(735b) Bayesian Uncertainty Quantification of Transferable Coarse-Grained Models of Monosaccharides

Carbohydrate-based materials are widely used in various applications due to their unique properties, such as solubility, biocompatibility, non-toxicity, biodegradability, and mechanical strength. To accelerate the design of novel carbohydrate-based glycomaterials with desired properties, the employment of accurate and robust all-atom and coarse-grained (CG) models used in molecular dynamics (MD) simulations is crucial. Specifically, ensuring the reliability of MD simulation predictions requires robust, True-Tightened force field (FF) parameters that not only consistently match physical predictions but even enhance our precision relative to what was known from experiments. In this work, we present transferable coarse-grained (CG) models of carbohydrates that can accurately predict their structural, chemical, and physical properties. To comprehensively assess and enhance the model's robustness by estimating uncertainty in both parameter and property predictions, we developed and employed a Bayesian Uncertainty Quantification (BUQ) framework. First, we used the Sobol sequence to perturb the final set of CG FF parameters and performed MD simulations to generate training data. This data was used to train a Gaussian Process Regression (GPR) model, which was used as a surrogate model during BUQ sampling. The BUQ sampling was performed using the Ensemble Slice Sampling (ESS) algorithm. The data obtained from the posterior distribution of parameters and properties was then used to evaluate the robustness of the parameters. Currently, we are employing these models to design novel glycomaterials for biomedical applications.