Discovering causal interactions using Bayesian network scoring and information gain | JoVE Visualize

Area of Science:

Genetics
Bioinformatics
Causal Inference

Background:

Standard statistical methods struggle with learning discrete causal relationships and interactions.
Existing methods for learning discrete interactions, like those in genome-wide association studies (GWAS), often fail to distinguish true interactions from non-interacting causes with strong individual effects.
The MBS-IGain algorithm uses Bayesian networks and information gain but requires marginal effects for interactions involving more than two causes.

Purpose of the Study:

To develop a novel algorithm, Exhaustive-IGain, that addresses the limitations of MBS-IGain by performing an exhaustive search.
To evaluate the performance of Exhaustive-IGain against MBS-IGain on simulated datasets, particularly for interactions lacking marginal effects.
To apply Exhaustive-IGain to real-world data to investigate clinical variable interactions affecting breast cancer survival.

Main Methods:

Developed Exhaustive-IGain, an algorithm that combines Bayesian network learning and information gain with an exhaustive search strategy.
Compared Exhaustive-IGain with MBS-IGain using simulated low-dimensional datasets.
Datasets were designed to include interactions with and without marginal effects, focusing on 3 and 4-cause interactions.

Main Results:

Exhaustive-IGain and MBS-IGain showed similar performance on datasets with marginal effects.
Exhaustive-IGain significantly outperformed MBS-IGain on datasets featuring 3 and 4-cause interactions without marginal effects.
Application of Exhaustive-IGain to breast cancer survival data yielded results consistent with expert clinical judgment.

Conclusions:

The combination of information gain and Bayesian network scoring, coupled with an exhaustive search, enables the discovery of higher-order interactions even without marginal effects.
Exhaustive-IGain demonstrates effectiveness and utility when applied to real biological and clinical datasets.
This method advances causal discovery for complex interactions in high-dimensional biological data.