Clustering is a type of unsupervised learning approach to identify underlying structure in multidimensional data, based on exploring the data alone (i.e. without labels as occurs in supervised learning). Ultimately, the structure uncovered in a solution is a hypothetical relationship of the data that may indicate meaning. This structure is different when the relationships of the data are perturbed, such as by transforming the data, or when alternate criteria are considered in the process (like using a different measure of distance between points or algorithms). Therefore, one could cluster many times, making such perturbations, to explore the space of solutions.
Kristen Naegle developed ensemble approaches to clustering of biological data in her Ph.D. work that demonstrated that one can infer function of tyrosine phosphorylation from quantitative measurements of the dynamic changes of network phosphorylation in cells in response to growth factor stimulation. During her post-doctoral work, Dr. Naegle went on to show that robustness in clustering was predictive of protein interactions and inferred novel interactions in the epidermal growth factor receptor network.
The Naegle lab has gone on to utilize these frameworks in collaborations with Valeria Cavalli, Linda Pike, and Paul Huang to explore a variety of biological problems from axonal degeneration to DDR2 signaling.
Additionally, team member Roman Sloutsky and Kristen Naegle posed frameworks for how to incorporate the noise that is inherent in biological data during the clustering process in order to understand how the relationships identified in clustering are altered when noise is considered. Team members Tom Ronan and Kristen Naegle wrote a review article on clustering and the unique approaches that ensemble approaches afford. This review was one of the most highly accessed articles of Science journals and made the home page.