Control of Gene Regulatory Networks Using Bayesian Inverse Reinforcement Learning

Area of Science:

• Systems Biology
• Computational Biology
• Bioinformatics

Background:

• Gene regulatory networks (GRNs) control gene expression, but their complex dynamics are challenging to manipulate.
• Existing methods for GRN control often require full knowledge of intervention costs, limiting practical application.
• Learning cost functions from experimental data is crucial for realistic GRN intervention strategies.

Purpose of the Study:

• To develop a novel Bayesian Inverse Reinforcement Learning (BIRL) approach for inferring immediate cost functions in Boolean GRNs.
• To address the challenge of unknown intervention costs in controlling gene expression states.
• To enable data-driven identification of undesirable genes and states within GRNs.

Main Methods:

• Utilized a Partially-Observed Boolean Dynamical System (POBDS) model for noisy gene expression measurements.
• Employed the Boolean Kalman Smoother (BKS) algorithm to infer hidden Boolean states from expression data.
• Combined BIRL with Q-learning for efficient quantification of the immediate cost function.

Main Results:

• Successfully demonstrated the BIRL approach on two GRN models: a melanoma WNT5A network and a p53-MDM2 network.
• The methodology effectively learned the cost of undesirable states and identified critical genes without prior cost information.
• Validated the performance of the state-feedback controller guided by the learned cost function.

Conclusions:

• The proposed BIRL methodology provides a robust framework for learning cost functions in partially observable Boolean GRNs.
• This data-driven approach enhances the ability to control gene expression by identifying and mitigating undesirable states.
• The findings have significant implications for precision medicine and synthetic biology applications involving complex gene regulatory systems.