Flexible model selection for mechanistic network models.

Area of Science:

• Computational Biology
• Network Science
• Statistical Modeling

Background:

• Network models are crucial in diverse fields, with statistical and mechanistic approaches being prominent.
• Mechanistic models offer advantages in incorporating domain knowledge and simulating interventions but often have intractable likelihoods.
• Model selection for mechanistic networks is less developed compared to statistical models.

Purpose of the Study:

• To propose a novel simulator-based procedure for mechanistic network model selection.
• To address the challenge of intractable likelihoods in mechanistic network models.
• To quantify the uncertainty associated with the selected mechanistic network model.

Main Methods:

• Developed a simulator-based model selection framework inspired by Approximate Bayesian Computation.
• Employed various learning algorithms, including the Super Learner, to enhance robustness.
• Leveraged the ability to forward simulate from mechanistic models to bypass likelihood computation.

Main Results:

• The proposed method accurately discriminates between competing mechanistic network models.
• The Super Learner approach reduces sensitivity to the choice of a specific learning algorithm.
• Demonstrated the framework's flexibility and wide applicability through simulation studies.

Conclusions:

• The simulator-based procedure provides a viable solution for mechanistic network model selection.
• The approach effectively handles models with intractable likelihoods, expanding research possibilities.
• Successfully applied the method to a yeast protein-protein interaction network, validating its practical utility.