Start of funding 01.01.2019

Scalable differential expression analysis for computational pipelines of the Human Cell Atlas

Prof. Dr. Fabian Theis
Technische Universität München
Dept. of Mathematik - Institute of computational biology (ICB)

Joshua Batson
Chan Zuckerberg Biohub



Several noise models and optimizers and implementations thereof have been proposed for differential expression analysis both for bulk and for single-cell RNA-seq. We present diffxpy, a software framework designed to integrate multiple noise models and optimizers for differential expression analysis which addresses three current issues in the field: 1) Computation speed and memory requirements on large data sets and single-cell-specific testing scenarios, 2) side-by-side comparison of noise models in a single framework (“model zoo”) to avoid confounding results with optimization failures and 3) a novel model to perform batch correction in cases of perfect confounding which integrates into multiple previously proposed models. Diffxpy is based on the high-performance computing framework tensorflow and provides a software basis that easily allows the integration of noise models so that it can easily adapt to changes in the field in the future. We collaborate in this project with the team of the Chan Zuckerberg BioHub in San Fransico to adapt the software and models to the needs of the Human Cell Atlas community and to organize its deployment to this large user base.