Quantitative spatial analysis of bacterial transcriptome and chromosome structural data with GRATIOSA: application to twin-supercoiled domain distribution

Abstract

While classical models of transcriptional regulation focus on transcription factors binding at promoters, gene expression is also influenced by chromosome organization. Understanding this spatial regulation strongly benefits from integrated and quantitative spatial analyses of genome-scale data such as RNA-Seq and ChIP-Seq. We introduce Genome Regulation Analysis Tool Incorporating Organization and Spatial Architecture (GRATIOSA), a Python package making such combined analyses more systematic and reproducible. While current software focuses on initial analysis steps (read mapping and counting), GRATIOSA proposes an integrated framework for subsequent analyses, providing a broad range of spatially resolved quantitative data analyses, comparisons, and representations. Several tutorials illustrate applications across diverse species for typical tasks involving RNA-Seq, ChIP-Seq, and processed Hi-C data. We also use the software to quantitatively assess the validity and extension of the twin-supercoiled domain model in Escherichia coli genome-wide transcription, using recent topoisomerase ChIP-Seq data. We show that topoisomerases are locally recruited specifically by the 40% most highly expressed transcription units, with magnitudes correlating with expression levels. The recruitment of topoisomerase I extends to around 10 kb upstream, whereas DNA gyrase is recruited at least 30 kb downstream of transcription units, with subtle requirements for each enzyme depending on the orientation and expression level.

Graphical Abstract

Graphical Abstract.

Introduction

In bacterial cells, the chromosome is highly compacted through various factors, including DNA supercoiling (SC) and nucleoid-associated proteins (NAPs). Together, they dynamically organize its structure through looping, bending, and twisting of the DNA [1, 2]. In turn, the conformation of the chromosome affects DNA transactions, in particular transcription. The physical arrangement of DNA within the cell thus represents an “analog” information source crucial for the regulation of gene expression, which complements the “digital” information encoded by the genetic sequence [3, 4], and yet it largely escapes current modeling schemes.

High-throughput data such as RNA-Seq, ChIP-Seq, and Hi-C, when combined, provide invaluable information on this relationship [5–7]. However, the analysis methodology should then differ from that inherited from classical, “digital” regulatory models based on transcription factors (TFs). In the latter, a regulator controls a gene depending on its sole promoter (“regulon”), independently from its genomic position and distance to neighboring genes. Software solutions currently used in the analysis of differential expression [8] and inference of regulatory networks [9] are all inspired by classical regulatory networks. Recently, combined software solutions were proposed for RNA-Seq and ChIP-Seq data, such as GenPipes [10], pyrpipe [11], and SnakeChunks [12], which facilitate the analysis of such data within a unified framework [downloading of raw sequencing reads, alignment, differential expression, peak calling, classification, and Gene Ontology (GO) enrichment tests], but also rely on such methodologies, and are therefore irrelevant to “analog” regulation. For example, numerous whole-genome analyses suggest that neighboring genes are co-expressed, even when they are unrelated in classical transcriptional regulatory networks [13], presumably because they belong to the same topological domain, and this feature is evolutionarily conserved [14]. Analyzing this spatial regulatory interaction obviously requires considering the position of each gene, which none of the previously mentioned software does. In turn, the chromosome conformation is organized by abundant proteins (NAPs, topoisomerases) that bind the genome extensively and with looser sequence specificity than TFs, requiring a quantitative (more/less) rather than qualitative (bound or unbound promoter) analysis.

New analysis approaches are therefore necessary, for which no software solutions are available. Studies addressing these questions currently rely on homemade numerical analysis scripts^, [14–17] with two major drawbacks: (i) specific skills and significant efforts are required for importing and comparing data of heterogeneous types and formats and (ii) reproducibility and standardization are more difficult to ensure. Genome Regulation Analysis Tool Incorporating Organization and Spatial Architecture (GRATIOSA) is a new Python package targeted to computational biologists, and addressing this issue by facilitating such quantitative and spatial analyses of expression/structural data. It provides a unified framework for the automatic import and combination of many usual data types, and subsequent quantitative and spatial analyses of these data along bacterial genomes. Here “spatial” refers to the linear organization (along the genome sequence), where each gene occupies a well-defined position (the extension to 3D localization in the cell is addressed later in the manuscript).

GRATIOSA does not replace usual RNA-Seq or ChIP-Seq pre-analysis tools (read quality check, mapping, differential expression, or peak calling, possibly using the integrated packages mentioned above), but rather facilitates the subsequent steps that previously required manual processing, i.e. the integration, comparison, and combined statistical analyses of these data (see a comparison to existing software in Supplementary Table S1). It takes advantage of the versatility of Python for data analysis and numerical processing thanks to numerous available libraries. To the best of our knowledge, no comparable packages were proposed before as open tools for the community. In this article, we present the structure of the software and describe several tutorials showcasing several typical tasks based on RNA-Seq, ChIP-Seq, and processed Hi-C data in three bacterial species, Escherichia coli, Dickeya dadantii, and Sinorhizobium meliloti (which harbors two megaplasmids). We then show the features of the software more comprehensively through a specific but important example of “analog” regulation: the quantitative interplay between gene expression and topoisomerase activity along the E. coli genome, which has never been analyzed in a global and systematic way.

