Case study 1: histone modifications in the human genome
Certain post-translational covalent modifications of histones are associated with gene expression [10
], with specific combinations known to act cooperatively [14
]. To demonstrate Segtools's functionality, we generated a segmentation from the ChlP-seq "peaks" (genomic loci exhibiting significantly elevated read count) for core histone H3 methylated at three different lysine residues (H3K4me3, H3K27me3, H3K36me3). The Broad Institute produced these data from the chronic myelogenous leukemia cell line K562 as part of the ENCODE Project [16
], and we downloaded them from the UCSC Table Browser [17
] on assembly NCBI36.
We compared the segmentation against GENCODE [18
] version 3c gene annotations and transcription start sites (TSSs). We classified a gene as active when the number of ENCODE Project RNA-seq [19
] reads per kilobase per million mapped reads (RPKM) in the gene exceeded the 75th percentile and as inactive when the gene had 0 RPKM. We classified TSS as active when it had at least 2 K562 cytosolic poly(A)+
CAGE tags mapped from the ENCODE Project CAGE data [20
], and as inactive when the TSS had 0 CAGE tags. First, we used flatten to create a segmentation in which the label for each segment corresponds to the combination of histone modifications with a peak at that segment. For example, the "4/27" label corresponds to regions spanned by both H3K4me3 and H3K27me3 ChlP-seq peaks. We then used aggregation
in "gene mode" to visualize the enrichment of each label around the 11,693 protein-coding GENCODE genes active in the K562 cell line. Consistent with previous studies, Figure shows the enrichment of H3K4me3 (4) around active transcription start sites in the first row, depletion of H3K27me3 (27) around active genes in the second row, and enrichment of H3K36me3 (36) in the bodies of actively-transcribed genes in the third row.
Then we created Figure , with nucleotide-frequency. It shows the increased frequency of CpG in all labels that include promoter-associated H3K4me3 (4) peaks.
Finally, we used overlap to explore each label's predictive power for protein-coding TSS activity. With precision (also known as the positive predictive value) of 70.2% and and recall (or sensitivity) of 54.2%, segments high in both H3K4me3 and H3K36me3 were most predictive of overlapped TSSs being active. Surprisingly, segments high in all three histone modifications were the next most predictive of TSS activity, with precision of 68.7% and recall of 20.1%, suggesting that the presence of the other two histone modifications compensates for the inhibitory effect of H3K27me3. Segments with H3K27me3 alone were the most predictive of inactive TSSs, with precision of 95.2% and recall of 30.7%, though segments also high in H3K36me3 spanned an additional 5.2% of the inactive TSSs with a precision of 83.6%. In general, Segtools analyses are quick and parallelize easily. For this case study, the flatten analysis, which operated on three segmentations consisting of around 61,000 segments spanning ~50% of the human genome, required only 15 s on a single 2.33 GHz Intel Xeon CPU. The nucleotide-transition command processed the 1.6 billion bases spanned by the segmentation in 4 min, the overlap command summarized the intersection between these segments and 73,000 transcription start sites in 17 s, and the aggregation aggregated the segmentation over 9,000 gene models in 2 min.
Case study 2: gene expression and local chromatin structure in the Plasmodium falciparum genome
We used Segtools to investigate the relationship between gene expression and local chromatin structure in Plasmodium falciparum
, the parasite responsible for the most lethal form of malaria. Le Roch et al. [21
] performed microarray expression assays in two time series across the Plasmodium
erythrocytic cell cycle, corresponding to cell cycle synchronization performed with a 5% D-sorbitol treatment (cell cycle D) and a temperature cycling incubator (cell cycle I). Recently, these data were complemented with cell cycle time series data from two assays that measure local chromatin structure [22
]: formaldehyde-assisted isolation of regulatory elements (FAIRE) [23
] and MNase-assisted isolation of nucleosomal elements (MAINE) [24
]. We used Segtools to investigate the extent to which the local chromatin profile varies as a function of gene expression.
