Machine-learned summary statistics for Bayesian inference of systems biology–model parameters: Opportunities and challenges

Abstract

Mechanistic systems biology models can capture complex dynamic interactions, but their accuracy often relies on parameter inference from high-dimensional, noisy data with corresponding intractable likelihoods. Approximate Bayesian computation (ABC) avoids likelihood evaluation by comparing simulated and observed data via low-dimensional summary statistics. However, effective selection of these summaries remains a limitation. Recent advances in machine learning offer algorithmic approaches to the selection of informative summaries, improving parameter identifiability, and reducing computational cost. Machine learning of summaries, however, introduces new challenges. We survey summary selection techniques for ABC, discuss how automated summaries can enhance parameter identifiability and inference efficiency, discuss algorithmic trade-offs in informativeness, tractability, and interpretability, and highlight strategies to ensure reliable inference. Through biological case studies, we review recently developed methods for selecting summaries. Finally, we outline challenges and future directions for leveraging machine-learned summaries to support ABC as a powerful and transparent tool for parameter inference in systems biology.