Estimation of sparse directed acyclic graphs for multivariate counts data.

Area of Science:

• Bioinformatics
• Statistical Genetics
• Computational Biology

Background:

• High-throughput sequencing data are count-based and often non-normally distributed.
• Accurate network inference is crucial for understanding complex biological systems like gene regulation.

Purpose of the Study:

• To develop a novel statistical framework for estimating sparse directed acyclic graphs (DAGs) from multivariate count data.
• To address the limitations of existing methods that assume normal distributions or use log-transformations.

Main Methods:

• Proposed an L1-penalized likelihood framework assuming a Poisson log-normal distribution for count data.
• Developed an efficient search algorithm using iterative optimization to estimate adjacency and variance matrices.
• Compared performance against multivariate normal, log-transformation, and rank-based PC methods.

Main Results:

• The proposed Poisson log-normal model significantly outperformed existing methods in simulations, especially for sparse or hub network structures.
• Demonstrated superior accuracy in estimating gene regulatory networks from real-world ovarian cancer data.

Conclusions:

• The L1-penalized Poisson log-normal model provides a robust and efficient approach for inferring network structures from high-throughput count data.
• This method enhances the analysis of biological networks, particularly in genomics and cancer research.