Joint estimation of causal effects from observational and intervention gene expression data

Area of Science:

• Computational Biology
• Systems Biology
• Genomics

Background:

• Transcriptomic data analysis is crucial for understanding gene regulatory networks.
• Current methods using graphical Gaussian models on observational data yield undirected graphs, limiting causal inference.
• There is a need for methods that can accurately estimate causal relationships among genes.

Purpose of the Study:

• To develop an improved method for estimating causal effects among genes.
• To integrate observational transcriptomic data with intervention data (gene knock-outs/knock-downs).
• To accommodate complex intervention designs, including partial or multiple gene perturbations.

Main Methods:

• Developed a Markov chain Monte Carlo algorithm within the causal Gaussian Bayesian networks framework.
• Employed a Mallows proposal model and analytical likelihood maximization for posterior sampling.
• Validated the method using simulated data and data from the DREAM 2007 challenge.

Main Results:

• The proposed algorithm demonstrated high accuracy in estimating causal effects, often surpassing alternative methods.
• Multiple gene knock-outs provided more informative data than single knock-outs.
• Observational data alone is insufficient for estimating causal gene orderings.

Conclusions:

• The novel algorithm effectively estimates causal regulatory relationships by incorporating intervention experiments.
• Intervention data significantly enhances the accuracy of causal effect estimation compared to observational data alone.
• The developed method offers a flexible and accurate approach for gene regulatory network inference.