or even hundreds of times, an LTS gene is expected to produce tens or even hundreds of times more mRNA molecules per tissue/organ than an STS gene with a similar expression level. If the compactness of highly expressed genes has evolved to minimize the energetic cost of transcription, the LTS genes should be more compact than the STS genes with similar expression levels. However, pairwise comparisons of LTS-STS gene pairs with similar expression levels (for details, see Methods) do not show significant differences in average intron length, total intron length, intron number, coding sequence (CDS) length or untranslated region (UTR) length between LTS genes and STS genes, in either humans or mice (Figure 1).

How large a difference in expression level is required to generate a significant difference in gene compactness? The genes analyzed above were divided on the basis of expression level, rather than the size of tissue/organ; genes in the top 30% quantile were considered to be highly expressed and those in the bottom 30% quantile were considered to be weakly expressed genes. As shown in Table 2, the introns and UTRs of highly expressed genes are significantly shorter than those of weakly expressed genes, but there is no significant difference in intron number or CDS length (Table 2). This result is in contrast to a previous study 6, but is in line with another study, which found that total exon length is much more weakly related to expression level than intron length 1. We suspect that the small number of genes analyzed in this study may have obscured a weak trend. One might expect that increasing the difference in expression level between highly expressed and weakly expressed genes (for example comparing genes in the top 10% quantile with those in the bottom 10% quantile) would reveal significant differences in intron number and CDS length. In fact, selecting 10% quantiles resulted in a much smaller number of genes being analyzed and, consequently, statistically less convincing results (data not shown). The difference in expression level between the top and bottom 30% quantiles of human genes or mouse genes is about 20 times (Table 2). As the expression value detected by microarray is linear with the concentration of target RNA (Affymetrix 2001, technical note, new statistical algorithms for monitoring gene expression on GeneChip^® probe arrays), this difference in expression level can reflect the difference in the concentrations of the target mRNAs.

The weight ratio of a large tissue/organ to a small tissue/organ is much larger than the ratio in mRNA abundance required producing a significant difference in average intron length, total intron length and UTR length. However, large differences in tissue/organ weights do not produce significant differences in intron length or UTR length (Figure 1). This result is unexpected on the basis of the energetic cost hypothesis.

We also qualitatively estimated the length and number of introns in genomes that may be selected against because of their energetic cost during transcription. In a highly expressed housekeeping gene (housekeeping genes are expressed in all cells in the human body, so their cumulative energetic burden is higher), let us assume that there is an intron with the threshold length (L) to trigger natural selection. Several studies have shown that most eukaryotic genes are expressed at the level of two or three copies of mRNA per cell 252627, so a gene that produces 30 mRNA copies in each cell can be viewed as a highly expressed gene. The median half-life of human mRNA is about 10 h, and fast decay mRNAs have half-lives of < 2 h 28. For a conservative estimation, we can assume that the gene needs to synthesize 30 mRNA copies every 2 h, that is, 360 mRNA copies per day, per cell. The expense of transcription is two ATP molecules per nucleotide. Therefore, transcription of the intron requires 360 × 2 L = 720 L ATP molecules per day in each cell. Estimates of the number of cells in an adult human body vary from 10¹³ to 10¹⁴ 29. For a conservative estimation of the energetic cost of gene transcription, we used the higher value, 10¹⁴ cells. As an adult human consumes about 200 mol of ATP per day 1830, the energy consumption of each human cell is (200 × 6.02 × 10²³)/10¹⁴ = 1.2 × 10¹² ATP molecules per day. It should be noted that this is a conservative estimation; the energy consumption involved in strenuous exercises (for example, mountain climbing) may be as much as 10 times more than that used when resting 18. The proportion of human daily energy consumption representing the energetic cost of the long putative intron of a highly expressed housekeeping gene (which can be considered as the coefficient of natural selection, S) is 720 L/(1.2 × 10¹²) = 6 L × 10^-10. The recent effective population size (Ne) of humans is ≤ 10⁴ 3132. According to S = 1/(2 Ne) as the margin above which natural selection is stronger than genetic drift, L = 1/(2 × 10⁴× 6 × 10^-10) = 8.3 × 10⁴ nt. In human genome, only 0.9% of introns are longer than this threshold. In principal, this estimation is applicable to the energetic cost of the transcription of a CDS or UTR.

