Individualized causal discovery with latent trajectory embedded Bayesian networks.

Area of Science:

• Computational Biology
• Causal Inference
• Network Science

Background:

• Causal discovery methods often assume homogeneous data, leading to bias in heterogeneous populations.
• Existing approaches struggle with identifying causal structures from purely observational, cross-sectional data due to Markov equivalence.

Purpose of the Study:

• To develop a novel causal Bayesian network model (BN-LTE) that addresses data heterogeneity.
• To enable precise causal network inference at the observation level, improving upon population-level estimations.
• To achieve unique identifiability of causal structures from observational data, leveraging causal effect heterogeneity.

Main Methods:

• Embedding heterogeneous samples onto a low-dimensional manifold.
• Constructing Bayesian networks conditional on the learned embedding.
• Proving unique identifiability of the BN-LTE model under mild assumptions.

Main Results:

• BN-LTE demonstrates superior performance in causal structure learning.
• The model accurately infers observation-specific gene regulatory networks from observational data.
• The framework enhances estimation resolution from population to observation level.

Conclusions:

• BN-LTE effectively handles population heterogeneity in causal discovery.
• The model offers a uniquely identifiable approach to inferring causal networks from observational data.
• This method advances the field of causal inference in complex biological systems.