Open Access

Daniel Hüebschmann, Nils Kurzawa, Sebastian Steinhauser, Philipp Rentzsch, Stephen Krämer, Carolin Andresen, Jeongbin Park, Roland Eils, Matthias Schlesner, Carl Herrmann

• Metazoans are crucially dependent on multiple layers of gene regulatory mechanisms which allow them to control gene expression across developmental stages, tissues and cell types. Multiple recent research consortia have aimed to generate comprehensive datasets to profile the activity of these cell type- and condition-specific regulatory landscapes across many different cell lines and primary cells. However, extraction of genes or regulatory elements specific to certain entities from these datasets remains challenging. We here propose a novel method based on non-negative matrix factorization for disentangling and associating huge multi-assay datasets including chromatin accessibility and gene expression data. Taking advantage of implementations of NMF algorithms in the GPU CUDA environment full datasets composed of tens of thousands of genes as well as hundreds of samples can be processed without the need for prior feature selection to reduce the input size. Applying this framework toMetazoans are crucially dependent on multiple layers of gene regulatory mechanisms which allow them to control gene expression across developmental stages, tissues and cell types. Multiple recent research consortia have aimed to generate comprehensive datasets to profile the activity of these cell type- and condition-specific regulatory landscapes across many different cell lines and primary cells. However, extraction of genes or regulatory elements specific to certain entities from these datasets remains challenging. We here propose a novel method based on non-negative matrix factorization for disentangling and associating huge multi-assay datasets including chromatin accessibility and gene expression data. Taking advantage of implementations of NMF algorithms in the GPU CUDA environment full datasets composed of tens of thousands of genes as well as hundreds of samples can be processed without the need for prior feature selection to reduce the input size. Applying this framework to multiple layers of genomic data derived from human blood cells we unravel mechanisms of regulation of cell type-specific expression in T-cells and monocytes.…