Genome-Wide Motif Statistics are Shaped by DNA Binding Proteins over Evolutionary Time Scales

The composition of a genome with respect to all possible short DNA motifs impacts the ability of DNA binding proteins to locate and bind their target sites. Since nonfunctional DNA binding can be detrimental to cellular functions and ultimately to organismal fitness, organisms could benefit from reducing the number of nonfunctional DNA binding sites genome wide. Using in vitro measurements of binding affinities for a large collection of DNA binding proteins, in multiple species, we detect a significant global avoidance of weak binding sites in genomes. We demonstrate that the underlying evolutionary process leaves a distinct genomic hallmark in that similar words have correlated frequencies, a signal that we detect in all species across domains of life. We consider the possibility that natural selection against weak binding sites contributes to this process, and using an evolutionary model we show that the strength of selection needed to maintain global word compositions is on the order of point mutation rates. Likewise, we show that evolutionary mechanisms based on interference of protein-DNA binding with replication and mutational repair processes could yield similar results and operate with similar rates. On the basis of these modeling and bioinformatic results, we conclude that genome-wide word compositions have been molded by DNA binding proteins acting through tiny evolutionary steps over time scales spanning millions of generations.

Popular Summary

Nonfunctional DNA binding by proteins can disrupt basic cellular processes such as transcription, gene regulation, replication, and mutational repair. By evolving their global composition of short DNA “words,” genomes could potentially reduce the frequency of nonfunctional binding. Here, we show that such an evolutionary process imposes global constraints in all genomes and operates via tiny steps over time scales spanning millions of generations.

In this study, we analyze DNA-binding proteins in 947 bacterial or archaeal genomes and the genomes of 75 eukaryotic species. We determine the global constraints that are set by the distinct complement of DNA-binding proteins present in each genome that are responsible for preserving ancient phylogenetic signals. Using a mathematical model, we show that a distinctive genomic signature of such constraints is preserved in the genomes of various divergent species, for example, between D. melanogaster (the common fruit fly) and M. musculus (the house mouse), species that diverged over 600 million years ago. Our analysis demonstrates that weak binding sites in genomes are preferentially avoided, a result that holds true across the domains of life. Put another way, we show that the global word composition of each genome has been molded by its DNA-binding proteins over the course of evolution.

The outcomes of our study, which reveal that a large number of small effects act collectively to maintain genomic binding landscapes over long evolutionary time scales, pave the way for investigations of how this general evolutionary mechanism impacts a wide range of cellular processes.

Figure 1

Mouse genomic binding landscape and in vitro DNA binding measurements. (a) Distribution of binding scores $b_i$ for the Mafk TF over all 8-mer words $i$. (b) Distribution of 8-mer word frequencies $f_i$ in mouse introns (black curve); $f_i$ are shown normalized with respect to expectation based on genome-wide nucleotide composition. Words are shown as separate histograms according to their CpG counts (colored bars). (c) Correlation of $b_i$ and $\log f_i$ for the Mafk TF in separate CpG categories. Color indicates density of points. (d) Correlation coefficients (Spearman’s $\rho$) of $b_i$ versus $f_i$ are shown for each mouse TF separately computed conditioned on word CpG content. Bars are shown only for statistically significant correlations, with $p$ value $<{10}^{-6}$. (e) Binding scores of words ($b_i$) are correlated with the average binding score of their mutational neighborhoods (${\tilde{b}}_i$); results shown for Mafk, $\rho (b,\tilde{b})=0.87$. (f) Correlation of $f_i$ and ${\tilde{f}}_i$ over all 8-mer words $i$ for mouse introns; words are colored according to CpG content, and frequencies are normalized as in (b) and (c).

Figure 2

Word-neighbor correlations are observed in genomes across all domains of life. (a) Word frequency plots of $f_i$ versus ${\tilde{f}}_i$ in exons of representative bacterial and eukaryotic genomes. Points correspond to all possible 6 bp words. Frequencies are shown relative to the null expectation from synonymous codon shuffling, where a value of 1 means the observed frequency is equal to the expected frequency. For eukaryotes, frequencies are computed across all chromosomes. (b) Word frequency plots from exons of individual human chromosomes (see Fig. S9 of Supplemental Material [32] for all data).

Figure 3

Cross-species comparison of intron word frequencies using principle components analysis (PCA). Each point in the PCA plots corresponds to a single chromosome projected on the two principal axes. Principal axes are computed using the 6-mer relative word frequency vectors (2-mer null model) of all chromosomes shown in each plot. See Table S3 of Ref. [32] for further species information.

Figure 4

Evolutionary model of global selection. (a)–(c) Solution of the model at mutation-selection balance. Plots in the $(\tilde{f},f)$ plane of the equilibrium word frequencies for independent, normally distributed selective coefficients $s_i\sim \mathcal{N}(0,{\sigma }^2)$, using the indicated values of $\sigma$. We numerically solve Eq. (2) to determine the equilibrium word frequencies. Color shows different bins of selective coefficients, and predicted equal-pressure lines [Eq. (4)] are drawn using each interval’s average pressure. Inset in (c) shows a zoomed view. Parameters are $k=6$ and $u={10}^{-4}$, and frequencies are shown relative to the neutral expectation $4^{-k}$. (d) Equilibrium for strongly correlated pressures on neighboring words. (e) Correlation is reduced by shuffling 50% of words’ pressures from (d).

Figure 5

Distribution of global selective coefficients in different genomes. (a) Selective coefficients of 6-mers in four eukaryotic species. Intron data are used and $\mathrm{\Delta }{s}_i$ values are given with respect to the 2-mer null model. (b) Selective coefficients of 6-mers measured in two bacterial species. Exon data are used and $\mathrm{\Delta }{s}_i$ values are given with respect to the synonymous codon-shuffling null model. In E. coli (bottom panels), an inset shows the bulk of the distribution, since a small number of words including known restriction sites have large negative coefficients. Pressure distributions are shown using a value of $c=1$ in Eq. (4), which corresponds to taking the average global fitness of a sequence to be one.