Determining the optimal number of independent components for reproducible transcriptomic data analysis

Kairov, Ulykbek; Cantini, Laura; Greco, Alessandro; Molkenov, Askhat; Czerwińska, Urszula; Barillot, Emmanuel; Zinovyev, Andreï

doi:10.1186/s12864-017-4112-9

Cited by 60 publications

(59 citation statements)

References 25 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Additionally, the CysB i-modulons demonstrated that i-modulons may represent the effects of multiple regulators, when the activities 400 of the regulators are highly correlated across the measured conditions. However, adding new data or increasing the dimensionality of the decomposition can decouple the regulators, splitting the i-modulon into its biological parts 30,44 .…”

Section: Discussionmentioning

confidence: 99%

Matrix factorization recovers consistent regulatory signals from disparate datasets

Sastry¹,

Hu²,

Heckmann³

et al. 2020

Preprint

View full text Add to dashboard Cite

15The availability of gene expression data has dramatically increased in recent years. This data deluge could result in detailed inference of underlying regulatory networks, but the diversity of experimental platforms and protocols introduces critical biases that could hinder scalable analysis of existing data. Here, we show that the underlying structure of the E. coli transcriptome, as determined by Independent Component Analysis (ICA), is conserved across multiple independent 20 datasets, including both RNA-seq and microarray datasets. We also show that echoes of this structure remain in the proteome, accelerating biological discovery through multi-omics analysis.We subsequently combined five transcriptomics datasets into a large compendium containing over 800 expression profiles and discovered that its underlying ICA-based structure was still comparable to that of the individual datasets. ICA thus enables deep analysis of disparate data 25 to uncover new insights that were not visible in the individual datasets.Publicly available datasets, such as the NCBI Gene Expression Omnibus (GEO) 1 and Array Express 2 , contain thousands of transcriptomics datasets that are often designed and 30 analyzed for a specific study. Historically, microarrays were the platform of choice for transcriptomic interrogation. Over the past decade, usage of RNA sequencing (RNA-seq) has surpassed microarrays due to its higher sensitivity and ability to detect new transcripts 3 . Public repositories for proteomics datasets have introduced additional reusable expression datasets 4,5 . 35Multiple consortia have performed extensive comparisons of expression levels across different microarray and RNA-seq platforms 6-8 . These studies showed that absolute gene expression levels cannot be accurately measured by either expression profiling technique, whereas relative abundances are consistent across a wide range of transcriptomics platforms, with appropriate quality controls. However, transcript levels alone cannot predict protein 40 expression levels 9 To further complicate matters, batch effects and technical heterogeneity continue to present significant challenges to successful integration of omics datasets 10 . Differential expression analysis is the most common analytical method applied to transcriptomics datasets. However, differential expression analysis is limited in dimensionality, 45 interpretability, and reproducibility; it can only be applied to pairs of experimental conditions, requires additional analysis to interpret large swaths of differentially expressed genes 11,12 , and is highly dependent on the quantification pipeline 13,14 . Alternatively, machine learning methods, especially matrix factorization 15,16 , have provided new tools for extracting low-dimensional biological information from large omics data. 50In particular, independent component analysis (ICA) has been shown to extract biologically significant gene sets from many transcriptomics datasets 17-22 and two proteomics datasets 23 . ICA outperformed 42 module detect...

show abstract

Section: Discussionmentioning

confidence: 99%

Matrix factorization recovers consistent regulatory signals from disparate datasets

Sastry¹,

Hu²,

Heckmann³

et al. 2020

Preprint

View full text Add to dashboard Cite

show abstract

“…ICA was applied as previously described (Biton et al, 2014), using stabilization, with an additional procedure for determining the optimal number of independent components (Kairov et al, 2017). In the ICA decomposition X = AS, X is the gene expression (sample vs. gene) matrix, A is the (sample vs. component) matrix describing the loadings of the independent components, and S is the (component vs. gene matrix) describing the weights (projections) of the genes in the components.…”

