The availability of biological data in massive scales continues to represent unlimited opportunities as well as great challenges in bioinformatics research. Developing innovative data mining techniques and efficient parallel computational methods to implement them will be crucial in extracting useful knowledge from this raw unprocessed data, such as in discovering significant cellular subsystems from gene correlation networks. In this paper, we present a scalable combinatorial sampling technique, based on identifying maximum chordal sub graphs, that reduces noise from biological correlation networks, thereby making it possible to find biologically relevant clusters from the filtered network. We show how selecting the appropriate filter is crucial in maintaining the key structures from the original networks and uncovering new ones after removing noisy relationships. We also conduct one of the first comparisons in two important sensitivity criteria -- the perturbation due to the vertex numbers of the network and perturbations due to data distribution. We demonstrate that our chordal-graph based filter is effective across many different vertex permutations, as is our parallel implementation of the sampling algorithm.
展开▼