The major differences between humans and mice are in their body sizes, their metabolic rates and their effective population sizes. We could not find an estimation of the number of cells in a mouse body. However we did find data on mass-specific metabolic rates 3334, from which we can estimate energy consumption per mouse cell by assuming that human and mouse cells do not differ greatly in mass. The mass-specific metabolic rate of mice is 0.0151 W/g and that of humans is 0.00118 W/g 34, so a mouse cell uses ~12.8 times more energy than a human cell. As estimated above, the energy consumption of each human cell is about 1.2 × 10¹² ATP molecules per day, so that of each mouse cell is about 1.5 × 10¹³ ATP molecules per day. The proportion of mouse daily energy consumption (S) representing the energetic cost of the long putative intron of a highly expressed housekeeping gene is (360 × 2 L)/(1.5 × 10¹³) = 4.8 L × 10^-11, where L is defined as described in the previous paragraph. Different sources of data on the effective population size of mice are not consistent 3536; we retained a higher value (Ne = 8.1 × 10⁵) for a conservative estimation. Thus, in mice, the threshold length of introns to trigger natural selection is L = 1/(2 × 8.1 × 10⁵× 4.8 × 10^-11) = 1.3 × 10⁴ nt. Similar to the situation in humans, only a small fraction of introns in the mouse genome (6.8%) are longer than this threshold.

Owing to a lack of the required information (such as mRNA decay rates), it is impossible to accurately estimate the burden of long introns in other vertebrates and invertebrates. Considering that the effective population size of vertebrates is only about 10⁴ 37, we suggest that long introns in highly expressed vertebrate genes are unlikely to be selected against. However, for invertebrates, with an effective population size of about 10⁶ 37, it would be too bold to give a rough estimation.

Benefiting from the extensive studies on yeast Saccharomyces cerevisiae, we also found enough data to estimate the energetic burden of a long intron in a unicellular eukaryote. A gene that produces 30 mRNA copies in each cell can also be viewed as a highly expressed gene in yeasts 252627. The median half-life of yeast mRNAs is about 21 min, and the 90th percentile of mRNA half-lives is 10 min 26. Conservatively, we assumed that such a gene would need to synthesize 30 mRNA copies every 10 min; that is, 30 × 24 × 60/10 = 4320 copies of mRNA every day. To transcribe a long intron, a yeast cell consumes 4320 × 2 L = 8340 L ATP molecules, where L is defined as previously. A yeast cell weighs 3.35 × 10^-11 g and the median value of yeast metabolic rates at eight different temperatures is 0.267 W/g 38, so the metabolic rate of a yeast cell is 8.9 × 10^-12 W, which can be convert to 1.39 × 10¹³ ATP molecules per day. The proportion of yeast daily energy consumption representing the energetic cost of the putative long intron in a highly expressed gene is 8640 L/(1.39 × 10¹³) = 6.2 L × 10^-10. The effective population size of yeasts is about 10⁷ 3739. Thus, in yeasts, the threshold length of introns to trigger natural selection is L = 1/(2 × 10^7× 6.2 × 10^-10) = 81 nt. Unlike the situation in humans and mice, 86.5% of the introns in the genome of S. cerevisiae are longer than this threshold length. The fractional energetic cost of long introns may be overestimated here; thus the extant long introns, even in highly expressed genes, may be not under negative selection. At least, this result is helpful to explain the fact that unicellular eukaryotes generally have much shorter introns than mammals, and it is consistent with a previous study, which showed that energy is a constraint on evolutionary changes in yeast gene expression 39. However, these estimations are at least seemingly contradictory to the observations that highly expressed genes have longer introns than weakly expressed genes in yeasts 4041. To reach a conclusion, further investigations are required.

Considered just from the point of view of the energetic cost of transcription, loss of entire introns may be favored in yeasts, but unlikely in mammals. On the other side, intron gain may be selected against in yeasts, but is most likely neutral, and thus, under genetic drift in mammals. This idea is consistent with the paucity of introns in yeast genes and the abundance of introns in animal genes 4243. Previously, the existence of different rates of intron loss in the evolution of different lineages was explained by differential retrotransposon activities 444546. We look forward to further evidence to determine whether selection to reduce energetic cost is a complementary explanation. In evolution, insertion of several nucleotides or various transposons into introns and deletion of short sequences from introns are much more frequent than gain and loss of entire introns. Considered just from the point of view of the energetic cost of transcription, the effects of common indels are negligible in mammals, but visible to natural selection in yeasts. This idea is similar to the theory of Lynch on the evolution of genome complexity 4748.

