Home - Explore - Tools - Help

User: Pass:

You are not logged in

UMR 7238 CNRS
Universite Pierre et Marie Curie

NuST@Genomique des Microorganismes

Learn by example: interpreting the results and choosing the parameters.

Contents

1 Introduction

2 Learn by example: linear aggregation analysis

). The statistically significant gene clusters obtained using a number of bins between a minimal (left sliding-bar) and maximal (right sliding-bar) value can be drawn on the circular chromosome.
As an illustrative case, a global bias of sigma-factor target positions on the chromosome has been recently shown (Sobetzko et al PNAS 109, E42-50 2012): the vegetative sigma factor 70 has a higher target density near the origin of replication, while the stationary-phase sigma factor 38 presents an opposite bias. This large-scale trend can be shown performing the linear aggregation analysis on the two common data sets "Genes regulated by the sigma factor 70 (from Regulon DB)" and "Genes regulated by the sigma factor 38 (from Regulon DB)" present in the internal database, using an appropriate (small) number of bins (from 4 to 8 in this example):

Since this is a large-scale bias, it becomes less clearly visible using a higher number of bins. In other cases, increasing the number of bins can highlight additional useful information. As discussed in (Scolari et al. Molecular BioSystems 7, 878-888 2011), clusters of genes whose expression is affected by deletion of either Fis or H-NS and upon changes in negative supercoiling, identified in different experiments, appear to overlap well with the Ter macrodomain at large scale, while the analysis at smaller scales shows a preferential localization towards the edges of the macrodomain. This is the type of results that can be extracted making full use of the multi-scale analysis offered by the linear aggregation tool. For example, performing the linear aggregation analysis on the common dataset "H-NS knockout sensitive genes under negative-supercoiling (from Blot et al 2006)" and representing the significative clusters from a large scale (4 bins) to a smaller scale (128 bins) shows both the statistical clustering in the Ter macrodomain and the preference for the Ter macrodomain edges:    TOP 3 Learn by example: multiple sliding window histograms Sliding window histograms allow a graphical visualization of gene densities along the genome. The resulting representation depends on the observation scale selected by the user by changing the number of bins with the sliding bar in the output page (described in detail here ). While one can be guided by the linear aggregation clusters found for a gene set, there is no general rule for choosing an appropriate observation scale for studying and comparing gene density profiles from different input lists. The selected observation scale should reflect the specific biological question of the user and the results need to be interpreted accordingly. However, as a first qualitative criteria, the number of bins should be sufficiently low to avoid a mere representation of gene and operon positions. In fact, if the plot appears as "spikes" alternated with zero density regions, and the number of genes in the density peaks is just few units or even one (provided the "Normalize?" option is not selected), the plot is just showing single gene positions and their operon organization, as in this example: Decreasing the number of bins can give a more relevant representation of possible gene aggregation at larger scales, for the same data set: A lower bin number makes the size of the sliding window used to evaluate the gene density larger. Hence, the plot typically looks smoother. Similar considerations hold if multiple sliding window histograms are used to compare gene density profiles of different gene lists. A number of bins that is too large could show the mere co-occurrence of the same gene/operon in the two lists. If the user is interested in such small-scale information it could be more appropriate to use the hypergeometric test implemented in NuST, which gives immediately the statistical significance of the intersection between the two lists (the hypergeometric test in NuST is described here ). However, two gene lists can have a null intersection while presenting similar gene density profiles if observed at the right scale. In this case, the multiple sliding window histograms can be a useful tool to explore a range of possible observation scales (changing the number of bins), visually inspecting the results for possible profile similarities. In order to perform a more quantitative of the correlation between two gene density profiles at a certain observation scale, the Pearson correlation analysis described in the following section can be used.    TOP 4 Learn by example: local Pearson correlation The Pearson correlation analysis allows the user to calculate the correlation between gene density profiles along the chromosome of two input gene lists. In particular, the global Pearson correlation (reported on the top-right of the output plot) and its local version (the curve in the output plot) can be evaluated (a description of the output format can be found here). This analysis is quite delicate, as both quantities are heavily influenced by the observation scale as the gene density profiles are dependent on the number of bins. For instance, two lists can show a relatively high global correlation on a large scale, but almost no correlation on smaller scales. This is the case for the two lists "Genes regulated by the sigma factor 70 (from Regulon DB)" and "Genes regulated by the sigma factor 38 (from Regulon DB)" present in the internal database. As recently discussed (Sobetzko et al PNAS 109, E42-50 2012), the target genes of this two sigma factors are preferentially located in a large region proximal to the replication origin (Ori) and the replication terminus (Ter) respectively. Therefore, anticorrelation between their density profiles is expected at large scales. However, since in these broad regions their genes are not positioned in mutually exclusive clusters, almost no correlation is observed on smaller scales. A a result, the global Pearson (given by the top-right legend in the plots) shows anticorrelation using 4 bins, but the signal is lost increasing the bin number: The local Pearson correlation represented in the plots shows the relative contributions of the different chromosomal regions to the value of the global Pearson correlation between the density profiles. Note that its absolute value does not have a statistical meaning on its own and it is not bounded between -1 and 1 (as the global Pearson correlation). However, it is a quantitative way to compare different chromosomal sectors and extrapolate the locally highly (anti)correlated regions where the gene densities concurrently deviate from their mean values, thus considerably contributing to the global correlation or anticorrelation depending on whether the deviations are on the same or opposite direction. As a second example of application of the local Pearson correlation analysis, we consider two sets of putative H-NS targets obtained in different groups using different experimental techniques: "Putative H-NS target genes in stationary phase from ChIP-seq experiments (from Kahramanoglou et al 2010)" and "Putative H-NS target genes from ChIP-chip experiments (from Oshima et al 2006)" in the common database. As expected, the two lists present both a significant intersection (tested using the hypergeometric test described here) and globally correlated gene density profiles: The chromosomal regions with high local Pearson coefficient correspond to genomic sectors where both experiments found coherently a higher-than-average or a lower-than-average density of targets, thus the regions predominantly determining the global positive correlation. Note however that within a region with good correlation, but where the values of the two densities that are compared are both close to their global average, the local Pearson will be smaller than in equally correlated regions that are far from the global averages.    TOP 4 Credits Vittore F Scolari (vittore.scolari_at_upmc.fr) Mina Zarei Matteo Osella Marco Cosentino Lagomarsino Genophysique / Genomic Physics Group UMR 7238 CNRS Universite Pierre et Marie Curie Paris, France Grant RGY-0069/2009-C