More Is Better: Recent Progress in Multi-Omics Data Integration Methods.
Huang S, Chaudhary K, Garmire LX. · 2017
Read More Is Better: Recent Progress in Multi-Omics Data Integration Methods. by Huang S, Chaudhary K, Garmire LX. (2017) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/europmc-pmc5472696.txt.
Introduction
Huang, Chaudhary, and Garmire present a review in Frontiers in Genetics on ways to join multi-omics data. Omics studies measure genes, proteins, and other layers of a cell at once. Each layer alone gives only part of the picture. Joining them well is a key challenge for precision medicine. The authors survey the tools built for that task.
The review covers unsupervised, supervised, and semi-supervised methods. It groups them by type, such as matrix factorization, Bayesian, and network-based tools. A final section looks at methods that predict patient survival. Readers get a broad guide to the field and its software. The authors stress that more data types can improve clinical prediction.
Catalog record
| CREATOR | Huang S, Chaudhary K, Garmire LX. |
|---|---|
| YEAR | 2017 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 55 KB |
| LICENSE | CC BY 4.0 |
| RIGHTS | CC BY |
| SOURCE | Europe PMC |
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
More Is Better: Recent Progress in Multi-Omics Data Integration Methods By Sijia Huang, Kumardeep Chaudhary, Lana X Garmire Published in: Frontiers in Genetics DOI: 10.3389/fgene.2017.00084 ABSTRACT Abstract Multi-omics data integration is one of the major challenges in the era of precision medicine. Considerable work has been done with the advent of high-throughput studies, which have enabled the data access for downstream analyses. To improve the clinical outcome prediction, a gamut of software tools has been developed. This review outlines the progress done in the field of multi-omics integration and comprehensive tools developed so far in this field. Further, we discuss the integration methods to predict patient survival at the end of the review. Keywords: multi-omics, integration, prognosis, prediction, precision medicine, supervised learning, unsupervised learning INTRODUCTION A new era of personalized medicine has arrived, which proposes an individualized health care model with tailored medical target treatment and management for each patient (Chin et al., 2011). Under this regime, not only clinical profiles of patients but also their molecular profiles are personally managed to drive for advanced treatment. Cancer studies that are focused on one-dimensional omics data have only provided limited information regarding the etiology of oncogenesis and tumor progression. To overcome this, tremendous efforts have been made to obtain multi-platform based genomic data from biospecimen. The Cancer Genome Atlas (TCGA) is by far the largest endeavor in the USA to collect and analyze the tumor specimens from over 10,000 cancer patients (Weinstein et al., 2013). Measurements of these specimens include tissue exome sequencing, copy number variation (CNV), DNA methylation, gene expression, and microRNA (miRNA) expression, as well as some physiological and clinical data such as race, tumor stage, relapse, and treatment response. However, relative to the genomic data of different levels that are available to the public, the clinical information is more limited. A scale-up of TCGA is the International Cancer Genome Consortium (ICGC), which provides the information of genomic, transcriptomic and epigenomic abnormalities, and somatic mutations over 50 different cancer types (Hudson et al., 2010). These consortia have created unprecedented opportunities to reveal underlying oncogenic molecular signatures beneath phenotypes. However, human genomes are complex and regulated at multiple levels, which can be manifested by various genomic assays mentioned above. While each of these assays offers a peek of the complex system, these events are rather interdependent (or interactive). Thus, when combining several different omics data to discover the coherent biological signatures, it is challenging to incorporate different biological layers of information to predict phenotypic outcomes (tumor/normal, early/late stage, survival, etc.). It is herein our goal to address the pressing and challenging issues for developing novel algorithms and theoretical methods for multi-omics data integration, in the hope to extract biologically meaningful information of clinical relevance. The outline of this review is as follows. First, we will discuss the unsupervised data integration algorithms. Among them, we will highlight matrix factorization methods, Bayesian methods, and network-based methods. Next, we will review in-depth the supervised data integration methods, including network-based