AN ISA ALGORITHM WITH UNKNOWN GROUP SIZES IDENTIFIES MEANINGFUL CLUSTERS IN METABOLOMICS DATA

摘要

Independent Subspace Analysis (ISA) denotes the task of linearlyseparating multivariate observations into statistically independentmulti-dimensional sources, where dependencies only exist withinthese subspaces but not between them. So far ISA algorithms havemostly been described in the context of known group sizes. Here, weextend a previously proposed ISA algorithm based on joint block diagonalizationof 4-th order cumulant matrices to separate subspacesof unknown sizes. Further automated interpretation of the demixedsources then requires a means of recovering the subspace structurewithin them, and we propose two distinct methods for this. We thenapply the method to a novel application field, namely clustering ofmetabolites, which seems to be well-fit to the ISA model. We areable to successfully identify dependencies between metabolites thatcould not be recovered using conventional methods.