DECO

DEcomposing heterogeneous Cohorts using Omic profiling data

FJ Campos-Laborie, JM Sanchez-Santos and J De Las Rivas1

1Bioinformatics and Functional Genomics Group. Cancer Research Centre (CiC-IBMCC, USAL/CSIC/IBSAL). Salamanca (Spain)

Abstract

We present a tutorial to use DECO, a bioinformatic method to explore and find differences in heterogeneous large datasets usually produced in biological or biomedical omic-wide studies. The method makes a comprehensive analysis of multidimensional datasets, usually consisting on a collection of samples where thousands of features have been measured with a high-throughput technology (genomic or proteomic techniques). The method finds the differences in the profiles of the features along the samples and identifies the associations between them, showing the features that best mark a given class or category as well as possible sample outliers that do not follow the same pattern of the majority of the corresponding cohort. The method can be used in a binary, multiclass or unsupervised mode, it allows the discovery of multiple classes or categories, and is quite adequate for patients stratification.

---

1 Variability and heterogeneity in high-dimensional data

Individual diversity and variability is one of the most complex issues to deal within high-dimensional studies of large populations, as the ones currently performed in biomedical analyses using omic technologies. DECO is a method that combines two main computational procedures: (i) a Recurrent-sampling Differential Analysis (RDA) that performs combinatorial sampling without replacement to select multiple sample subsets followed by a differential analysis (LIMMA); and (ii) a Non-Symmetrical Correspondence Analysis (NSCA) on differential events, which would improve the characterization and assigment of features and samples in a common multidimensional space. This second step combines in a single statistic (h-statistic) both the feature-sample changes detected and the predictor-response information provided by NSCA.

The statistical procedure followed in both parts of the method are detailed in the original publication [1], but this brief vignette explains how to use DECO to analyze multidimensional datasets that may include heterogeneous samples. The aim is to improve characterization and stratification of complex sample series, mostly focusing on large patient cohorts, where the existence of outlier or mislabeled samples is quite possible.

Thus, DECO can reveal exclusive associations between features and samples based in specific differential signal and provide a better way for the stratifycation of populations using multidimensional large-scale data. The method could be applied to data derived from different omic technologies since LIMMA has been demonstrated to be applicable to non-transcriptomic data. Along the vignette, we used genome-wide expression data obtained with microarrays or with RNA-seq (either for genes, miRNAs, ncRNAs, etc).