Ensemble Clustering of PPI Networks

CREATED: 200802260846 ** Ensemble clustering Speaker: Srinivasan Parthasarathy, Dept CSE, Ohio State University *** Challenges in analysing PPI networks

• noisy data
• existence of hub nodes
• proteins can be multi-faceted
• data integration issues ** sources: 2-hybrid, mass spect, genetic co-occurrence ** targets: Y2H, mass spec - target binding, genetic co-occurrence - target functional ** weaknesses *** Ensemble Clustering
• consensus of many clustering
• errors should no correlate, methods should be diverse
• much work done in context of classification *** Ensemble clustering on PPI
• based methods, handling influence of noise and hubs
• scalability
• soft clustering *** Framework
• x distance metric, y clustering algorithms -> xy arrangements
• Similarity metrics ** clustering coefficient based (edge oriented, local, targets FP) ** edge betweenness based (edge oriented, global, targets FP) ** neighborhood based (local, node based, targets FN and FP)
• Clustering algorithms ** kMetic, favors balanced clusters ** Repeated bisection, hierarchal clustering ** Direct k-way partitioning, favors globular shapes *** PCA-based consensus

purification

** prune away clusters with high intra-cluster distance

dimensionality reduction

** construct incidence matrix and apply logistic PCA, each protein as a vector of 0/1 which indicates which cluster it is part of

consensus clustering

** agglomerative hierarchical clustering, recursive bisection algorithm *** Validation Metrics

• topological measure: modularity, $\sum_i (d_{ii} - (\sum_j d_{jj})^2)$
• information theoretic
• gene ontology annotations for each cluster of proteins
• P-value to measure statistical significance of clusters
• Clustering Score *** Experiment setup
• Dataset - Database of Interacting Proteins
• Other methods ** MCLA, ensemble method ** Molecular Complex Detection (MCODE) ** Markov Cluster Algorithm (MCL)
• Comparison with other ensemble methods
• Comparison with domain specific method to find dense regions
• Comparison with hub duplication method for soft clustering
• Specific comparison using CKA1 hub protein *** Future work
• incorporate directionality
• scalability
• temporal information about interaction is not used
• graphical models