Henrik Mannerström

A probabilistic generative model for quantification of DNA modifications enables analysis of demethylation pathways

We present a generative model, Lux, to quantify DNA methylation modifications from any combination of bisulfite sequencing approaches, including reduced, oxidative, TET-assisted, chemical-modification assisted, and methylase-assisted bisulfite sequencing data. Lux models all cytosine modifications (C, 5mC, 5hmC, 5fC, and 5caC) simultaneously together with experimental parameters, including bisulfite conversion and oxidation efficiencies, as well as various chemical labeling and protection steps. We show that Lux improves the quantification and comparison of cytosine modification levels and that Lux can process any oxidized methylcytosine sequencing data sets to quantify all cytosine modifications. Analysis of targeted data from Tet2-knockdown embryonic stem cells and T cells during development demonstrates DNA modification quantification at unprecedented detail, quantifies active demethylation pathways and reveals 5hmC localization in putative regulatory regions.

SCHiRM: Single Cell Hierarchical Regression Model to detect dependencies in read count data

Motivation Modern single cell RNA sequencing (scRNA-seq) technologies have made it possible to measure the RNA content of individual cells. The scRNA-seq data provide us with detailed information about the cellular states but, despite several pioneering efforts, it remains an open research question how regulatory networks could be inferred from these noisy discrete read count data. Results Here, we introduce a hierarchical regression model which is designed for detecting dependencies in scRNA-seq and other count data. We model count data using the Poisson-log normal distribution and, by means of our hierarchical formulation, detect the dependencies between genes using linear regression model for the latent, cell-specific gene expression rate parameters. The hierarchical formulation allows us to model count data without artificial data transformations and makes it possible to incorporate normalization information directly into the latent layer of the model. We test the proposed approach using both simulated and experimental data. Our results show that the proposed approach performs better than standard regression techniques in parameter inference task as well as in variable selection task. Availability An implementation of the method is available at https://github.com/jeintos/SCHiRM. Contact jukka.intosalmi@aalto.fi, harri.lahdesmaki@aalto.fi