Network-assisted data analysis

Transcription

1 Network-assisted data analysis Bing Zhang Department of Biomedical Informatics Vanderbilt University

2 Protein identification in shotgun proteomics Protein digestion LC-MS/MS Protein assembly Database search 2 BMIF310, Fall 2009

3 Protein assembly and classification Background a b Zhang, et al. J Proteome Res 6:3549, BMIF310, Fall 2009

4 Network-assisted protein identification: motivation Current protein assembly pipelines treat proteins as individual entities. Biologically interesting proteins may be eliminated due to insufficient experimental evidence. Most biological functions arise from interactions among proteins. Can we use protein interaction network information to improve protein identification? Hypothesis: an eliminated protein is more likely to be present in the original sample if it lies close to other confidently identified proteins in the protein interaction network. 4 BMIF310, Fall 2009

5 Do confidently identified proteins cluster together? Li et al. Mol Syst Biol,5:303, BMIF310, Fall 2009

6 Neighborhood majority voting Direct neighbors (local) Iterative majority voting (global) 6 Local Global BMIF310, Fall 2009

7 Module-based analysis Direct neighbor voting Module-based 7 BMIF310, Fall 2009

8 Method evaluation: Cross validation for proteins with known labels Yeast Yeast cell culture dataset YPIN, 5,666 nodes; 126,126 edges Results: sensitivity of 56% at the specificity of 90%. CEA is more accurate and robust than other methods Mouse Mouse organ datasets: brain, lung, placenta MPIN1: 9,776 nodes; 69,470 edges MPIN2: 12,271 nodes; 236,675 edges Results: sensitivity of ~45% at the specificity of 90% Li et al. Mol Syst Biol,5:303, BMIF310, Fall 2009

9 Method evaluation: Supporting evidence for rescued proteins Mouse-MPIN1 Proteins rescued Brain: 171 (12%) Lung: 181 (10%) Placenta: 156 (11%) Independent evidences Microarray EST library Publication Ratio of support ( 2 evidences) Li et al. Mol Syst Biol,5:303, BMIF310, Fall 2009

10 Application: Breast cancer data set (normal vs tumor) Rescued proteins Normal: 139 (23%) Tumor: 95 (8%) Rescued cancer-related proteins Ctnnb1 Top1 Cancer specific sub-networks Wnt signaling pathway Cell adhesion Apoptosis Li et al. Mol Syst Biol,5:303, BMIF310, Fall 2009

11 Extended hypothesis Hypothesis: an eliminated protein is more likely to be present in the original sample if it lies close to other confidently identified proteins in the protein interaction network. Extended hypothesis: Proteins lie close to each other in a protein interaction network are more likely to share common attributes. Protein expression Protein function Disease association 11 BMIF310, Fall 2009

12 Protein function prediction: motivation Gene Ontology annotation coverage in different organisms Sharan et al. Mol Syst Biol, 3:88, BMIF310, Fall 2009

13 Network-based protein function prediction Proteins that lie closer to one another in a protein interaction network are more likely to have similar function and involve in similar biological process. GO semantic similarity Sharan et al. Mol Syst Biol, 3:88, BMIF310, Fall 2009

14 Disease gene prioritization: motivation Most common genetic disorders may result form variants in many genes, each contributing only weak effects to the disease Quantitative trait locus mapping and genome-wide association studies identify large set of candidate disease genes Assessing and selecting genes for validation is nontrivial Causal genes for the same disease are often involved in a few biological processes or molecular pathways Limb-girdle muscular dystrophy: many disease genes involve in the dystrophy complex Fanconi anemia: 5 out of the 10 disease genes identified in a genetic study function in a nuclear complex 14 BMIF310, Fall 2009

15 Network-based disease causal gene prioritization: method Kohler et al. Am J Hum Genet. 82:949, BMIF310, Fall 2009

16 Network-based disease causal gene prioritization: result Kohler et al. Am J Hum Genet. 82:949, BMIF310, Fall 2009

17 Disease biomarker identification: gene-level approach Scoring individual genes to generate a marker set Use gene expression matrix to train a classifier High variability: two breast cancer biomarker sets of ~70 genes each shared only 3 genes in common Doesn t improve mechanistic understanding of cancer Chuang et al. Mol Systems Biol, 3:140, BMIF310, Fall 2009

18 Disease biomarker identification: pathway-level approach Scoring known pathways by the coherency of expression changes among their member genes More robust More sensitive Better interpretation The majority of human genes have not yet been assigned to a definitive pathway 18 BMIF310, Fall 2009

19 Disease biomarker identification: network approach Overlay gene expression data on the protein interaction network Search for subnetworks that are highly discriminative of metastasis Use subnetwork activity matrix to train a classifier Chuang et al. Mol Systems Biol, 3:140, BMIF310, Fall 2009

20 Subnetwork markers are more robust Chuang et al. Mol Systems Biol, 3:140, BMIF310, Fall 2009

21 Subnetwork markers are more informative 47.3% and 65.4% subnetwork markers showed functional enrichment 66 and 153 subnetworks corresponded to the major events that have been implicated in the progression of cancer Signaling of cell growth and survival Cell proliferation and replication Apoptosis Cell and tissue remodeling Circulation and coagulation Metabolism Detecting important non-differentially expressed genes TP53, KARS, HARS, ERBB2, PIK3CA Chuang et al. Mol Systems Biol, 3:140, BMIF310, Fall 2009

22 Summary Guilt by association: Proteins lie close to each other in a protein interaction network are more likely to share common attributes. Protein expression Protein function Disease-related proteins Network-based disease biomarkers 22 BMIF310, Fall 2009

23 Key references Li et al. Network-assisted protein identification and data interpretation in shotgun proteomics. Molecular Systems Biology, 5:303, 2009 Ideker et al. Protein networks in disease. Genome Research, 18:644, 2008 Sharan et al. Network-based prediction of protein function. Molecular Systems Biology, 3:88, BMIF310, Fall 2009