High-level organization of isochores into gigantic superstructures in the human genome

Human DNA shows a complex structure with compositional features at many scales; the isochores—long DNA segments ($~{10}^5$ bp) of relatively homogeneous guanine-cytosine (G + C) content—are the largest well-documented and well-analyzed compositional structures. However, we report here on the existence of a high-level compositional organization of isochores in the human genome. By using a segmentation algorithm incorporating the long-range correlations existing in human DNA, we find that every chromosome is composed of a few huge segments ($~$ ${10}^7$ bp) of relatively homogeneous G + C content, which become the largest compositional organization of the genome. Finally, we show evidence of the biological relevance of these superstructures, pointing to a large-scale functional organization of the human genome.

Figure 1

(a) Autocorrelation function $C(d)$ vs distance $d$ (in isochores) obtained from the series of the G + C content of the isochores[6] in human chromosomes 1 (bottom+left axes) and 21 (top+right axes). The slowly decreasing behavior indicates a strong correlation of isochore G + C content up to large distances, showing that isochores are organized in large clusters of similar composition. The typical cluster sizes are 400 and 80–85 isochores in chromosomes 1 and 21, respectively. As the isochore average size is $1.0\times {10}^5$ bp (Chr 1), and $1.5\times {10}^5$ bp (Chr 21), we find $\langle s\rangle$ values of about 40 and 12 Mb, respectively. (b) DNA walks obtained for the p-arm of human chromosome 1 (bottom+left axes) and the q-arm of human chromosome 21 (top+right axes). In both cases large regions of almost constant slope (i.e., with well-defined G + C content) can be seen. The segmentation algorithm we introduce below splits these two sequences into three regions with definite G + C content, with average sizes of 40 Mb (Chr 1p) and 11.5 Mb (Chr 21q), in close agreement with their $\langle s\rangle$ values.

Figure 2

(a) A random ($\alpha =0.5$) numeric DNA sequence $N(i)$ of 512 bp and a G + C content of $0.5$. (b) A numeric LRC ($\alpha =0.75$) DNA sequence $N(i)$ of 512 bp with a G + C content of 0.5. (c) The profile of the $D_{\mathrm{JS}}$ values at any position in the sequences of parts (a) and (b). In both cases, the position where $D_{\mathrm{JS}}$ reaches its maximum is marked with an arrow, and the $D_{\mathrm{JS},max}$ value itself with a dashed line.

Figure 3

Behavior of percentiles 95 and 99 ($p$ values 0.05 and 0.01) of the distribution $p_{\ell ,\alpha }(D_{\mathrm{JS},max})$ vs the sequence length $\ell$ for synthetic LRC DNA sequences of different $\alpha$ values. Solid lines show power-law fittings.

Figure 4

(a) A Gaussian wavelet plot of the G + C content of Chr 21q. The wavelet scale varies in the range $({10}^4,1.25\times {10}^6)$ (bp) (left axis). The chromosome coordinates are given by the bottom axis of part (b). (b) Walk of Chr 21q (left+bottom axes), and the superstructures (right+bottom axes) obtained with our algorithm with $p_0=0.01$. (c) A zoom of a small region of the walk of part (b) marked with a box (left+bottom axes) and the isochores[6] contained in that region with their G + C contents shown in the right axis.

Figure 5

The cumulative probability $P(m)$ obtained for genes pairs co-located in the same segment of different segmentations: Superstructures (SS) and three different sets of isochores (C[6], I[4], B[5]), as well as random (R) gene-pairs used for comparison.

Figure 6

The average semantic similarity $\langle m\rangle$ for pairs of genes co-located in the same segment of different segmentations: superstructures (SS) and three different sets of isochores (C[6], I[4], B[5]), as well as random (R) gene pairs used for comparison.