Prognostic transcriptional association networks: a new supervised approach based on regression trees

Abstract

Motivation: The application of information encoded in molecular networks for prognostic purposes is a crucial objective of systems biomedicine. This approach has not been widely investigated in the cardiovascular research area. Within this area, the prediction of clinical outcomes after suffering a heart attack would represent a significant step forward. We developed a new quantitative prediction-based method for this prognostic problem based on the discovery of clinically relevant transcriptional association networks. This method integrates regression trees and clinical class-specific networks, and can be applied to other clinical domains.

Results: Before analyzing our cardiovascular disease dataset, we tested the usefulness of our approach on a benchmark dataset with control and disease patients. We also compared it to several algorithms to infer transcriptional association networks and classification models. Comparative results provided evidence of the prediction power of our approach. Next, we discovered new models for predicting good and bad outcomes after myocardial infarction. Using blood-derived gene expression data, our models reported areas under the receiver operating characteristic curve above 0.70. Our model could also outperform different techniques based on co-expressed gene modules. We also predicted processes that may represent novel therapeutic targets for heart disease, such as the synthesis of leucine and isoleucine.

Availability: The SATuRNo software is freely available at http://www.lsi.us.es/isanepo/toolsSaturno/.

Fig. 1.

Schematic view of the proposed method. The first step involves building clinically relevant gene association networks from gene expression data of patients with the same clinical category. These networks are built based on the linear models generated by the model tree induction algorithm called M5P (Witten and Frank, 2005), an extension of regression tree algorithm. The second step involves predicting the clinical category of a new patient through the inferred networks. The prediction is based on the relative error between the true and predicted gene expression values of those genes involved in the inferred networks.

Fig. 2.

Clinically relevant gene association networks obtained from the heart dataset. Both networks were built from a microarray with 15 307 genes and the forest of trees was pruned using the threshold value θ = 15. Furthermore, the representative accuracy of these prognostic transcriptional networks was 72% (LOOCV).