Materials and methods

Framework

GRATIOSA is written in several Python source files, each defining one of the main object classes (Genome.py, Gene.py, GO.py, Transcriptome.py, Chipseq.py, etc.) with their attributes (e.g. a Genome object contains Genes objects and a genomic sequence) and methods (e.g. load_annotation, load_seq, etc.). In addition, some miscellaneous functions for each class are defined in a separate source file (e.g. intermediate functions for reading a gff file are defined in useful_functions_genome.py). Finally, three additional program files contain intermediate functions for statistical tests (stat_analysis.py) and graphical representations of genomic data or statistical results (plot_genome.py and plot_stat_analysis.py, respectively). All methods and functions rely on the packages NumPy [18], Matplotlib [19], pandas [20], pysam, SciPy, and Statsmodels. The package can be loaded in all systems, but functions involving input/output methods were written for UNIX environments (Linux, MacOS) and are not fully supported in Windows (e.g. file opening/writing/copying). Three main types of interdependent objects can be distinguished: the Genome object, the elementary objects (Gene, TU, TSS, TTS objects), and the objects containing experimental data (Transcriptome, ChIP-Seq, Hi-C objects). An elementary object represents a single gene or a single transcription unit (TU), while the Genome object contains the annotation of all elementary objects (i.e. all genes, all TUs). Similarly, the elementary object may have attributes such as expression level or average ChIP-Seq signal for its segment, but the overall expression levels and genomic signals are attributes of the Transcriptome and ChIP-Seq objects, respectively (Supplementary Fig. S1).

Experimental data preprocessing

GRATIOSA can import experimental data formatted as continuous signals (i.e. one value per genomic position), discrete sites (i.e. list of genomic positions), or gene-wise values (i.e. one value per gene). The data files follow usual formats generated with standard external tools such as Bowtie2, DESeq2, MACS, or deepTools (Table 1). Functional enrichment analysis using GO terms is also performed using standard description files, typically an annotation file for the analyzed species and a .obo file describing all GO terms, which can be automatically downloaded from the GO website [21]. The GO annotation file of the analyzed species (typically generated by software such as Blast2GO) contains at least two columns, with each row relating one gene with one GO (with multiple lines for each gene).

Table 1.

Summary table of the data types analyzed by the package (mostly tabular files), with their content information (rows and columns) and typical software used to obtain them

Data type	Rows	Columns	Standard format	Commonly used packages
RNA-Seq reads	–	–	bam	bowtie2 [22]
RNA-Seq coverage	Positions	Coverage	bedgraph, wig	Bedtools [23]
RNA-Seq expression	Genes	Locus, expression level (RPKM or FPKM)	tab	Cufflinks, HTSeq [24]
RNA-Seq differential expression	Genes	Locus, fold change, P-value	csv	limma [25], DESeq2 [8]
Microarray differential expression	Genes	Locus, fold change, P-value	csv	limma [25]
ChIP-Seq coverage	Positions	Bin start, bin end, coverage	bedgraph, wig	deepTools [26]
ChIP-Seq peaks	Sites	Position, value	bed	MACS [27]
Hi-C border positions	Sites	Position, score	csv, txt	Chromosight [28], HiCDB [29]
list of genomic position data	Sites	Position, value	csv

The last row describes the import of diverse dataset types describing lists of discrete genomic sites (e.g. experimental or predicted protein binding sites, list of other genomic features, etc.), and the previous row (Hi-C border positions) is a special case of these. RPKM/FPKM: reads/fragments per kilobase per million reads.

Database organization

For GRATIOSA analysis, the user-provided data must be organized in a database with a fixed structure: each directory represents one organism (with a reference genome) and contains multiple subdirectories (Supplementary Fig. S2). The annotation subdirectory contains the annotation, usually in gff3 format. All other subdirectories have a similar structure, each containing experimental data files and an info file (see the “Experimental data preprocessing” method). The chipseq subdirectory is further subdivided, to distinguish between two types of data (discrete peaks or continuous signals).

Data import

GRATIOSA can handle any genome described by two standard annotation files, providing the reference sequence (.fasta format) and genome annotation (.gff3 format), respectively. This condition includes some genomes with missing nucleotides (if an annotation is available) and with several chromosomes and/or plasmids. These files can be automatically downloaded from NCBI. All other datasets must be provided by the user. For experimental data present in multiple files (e.g. different experimental conditions for gene expression), the import function uses an “info” file, i.e. a tab-separated text file containing the list of data files that will be loaded for analysis and some descriptors for each file (Supplementary Fig. S3). This method enables the automatic selection and import of data from various sources or softwares, whether in standard or custom formats. For example, in the case of the expression directory describing genome-wide expression levels in various conditions (microarray intensities or RPKM/FPKM for RNA-Seq data, indicating the number of reads/fragments per kilobase per million reads), the info file contains a list of rows (one per experimental condition to be loaded), each indicating the condition name, associated file name, column index for gene names/tags, column index for expression data, etc. (see documentation and tutorials for detailed instructions).

Depending on the type of data to be analyzed, the corresponding object needs to be initialized first (here Transcriptome), and then data import is achieved using the appropriate method (e.g. load_expression to import expression data automatically onto the Transcriptome object; see Supplementary Fig. S3). Intermediate functions (e.g. useful_function_Genome.read_seq) are used to read data files, and might be called by the user in unusual cases (e.g. files with very custom formats).

Data processing

Once data are loaded (typically as Python lists, NumPy arrays, or Pandas dataframes), new attributes can be either added manually or computed from existing ones. For example, ChIP-Seq signals can be binned, smoothed, or averaged. For binning, the user defines a bin size and method (by default, average signal within the bin); for smoothing, the size of the sliding window; for averaging, the list of different signals to be used (usually replicates). Raw or processed signals can be automatically mapped to Gene objects by computing the average signal over the gene region. Similarly, various attributes (such as orientation, expression level or fold change, protein occupancy) can be added to Gene objects based on imported Genome, Transcriptome, or ChIP-Seq attributes (Supplementary Fig. S1, attributes written in italics). An orientation can be defined for a gene or intergenic position, depending on the direction of neighboring genes (tandem, divergent, convergent). Before performing statistical analyses, quantitative data can be classified into qualitative categories, based on user-defined thresholds or on quantiles (of equal or custom sizes).

