Multi‐Omics Factor Analysis—a framework for unsupervised integration of multi‐omics data sets

Molecular Systems Biology, Journal Year: 2018, Volume and Issue: 14(6)

Published: June 1, 2018

Method20 June 2018Open Access Transparent process Multi-Omics Factor Analysis—a framework for unsupervised integration of multi-omics data sets Ricard Argelaguet orcid.org/0000-0003-3199-3722 European Molecular Biology Laboratory, Bioinformatics Institute, Hinxton, Cambridge, UK Search more papers by this author Britta Velten orcid.org/0000-0002-8397-3515 Laboratory (EMBL), Heidelberg, Germany Damien Arnol orcid.org/0000-0003-2462-534X Sascha Dietrich orcid.org/0000-0002-0648-1832 Heidelberg University Hospital, Thorsten Zenz orcid.org/0000-0001-7890-9845 German Cancer Research Center (dkfz) and National Tumor Diseases (NCT), & Hematology, Hospital Zurich Zurich, Switzerland John C Marioni orcid.org/0000-0001-9092-0852 Cambridge Wellcome Trust Sanger Florian Buettner Corresponding Author [email protected] orcid.org/0000-0001-5587-6761 Helmholtz Zentrum München–German Environmental Health, Institute Computational Biology, Neuherberg, Wolfgang Huber orcid.org/0000-0002-0474-2218 Oliver Stegle orcid.org/0000-0002-8818-7193 Information Argelaguet1,‡, Velten2,‡, Arnol1, Dietrich3, Zenz3,4,5, Marioni1,6,7, *,1,8, *,2 *,1,2 1European 2European 3Heidelberg 4German 5Germany 6Cancer 7Wellcome 8Helmholtz ‡These authors contributed equally to work *Corresponding author. Tel: +49 89 23742560; E-mail: 6221 387 8823; 3878190; Systems (2018)14:e8124https://doi.org/10.15252/msb.20178124 PDFDownload PDF article text main figures. Peer ReviewDownload a summary the editorial decision including letters, reviewer comments responses feedback. ToolsAdd favoritesDownload CitationsTrack CitationsPermissions ShareFacebookTwitterLinked InMendeleyWechatReddit Figures Info Abstract Multi-omics studies promise improved characterization biological processes across molecular layers. However, methods resulting heterogeneous are lacking. We present Analysis (MOFA), computational method discovering principal sources variation in sets. MOFA infers set (hidden) factors that capture technical variability. It disentangles axes heterogeneity shared multiple modalities those specific individual modalities. The learnt enable variety downstream analyses, identification sample subgroups, imputation detection outlier samples. applied cohort 200 patient samples chronic lymphocytic leukaemia, profiled somatic mutations, RNA expression, DNA methylation ex vivo drug responses. identified major dimensions disease heterogeneity, immunoglobulin heavy-chain variable region status, trisomy chromosome 12 previously underappreciated drivers, such as response oxidative stress. In second application, we used analyse single-cell data, identifying coordinated transcriptional epigenetic changes along cell differentiation. Synopsis (MOFA) is discovery when omics assays same broadly applicable approach integration. inferred latent represent underlying Factors can be or data-type specific. model flexibly handles missing values different types. an application Chronic Lymphocytic Leukaemia, discovers low dimensional space spanned known clinical markers profiles from single-cells, recovers differentiation trajectories identifies between transcriptome epigenome. Introduction Technological advances increasingly layers probed parallel, ranging genome, epigenome, transcriptome, proteome metabolome phenome profiling (Hasin et al, 2017). Integrative analyses use information these deliver comprehensive insights into systems under study. Motivated this, domains, cancer biology (Gerstung 2015; Iorio 2016; Mertins Genome Atlas Network, 2017), regulatory genomics (Chen 2016), microbiology (Kim 2016) host-pathogen (Soderholm 2016). Most recent technological have also enabled performing at level (Macaulay Angermueller Guo 2017; Clark 2018; Colomé-Tatché Theis, 2018). A common aim applications characterize samples, manifested one several (Ritchie 2015). particularly appealing if relevant not priori, hence may missed consider single modality targeted approaches. basic strategy testing marginal associations prominent example quantitative trait locus mapping, where large numbers association tests performed genetic variants gene expression levels (GTEx Consortium, 2015) marks While em-inently useful variant annotation, inherently local do provide coherent global map differences kernel- graph-based combine types similarity network (Lanckriet 2004; Wang 2014); however, it difficult pinpoint determinants graph structure. Related there exist generalizations other clustering reconstruct discrete groups based on (Shen 2009; Mo 2013). key challenge sufficiently addressed approaches interpretability. particular, would desirable drive observed These could continuous gradients, clusters combinations thereof. Such help establishing explaining with external phenotypes covariates. Although factor models address been proposed (e.g. Meng 2014, Tenenhaus 2014; preprint: Singh 2018), either lack sparsity, which reduce interpretability, require substantial number parameters determined using computationally demanding cross-validation post hoc. Further challenges faced existing scalability larger sets, handling non-Gaussian modalities, binary readouts count-based traits. Results statistical integrating fashion. Intuitively, viewed versatile statistically rigorous generalization component analysis (PCA) data. Given matrices measurements partially overlapping interpretable low-dimensional representation terms (Fig 1A). thus facilitating gradients subgroups loadings sparse, thereby linkage most features. Importantly, what extent each unique 1B), revealing Once trained, output range visualization, classification space(s) factors, well automated