Using Unlabeled Data to Discover Bivariate Causality with Deep Restricted Boltzmann Machines.

Area of Science:

• Microbiology
• Computational Biology
• Causal Inference

Background:

• Determining if treatments alter gut flora and metabolism is crucial.
• Causal inference from non-temporal observational data, especially discrete data, presents significant challenges.
• Existing methods for continuous data are computationally intensive, and some categorical data methods require extensive hyper-parameter tuning.

Purpose of the Study:

• To develop a novel, efficient causal inference method for discrete observational data.
• To address limitations of current state-of-the-art methods in terms of computational complexity and applicability to categorical data.
• To investigate drug confounding in human metagenome data.

Main Methods:

• A novel causal inference method based on the independence assumption between P(X) and P(Y|X).
• A semi-supervised approach estimating conditional probability P(Y|X) from labeled data and marginal probability P(X) from abundant unlabeled data.
• Validation on standard cause-effect pairs and biological network reconstruction benchmarks.

Main Results:

• The proposed method demonstrates high accuracy and computational efficiency compared to existing state-of-the-art techniques.
• Experimental results on biological network reconstruction show competitive performance.
• Successful application to a novel medical task analyzing drug confounding in human metagenome.

Conclusions:

• The novel semi-supervised causal inference method is effective for discrete observational data.
• The approach offers a computationally efficient and accurate alternative for complex biological and medical data analysis.
• This method advances the study of treatment effects on gut microbiome and metabolism.