Likelihood free inference for Markov processes: a comparison.

Area of Science:

• Computational Biology
• Statistical Inference
• Systems Biology

Background:

• Bayesian inference with intractable likelihoods is crucial in modern statistical modeling.
• Approximate Bayesian Computation (ABC) and likelihood-free Markov chain Monte Carlo (MCMC) are common but computationally intensive methods.
• Stochastic kinetic models in systems biology often feature intractable likelihoods, necessitating simulation-based inference.

Purpose of the Study:

• To compare the computational cost and inferential efficiency of ABC and likelihood-free MCMC.
• To evaluate these methods for parameter inference in stochastic kinetic models.
• To assess the impact of different observation regimes and measurement error on inference accuracy.

Main Methods:

• Comparative analysis of Approximate Bayesian Computation (ABC) and likelihood-free Markov chain Monte Carlo (MCMC).
• Application to parameter inference for stochastic kinetic models, including Lotka-Volterra and Schlögl systems.
• Simulation-based inference under varying observation scenarios (full/partial species, time course data) and measurement error assumptions.

Main Results:

• Both ABC and likelihood-free MCMC present trade-offs between computational expense and inferential efficiency.
• The choice of observation regime significantly impacts the performance and accuracy of both inference methods.
• Measurement error introduces challenges, particularly for models exhibiting complex dynamics like the Schlögl system.

Conclusions:

• Understanding the computational costs and efficiencies of ABC and likelihood-free MCMC is vital for selecting appropriate inference strategies.
• Careful consideration of data availability and quality (observation regimes, measurement error) is necessary for reliable parameter inference in systems biology.
• Further research can optimize these simulation-based methods for complex biological models.