Data integration lab_meeting
Download as PPTX, PDF1 like202 views
This document outlines several unsupervised methods for integrating multi-omic data, including matrix factorization methods (iCluster+), Bayesian methods (BCC), network-based methods (SNF), and multiple kernel learning (rMKL-LPP). It provides details on the procedures and assumptions of each method and discusses applications to discovering subtypes in cancers like breast cancer and glioblastoma multiforme.
Data integration lab_meeting
- 1. Liangqun Lu 2018 - 04 - 25
- 2. Outline ● Background on Data Integration ○ Biological regulation ○ Omic data integration objectives ○ Data Integration Challenges ● Unsupervised methods and Application ○ Matrix factorization methods (iCluster+ ) ○ Bayesian methods (BCC) ○ Network-based methods (SNF) ○ Multiple Kernel Learning and Multi-Step Analysis (rMKL-LPP) 2
- 3. Biological regulation ● Central dogma 3
- 4. Gene Regulatory Network Regulatory elements ● Receptors ● Transcriptional factors ● Inhibitory factors ● Cis-trans element Source: https://en.wikipedia.org/wiki/Gene_regulatory_network 4
- 5. Rich data 5
- 6. Single omic study ● One-dimension data explains the diagnostics and progression for complex disorders ● Information is limited ● Different layers of biological system are relevant and dependent 6
- 7. Omic data integration objectives ● Promoting precise medicine from big data ● Multiview investigation on the completeness and complexity of the biological system ● Discover hidden biological regularities ● Make use of complementary information and discover biomarkers for diagnosis, progression and treatment in human diseases 7
- 8. Data Integration Challenges (From Computational) ● Data integration is broad ● Data heterogeneity ● Data unification ● Data noise and bias ● Data integration and dimensionality reduction 8
- 9. 9
- 10. Unsupervised classification ● Matrix factorization methods (iCluster and iCluster+ ) ○ Assumption: common latent variable in different data ● Bayesian methods (Bayesian consensus clustering) ○ Assumption: assumptions on data distribution and data correlation ● Network-based methods (SNF) ○ Assumption: samples relationship can be enhanced from complementary multiple omic data ● Multiple Kernel Learning and Multi-Step Analysis (rMKL-LPP) ○ Assumption: pattern in a lower dimensional and integrative subspace 10
- 11. Data Integration for subtype discovery ● Data Source ○ Gene expression; DNA methylation; gene mutation ● Procedures ○ Data fusion -- Clustering -- Evaluation ● Biological interpretation ○ Molecular alterations ○ Survival outcome ○ Response to therapies 11
- 12. 12
- 14. Procedure ● Data Fusion and K-means model selection ○ EM algorithm to obtain maximum likelihood estimates ■ E-step provides a simultaneous dimension reduction ■ M-step is to update the parameter estimates ● Evaluation ○ Proportion of deviance -- POD (d/n^2) ○ Smaller, stronger cluster separability ○ Determine cluster number and lasso parameter λ 15
- 15. Application on breast cancer 16
- 16. Summaries ● The joint latent variable model is completely scalable to include additional data types ● iCluster have been applied to discover subtypes at breast cancer and glioblastoma multiforme (GBM) ● iCluster+ makes different modeling assumptions on data types: binary, continuous, categorical, and sequential data 17
- 17. Similarity Network Fusion (SNF) 18
- 18. SNF data fusion 1. Calculate sample similarity W in each omic dataset using (1) 2. Calculate normalized weight matrix P from W using (2) 3. Use K nearest neighbors (KNN) to calculate local affinity matrix S through the formulas (3) from W. P carries the full information about the similarity of each patient to all others whereas S only encodes the similarity to the K most similar patients for each patient. 4. Network fusion process: for 2 datasets, P1, S1 and P2, S2 can be calculated, then iteratively update P1 and P2 for t steps using (4) and (5); for more than 2 datasets, update the Ps using (5) 5. Obtain the overall fused matrix P by averaging the updated single Ps 19
- 19. Spectral Clustering Input X (n x n sample similarity matrix) and k clusters Goal subgroups in a graph with disjoint cliques Procedures: 1. Compute the normalized Laplacian L 2. Compute the first eigenvectors u and eigenvalues for L 3. Let U be the matrix containing eigenvectors u as columns 4. Form the matrix T from U by normalizing the rows to norm 1 5. Cluster the points with k-means into clusters C1, ..., Ck 20
- 20. Application: GBM subtype discovery Evaluations: 1. P value in Cox log-rank test 2. Silhouette score 21
- 21. Summaries ● SNF can construct sample sample network by integrating multiple datasets ● SNF can be expanded to include more datasets and be applied in more questions 22
- 22. Bayesian Consensus Clustering ● An integrative statistical model that permits a separate clustering of the objects for each data source. ● These separate clusterings adhere loosely to an overall consensus clustering ● BCC do simultaneous estimation of both the consensus clustering and the source-specific clusterings 23
- 23. Procedures ● Dirichlet mixture model to accommodate multiple data (X) ● Probability of belonging to one cluster ● Estimation ○ Gibbs sampling procedure to approximate the posterior distribution ○ Markov chain Monte Carlo (MCMC) proceeds by iteratively sampling ● Choose K based on highest mean adjusted adherence 24
- 24. Application on breast cancer ● RNA gene expression (GE) data for 645 genes. ● DNA methylation (ME) data for 574 probes. ● miRNA expression (miRNA) data for 423 miRNAs. ● Reverse phase protein array (RPPA) data for 171 proteins. 25
- 25. 26
- 26. Summaries 1. BCC model assumes a simple and general dependence between data sources. 2. BCC models both an overall clustering and a clustering specific to each data source, with advantages over traditional methods in terms of modeling uncertainty and the ability to borrow information across sources. 3. BCC is suitable to work on multisource biomedical data, as well may be used to compare clusterings from different statistical models for a single homogeneous dataset. 27
- 27. Regularized Multiple Kernel Learning Locality Preserving Projections (rMKL-LPP) 28 ● It is an extension of the current multiple kernel learning with dimensional reduction (MKL-DR) method, where the data are projected into a lower dimensional and integrative subspace. ● A regularization term is added to avoid overfitting during the optimization procedure, and it allows using several different kernel types. ● The Locality Preserving Projections (LPP) is applied to conserve the sum of distances for each sample’s k-Nearest Neighbors.