Computational resources

GRATIOSA is designed to run on a standard desktop/laptop computer, coming after the most computationally intensive steps in sequencing data analysis (especially read mapping). Usual operations take less than a few seconds with <1 Gb memory load. Most data types (annotation, gene-wise expression values) have a low (a few Mb) memory load. Data types relying on genomic coverage (RNA-Seq, ChIP-Seq) are usually larger but, because of the small size of bacterial genomes, are also easily tractable (usually <100 Mb; optionally, input tabular files can be automatically converted into binary NumPy array files that are lighter and faster to load).

Statistical analysis and outputs

The package offers three standard statistical tests (Supplementary Table S2): an enrichment test (hypergeometric, with false discovery rate correction), particularly useful for functional enrichment analysis; a proportion test (based on the z-score, equivalent to a chi-square test), e.g. for comparing the proportion of differentially expressed genes in different conditions or subsets of the genome; and two standard tests for quantitative data (Student’s t-test and Wilcoxon–Mann–Whitney nonparametric rank sum test), e.g. to compare the average expression level or fold change among different sets of genes. These tests are implemented using common Python modules, and their detailed results are saved in a csv file. Correlation tests and many other statistical tests can be easily carried out with standard NumPy/SciPy functions.

Graphical functions

Statistical results can be visualized as bar plots generated using the matplotlib.pyplot.bar module, with error bars representing 95% confidence intervals, and optionally, an annotation of these bar plots with brackets and stars according to their significance level (* for P-value <.05, ** for P-value <.01, and *** for P-value <.001). Other graphical functions are implemented in the package to visualize signals, such as ChIP-Seq and RNA-Seq coverage within their genomic context (Fig. 1). RNA-Seq coverage corresponds to the number of reads for each genomic position, on each strand. Depending on the size of the visualization window (from a few genes to the entire genomes), genes on each strand can be displayed/annotated.

Figure 1.

Graphical overview of the GRATIOSA package. The left section (database) presents the different types of data processed by the package. The right section shows the main steps and outputs. For a more detailed description, see Supplementary Fig. S3.

Processing of E.coli topoisomerase and expression signals

For most analyses shown in this article, we relied on the E. coli W3110 topoisomerase signals previously published by Sutormin et al., together with FPKM values of annotated TUs [5, 16]. The mapping of gyrase cleavage sites was obtained using a method called Topo-Seq; we used the datasets obtained with ciprofloxacin as gyrase poison (GEO: GSE117186) [5]. Topoisomerase I binding sites were identified with ChIP-Seq (GEO: GSE181915) [16], with or without rifampicin treatment (an inhibitor of transcription). For each sample, the coverage was first normalized by its mean, and the log₂ ratios between immunoprecipitation (IP) and control (mock) coverage were computed. The log₂ ratios were then normalized (by their mean) and the resulting values were saved in bedgraph format compatible with GRATIOSA. The replicates of topoI and gyrase signals were averaged for each genomic position (using the Chipseq.load_signals_average method), representing the log₂ occupancy by each protein relative to the control data. The signals were then separated into macroscale and microscale contributions (see the “Results” section): the macroscale contribution was obtained by smoothing each signal along 200-kb sliding windows (with useful_functions_Chipseq.smoothing); the microscale contribution was obtained by subtracting the macroscale contribution from the raw signal. For expression analyses, we generated approximate expression signals (per genomic position) from FPKM expression values provided [16]. Pearson correlation coefficients for signal autocorrelation were calculated using the scipy.stats.pearsonr function by comparing each signal with the same signal shifted by genomic position.

Results

The GRATIOSA package

GRATIOSA is written in Python and runs on a standard desktop or laptop computer under a Unix environment (Linux preferred, MacOS supported, Windows not fully supported; see the “Materials and methods” section for details). It facilitates the integration and comparison of bacterial genomic data types using standardized statistical analyses, and is particularly designed to quantitatively investigate the spatial relationship between transcriptional regulation and chromosome organization. It can handle any bacterial species described by a reference sequence and genomic annotation (including multiple chromosomes or plasmids). Additionally, annotations of TUs and GO can be imported if available.

Two basic types of experimental datasets can be analyzed with the package: expression data (transcriptomics) and structural data (ChIP-Seq or related technologies), and both can be represented either in a continuous (RNA-Seq or ChIP-Seq coverage) or in a discrete manner (gene expression levels, positions of ChIP-Seq peaks), following preprocessing with standard software. A summary table of usual data types accepted by GRATIOSA is provided in Table 1, and a comparison to existing software solutions in Supplementary Table S1.

An overview of the package is shown in Fig. 1, and more detailed technical information is given in the “Materials and methods” section. The simplest method to import data with GRATIOSA is to organize the data files (in usual formats) following a predefined structure, classified by organism and data type (Supplementary Fig. S2; the software can generate it automatically for a new species, based on a NCBI accession number). Data files can be accompanied by an information file enabling automatic selection and import (e.g. a subset of experimental conditions for a particular analysis of expression data). This basic structure can be used to import many other and/or custom types of data, e.g. predicted protein binding (discrete) sites or (continuous) signals (from sequence motifs), miscellaneous annotated features, positions of Hi-C domain borders, etc.

The GRATIOSA package is typically accessed through a Jupyter notebook, and data analysis is executed as Python commands, in three major steps (Fig. 1). Examples are provided as tutorials and will be described in the next section.

First, different objects (Genome, Transcriptome, ChIP-Seq, etc.; see Supplementary Fig. S3, gray) are initialized, depending on the user’s needs. The required data are then loaded as object attributes (Supplementary Figs S1 and S3, yellow) and processed by the software (Fig. 1).