Section: Exploratory Analysis Of Scrna-seq Datamentioning

confidence: 99%

Transcriptional programs define intratumoral heterogeneity of Ewing sarcoma at single cell resolution

Aynaud

Mirabeau

Gruel

et al. 2019

Preprint

Self Cite

View full text Add to dashboard Cite

EWSR1-FLI1, the chimeric oncogene specific for Ewing sarcoma (EwS), induces a cascade of signaling events leading to cell transformation. However, it remains elusive how genetically homogeneous EwS cells can drive heterogeneity of transcriptional programs. Here, we combined independent component analysis of single cell RNAsequencing data from diverse cell types and model systems with time-resolved mapping of EWSR1-FLI1 binding sites and of open chromatin regions to characterize dynamic cellular processes associated with EWSR1-FLI1 activity. We thus defined an exquisitely specific and direct, super-enhancer-driven EWSR1-FLI1 program. In EwS tumors, cell proliferation was associated with a well-defined range of EWSR1-FLI1 activity; moreover, cells with a high EWSR1-FLI1 activity presented a strong oxidative phosphorylation metabolism. In contrast, a subpopulation of cells from below and above optimal EWSR1-FLI1 activity was characterized by increased hypoxia. Overall, our study reveals sources of intratumoral heterogeneity within Ewing tumors.

show abstract

“…When applying ICA, the target dimension (number of components) is fixed before numerical optimisation and the algorithm can converge to different projections that share only a subset of components. In keeping with the idea developed by [20], we choose the target dimension K = 40 after examining the stability of the components and only stable components were used. Here, we used averagelink clustering based on absolute Pearson correlation coefficient (r) between columns of the source matrix (dimension 1,512×K) and a cut-off |r| ≥ 0.8 to compare components between runs.…”

Section: Building Covariates For Motif Discoverymentioning

confidence: 99%

“…Two main novelties of the proposed model are to allow overlaps between motif occurrences and to incorporate covariates summarising expression profiles into the probability of occurrence in a given promoter region. Covariates can correspond to the positions of the genes on an axis such as obtained by PCA [18] or ICA [19,20] but we also show how to use positions in a hierarchical clustering trees [21,8]. All the parameters are estimated in a Bayesian framework using a dedicated trans-dimensional MCMC algorithm.…”

Section: Introductionmentioning

confidence: 99%

Statistical modelling of bacterial promoter sequences for regulatory motif discovery with the help of transcriptome data: application to Listeria monocytogenes

Sultan

Fromion

Schbath

et al. 2019

Preprint

View full text Add to dashboard Cite

Automatic de novo identification of the main regulons of a bacterium from genome and transcriptome data remains a challenge. To address this task, we propose a statistical model of promoter DNA sequences that can use information on exact positions of the transcription start sites and condition-dependent expression profiles. Two main novelties are to allow overlaps between motif occurrences and to incorporate covariates summarising expression profiles (e.g. coordinates in projection spaces or hierarchical clustering trees). All parameters are estimated using a dedicated trans-dimensional Markov chain Monte Carlo algorithm that adjusts, simultaneously, for many motifs and many expression covariates: the width and palindromic properties of the corresponding position-weight matrices, the number of parameters to describe position with respect to the transcription start site, and the choice of relevant expression covariates. A data-set of transcription start sites and expression profiles available for the Listeria monocytogenes is analysed. The results validate the approach and provide a new global view of the transcription regulatory network of this important model food-borne pathogen. A previously unreported motif that may play an important role in the regulation of growth was found in promoter regions of ribosomal protein genes.

show abstract

Determining the optimal number of independent components for reproducible transcriptomic data analysis

Cited by 60 publications

References 25 publications

Matrix factorization recovers consistent regulatory signals from disparate datasets

Matrix factorization recovers consistent regulatory signals from disparate datasets

Transcriptional programs define intratumoral heterogeneity of Ewing sarcoma at single cell resolution

Statistical modelling of bacterial promoter sequences for regulatory motif discovery with the help of transcriptome data: application to Listeria monocytogenes

Contact Info

Product

Resources

About