models, multiple kernel learning methods, and multi-step analysis based models. Subsequently, we will elaborate semi-supervised data integration methods. Finally, we will discuss the advancement of data integration methods for the aim of prognosis prediction and the biological insights underneath the data integration methods. UNSUPERVISED DATA INTEGRATION Unsupervised data integration refers to the cluster of methods that draw an inference from input datasets without labeled response variables. The different approaches under the umbrella of unsupervised data integration are presented in Figure 1 and Table 1. We have categorized them below into five areas: matrix factorization methods, Bayesian methods, network-based methods and multiple kernel learning, and multi-step analysis. [Figure: Unsupervised data integration methodology.] [Table: Summary of data integration tools.] Matrix factorization methods Joint Non-negative Matrix Factorization (NMF) The most straightforward method for unsupervised data integration falls into the matrix factorization category, which focuses on the projection of variations among data sets onto dimension-reduced space (Lee and Seung, 2001). Zhang et al. proposed NMF framework for multi-omics data integration (Zhang et al., 2011, 2012). This method is based on decomposing a non-negative matrix into non-negative loadings and non-negative factors: [formula: (1)min||X-WH||2,W≥0,H≥0] where X is the matrix of mRNA transcriptome, methylome, or other omics data that has M × N dimensions, W is the common factor for M × K dimension matrix and H is the K × N dimension coefficient matrix. Rather than simple correlation, the rationale is to project data onto common basis space, so that one can detect the coherent patterns among data, by examining the elements having significant z-scores. However, NMF is quite time-consuming and requires bulk memory space. For NMF, it is worth noting that not only it requires non-negative input matrices, but also proper normalization step for these input data sets as they have quite different distributions and variability. iCluster Like NMF, iCluster (Shen et al., 2009, 2012) assumes a regularized joint latent variable, which is similar to W in NMF but without non-negative constraints. H is the loading factor (coefficient), the imposed sparsity with different types of penalty functions for various data types. iCluster uses E to represent the error/noise term, and the underlying decomposition equation is: [formula: (2)X=WH+E] iCluster+ The upgraded iCluster+ expands iCluster by making the assumption of different modeling approaches for the relationships of X and W within different data platforms. It allows for diverse data types including binary, continuous, categorical, and sequential data with different modeling assumptions including logistic, normal linear, multilogit, and Poisson distributions (Mo et al., 2013). The common latent variable vector W represents the underlying driving factors that can be used for disease subtype assignment. Least absolute shrinkage and selection operator (LASSO) penalty is introduced to address the sparsity issue in H (Tibshirani, 1994). Since this approach requires high computational complexity, it is necessary to preselect the features critical for clustering results (Wang et al., 2014; Speicher and Pfeifer, 2015). Both iCluster and iCluster+ do not require non-negative input data, unlike NMF. Joint and Individual Variation Explained (JIVE) Another variation of NMF category is Joint and Individual Variation Explained (JIVE) method. JIVE decomposes the original data of each layer into three parts, including an approximation of joint variation across data types, approximation of specific structured variation for each data type, and residual noise. In other words, JIVE factors the original data input matrix (gene expression etc.) into two lower ranked representative portions Wc (shared factor) and Ws (data-specific factor), dependent on Hc and Hs (Lock et al., 2013). H matrix is contributed from one sub-matrix Hc common for all data types, and the other sub-matrix Hs specific to each data type. [formula: (3)X=WcHc+WsHs+E] It should be noted that there can be separate loading factors (Hc and Hs) for the shared factor and data-specific factor (Wc and Ws). The ranks of the two loading factors can be different. An application of JIVE on gene expression data and microRNA data on Glioblastoma (GBM) samples provided information to better characterize samples into different subtypes and strong clues for associations between each input layer (gene expression and microRNA). Based on PCA for factorization, JIVE suffers from outliers, thus the robustness of JIVE is a major concern. L1 penalties are also placed to reduce the dimensions in JIVE, giving