Selecting suitable SBI algorithms for real-world parameter estimation

Determine a principled method for selecting a suitable simulation-based inference (SBI) algorithm to estimate model parameters from real-world observational datasets in computational biology, where likelihoods are intractable and models may be misspecified. The selection procedure should enable practitioners to choose among available SBI approaches (e.g., ABC, BSL, NPE, SNLE and robust variants) in a way that balances computational efficiency with estimation accuracy under realistic data conditions.

Background

The paper surveys statistical and neural SBI methods used when likelihoods are intractable, including Approximate Bayesian Computation (ABC), Bayesian Synthetic Likelihood (BSL), Neural Posterior Estimation (NPE), and Neural Likelihood Estimation (NLE). Real-world biological datasets often exhibit model misspecification, noise, and lack ground-truth posteriors, complicating method selection.

The authors note that while neural SBI can be highly simulation-efficient, it may yield biased estimates with real-world data, whereas statistical SBI can be more accurate given sufficient simulations. They explicitly highlight that the community lacks clear, validated guidance for choosing an appropriate SBI algorithm in real-world contexts.