LEARNING LOCAL DIRECTED ACYCLIC GRAPHS BASED ON MULTIVARIATE TIME SERIES DATA.

Area of Science:

• Computational Biology
• Genomics
• Network Inference

Background:

• Multivariate time series (MTS) data, like gene expression, are crucial for understanding dynamic biological systems.
• These datasets reveal causal dependencies among variables, essential for biological pathway elucidation.
• Existing methods for learning causal structures from MTS data have limitations in efficiency and accuracy.

Purpose of the Study:

• To introduce a computationally efficient algorithm for learning directed acyclic graphs (DAGs) from MTS data.
• To focus on accurately reconstructing the local causal structure around a target variable.
• To extend existing causal discovery algorithms to handle dependent observations in time series.

Main Methods:

• Developed the time series PCD-PCD (tsPCD-PCD) algorithm, an iterative approach learning parents, children, and descendants.
• Utilized the temporal order of variables to orient edges in the graph.
• Employed composite likelihood ratio tests (CLRTs) for conditional independence testing, with analysis of asymptotic distributions.

Main Results:

• The tsPCD-PCD algorithm is guaranteed to recover the true DAG structure under specific conditions (faithfulness, correct hypothesis rejection).
• Simulation studies demonstrated the validity and good performance of CLRTs, even with small sample sizes.
• The tsPCD-PCD algorithm showed superior performance in recovering local graph structures compared to the original PCD-PCD algorithm.

Conclusions:

• The tsPCD-PCD algorithm provides an efficient and accurate method for causal discovery in MTS data.
• The approach is validated through simulations and demonstrated on real-world gene expression data.
• This method advances the analysis of dynamic biological systems by improving causal network inference.