annotation (gene set) enrichment analysis, 1B). Figure 1. Analysis: overview Model overview: takes M input (Y1,…, YM), modality, co-occurrent but features necessarily related differ numbers. decomposes matrix (Z) weight matrices, (W1,.., WM). White cells correspond zeros, i.e. inactive features, whereas cross symbol denotes values. fitted queried (i) variance decomposition, assessing proportion explained (ii) semi-automated inspection (iii) visualization (iv) values, assays. Download figure PowerPoint Technically, builds upon group (Virtanen 2012; Khan Klami Bunte Zhao Leppäaho Kaski, adapted requirements (Materials Methods): fast inference variational approximation, sparse solutions interpretation, efficient flexible combination likelihood enables diverse binary-, count- continuous-valued relationship previous Virtanen 2013; Remes Hore Leppáaho 2017) discussed Materials Methods Appendix Table S3. implemented well-documented open-source software comes tutorials workflows domains Methods). Taken together, functionalities powerful tool disentangling studies. validation comparison simulated First, validate MOFA, its generative model, varying views, models, Methods, S1). found was able accurately dimension, except settings high proportions (Appendix Fig account observations fit simulating count Figs S2 S3). compared two reported integration: GFA (Leppäaho iCluster (Mo Over simulations, tended infer redundant S4) were less accurate recovering patterns activity views S5). than EV1). For example, training CLL next, required 25 min versus 34 h 5–6 days iCluster. Click here expand figure. EV1. Scalability iClusterTime (red), (blue) (green) function K, D, N M. Baseline = 3, K 10, D 1,000 100 5% Shown average time 10 trials, error bars denote standard deviation. only shown lowest all training. Application leukaemia study (CLL), combined mutation (Dietrich 2A). Notably, nearly 40% some types; value scenario uncommon studies, designed cope Methods; configured order accommodate 2. A. Study Data rows (D features) (N) columns, grey bars. B, C. (B) Proportion total (R2) assay (C) cumulative explained. D. Absolute top 1 2 Mutations E. Visualization colours IGHV status tumours; shape colour tone indicate status. F. Number enriched Reactome per (FDR < 1%). categories pathways defined S2. (minimum 2% least type; robust algorithm initialization subsampling S6 S7). largely orthogonal, capturing independent S6). Among these, active assays, indicating broad roles 2B). contrast, 3 5 4 only. Cumulatively, 41% 38% mRNA 24% 2C). trained excluding probe their redundancy, finding still recovered, while others dependent type S8). 2013), consistent instances S9). important reveals axis attributed stress As part pipeline, provides strategies identify aetiology weights aligned (IGHV), 2D E). Thus, correctly them (Zenz 2010; Fabbri Dalla-Favera, marker associated 1, surrogate state tumour's origin activation B-cell receptor. practice generally considered (Fabbri our results complex substructure 3A, S10). At current resolution, three subgroup Oakes al (2016) Queiros (2015) S11), although suggestive evidence continuum. connected S12 S13), genes linked (Vasconcelos 2005; Maloum Trojani Morabito Plesingerova 3B C) drugs target kinases receptor pathway 3D 3. Characterization Beeswarm plot corresponding 3-means (LZ), intermediate (IZ) (HZ). largest absolute Plus minus symbols right sign loading. Genes highlighted orange described prognostic Heatmap (B). weights, annotated category. Drug curves stratified (A). Despite importance, accounted 20% suggesting existence heterogeneity. One 5, revealed tagged senescence (Figs 2F EV2A), heat-shock proteins (HSPs; EV2B C), essential protein folding up-regulated conditions (Srivastava, 2002; Åkerfelt 2010). HSP cancers tumour survival (Trachootham 2009), far family has received little attention context CLL. Consistent strongest stress, reactive oxygen species (ROS), damage apoptosis EV2D EV2. (oxidative factor) 5. Colours TNF, inflammatory marker. Gene (t-test, six Samples ordered Scaled loading, captured 9% suggested aetiologies immune T-cell signalling 2F), likely due composition samples: comprised mainly B cells, possible contamination T monocytes S14). 11% samples' general sensitivity (Geeleher S15). imputes Next, explored annotations, missing, mis-annotated inaccurate, since they frequently imperfect surrogates (Westra 2011). Since biomarker impacting care, assessed consistency 176 out patients, agreement further allowed classifying patients lacked clinically measured EV3A B). Interestingly, assigned label. Upon nine cases showed signatures, borderline classification; remaining clearly discordant EV3C D). Additional whole exome sequencing confirmed outliers within EV3E F). EV3. Prediction denoting predicted labels Pie chart showing imputed Sample-to-sample correlation ONO-4509 (not included data): Boxplots viability ONO-4509. middle; left right, viabilities M-CLL U-CLL shown, respectively. panels show concentrations tested. Boxes first third quartiles value. Whole mutations y-axis, separately labelled. incomplete problem high-throughput ability fill entire both tasks, yielded predictions established strategies, feature-wise mean, SoftImpute (Mazumder 2010) k-nearest neighbour (Troyanskaya 2001; EV4, S16), GFA, especially case S17). EV4. Imputation A, B. Considered SoftImpute, mean (Mean) (kNN). averages squared (MSE) 15 experiments increasing fractions considering (A) random random. Error plus error. Latent predictive outcomes Finally, utility predictors outcomes. Three significantly next treatment (Cox regression, FDR 1%, 4A B): origin, Factors, 7 8, chemo-immunotherapy prior collection (P 0.01, t-test). captures del17p TP53 oncogenes (Garg Fluhr S18), 8 WNT S19). 4. Relationship Association univariate Cox regression 174 (96

