Hierarchical Bayesian calibration of mesoscopic models for ultrasound contrast agents from force spectroscopy data

Abstract

Ultrasound-guided drug and gene delivery (USDG) is a promising non-invasive approach for targeted therapeutic applications. Mechanical properties of encapsulated microbubbles (EMBs), which serve as contrast agents, strongly affect their specific interactions with ultrasound and are thus critical to the success and efficiency of USDG. Accurate calibration of high-fidelity particle-based models of EMB capsid mechanics is computationally challenging because direct Bayesian inference with dissipative particle dynamics (DPD) is prohibitively expensive. We employ a surrogate-accelerated Bayesian calibration workflow that combines deep neural network (DNN) surrogates, transitional Markov chain Monte Carlo sampling, and hierarchical regularization across EMB diameters. Using this framework, we develop two data-informed DPD models of commercial EMB agents, i.e., Definity and SonoVue, and perform inference of force field parameters based on published compression experiments for Definity and indentation experiments for SonoVue, each spanning three distinct diameters. The inferred posteriors show that key model parameters, such as the stretching stiffness and bending modulus, are consistently constrained by the available data. The presented methodology can be used to derive bespoke, data-informed models for a wide range of ultrasound contrast agents, including encapsulated gas vesicles, EMBs with diverse capsids consisting of lipids, proteins, or polymers, and functionalized with ligands.