Interpretable AI for inference of causal molecular relationships from omics data

Area of Science:

• Bioinformatics
• Computational Biology
• Systems Biology

Background:

• Discovering molecular relationships in high-dimensional data is a key challenge in bioinformatics.
• Current machine learning and feature attribution models lack causal interpretation for biological networks.
• Understanding gene regulatory networks is crucial for deciphering complex diseases.

Purpose of the Study:

• To develop a causally interpretable method for identifying gene regulatory relationships.
• To introduce CIMLA (Counterfactual Inference by Machine Learning and Attribution Models) for analyzing differences in gene regulatory networks.
• To apply CIMLA to identify potential regulators in Alzheimer's disease.

Main Methods:

• Leveraging feature attribution models to estimate causal quantities reflecting direct variable influence.
• Proposing a precise definition for gene regulatory relationships based on counterfactual inference.
• Benchmarking CIMLA against leading methods using simulated data for robustness and accuracy.

Main Results:

• CIMLA demonstrates robustness to confounding variables and superior accuracy compared to existing methods.
• The tool successfully identified potential gene regulatory network differences between biological conditions.
• Analysis of an Alzheimer's disease dataset revealed several novel potential AD regulators.

Conclusions:

• CIMLA provides a causally interpretable framework for dissecting gene regulatory networks.
• The tool enhances the identification of condition-specific molecular relationships.
• CIMLA holds promise for advancing the understanding and treatment of complex diseases like Alzheimer's disease.