Language: Английский

---

Guidelines for the use of flow cytometry and cell sorting in immunological studies (second edition)

European Journal of Immunology, Journal Year: 2019, Volume and Issue: 49(10), P. 1457 - 1973

Published: Oct. 1, 2019

These guidelines are a consensus work of considerable number members the immunology and flow cytometry community. They provide theory key practical aspects enabling immunologists to avoid common errors that often undermine immunological data. Notably, there comprehensive sections all major immune cell types with helpful Tables detailing phenotypes in murine human cells. The latest techniques applications also described, featuring examples data can be generated and, importantly, how analysed. Furthermore, tips, tricks pitfalls avoid, written peer-reviewed by leading experts field, making this an essential research companion.

Language: Английский

---

Mediterranean diet intervention in overweight and obese subjects lowers plasma cholesterol and causes changes in the gut microbiome and metabolome independently of energy intake

Gut, Journal Year: 2020, Volume and Issue: 69(7), P. 1258 - 1268

Published: Feb. 19, 2020

Objectives This study aimed to explore the effects of an isocaloric Mediterranean diet (MD) intervention on metabolic health, gut microbiome and systemic metabolome in subjects with lifestyle risk factors for disease. Design Eighty-two healthy overweight obese a habitually low intake fruit vegetables sedentary participated parallel 8-week randomised controlled trial. Forty-three participants consumed MD tailored their habitual energy intakes (MedD), 39 maintained regular diets (ConD). Dietary adherence, parameters, were monitored over period. Results Increased adherence MedD group successfully reprogrammed subjects’ fibre animal proteins. Compliance was confirmed by lowered levels carnitine plasma urine. Significant reductions cholesterol (primary outcome) faecal bile acids occurred compared ConD group. Shotgun metagenomics showed changes that reflected individual increase gene richness who reduced inflammation intervention. The led increased fibre-degrading Faecalibacterium prausnitzii genes microbial carbohydrate degradation linked butyrate metabolism. dietary urinary urolithins, acid insulin sensitivity co-varied specific taxa. Conclusion Switching while maintaining blood caused multiple are relevant future strategies improvement health.

Language: Английский

---

Computational principles and challenges in single-cell data integration

Nature Biotechnology, Journal Year: 2021, Volume and Issue: 39(10), P. 1202 - 1215

Published: May 3, 2021

Language: Английский

---

MOGONET integrates multi-omics data using graph convolutional networks allowing patient classification and biomarker identification

Nature Communications, Journal Year: 2021, Volume and Issue: 12(1)

Published: June 8, 2021

Abstract To fully utilize the advances in omics technologies and achieve a more comprehensive understanding of human diseases, novel computational methods are required for integrative analysis multiple types data. Here, we present multi-omics method named Multi-Omics Graph cOnvolutional NETworks (MOGONET) biomedical classification. MOGONET jointly explores omics-specific learning cross-omics correlation effective data We demonstrate that outperforms other state-of-the-art supervised approaches from different classification applications using mRNA expression data, DNA methylation microRNA Furthermore, can identify important biomarkers related to investigated problems.