- 28. Procedures ● Data fusion ○ rMKL-LPP ○ Optimization ○ integrated kernel matrix ● Clustering ○ K-means ○ Mean silhouette width used to optimize number of clusters ● Evaluation ○ Silhouette score and cross validation (Rand index) 29
- 29. Applications in 5 cancers 1. Comparison to state-of-the-art (SNF) 2. Robustness analysis 3. Comparison of clusterings to established subtypes 4. Clinical implications from clusterings 30 5 cancers 1. glioblastoma multiforme (GBM) -- 213 samples 2. breast invasive carcinoma (BIC) -- 105 samples 3. kidney renal clear cell carcinoma (KRCCC) -- 122 samples 4. lung squamous cell carcinoma (LSCC) -- 106 samples 5. colon adenocarcinoma (COAD) -- 92 samplesDatasets: gene expression, DNA methylation and miRNA expression data
- 30. 1. Comparison to state-of-the-art 31
- 31. 2. Robustness analysis 32 Fig. 2. Robustness of clustering for leave-one-out datasets measured using Rand index. Fig. 3. Robustness of clustering for leave- one-out cross-validation applied to reduced sized datasets measured using Rand index.
- 32. 3. Comparison of clusterings to established subtypes 33
- 33. 4. Clinical implications from clusterings 34 GBM: ● 94 of 213 were treated with Temozolomide
- 35. Summaries 1. rMKL-LPP found subtypes with more interesting log-rank test compared to the state-of-the-art method 2. Several kernel matrices per data type can improve performance burdance, remove the burden of selecting the optimal kernel matrix and have fair stability 3. rMKL-LPP compared to unregularized MKL-DR remains stable also for small datasets 4. The application at GBM shows to capture this diverse information within one clustering 36
- 36. References 1. Huang, S., Chaudhary, K. & Garmire, L. X. More Is Better: Recent Progress in Multi-Omics Data Integration Methods. Front. Genet. 8, 84 (2017). 2. Wang, B. et al. Similarity network fusion for aggregating data types on a genomic scale. Nat. Methods 11, 333–337 (2014). 3. Shen, R., Olshen, A. B. & Ladanyi, M. Integrative clustering of multiple genomic data types using a joint latent variable model with application to breast and lung cancer subtype analysis. Bioinformatics 25, 2906–2912 (2009). 4. Shen, R. et al. Integrative subtype discovery in glioblastoma using iCluster. PLoS One 7, e35236 (2012). 5. Mo, Q. et al. Pattern discovery and cancer gene identification in integrated cancer genomic data. Proc. Natl. Acad. Sci. U. S. A. 110, 4245–4250 (2013). 6. Speicher, N. K. & Pfeifer, N. Integrating different data types by regularized unsupervised multiple kernel learning with application to cancer subtype discovery. Bioinformatics 31, i268–75 (2015). 7. Lock, E. F. & Dunson, D. B. Bayesian consensus clustering. Bioinformatics 29, 2610–2616 (2013). 37
Editor's Notes
- #11: The main advantage of Bayesian methods in data integration is that they can make assumptions not only on different types of data sets with various distributions but also on the correlations among data sets.
- #16: estimating the number of clusters K and the lasso parameter λ.
- #17: (C) Model selection based on POD measure. A four-cluster sparse solution (λ = 0.2) was chosen.
- #21: Spectral clustering is suitable for graph clustering
- #29: It is an extension of the current multiple kernel learning with dimensional reduction (MKL-DR) method MKL-DR: https://pdfs.semanticscholar.org/1cd3/bbae54b217843870fdc771d727b6043225b8.pdf
- #33: Fig. 2. Robustness of clustering for leave-one-out datasets measured using Rand index. Each patient is left out once in the dimensionality reduction and clustering procedure and afterwards added to the cluster with the closest mean based on the learned projection for this data point, which is given by projðxiÞ ¼ AT Ki b. The resulting cluster assignment is then compared with the clustering of the whole dataset. The error bars represent one standard deviation Fig. 3. Robustness of clustering for leave-one-out cross-validation applied to reduced sized datasets measured using Rand index. For each cancer type, we sampled 20 times half of the patients and applied leave-one-out cross-validation as described in Section 3.4. The error bars represent one standard deviation
- #36: The results are very similar to those found by Noushmehr et al. (2010) for their identified G-CIMP positive subtype. In addition, we found the set of underexpressed genes to be highly enriched for processes associated to the immune system and inflammation [cf. Table 3 (column 2)]. Since chronic inflammation is generally related to cancer progression and is thought to play an important role in the construction of the tumor microenvironment (Hanahan and Weinberg, 2011), these downregulations might be a reason for the favorable outcome of patients from this cluster.