Once imported, the user can manipulate these data easily (mainly as NumPy arrays or Pandas dataframes), either manually (for custom computations) or using the tools implemented in the package (described in Supplementary Fig. S3, green). For example, attributes associated with genomic positions (“signals”) can be mapped onto genes (e.g. the average ChIP-Seq coverage is computed for each gene). Signals can be smoothed. Genes can be classified according to their orientation (with respect to neighbors) or length. Thanks to the easily manipulable formats, the user can perform additional custom processing to modify or add attributes, and manually import external data. A graphical exploration of the signals can be performed at the end of the first or second step using a graphical function that automatically plots signals with the annotated genome. Finally, to prepare for statistical analysis, quantitative data (expression level, protein binding strength, etc.) can also be classified using thresholds or class sizes, e.g. in expression quantiles.

The last step is the statistical analysis (Fig. 1 and Supplementary Fig. S3, blue) with enrichment or proportion tests (for qualitative attributes) or Student’s t-tests (or a corresponding nonparametric test, for quantitative comparisons). Results of all tests are exported as tables (in csv format) and can be visualized as annotated bar plots created with graphical functions included in the package (Supplementary Fig. S3, purple).

Tutorials: a few typical analyses

Before executing any tutorial, the user must first define two system variables, indicating the location of (i) the GRATIOSA software and (ii) the database. For testing purposes, the user can download the package from GitHub (https://github.com/sammeyer2017/GRATIOSA) and follow instructions to work in the local directory. For long-term use, we recommend installing the code in permanent directories through system tools (e.g. PyPI) and defining system environment variables.

Simple RNA-Seq analysis

The tutorials Ecoli_RNASeq and Dickeya_RNASeq show similar basic analyses of RNA-Seq data, based on different data sources. For E. coli (annotation taken from NCBI), the tutorial code imports a processed RNA-Seq coverage data file (from NCBI GEO) and an RNA-Seq differential expression file (obtained from [30]). For D. dadantii, the annotation file was manually created; the RNA-Seq coverage is automatically computed by the software from BAM mapped read files in two experimental conditions, and differential expression is imported after external processing with DESeq2. In both species, the coverage is plotted along a chosen window (similar to a genome browser) and a few lines of code are sufficient to compute and plot a statistical relation between gene orientations and their transcriptional response.

Semiautomatic creation of a database for a complex genome

The tutorial Sinorhizobium_complex_genome describes the automatic creation of the formatted directories and import of sequence/annotation files for Sinorhizobium meliloti (from NCBI accession GCA_000006965.1, containing one circular chromosome and two megabase-long plasmids), and the subsequent import of gene expression values in a wild-type as well as mutant strains (from a processed tabular file obtained in [31]). Statistical comparisons of absolute and differential expression levels between chromosomes/plasmids, and between different gene orientations in each chromosome/plasmid, are carried in a few lines of code.

Visualization and analysis of ChIP-Seq data

The tutorial ChIP-Seq_plot shows the import of ChIP-Seq coverage datasets in E. coli, averaging of replicates, and visualization over a genomic region. The tutorial ChIP-Seq_orientation shows how this average signal can be statistically compared between different regions of the chromosome, in particular here between convergent and divergent genes.

Analysis of processed Hi-C datasets

The tutorial Ecoli_HiC_borders shows how the package can be used to analyze a broader range of data: in this case, to characterize the regions located at the borders between chromatin interaction domains obtained from a Hi-C map in E. coli. The borders are loaded from a tabular file generated by Chromosight [28], as well as a gene expression dataset. In a few lines of code, we find a statistical enrichment in highly expressed genes at the borders, as observed before [32, 33], associated with several GO terms. We also find a significant enrichment in convergent genes, which may be surprising since these genes are usually weakly expressed. This new result can be related to previous observations relating Hi-C borders with regions of positive supercoils (see the “Discussion” section and [34]).

Detailed analysis of topoisomerase binding in E. coli

Finally, we illustrate a more complex use of the software in the following case study of topoisomerase binding in E. coli, for which GRATIOSA allowed a drastic reduction in code writing and an improvement in reproducibility. The codes used for all upcoming figures are given in a Jupyter notebook deposited in GitHub (article case study).

Characterization of E.coli twin-supercoiled domains in vivo

DNA SC, i.e. torsion-induced deformation of the DNA double helix, besides its essential role in bacterial chromosome compaction, is recognized as a global transcriptional regulator [35–37]. The superhelical level is controlled globally by the action of topoisomerases, mostly topoisomerase I (topoI) and gyrase [38], but is also strongly affected by local transcription, according to the “twin-supercoiled domain” [39], leading to the accumulation of negative (resp. positive) supercoils behind (resp. ahead of) the elongation complex. However, until recently, there was limited information on twin domains in the chromosome in vivo.

Since topoI preferentially binds to negatively supercoiled regions (where DNA forms transient denaturation bubbles), and DNA gyrase to positive supercoiled regions (forming plectonemes), mapping the activity of these two enzymes provides spatially resolved information on the supercoiling distribution. Recently, Sutormin et al. used ChIP-Seq of topoisomerase I and a ChIP-Seq derived assay mapping cleavage intermediates of gyrase [5, 16], to analyze the genome-wide binding (resp. cleavage) distributions of these two enzymes in E. coli W3110, thus providing invaluable insights into the presence, amplitude, and extension of twin-supercoiled domains in vivo. These authors found enrichments of topoI upstream and of gyrase downstream of the 200 most highly expressed transcription units (HETUs) out of ∼2100 annotated TUs in E. coli [16], in agreement with the twin-supercoiled model.

We used GRATIOSA to characterize these twin domains more systematically and quantitatively. Is this behavior specific to that small minority of highly expressed genes, meaning that other genes do not locally recruit topoisomerases? How many genes exhibit a local enrichment in topoisomerase recruitment? To what distance does it extend? Can it be mapped to the formation of topological domains?