Our analysis consisted of three steps. First, we identified genes that were significantly expressed in each of the three primary stages of the erythrocytic cycle: ring, trophozoite and schizont. To do so, we applied the statistical criterion from [21
], and we required that the gene be expressed either in the "early" or "late" gene expression experiment for the given stage. This procedure was carried out separately for the two cell cycle data sets (D and I). Second, we used a previously curated set of transcription start sites (TSSs) [26
] to identify genes with a single, known TSS, and then we labeled these TSSs with one of eight labels (R, S, T, RS, RT, ST, RST, 0) indicating the stages during which the gene is expressed. This labeling was accomplished by creating a BED file for each stage and then using flatten
to merge the separate files into a single segmentation. The flattening was carried out separately for each cell cycle data set, resulting in two distinct labelings. Third, we applied several Segtools commands to each of the two segmentations, using a Genomedata archive that contained the FAIRE and MAINE data.
Figure shows the results of applying length-distribution. Because we selected a 200 bp window around each TSS, the percent coverage by "Segments" or "Bases" is identical so we specified -no-segments to only plot the base coverage. The figure shows that a large proportion (47%-48%) of genes with known TSSs are expressed in all three stages of the erythrocytic cycle, and only a small proportion (10%-13%) are expressed, or at least accessible to transcription factors, exclusively in a single stage. This observation is consistent across the two cell cycles. Altogether, the data indicates that only a small proportion of the genes can be expressed in a stage specific manner.
Figure shows the distribution of MAINE and FAIRE values over the course of the erythrocytic cell cycle as a function of different gene expression classes, produced using signal-distribution. Each cell corresponds to one expression label and one time point. The color of each cell indicates the strength of the MAINE or FAIRE signal in TSSs with the corresponding label. Each row of the plot is linearly scaled so that the minimum and maximum values are 0 and 1, respectively. Horizontal lines within the plot indicate the magnitude of the standard deviation in a given cell, relative to all other cells. Rows and columns have been ordered using the hierarchical clusterings shown on the top and right of each heat map. These two plots exhibit several intriguing features.
First, we note that the hierarchical clusterings shown along the right edge of both panels indicate that the FAIRE measurements at the end of the erythrocytic cycle (hr36) most closely resembles MAINE measurements (at hours 12, 18 and 30 in cell cycle D and hours 6, 24 and 30 in cell cycle I). This observation — that the FAIRE measurement of open chromatin at hr36 resembles measurements of closed chromatin — is consistent with the model proposed by Ponts et al., in which the parasite strongly compacts its chromatin in preparation for egress from the red blood cell at the end of the erythrocytic cycle. Second, we note that the genes expressed exclusively at the beginning of the cell cycle (R - ring stage) show an extremely strong and complementary pattern to genes expressed during the middle of the cell cycle (T - trophozoite stage). This pattern is particularly strong in cell cycle D (panel A), but also appears in cell cycle I (panel B). Apparently, ring-specific genes exhibit closed chromatin around their TSSs, whereas trophozoite-specific genes exhibit open chromatin around their TSSs. This pattern is consistent across nearly the entire cell cycle, with the possible exception of hr36, suggesting that local chromatin structure may contribute to stage-specific gene expression, but that local chromatin dynamics may not be the only mechanism regulating gene expression.
Overall, the figure shows relatively little correlation between the time at which a gene is expressed and changes in local chromatin structure. Canonically, time points 0, 6 and 12 of the MAINE/FAIRE data correspond to the ring stage, time points 18 and 24 correspond to the trophozoite, and time points 30 and 36 correspond to schizont. The absence of a strong correlation between time of expression and the degree of local chromatin compaction suggests that, though Ponts et al. have clearly demonstrated that local chromatin structure changes over the course of the erythrocytic cycle, the current analysis does not support a model in which the degree of chromatin compaction around the TSS directly correlates with the expression of the gene. Apparently, a more complex model that integrates additional types of data, such as transcription factor binding and histone modification profiles, is required to fully understand Plasmodium's unusual gene expression machinery.