The spaceMap R package constructs de novo networks from heterogeneous data types in a high-dimensional context by applying a novel conditional graphical model. The underlying statistical model is motivated by applications in integrative genomics, where two (or more) -omic data profiles are modeled jointly to discover their interactions. spaceMap is particularly effective in learning networks in diverse applications that exhibit hub topology. In an application to breast cancer tumor data, spaceMap learned regulatory networks which generated novel hypotheses of cancer drivers and confirmed known risk factors. In addition to learning network structure, an accompanying network analysis toolkit is provided. The toolkit has been developed with genomics in mind---but can be adapted for other applications ---and maps scientic domain knowledge onto networks. The image below provides an overview of how to fit spaceMap, perform network analysis in an integrative genomic context, and visualize important network signatures.
The R package spacemap is available from GitHub. Please see the most recent beta release for installation.
We built model selection and model aggregation tools into spaceMap for learning networks robust to the challenges of integrative genomics data due to high-dimensionality and complex noise structure. High throughput -omic experiments often have small sample sizes and the signal-to-noise can be low. Models applied to the resulting data can be prone to over-fitting and high variability. By a procedure called CV.Vote based on K-fold cross validation, one can learn model tuning parameters that balance the trade-off between power and false discovery rate (FDR). CV.Vote reports a network where edges must be present in a majority of networks fit under the training sets. Please see Model Tuning under the Vignettes tab for an example of how to tune models with cross validation. Specific usage details of CV.Vote are documented in the function cvVote.
Making your network robust to over-fitting can be further enhanced through bootstrap aggregation---called Boot.Vote---especially when sample size relative to dimensionality is a real concern. This procedure learns networks on bootstrap replicates of data under CV-selected tuning parameters. Only those edges with majority representation among these networks are reported in the final network. Specific usage details and an example of Boot.Vote are documented in the functions bootEnsemble and bootVote.
Several simulation experiments have been used to evaluate spaceMap in learning networks with prominent hub topology---a commonly encountered feature of biological networks. The GitHub repository sim-spacemap contains code affiliated with these simulations featuring:
- simulation of hub network topology
- data generation according to a given network topology
- fitting of spaceMap and other graphical models to the data
- evaluation of fitted networks relative to true network topology
Documentation of these simulation codes will be forthcoming.
Once a network has been learned, interpretation can be challenging without a proper set of tools. Thus, we included a network analysis toolkit as part of the spaceMap analysis pipeline to facilitate network interpretation with special focus on integrative genomic applications. Provided an annotation of nodes (e.g. gene coordinates, functional description) by the user, our toolkit enables:
- identification of cis/trans regulatory information
- prioritization of hub nodes
- module analysis (community detection)
- pathway enrichment analysis (GO/KEGG)
All these features are reported through structured tables and are easily incorporated into technical reports in a variety of formats. Moreover, the network analysis tool kit integrates the results into a network file (e.g. .graphml format) that can be exported to existing tools such as the Cytoscape ecosystem. In this sense, spaceMap is not just a model; rather it is a tool for deriving meaning from integrative genomics data.
This toolkit is illustrated through the analysis of The Breast Cancer Proteogomics Landscape Study data, which is hosted on the GitHub repository neta-bcpls.