We started from the raw data and computed coverage distributions for the different samples using standard methods (see the “Materials and methods” section). All data import and most classification steps could then be achieved with few lines of code with GRATIOSA, allowing the user to focus on the specific analyses and plots required to answer the questions above.

Distribution of supercoils around transcription units

We first wished to know how many genes were associated with detectable twin domains. We classified all TUs by expression level and analyzed the average gyrase and topoI coverages (using log₂-normalized ratios of IP signals versus mock, with an average level of 0), from 10 kb before TU start to 10 kb after TU end (Fig. 2).

Figure 2.

TopoI binding (top) and gyrase cleavage (bottom) signals around TUs in E. coli (ChIP-Seq data from [5, 16]). Shown signals are normalized log₂ ratios between IP and control coverage. TUs were classified based on their expression level, with 200 TUs in each class of HETUs (around one decile each: blue, cyan, green, orange), one class with all other protein-coding TUs (red), and one class with strongly expressed TUs encoding stable RNAs (ribosomal or transfer RNAs; black). By construction, both signals have a genomic average of 0. The bottom right inset shows the average expression level of each expression class, except for stable RNAs (strongest expression but FPKM do not reflect their expression level). TU_s and TU_e indicate the start and end of TUs.

The topoI binding signal (Fig. 2, top) exhibits a clear gradual increase over 10 kb upstream of the TU start, with a pronounced peak at the TU start. Following the TU start, the signal quickly decreases, and drops close to zero by the end of the TU (except for stable RNAs in black). Within the TU, higher expression levels are associated with higher topoI binding density, consistent with the known colocalization and interactions between topoI and RNAP [16, 40, 41]. Within stable RNA (transfer RNA and ribosomal RNA) operons (black curve), topoI binding appears relatively flat, potentially indicating an ongoing topoI recruitment throughout the TU, in contrast to all other categories (even strongly expressed). This observation suggests a specific extensive requirement for topoI for those operons. Upstream of the transcription start site, the density of topoisomerase I binding is also markedly dependent on TU expression levels. This enrichment is likely not due to the interaction of topoI with RNA polymerase, but reflects the negative supercoils generated during elongation, in clear agreement with the twin-supercoiled domain model [39] and as already identified for the 200 HETUs [16]. However, Fig. 2 shows that this phenomenon affects not only this small minority of genes (around 10%), but many more genes, classified into groups of genes of similar size. It extends remarkably far: up to 10 kb upstream of the TU for the top 400 most HETUs, up to 8 kb for the next 200 TUs, and up to 6 kb for the following 200 TUs. Thus, while previous conflicting values were reported before for the extension of the domains (see the “Discussion” section), our analysis shows that this extension (as apparent from ChIP-Seq data) is dependent on the TU expression level. In the 1300 (∼60%) weakly expressed TUs (red), no enrichment is observed. After the end of the TU (downstream), the topoI binding is slightly weaker than the genomic average (negative levels for all expression classes), as would be expected from the twin-domain model if positive supercoils inhibit topoI binding.

The gyrase signal (Fig. 2, bottom) exhibits significant differences compared to that of topoI. As expected again, the 10 kb upstream of TUs are less cleaved by gyrase than the genomic average, possibly due to the negative supercoils generated by transcription, with a particularly strong depletion at the beginning of the TU. Within the TU, we observe no enrichment in gyrase binding, suggesting that the enzymes are not required in the immediate vicinity of RNAP (in contrast to topoI). After the TU, the signal increases sharply, then slowly decreases but remains strongly enriched after 10 kb behind the TU in all highly expressed classes, and thus extends further and more uniformly than the topoI enrichment before the TU. This enrichment is clearly increasing with expression strength, whereas weakly expressed TUs exhibit no enrichment. The apparent periodicity of the stable ribosomal and transfer RNAs (rtRNA) signal likely arises from the low number of these TUs, coupled to the approximate regularity in the size of most TUs.

We wish to emphasize that, because of the high density of bacterial genomes, the 10-kb regions flanking the central TU contain many other genes (one TU every 2 kb in average, with short intergenic regions), which also contribute to the observed topoisomerase binding patterns. The obvious regularity of the observed patterns suggests a strong impact of the central one under investigation (whereas the neighboring TUs differ by their positions, expression strengths, and orientations), but it does not constitute direct evidence that supercoils actually propagate over >10 kb from the center, especially if the neighbors’ expression is correlated to the central TU’s and co-directionally oriented. Further correlation analyses are developed in the upcoming paragraphs, and their interpretation is addressed in the “Discussion” section.

Role of gene expression strength and orientation in shaping local supercoiling distributions

Since Fig. 2 clearly suggests an effect of the gene expression level on topoisomerase recruitment, we systematically investigate this correlation by examining the total binding signals around all TUs separated in expression deciles, in the 10 kb upstream of the TUs (Fig. 3, left panel), within the TUs (Fig. 3, middle panel, expression levels are indicated in gray), and in the 10 kb downstream (Fig. 3, right panel).

Figure 3.

Average topoI binding (gold) and gyrase cleavage (black) signals in the 10 kb before (left), inside (middle), and in the 10 kb after (right) TUs, depending on the expression level. TUs are classified in expression deciles, and the average expression level (FPKM) of each decile is shown in gray in the middle panel; operons of stable RNAs are classified separately. Mind the breaks in the vertical axes. Error bars indicate 95% confidence intervals.

