Subspace clustering identifies multiple feature subspaces embedded in a dataset together with the underlying sample clusters. When applied to omic data, subspace clustering is a challenging task, as additional problems have to be addressed: the curse of dimensionality, the imperfect data quality and cluster separation, the presence of multiple subspaces representative of divergent views of the dataset, and the lack of consensus on the best clustering method. First, we propose a computational method discover to perform subspace clustering on tabular high dimensional data by maximizing the internal clustering score (i.e. cluster compactness) of feature subspaces. Our algorithm can be used in both unsupervised and semi-supervised settings. Secondly, by applying our method to a large set of omic datasets (i.e. microarray, bulk RNA-seq, scRNA-seq), we show that the subspace corresponding to the provided ground truth annotations is rarely the most compact one, as assumed by the methods maximizing the internal quality of clusters. Our results highlight the difficulty of fully validating subspace clusters (justified by the lack of feature annotations). Tested on identifying the ground-truth subspace, our method compared favorably with competing techniques on all datasets. Finally, we propose a suite of techniques to interpret the clustering results biologically in the absence of annotations. We demonstrate that subspace clustering can provide biologically meaningful sample-wise and feature-wise information, typically missed by traditional methods.
Optimization algorithm for omic data subspace clustering
