Model-Based Causal Discovery for Zero-Inflated Count Data.

Area of Science:

• Statistics
• Bioinformatics
• Genomics

Background:

• Zero-inflated count data are prevalent across scientific disciplines, including social science, biology, and genomics.
• Existing causal discovery methods struggle to accommodate the excess zeros and overdispersion common in multivariate count data.

Purpose of the Study:

• Propose a novel zero-inflated generalized hypergeometric directed acyclic graph (ZiG-DAG) model for causal inference from observational zero-inflated count data.
• Develop a flexible framework capable of modeling diverse zero-inflated count data types and accommodating both linear and nonlinear causal relationships.

Main Methods:

• The ZiG-DAG model leverages a generalized hypergeometric probability distribution family for flexible data modeling.
• Causal structure identifiability is proven using a general technique applicable to count data.
• Score-based algorithms are employed for efficient causal structure learning.

Main Results:

• The proposed ZiG-DAG model demonstrates superior performance in discovering causal structures from observational zero-inflated count data compared to state-of-the-art methods.
• Extensive synthetic experiments and a real-world dataset with known ground truth validate the model's effectiveness.
• The method successfully reverse-engineered a gene regulatory network from single-cell RNA-sequencing data, showcasing practical utility.

Conclusions:

• The ZiG-DAG model offers a robust and flexible approach for causal discovery from complex zero-inflated count data.
• The identifiability proof and developed algorithms provide a strong foundation for future causal inference research in this domain.
• The model's application in bioinformatics highlights its potential for unraveling biological networks and driving scientific discovery.