Language: Английский

---

Machine Learning and Integrative Analysis of Biomedical Big Data

Genes, Journal Year: 2019, Volume and Issue: 10(2), P. 87 - 87

Published: Jan. 28, 2019

Recent developments in high-throughput technologies have accelerated the accumulation of massive amounts omics data from multiple sources: genome, epigenome, transcriptome, proteome, metabolome, etc. Traditionally, each source (e.g., genome) is analyzed isolation using statistical and machine learning (ML) methods. Integrative analysis multi-omics clinical key to new biomedical discoveries advancements precision medicine. However, integration poses computational challenges as well exacerbates ones associated with single-omics studies. Specialized approaches are required effectively efficiently perform integrative acquired diverse modalities. In this review, we discuss state-of-the-art ML-based for tackling five specific analysis: curse dimensionality, heterogeneity, missing data, class imbalance scalability issues.

Language: Английский

---

MOLI: multi-omics late integration with deep neural networks for drug response prediction

Bioinformatics, Journal Year: 2019, Volume and Issue: 35(14), P. i501 - i509

Published: June 6, 2019

Historically, gene expression has been shown to be the most informative data for drug response prediction. Recent evidence suggests that integrating additional omics can improve prediction accuracy which raises question of how integrate omics. Regardless integration strategy, clinical utility and translatability are crucial. Thus, we reasoned a multi-omics approach combined with datasets would relevance.

Language: Английский

---

State of the Field in Multi-Omics Research: From Computational Needs to Data Mining and Sharing

Frontiers in Genetics, Journal Year: 2020, Volume and Issue: 11

Published: Dec. 10, 2020

Multi-omics, variously called integrated omics, pan-omics, and trans-omics, aims to combine two or more omics data sets aid in analysis, visualization interpretation determine the mechanism of a biological process. Multi-omics efforts have taken center stage biomedical research leading development new insights into events processes. However, mushrooming myriad tools, datasets, approaches tends inundate literature overwhelm researchers field. The this review are provide an overview current state field, inform on available reliable resources, discuss application statistics machine/deep learning multi-omics analyses, findable, accessible, interoperable, reusable (FAIR) research, point best practices benchmarking. Thus, we guidance interested users domain by addressing challenges underlying biology, giving toolset, common pitfalls, acknowledging methods’ limitations. We conclude with practical advice recommendations software engineering reproducibility share comprehensive awareness for end-to-end workflow.

Language: Английский

---

MicrobiomeAnalyst 2.0: comprehensive statistical, functional and integrative analysis of microbiome data

Nucleic Acids Research, Journal Year: 2023, Volume and Issue: 51(W1), P. W310 - W318

Published: May 11, 2023

Abstract Microbiome studies have become routine in biomedical, agricultural and environmental sciences with diverse aims, including diversity profiling, functional characterization, translational applications. The resulting complex, often multi-omics datasets demand powerful, yet user-friendly bioinformatics tools to reveal key patterns, important biomarkers, potential activities. Here we introduce MicrobiomeAnalyst 2.0 support comprehensive statistics, visualization, interpretation, integrative analysis of data outputs commonly generated from microbiome studies. Compared the previous version, features three new modules: (i) a Raw Data Processing module for amplicon processing taxonomy annotation that connects directly Marker Profiling downstream statistical analysis; (ii) Metabolomics help dissect associations between community compositions metabolic activities through joint paired metabolomics datasets; (iii) Statistical Meta-Analysis identify consistent signatures by integrating across multiple Other improvements include added multi-factor differential interactive visualizations popular graphical outputs, updated methods prediction correlation analysis, expanded taxon set libraries based on latest literature. These are demonstrated using dataset recent type 1 diabetes study. is freely available at microbiomeanalyst.ca.

Language: Английский

---

Artificial intelligence and machine learning‐aided drug discovery in central nervous system diseases: State‐of‐the‐arts and future directions

Medicinal Research Reviews, Journal Year: 2020, Volume and Issue: 41(3), P. 1427 - 1473

Published: Dec. 9, 2020

Abstract Neurological disorders significantly outnumber diseases in other therapeutic areas. However, developing drugs for central nervous system (CNS) remains the most challenging area drug discovery, accompanied with long timelines and high attrition rates. With rapid growth of biomedical data enabled by advanced experimental technologies, artificial intelligence (AI) machine learning (ML) have emerged as an indispensable tool to draw meaningful insights improve decision making discovery. Thanks advancements AI ML algorithms, now AI/ML‐driven solutions unprecedented potential accelerate process CNS discovery better success rate. In this review, we comprehensively summarize AI/ML‐powered pharmaceutical efforts their implementations area. After introducing AI/ML models well conceptualization preparation, outline applications technologies several key procedures including target identification, compound screening, hit/lead generation optimization, response synergy prediction, de novo design, repurposing. We review current state‐of‐the‐art AI/ML‐guided focusing on blood–brain barrier permeability prediction implementation into neurological diseases. Finally, discuss major challenges limitations approaches possible future directions that may provide resolutions these difficulties.

Language: Английский

---