The frequencies of individual nucleotides exhibit significant fluctuations across eukaryotic genes. In this paper, we investigate nucleotide variation across an averaged representation of all known human genes. Such a representation allows us to average out random fluctuations that constitute noise and uncover remarkable systematic trends in nucleotide distributions, particularly near boundaries between genetic elements—the promoter, exons, and introns. We propose that such variations result from differential mutational pressures and from the presence of specific regulatory motifs, such as transcription and splicing factor binding sites. Specifically, we observe significant GC and TA biases (excess of G over C and T over A) in noncoding regions of genes. Such biases are most probably caused by transcription-coupled mismatch repair, an effect that has recently been detected in mammalian genes. Subsequently, we examine the distribution of all hexanucleotides and identify motifs that are overrepresented within regulatory regions. By clustering and aligning such sequences, we recognize families of putative regulatory elements involved in exonic and intronic splicing control, and 3' mRNA processing. Some of our motifs have been identified in prior theoretical and experimental studies, thus validating our approach, but we detect several novel sequences that we propose as candidates for future functional assays and mutation screens for genetic disorders.

Although histones can form nucleosomes on virtually any genomic sequence, DNA sequences show considerable variability in their binding affinity. We have used DNA sequences of Saccharomyces cerevisiae whose nucleosome binding affinities have been experimentally determined (Yuan et al. 2005) to train a support vector machine to identify the nucleosome formation potential of any given sequence of DNA. The DNA sequences whose nucleosome formation potential are most accurately predicted are those that contain strong nucleosome forming or inhibiting signals and are found within nucleosome length stretches of genomic DNA with continuous nucleosome formation or inhibition signals. We have accurately predicted the experimentally determined nucleosome positions across a well-characterized promoter region of S. cerevisiae and identified strong periodicity within 199 center-aligned mononucleosomes studied recently (Segal et al. 2006) despite there being no periodicity information used to train the support vector machine. Our analysis suggests that only a subset of nucleosomes are likely to be positioned by intrinsic sequence signals. This observation is consistent with the available experimental data and is inconsistent with the proposal of a nucleosome positioning code. Finally, we show that intrinsic nucleosome positioning signals are both more inhibitory and more variable in promoter regions than in open reading frames in S. cerevisiae.