The first alternate hypothesis is the time cost hypothesis. RNA polymerase II can elongate only about 20–40 nt per second 149. Recent evidence indicates that elongation, instead of RNA polymerase II recruitment, may be the predominant rate-limiting event in gene activation 5051. Therefore, gene length should have an important impact on the duration of gene expression. To be completely transcribed, a large gene in the human genome, such as DMD (2.3 Mb), requires 16 hours 49, a medium-sized gene (for example, TUBE1, 16.7 Kb) requires about 7–14 minutes, and a small gene (for example, HBA2, 834 bp) requires only about 20–40 seconds. Seoighe et al. 3 argued that the time required to transcribe multiple copies of mRNA is not a multiple of the transcription period of the first copy, because one template can be transcribed by several polymerases simultaneously 14. Assuming a normal elongation rate of 0.03 seconds per nucleotide, the completion of the transcription of the first copy of a gene with L nt requires 0.03 L seconds. Assuming that there are k polymerases attached to the same template simultaneously, the completion of an additional copy of this transcript requires an additional 0.03 L/k seconds. Thus, the completion of the transcription of n copies of an mRNA requires T_n = 0.03 L (1 + (n-1)/k) seconds. Apparently, if n <<k, T_n≈ 0.03 L, gene length and transcript copy number are not related. However, in highly expressed genes, n is unlikely to be much smaller than k; thus, both gene length (L) and transcript copy number (n) contribute to the duration of transcription. To produce a large number of transcripts in a limited period of time, natural selection may decrease L or increase k. Unfortunately no genome-wide data on the values for k are now available in animals.

On the other side of the same coin, the time taken to transcribe introns has long been proposed to contribute to the timing mechanisms during development 525354. An extension of this hypothesis is that long introns may be maintained in some genes to reduce the number of mRNA products in the otherwise too-long time during which the genes are activated.

Another alternate hypothesis is that short genes may experience lower frequencies of abortive transcription and/or erroneous splicing than long genes. Successful transcription requires the polymerase to be stably associated with the DNA template during the elongation process. However, in some cases, the RNA-DNA duplex may not be stable enough to avoid abnormal pausing and arrest of elongation 55. In a study of the human DMD gene, Tennyson et al. 49 found that 30–40% of transcription events were terminated or stopped at premature sites. Recently, Guenther et al. 50 found that many genes that have experienced transcription initiation do not produce complete transcripts. The short lengths of highly expressed genes may lead to a decreased possibility of a gene containing such sequences that are difficult to transcribe and cause abortion of elongation. In addition, evidence shows that long introns increase the frequency of erroneous splicing of nearby exons 56.

Long introns (and long UTRs) in highly expressed genes may also be selected against because of the crowding of active genes in a restricted interchromatin compartment 57.

A slightly more speculative and seemingly less likely hypothesis is that long introns are selected for in weakly expressed genes to avoid DNA damage resulting from transcriptional R-loops 658. The fact that mRNA lengths have a similar correlation with expression levels as intron lengths 169 negates this hypothesis.

In addition, there is also the possibility that highly expressed genes are compact because their epigenetic regulation is relatively simple, as suggested by the "genome design" hypothesis 10. Although there is some evidence against this idea, indicating that the lengths of intergenic spacers rather than those of introns are correlated with the complexity of epigenetic regulation 659, there is also evidence supporting it 6061626364.

In contrast to the observations that highly expressed genes have short introns in animals, P. patens and the pollen of A. thaliana, highly expressed genes were found to have longer introns than weakly expressed genes in unicellular organisms, the sporophytes of A. thaliana and Oryza sativa, and, at least, the vegetative stage of the slime mould Dictyostelium discoideum [404165, Y.F. Huang and D.K. Niu, unpublished results from analyzing the data from 66]. To date, there has been no satisfactory explanation for this difference 465. Perhaps, the compact genomes and compact genes in large genomes have lost most of their nonfunctional sequences; thus, most of the retained intronic sequences have regulatory functions in gene expression 67686970. Surprisingly, a weak, but significant negative correlation of mRNA length (and protein length) with expression level was found in all studied organisms 125671727374, which is also generally explained by minimizing the energetic cost of gene expression. In light of this study, we suggest other potential reasons for the short introns of highly expressed genes: to minimize the duration of gene expression, or to reduce the frequencies of abortive transcription and/or erroneous splicing. However, we do not wish to completely discount the energetic cost hypothesis for mRNA compactness, because we have insufficient data on protein abundance (note that translation is also an expensive process).