Within TUs, the average density of topoI (Fig. 3, middle panel, gold curve) exhibits a moderate increase in the first deciles of TU expression levels, which becomes much more pronounced in the last four deciles (i.e. the 40% most HETUs). In the 10 kb region upstream of the TUs (Fig. 3, left panel), the expression level dependence is very low in the first six deciles, whereas topoI binding steadily increases in the last four deciles. After the TUs (Fig. 3, right panel), as observed in Fig. 2, the topoI binding signal is weakly negative for all expression levels, suggesting a weak accumulation of positive supercoils. As for gyrase (Fig. 3, black curves), its recruitment before and within TUs is relatively constant, even in the most expressed ones. Before the TUs, the signal is slightly negative, compatible with the presence of negative supercoils repelling gyrase. After the TUs, a gradual enrichment of gyrase cleavage with higher expression is observed, especially once again in the last four deciles, and quite limited before. Note that all observed enrichments vanish upon transcription inhibition by rifampicin (Supplementary Fig. S5).

These observations confirm (i) the in vivo relevance of the twin-supercoiled domain model, not only for a small minority of highly expressed genes as observed before [6, 7, 16], but for a large fraction of the genome; (ii) topological constraints affect precisely the 40% most HETUs, with a clear and steady increase in the recruitment of topoisomerases depending on the expression level. The 60% remaining weaker TUs are not affected in a detectable way (with the sensitivity of ChIP-Seq), presumably because they generate moderate mechanical distortions that might be (at least partly) dissipated without local recruitment of topoisomerases; and (iii) within TUs, topoI is recruited in the wake of the elongating RNA polymerase [16], especially close to the promoter, but gyrase is not enriched.

Previous studies have shown that, because of the asymmetry of transcription-induced supercoils, local gene orientation is important for the transcription–SC coupling [15, 16, 42–45]. The package has built-in functions to classify intergenic regions depending on the orientation of their neighbors (divergent, convergent, tandem), and quantitatively compare how they recruit topoisomerases. We observed (Fig. 4) an enrichment of topoI and a depletion of gyrase between divergent genes compared to convergent genes, as already noted by Sutormin etal., following expectations of the twin-supercoiled domain model [16]. However, this difference could arise not only from transcription-induced supercoils, but also from static effects such as strong differences in base sequence (divergent regions are more enriched in AT-rich promoters; Supplementary Fig. S4). Indeed, the difference was still present, to a lesser extent, after treatment by the transcription inhibitor rifampicin (Fig. 4B and E). By normalizing the untreated coverage (Fig. 4A and D) with the coverage after rifampicin treatment (Fig. 4B and E), we erase out any static effect and obtain the direct contribution of adjacent transcription. As expected, we observe a differential and opposite recruitment of topoisomerases, specifically induced by transcriptional activity depending on the local gene orientation (Fig. 4C and F). However, the effect of transcription is not the same on the two enzymes. For topoI, the effect is symmetrical, with divergent regions exhibiting enrichment and convergent regions exhibiting depletion, while tandem regions are close to 0 (i.e. the genomic average). In contrast, for gyrase, transcription induces no specific enrichment in convergent regions (close to genomic average), whereas tandem and divergent regions are depleted. Remarkably, this difference suggests that transcription of convergent genes does not lead to a significant buildup of positive supercoils and subsequent recruitment of gyrase between them, whereas the negative supercoils upstream of TUs efficiently reduce its binding in tandem and especially divergent regions.

Figure 4.

TopoI binding site (gold, top row) and gyrase cleavage site (black, bottom row) densities in intergenic regions, depending on the orientation of surrounding genes. Each bar represents the average normalized log₂ ratios between IP and mock for a specific orientation, computed after binning intergenic positions into 100-bp bins. Curves in panels (A) and (D) were obtained without treatment; in panels (B) and (E) after treatment with rifampicin, a transcription inhibitor; and in panels (C) and (F) show the difference between untreated and rifampicin-treated samples. Error bars are 95% confidence intervals.

Asymmetrical extension of twin domains

Figure 2 showed that topoisomerase recruitment was affected at least up to 10 kb before and after TUs, close to the most commonly accepted size for topological domains in E. coli [46]. We now look at the extension of topoisomerase binding correlations, possibly reflecting these domains. The binding autocorrelation curves (computed along the whole genome) are shown in Fig. 5, and reflect characteristic extensions of identified binding regions for both enzymes. The correlation is strong below 1 kb (the typical gene length), for both topoisomerases and for gene expression patterns, as expected if short-scale binding variations are directly induced by transcription. This is directly confirmed by the sharp drop in correlation of both binding profiles in the absence of transcription (thin dashed lines), at distances reflecting typical lengths of ChIP-Seq fragments (a few hundred bases). With active transcription, the correlation reduces with distance, but in a very different manner for the two enzymes: the characteristic length scale of the topoI binding distribution is 10 kb, above which it is entirely decorrelated; in contrast, at that scale, the correlation in gyrase binding is still quite strong, and vanishes only a full order of magnitude above, after 100 kb. These two separate length scales were already described in earlier works, and correspond to two different levels of genome organization: (i) the 10-kb (hereafter quoted “micro”) scale is that of TUs and their immediate neighboring interactions, possibly within transient topological domains [46], leading to strong and evolutionarily conserved co-expression [14] and (ii) the larger 100–200-kb scale (“macro”), already observed in transcription and gyrase binding signals in E. coli [13] as well as in 3D conformation maps [1, 32]. The curve of gene expression (gray) confirms this separation, with the correlation mostly dropping at the microscale, but a small (∼10%) contribution still observed up to the macroscale. The latter contribution shows that the global architecture of the chromosome (involving NAPs and other factors) involves large-scale patterns of gyrase binding and gene expression. Accordingly, computing the cross-correlation of topoisomerase binding with gene expression (Supplementary Fig. S7) exhibits maxima at length scales of a few dozen kb (topoI) versus 200 kb (gyrase), with the latter being still present but weaker after rifampicin treatment.

Figure 5.

Pearson’s autocorrelation curves of gyrase cleavage (in black), topoI binding (gold), or expression signal (gray) (as a function of distance). These curves reflect the typical distances where each signal remains correlated along the chromosome. The corresponding curves after rifampicin treatment are shown as thin dashed lines for the two topoisomerases. The vertical lines indicate the distinct characteristic length scales for the two enzymes.

