Abstract
We introduce a tensor-based clustering method to extract sparse,low-dimensional structure from high-dimensional, multi-indexed datasets. Thisframework is designed to enable detection of clusters of data in the presenceof structural requirements which we encode as algebraic constraints in a linearprogram. Our clustering method is general and can be tailored to a variety ofapplications in science and industry. We illustrate our method on a collectionof experiments measuring the response of genetically diverse breast cancer celllines to an array of ligands. Each experiment consists of a cell line-ligandcombination, and contains time-course measurements of the early-signallingkinases MAPK and AKT at two different ligand dose levels. By imposingappropriate structural constraints and respecting the multi-indexed structureof the data, the analysis of clusters can be optimized for biologicalinterpretation and therapeutic understanding. We then perform a systematic,large-scale exploration of mechanistic models of MAPK-AKT crosstalk for eachcluster. This analysis allows us to quantify the heterogeneity of breast cancercell subtypes, and leads to hypotheses about the signalling mechanisms thatmediate the response of the cell lines to ligands.