Highly differentiated brain structures with distinctly different phenotypes are closely correlated with the unique combination of gene expression patterns. Using a genome-wide in situ hybridization image dataset released by Allen Mouse Brain Atlas, we present a data-driven method of dictionary learning and sparse coding. Our results show that sparse coding can elucidate patterns of transcriptome organization of mouse brain. A collection of components obtained from sparse coding display robust region-specific molecular signatures corresponding to the canonical neuroanatomical subdivisions including fiber tracts and ventricular systems. Other components revealed finer anatomical delineation of domains previously considered homogeneous. We also build an open-access informatics portal that contains the detail of each component along with its ontology and expressed genes. This portal allows intuitive visualization, interpretation and explorations of the transcriptome architecture of a mouse brain.
Keywords: Data-driven gene clustering; Mouse brain anatomy; Sparse coding; Transcriptome.