We wished to quantify the distance of propagation of transcription-induced supercoils. Figure 6 shows the cumulative topoisomerase binding (area under the binding profile) between the TU start/end and a remote position up to a distance of 50 kb: a flat curve thus indicates a local binding intensity similar to the genomic average (and thus presumably no influence of the TUs under consideration), whereas an increase (resp. decrease) indicates a stronger (resp. weaker) binding, possibly due to transcription.

Figure 6.

Extension of the transcriptional-induced topological constraints. Top and bottom panels show cumulative signals (i.e. total binding starting from the TU) for topoI and gyrase, respectively, in the 50 kb before TU start (left panels) and after TU end (right panels). TUs were classified based on their expression level (same color legend as in Fig. 2). The plotted lines show the average area under the curve of the microscale signal (i.e. after subtracting the macroscale contribution, especially for gyrase, up to 50 kb). All signals vanish upon rifampicin treatment (Supplementary Fig. S8).

As already observed, topoI binding increases with the expression level in the first 10 kb before the TU, and the extension of the transcription-induced negative supercoils also tends to increase (Fig. 6A). For stable RNAs (black curves), the signal reaches 25–30 kb, consistent with the size identified around the ribosomal operons by psoralen binding [6]. After the TU (Fig. 6B), the depletion in topoI binding (negative values) mostly occurs in the first 10–20 kb, and slightly increases with the expression level. For gyrase, no specific signal is detected upstream of the TU start in panel (C), except for stable RNAs, where the enzyme appears to be repelled over at least 50 kb (possibly reflecting strong negative supercoils accumulating there, but also possibly the low number and proximity of these TUs). Downstream of the TU (Fig. 6D), a strong enrichment in gyrase cleavage is observed, which increases with the expression level, but in contrast to topoI, it extends way beyond the 10-kb region, from around 20 kb for moderately expressed TUs (orange) to around 40 kb for strongly expressed ones (dark blue) and even beyond for stable RNAs. Thus, positive supercoils induced by transcription may extend as far as several dozen kilobases. All observed features (for both enzymes) vanish upon rifampicin treatment (Supplementary Fig. S8), confirming that they result from transcriptional activity. We also checked that neighboring TUs exhibit similar expression levels regardless of the class of the central one (Supplementary Fig. S9), suggesting that the latter likely is the main driver of the far-reaching supercoils affecting gyrase recruitment, as developed further in the “Discussion” section.

Discussion

A software solution for integrative spatial analysis of genomic data

The interplay between chromosome conformation and gene expression is a complex and intricate problem, requiring dedicated conceptual and technical tools. While classical regulatory networks are usually analyzed in a qualitative way (does a regulatory protein bind at a promoter?), the “analog” layer of transcriptional regulation can be only understood through quantitative analysis (does it bind more or less?), considering spatial (1D or 3D) relations between genes, and combining different actors (architectural proteins, SC, expression, etc.). GRATIOSA facilitates such analyses, by automating the formatting and import of various data into a uniform framework, providing a range of standard tests and visualization tools, and facilitating quantitative analyses of different data types with various Python tools and libraries. This is illustrated by several tutorials provided with the software, and more thoroughly in the example of topoisomerase distributions, where the basic steps of the analysis are standardized, and the programming effort can be dedicated to the design of specific statistical tests and graphs (see notebooks in Supplementary data). While the original analyses of these data [5, 16] were restricted to a small minority of highly expressed genes exhibiting the strongest effects, GRATIOSA provided the ideal framework to extend these analyses and characterize for the first time the twin domains in vivo in a systematic manner.

The present version of the code focuses on the spatial organization along the 1D genomic sequence (linear), based on well-defined gene positions, which are usually disregarded in regulatory analyses. A natural extension of the software is the inclusion of the 3D organization of the chromosome through Hi-C data, with substantial additional issues to consider: (i) the relation between Hi-C contact densities and 3D conformation (both static and dynamic) is nontrivial and relies on complex modeling and (ii) the extension of the package to automated 3D analyses will involve significantly more complex functions. The present version of the software already offers limited solutions for Hi-C data, through a prior analysis by dedicated software [28] providing lists of genomic positions with specific identified features (e.g. borders between chromatin interaction domains, or loops), which can then be imported by GRATIOSA and treated as unidimensional features, as shown in a tutorial.

A quantitative description of twin-supercoiled domains in the chromosome

We used GRATIOSA to quantitatively analyze the distribution of twin-supercoiled domains in vivo. This phenomenon has been postulated [39] and validated numerous times in the cell, either with plasmid-borne [44, 47, 48] or with chromosome-borne expression [33, 49], in native or topoisomerase mutant strains. However, these results almost always involved artificial constructions with strong inducible promoters, and the observed effect was strongly affected by co-transcriptional attachment to the cell membrane [49]. In contrast, there is only limited evidence of the extension and physiological relevance of the phenomenon for native genes. The mapping of negative supercoils by psoralen showed an asymmetrical enrichment/depletion of 25 kb length specifically around strongly expressed ribosomal operons [6], but no systematic assessment of the entire genome, possibly due to a strong noise level. Similarly, the mapping of positive supercoils through GapR ChIP-Seq [7] showed an accumulation downstream of the 10% most expressed genes over at least 4 kb. TopoI and gyrase enrichment were previously reported also for 10% highly expressed genes, with extensions of 12–15 kb [5, 16]. Whether most bacterial genes experience topological constraints and require the local recruitment of topoisomerases to be efficiently transcribed thus has remained unclear. Regarding spatial extension, while it was shown that transcription-induced supercoils can reach several kilobases [49, 50], there is as yet no upper bound, nor a clear identification of the nature of barriers. The most accepted value for the size of topological domains is 10–20 kb, obtained by analyzing the transcriptional response of 300 supercoiling-sensitive genes upon local relaxation by double-stranded breaks coupled to high-resolution microscopy [46]. Yet these domains are not necessarily identical to twin-supercoil domains, for instance if they contain several transcription units in different orientations. The maps of topoisomerase binding/activity analyzed here thus provided a new invaluable source of information to characterize transcription-induced supercoil distributions along the whole genome, and generate new quantitative results.

