A hybrid constrained continuous optimization approach for optimal causal discovery from biological data.

Area of Science:

• Causal inference and graph theory
• Computational biology and bioinformatics

Background:

• Causal effect discovery from observational data is crucial for scientific prediction.
• Existing causal discovery algorithms often fail on real-world biological datasets due to unmet data requirements.
• Benchmarking on synthetic data limits the evaluation of algorithms on realistic scenarios.

Purpose of the Study:

• To construct a large-scale, real-life dataset with known causal truth for evaluating causal discovery methods.
• To comprehensively benchmark existing causal discovery algorithms on this dataset.
• To propose and validate a novel hybrid algorithm for accurate causal effect estimation.

Main Methods:

• Construction of a large-scale, real-life biological dataset.
• Comprehensive benchmarking of various causal discovery algorithms, including the PC algorithm.
• Development and implementation of PC-NOTEARS (PCnt), a hybrid algorithm integrating PC output with NOTEARS optimization.

Main Results:

• The PC algorithm demonstrated high accuracy in estimating causal structure and direction.
• PC-NOTEARS (PCnt) achieved superior performance across structural and effect size metrics.
• PCnt effectively combines the strengths of the PC algorithm and NOTEARS for accurate causal effect estimation.

Conclusions:

• PC-NOTEARS (PCnt) represents a significant advancement in causal discovery from observational data.
• The developed dataset and benchmarking provide valuable resources for the causal inference community.
• Accurate causal effect estimation is now more feasible using PCnt on real-world biological data.