GeneBayesNet

To help demonstrate how to use this applet, we provide an archive containing the necessary files to both run an inference and compare the result with a known truth. Download the archive here. This archive contains an example of data sets to run the inference engine, as well as the true network to compare with the output.

Namely:

• Web50-001-expression.txt expression data for 500 individuals for 50 genes named from G01 to G50.
• Web50-001-marker.txt marker data for the same population.
• Web50-001_truth.txt the known gene network, to compare with the results of the inference.

We used artificial networks described in (Mendes et al., Bioinformatics 2003) and simulated genes behavior using non linear ordinary differential equations for each gene as described in (Hoeschele et al., Genetics 2008). We used the COPASI software from copasi.org to define the steady state measurement of each gene expression level for each individual of a RIL population in order to get samples with large genetic diversity and so varying gene-activities between individuals.

First, the expression data need be discretized to use the bayesian inference engine. To handle continuous data, we provide different ways to discretize the data.

If the expression data is already discretized, this step can be skipped by choosing the item "None".

The marker observation data is then appended to the quantified expression data to create the full dataset that the bayesian inference engine will use. The information about cis or trans markers, as well as the distinction between genes and markers is encoded in a set of forced and forbidden edges in the network structure. All this data, along with the defined settings, is finally passed to the inference engine.

The inference engine used in this applet is BaNJO. After running, BaNJO produces a report that contains, amongst other things, a summary of the configuration, the number of network structures explored, and the structure of the most likely structure it found. This report is saved in a text file in the selected output directory, and directly used by the applet to produce an interactive view of the network, as the final step of the run.

1. Fill in the project settings
   • Metadata (optional)
     • Project name short description text, will appear in the report filename, may be empty.
     • Dataset name short description text, will appear in the report filename, may be empty.
     • Notes free text, will appear inside the report header.
   • Input data (mandatory)
     • Gene observations expression data matrix, with variables in columns and variable names in the first row.
       (See provided example file)
     • Marker observations marker data matrix, same size and layout as the expression data file.
       (See provided example file)
     • Output directory where the output files will be created
2. Select the network scoring metric
   • BDEu Bayesian Dirichlet criterion with Markov equivalence and uniform prior on the parameters.
     You may also configure the value of tha α parameter for this scoring method.
   • BIC Bayesian Information Criterion (equivalent to the Minimum Description Length criterion).
   • fNML factorized Normalized Maximum Likelihood (Silander et al., JAR 2010).
3. Tune the search algorithm parameters
   • Uniform in-degree prior (γ) a sparsity prior local to each node. The value 0 corresponds to an uniform prior on the space of DAGs. The value 1 corresponds to an uniform node in-degree prior.
   • Maximum number of parents limit the in-degree of the nodes (high in-degree means much bigger memory consumption).
   • CPU Time limit force the search to stop after a certain amount of time (default 15 minutes).
   • Cache level trade memory usage for CPU consumption (level 2 caches more).
4. Select the discretization strategy for the expression data
   • Gaussian Mixture Model approximates the observed density with a variable number of gaussians.
   • Boxplot discretize into three categories (underexpressed, normal, and overexpressed) based on 0.25 and 0.75 quantile values (q25 and q75). An expression level lower than max(q0, q25 - ((q75 - q25) * χ)) is considered underexpressed, and values higher than min(q100, q75 + ((q75-q25) * χ)) are considered overexpressed.
   • Lognormdiscretize into three categories (underexpressed, normal, and overexpressed) using a threshold value χ in logarithmic (base 2) scale to the ratio between measured expressions and average expression levels. A ratio lower than 2^-χ is considered underexpressed, and values higher than 2^χ are considered overexpressed.
   • Quantiles convert into the specified number of states using quantiles.
   • Intervals convert into the specified number of states using intervals.
   • None use this if your expression data is already discrete.
5. Press the Run button!

The interactive visualization is powered by the JUNG library, using a custom graph layout based on the classic attraction-repulsion scheme. A visualization window will open automatically after a successful inference.

Following our one gene-one marker paradigm, marker nodes in the resulting network are folded into their corresponding gene nodes.

1. Select the graph to be viewed
   • Inferred graph the result of a run (BaNJO report), or a graph in SIF, graphviz, or DReAM format.
   • Original dataset required because most of the formats don't include the names of the variables.
   • (Optionally) Known truth a second graph structure file, corresponding to the known truth to compare the inferred graph with.
2. Press the view button!
3. The visualization window is split in two parts:

On the left, a toolbox containing, from top to bottom:

• The satellite view of the whole graph, in which you can move the view and zoom in and out.
• Buttons to start and stop the dynamic layout.
• The legend of the graph display.
• A summary of some of the graph properties.
• Two sliders to filter out nodes or edges:
  • By influence score: only display edges with a score greater then or equal to the threshold,
  • By degree: only display edges (and their neighbours if corresponding box is checked) with an in+out degree greater than or equal to the threshold.
• A searchbar to quickly find nodes by entering part of their name.

On the right, the graph view, where:

• The layout can be manually edited by dragging nodes with the mouse. Multiple selection is possible by shift-clicking a node.
• The view can be dragged while pressing the left mouse button.
• A right-click on a node or an edge will bring up a contextual menu from which you can display the discretization histogram or a subgraph rooted on a node
4. If you provide a known truth,

then three tabbed views are created on the right-hand side : the infered (or first) network, the true (or second) network, and a "diff" view where the true positives, false positives, and false negatives are shown with distinct colors.

The summary in the toolbox then also includes Precision and Recall measurements (updated when the influence score threshold is modified), and a button appears to display the PRC.