A powerful framework for an integrative study with heterogeneous omics data: from univariate statistics to multi-block analysis

Abstract

High-throughput data generated by new biotechnologies require specific and adapted statistical treatment in order to be efficiently used in biological studies. In this article, we propose a powerful framework to manage and analyse multi-omics heterogeneous data to carry out an integrative analysis. We have illustrated this using the mixOmics package for R software as it specifically addresses data integration issues. Our work also aims at applying the most recent functionalities of mixOmics to real datasets. Although multi-block integrative methodologies exist, we hope to encourage a more widespread use of such approaches in an operational framework by biologists. We have used natural populations of the model plant Arabidopsis thaliana in this work, but the framework proposed is not limited to this plant and can be deployed whatever the organisms of interest and the biological question may be. Four omics datasets (phenomics, metabolomics, cell wall proteomics and transcriptomics) were collected, analysed and integrated to study the cell wall plasticity of plants exposed to sub-optimal temperature growth conditions. The methodologies presented here start from basic univariate statistics leading to multi-block integration analysis. We have also highlighted the fact that each method, either unsupervised or supervised, is associated with one biological issue. Using this powerful framework enabled us to arrive at novel conclusions on the biological system, which would not have been possible using standard statistical approaches.

Integrative multi-omics analysis of childhood aggressive behavior

“…We explored how sensitive our results are to different specifications of the design matrices and found that these resulted in similar predictive abilities, but selected different numbers of only partially overlapping PGSs and CpGs (Appendices C and D). In line with expectations, the predictive ability of the null design matrix was slightly better than for the empirical design matrix, reflecting that such a model focuses on selecting discriminatory variables (Singh et al 2019;Duruflé et al 2020). We expected that the model with the full design model sacrificed predictive accuracy to select discriminatory variables that are also highly correlated, but this model had the lowest classification error rate as compared to the models with an empirical or null design matrix in the test and clinical data.…”

“…A 'null' design matrix denotes weak or no correlations among omics blocks by setting values close to or equal to zero. The full design matrix optimizes correlations among the omics blocks, while the null design matrix optimizes the discrimination between samples (Rohart et al 2017;Singh et al 2019;Duruflé et al 2020). We can also specify a design matrix with the empirical correlations among the omics blocks.…”

“…We employed an analytical design comprising three phases: 1) single-omics analyses; 2) pairwise cross-omics analyses; and 3) multi-omics analyses (Fig. 1) (Duruflé et al 2020). First, we built singleomics biomarker panels in the twin cohort, with 70% of the twin data for model training, 30% of the twin data for model testing, and an independent clinical cohort for follow-up.…”

Identification of Tumor Mutation Burden, Microsatellite Instability, and Somatic Copy Number Alteration Derived Nine Gene Signatures to Predict Clinical Outcomes in STAD

“…The most commonly used integration tools include “mixOmics,” “tRanslatome,” “R.JIVE,” and “iClusterPlus”. First, “mixOmics” was a powerful framework with four kinds of datasets (metabolomics, phenomics, cell wall proteomics, and transcriptomics) ( Durufle et al, 2021 ). Then, a deep neural network named “tRanslatome” was proposed which can predict the protein structure from input amino acid sequences but not for disordered proteins ( Du et al, 2021 ).…”

Integrative multi-omics analysis of genomic, epigenomic, and metabolomics data leads to new insights for Attention-Deficit/Hyperactivity Disorder

“…Associations and correlations across omics levels have been reported between the genome and epigenome (van Dongen et al, 2016; Min et al, 2021), the genome and the metabolome (Hagenbeek, Pool, et al, 2020) and the epigenome and the metabolome (Gomez-Alonso et al, 2021). To optimally analyze multi-omics data in association and etiological studies, dedicated statistical treatment of simultaneous omics influences is required (Durufle et al, 2020), with simultaneous modeling completing approaches in which separate analyses are applied to each omics level. Such innovative multi-omics analyses may result in novel insights and uncover new biological pathways underlying traits and diseases (Rajasundaram & Selbig, 2016).…”