Inferring structure and parameters of stochastic reaction networks with logistic regression

Area of Science:

• Systems Biology
• Computational Biology
• Epidemiology

Background:

• Modeling chemical reaction networks is complex due to challenges in identifying network structure and estimating reaction parameters.
• Stochastic reaction systems require robust methods for inferring network properties from observed data.

Purpose of the Study:

• To develop likelihood-based methods using multinomial logistic regression for inferring stoichiometries and network connectivity in stochastic reaction systems.
• To demonstrate the application of these methods in recovering network structures of various models and analyzing real-world epidemic data.

Main Methods:

• Utilized multinomial logistic regression on full time-series trajectories of stochastic reaction systems.
• Applied Bayesian logistic regression combined with differential-equation modeling for partially observable systems, exemplified by SIR models.

Main Results:

• Successfully recovered stoichiometric coefficients and network structures in models like Togashi-Kaneko, heat-shock protein networks, and SIR models.
• Demonstrated reliable recovery of core SIR parameters from synthetic COVID-19-like epidemic data, even with partial observability.

Conclusions:

• Likelihood-based tools, such as logistic regression, provide meaningful mechanistic insights into chemical reaction networks.
• The developed framework is effective for analyzing both synthetic and real-world epidemic dynamics, including complex scenarios like COVID-19 outbreaks.