Improving Bayesian Networks for single cell transcriptomic data

Identification: Chen, Shuonan


Description

Improving Bayesian Networks for single cell transcriptomic data

Shuonan Chen1, Jessica Mar1*

1Department of Systems & Computational Biology, Albert Einstein College of Medicine

*Corresponding author

A fundamental fact in biology states that genes do not operate in isolation, and yet, network-based methods that infer multi-gene effects for single cell transcriptomic data have been slow to emerge. Current network approaches for population cells are typically based on co-expression networks that fail to capture cellular heterogeneity. Our study investigates the utility of a Probabilistic Graphical Model (PGM) for representing single cell gene regulatory networks. Since these approaches capture the uncertainty in gene-gene relationships, they may provide deeper insights into cell variability and the transcriptional regulatory processes. As a test of standard practice, we selected five commonly-used network methods and investigated their capacity to infer gene regulatory networks for single cell expression data. Standard evaluation metrics against an appropriate reference showed that all these methods consistently performed poorly, while for in silico simulated data, variable performance was observed amongst them. These results suggest that optimized methods for more specific network representation are needed for single cell data.

Bayesian networks (BNs) are flexible PGMs that represent a promising method to model gene networks in single cells. However, just like other network modeling methods, learning BNs from continuous datasets usually is heavily based on assumptions of Normality in the distribution of the data. This assumption is restrictive since single cell data follows a range of distributions, including those that are non-Normal. To address these single cell-specific challenges, our work focuses on new modifications to the current BNs learning algorithms, starting with incorporating mixture models for gene expression distributions, to expand the Normality assumption. Our goal is to derive new approaches that yield a more accurate and interpretable representation of gene-gene relationships in single cells, and to detect new interactions and cellular sub-states.

Credits

Credits: None available.

You must be logged in and own this product in order to post comments.