LifeStyle-Specific-Islands (LiSSI): Integrated Bioinformatics Platform for Genomic Island Analysis

2.1 LiSSI Input Data Processing

LiSSI has an intuitive and straight-forward GUI layout using a wizard dialog guiding the user from selecting and retrieving the genome annotation files to setting the analysis parameters. The result plots and tables are organized in a separate tab.

Step 1 – Load genomes. The first step is to load a set of genome annotations of representatives for each of the lifestyles of interest. With LiSSI, we implemented three options: “Select from local folder”, “Download from GenBank” or a combination of both. One may now either select local directories storing files in GenBank format, or select a list of genomes/species for a list of all fully sequenced genomes available at NCBI.

Step 2 – Select genomes. Here, one has to confirm the selection of genomes using some basic information downloaded in the previous step.

Step 3 – Set parameters. The third and final step is displayed in Figure 2. The user is asked to set the parameters 1) for homology detection using BLAST [15] and Transitivity Clustering [16], 2) for island detection using Gecko [17], and 3) for the Random Forest classifier [18], [19]. Note that one may also load intermediate results from previous runs or external analyses here. The LiSSI web site provides detailed tutorials guiding through this process step by step. All preset/standard parameters have been selected based on statistical considerations published in previous work (homology detection, refer to [20] or on our experience with the use case data presented below [island detection and classification]).

Figure 2:

LiSSI software layout. Top-left the selection panel with a tab for Analysis and Results; Top-right the main panel, where all instructions and results will be displayed; and Bottom, the progress panel.

2.2 LiSSI Output Data Analysis

All intermediate results of LiSSI can be inspected individually. We briefly outline them in the following

Homology detection. A corresponding tab summarizes the results found during the homology detection step. It is divided in “Summary” and “Distribution”. In the first, one may find basic information about the homology detection process, such as time required for BLAST and Transitivity Clustering. In the latter, one can find a histogram with the cluster size distribution.

Classification. This tab provides all figures created during the classification process. The “Joint Distribution” depicts the distribution of the genetic features (either homologous genes or islands, according to user selection) among the two lifestyles under investigation. The remaining tabs hold the receiver operating characteristic (ROC) plots depicting the classification performance on three data-sets: 1) the full data set, 2) the data set with a bias towards class “non-pathogenic” (i.e., all features that were mainly found in non-pathogenic organisms), and 3) the data-set with a bias towards class “pathogenic”. Each of the plots displays the classification performance using real labels (dark-blue solid line) and using random labels (light-blue dashed line). One would expect a significant drop in performance on random labels, thus a dashed ROC curve being well below the solid one. If this is not the case, one can learn classifier models on randomly labeled data that is as good as the classifier learned from real-label data; rendering it useless. In addition, the distribution of AUC (Area Under Curve) values of the ROC curve for the distinct runs are represented as box-plots. See Figure 3 for an example.

Figure 3:

LiSSI: Pathogenicity classification performance results. A Receiver Operating Characteristics (ROC) plots shows the performance of the classification models using genes as features to distinguish pathogenic actinobacteria from non-pathogenic ones. The data was evaluated five times using different 5-fold cross-validation sets to assess the robustness of the classifiers. The real label classifier curves are presented as dark-blue solid lines, while the random label classifiers are depicted as light-blue dashed lines (the ones close to the baseline). The variation of the AUCs (area under curve) in the cross-validation was included in the figure as a box-plot (bottom right). The numbers below each box-plot are the lower and upper quartiles.

Feature Selection. Here, LiSSI presents the decision trees generated subsequent to feature selection. Similarly to the previous tab, it shows figures for the full data set and for each of the two bias directions. By clicking on the nodes one may retrieve more information about the respective genetic feature (homologous genes or islands) or run follow up analysis (Pfam or BLAST). See Figure 4 for an example.

Figure 4:

Decision tree created using the most discriminative non-pathogenicity islands. Our classification pipeline (see text) selected the above island as the most representative island for non-pathogens. Nodes containing an identifier represent a genetic feature, in this case, an island. By clicking on the nodes it is possible to visualize the island structure, as well as the gene content and the respective genomes.

3.1 Pathogenicity

Bacteria can be roughly divided according to pathogenicity classes into non-pathogens and pathogens. This subdivision is an oversimplification of what might actually be a continuum, where the bacteria classification depends not only on intrinsic characteristics but on the host and environmental factors [22].

