Maximizing biomarker discovery by minimizing gene signatures

Abstract

Background: The use of gene signatures can potentially be of considerable value in the field of clinical diagnosis. However, gene signatures defined with different methods can be quite various even when applied the same disease and the same endpoint. Previous studies have shown that the correct selection of subsets of genes from microarray data is key for the accurate classification of disease phenotypes, and a number of methods have been proposed for the purpose. However, these methods refine the subsets by only considering each single feature, and they do not confirm the association between the genes identified in each gene signature and the phenotype of the disease. We proposed an innovative new method termed Minimize Feature's Size (MFS) based on multiple level similarity analyses and association between the genes and disease for breast cancer endpoints by comparing classifier models generated from the second phase of MicroArray Quality Control (MAQC-II), trying to develop effective meta-analysis strategies to transform the MAQC-II signatures into a robust and reliable set of biomarker for clinical applications.

Results: We analyzed the similarity of the multiple gene signatures in an endpoint and between the two endpoints of breast cancer at probe and gene levels, the results indicate that disease-related genes can be preferably selected as the components of gene signature, and that the gene signatures for the two endpoints could be interchangeable. The minimized signatures were built at probe level by using MFS for each endpoint. By applying the approach, we generated a much smaller set of gene signature with the similar predictive power compared with those gene signatures from MAQC-II.

Conclusions: Our results indicate that gene signatures of both large and small sizes could perform equally well in clinical applications. Besides, consistency and biological significances can be detected among different gene signatures, reflecting the studying endpoints. New classifiers built with MFS exhibit improved performance with both internal and external validation, suggesting that MFS method generally reduces redundancies for features within gene signatures and improves the performance of the model. Consequently, our strategy will be beneficial for the microarray-based clinical applications.

Figure 1

Analysis workflow. This figure illuminates the general outline of the whole process.

Figure 2

Heatmaps for gene signatures on validation dataset. (a) Heatmap for BR_D_Model; (b) Heatmap for BR_E_Model. Each column represents a sample in the dataset, and each row represents a gene in the gene signature. Note that the end row is endpoint status.

Figure 3

Performances of original and swap models based on classification algorithm level similarity analysis. a) Endpoint D original models; b) Endpoint E original models; c) Endpoint D swap models; and d) Endpoint E swap models. Coordinate axes are MCC (internal validation), Val_MCC (external validation) and MCC_Std (internal validation standard deviation). Each classification algorithm is represented by a different color. The radius of each sphere is related to the number of model features, within a range of 50-1 000. The blue stars are our own models, while spheres are models from which our models were developed.