While a statistical relation between gene orientation and topoisomerase recruitment has been shown before [16, 42], this analysis allowed to decouple the specific effect of transcription versus static effects (e.g. strong differences in A/T% shown in Supplementary Fig. S6) and provides additional results (Fig. 4C–F). TopoI is not required (at the average expression level) at tandem genes, but is strongly attracted by supercoils induced by divergent promoters. This observation entirely matches early results in topA mutants of Salmonella enterica [48]. In contrast, while gyrase is repelled from divergent regions by negative supercoils, it is (surprisingly) not specifically attracted by the expression of convergent genes (Fig. 4F). At least in the analyzed experimental conditions, transcription is thus likely not hampered by accumulating positive supercoils downstream of most genes, even convergent ones. This observation is consistent with two recent unexpected results: the expression of a single gene is dependent on the distance to a topological barrier on the upstream but not the downstream side [51], and gyrase inhibition does not usually inhibit but rather activate convergent genes [15]. Together, these coherent results suggest that RNAP stalling (because of positive supercoils) is likely not an extensive phenomenon in vivo, but may rather be limited to a few highly expressed genes or specific loci.

Twin-supercoil domains (as reported by enrichment of topoisomerase binding/activity) are detected specifically in the 40% highly expressed fraction of the genome. Besides, they exhibit a clear and increasing magnitude with the expression strength (Fig. 3). For topoI, the propagation distance equally depends on the strength, ranging from 6 to 10 kb, whereas for gyrase, the distance reaches around 30 kb (Fig. 6). For the 60% remaining genes, the absence of detectable signal suggests that the weak transcription level can be sustained without local recruitment of topoisomerases (at the level of sensitivity of the ChIP-Seq technique). This asymmetry in the distance of propagation of negative versus positive supercoils is somewhat surprising, as well as the large value for the latter compared to the usually reported size of topological domains (10–20 kb) obtained from very different data [46]. These observations may be related to the different physical effects of negative versus positive supercoils: negative supercoils mostly induce local structural transitions that strongly affect DNA functions (DNA denaturation, which in turn attracts topoI, but also B-Z transitions, cruciforms, etc.) as well as detrimental R-loops [52, 53], whereas positive supercoils mostly turn into writhing, likely less impacting transcription and other processes as noted above. Thus, one explanation is that gyrase does not remove all positive supercoils, but leaves a fraction of them diffusing further from most transcription units without major harm. The associated level of positive supercoiling might be considerably milder than the relaxation induced by double-strand breaks used to estimate the size of topological domains [46].

While the diffusing distances reported here were obtained by averaging hundreds of transcription units, we further looked whether they could also be observed at the level of individual ones. We plotted the distribution of topoI/gyrase enrichment, ordered by transcription level or gyrase enrichment (Supplementary Fig. S10A and B). The graphs do not exhibit striking systematic features, but rather a remarkable heterogeneity of behaviors. In particular, topoI enrichment is not always correlated with gyrase enrichment, showing the influence of factors beyond local transcription, as further discussed below. This heterogeneity is consistent with the diversity of behaviors reported in the literature relating supercoiling with transcription, and highlights the difficulty to identify systematic regulatory patterns based on individual genes, justifying the statistical approach developed here.

The spatial analysis exhibited a different chromosomal organization of topoI and gyrase recruitment. At a characteristic length scale of ∼10 kb, the recruitment of both enzymes is dominated by local transcription units, with an asymmetrical organization in downstream versus upstream regions (Fig. 6). However, gyrase also presents a larger-scale organizational pattern, with a characteristic length scale of ∼200 kb, related to the global organization of the chromosome. The shorter spatial scale is remarkably close to that previously proposed for topological domains. The latter scale was already identified in early analyses of transcription and gyrase binding patterns [13], as well as 3D Hi-C maps (“chromatin interaction domains”) [1, 32, 54]. The underlying mechanisms remain ill-defined, but presumably implicate a complex interplay of architectural proteins, transcription factors, and replication-related factors [1]. Accordingly, while transcription inhibition entirely abolishes the distribution of topoI, the large-scale distribution of gyrase is partly retained (Supplementary Fig. S7). Interestingly, a recent study also exhibited a statistical relation between Hi-C contact density and GapR binding (indicative of positive supercoils), but not topoI (indicative of negative supercoils) [17]. Since GapR and gyrase are expected to bind similar positively supercoiled regions, that result is fully consistent with our observation of a predominant relation of gyrase binding with large-scale chromosome structure.

Supplementary data

Supplementary data is available at NAR online.

Conflict of interest

None declared.

Funding

Funding was provided by the Agence Nationale de la Recherche grants ANR-18-CE45-0006-01 and ANR-24-CE95-0003-01. Funding to pay the Open Access publication charges for this article was provided by Agence Nationale de la Recherche grant ANR-24-CE95-0003-01.

Data availability

GRATIOSA is an open-access Python package. It can be automatically installed via PyPI (typically with the command pip install GRATIOSA), and is accessible on GitHub with tutorials, scripts, and datasets required to reproduce our analyses: https://github.com/sammeyer2017/GRATIOSA. The code has also been deposited in Zenodo: https://doi.org/10.5281/zenodo.15223250. Documentation is available at https://gratiosa.readthedocs.io/en/latest/Presentation.html.