Pathogens have a higher tendency to gene loss which leads to smaller genomes when compared to non-pathogens, which is commonly explained by the metabolic abundance provided by the host. That eventually leads to the loss of metabolic genes that are no longer under selective pressure [23]. Alternatively, the lack of certain metabolic pathways might actually enhance bacterial virulence [24], [25].

In contrast to pathogens, non-pathogens are exposed to a constantly changing environment, where they need to quickly adapt to extreme changes in salinity or sunlight exposure. Furthermore, their survival depends on the ability to metabolize several sources of nutrients [25], [26], [27], [28]. Therefore, it is expected that these organisms possess a large genetic arsenal ensuring their survival in ever changing environmental conditions.

We applied LiSSI to first find homologous gene sets and genomic islands. We observe homologous genes exclusively found in either pathogens or non-pathogens. We also see islands exclusively found in one of the two classes, PA or NP. We found 375,427 distinct homologous genes, where 317,751 were mainly present in non-pathogens and 57,676 in pathogens. The situation is the opposite for islands. Most of the 465 islands are mainly present in pathogens (386); the remaining 79 are mainly present in non-pathogens. Note that there is no island that is present in more than 35 % of the either, non-pathogens or pathogens.

Afterwards, we applied LiSSI to learn random forest classifiers and decision trees on the best extracted features. The results vary heavily depending on whether we use genes or islands as features. When using genes, the classifiers show good performance for both non-pathogen bias (AUC¯ = 94.16 %) and pathogen bias (AUC¯ = 93.91 %), as well as for the classifiers using the full data set (AUC¯ = 94.81 %). It indicates that we can find gene sets specific for pathogens, as well as gene sets specific for non-pathogens. On the other hand, when using islands as features, the scenario is fairly different. The overall classification performance dropped massively: non-pathogen bias performs poorly (AUC¯ = 63.61 %) and pathogen bias as well although not as bad (AUC¯ = 88.84 %). See Supplementary File 2 for more details.

3.2 Oxygen Consumption

The presence of atmospheric oxygen is a limiting factor for bacterial growth; specifically, oxygen levels cannot exceed those found in a bacterium’s native habitat [29]. Above these levels bacteria are subject to decrease in population growth and ultimately death due to the harmful effects of oxidation caused by superoxide and hydrogen peroxide in cellular components [29], [30]. During oxidative stress, lipids are the major target, leading to alterations in membrane fluidity and potentially disrupting membrane-bound proteins. Further, modifications in proteins can lead to conformational changes and consequently loss of function. Another main target of oxidative stress is the DNA, leading to single- or double-strand breaks and, in extreme cases, blocking replication by cross-linking the DNA to other molecules [31], [32]. Regarding oxygen tolerance, bacteria can be roughly divided into aerobes and anaerobes.

Aerobes are defined as organisms that require atmospheric oxygen conditions (roughly 20 %) to achieve optimal growth. The overhead associated with an oxidative environment is compensated by enabling aerobic respiration, a pathway substantially more efficient than fermentation [33]. Aside from the presence of a metabolic pathway that can use oxygen as the final electron acceptor, other features are ubiquitous among these organisms, such as enzymes that degrade peroxide (catalases and peroxidases) [29], [34]. Other metabolic features are also expected to be found to prevent oxidative agents formation, plus, mechanisms to repair oxidative damage and eliminate damaged molecules [30].

Anaerobic organisms are defined as organisms that can tolerate only low amounts of atmospheric oxygen and are not capable of performing cellular respiration. Organisms of this class lack the mechanisms for cellular respiration and to protect the cellular components against oxidative damage [35]. It is not clear which genes might be either exclusive or essential for this class of organism [36].

Comparing the two classes using LiSSI, we found 335,532 distinct homologous genes, where 198,529 were mainly present in aerobes and 28,974 in anaerobes. Similarly, most of the 181 identified islands were mainly present in aerobes (107); the remaining 74 were mainly present in anaerobes. Again, the classification results differ between the gene-based and the island-based procedure. Using homologous gene sets, the classifiers had good performance for both biases: aerobe bias (AUC¯ = 92.48 %) and anaerobe bias (AUC¯ = 99.15 %), as well as for the classifier using the full data set (AUC¯ = 95.15 %). On the other hand, using islands as features, the scenario is different. The classification performance dropped significantly: aerobe bias achieved an AUC¯ of 66.51 %, and the anaerobe bias yielded an AUC¯ of 78.26 %. See Supplementary File 2 4 for more details.