Automatic forward model parameterization with Bayesian inference of conformational populations

Area of Science:

• Computational Chemistry
• Molecular Dynamics
• Biophysics

Background:

• Accurate theoretical predictions of molecular structural ensembles rely on precise forward models (FMs).
• Existing methods struggle to reconcile simulated molecular dynamics with experimental data, especially with sparse or noisy observations.
• Refining empirical FM parameters is crucial for improving the accuracy of these predictions.

Purpose of the Study:

• To enhance the Bayesian Inference of Conformational Populations (BICePs) algorithm for optimizing forward model (FM) parameters.
• To introduce and evaluate novel methods for refining empirical FM parameters.
• To facilitate force field validation and the training of neural network-based FMs.

Main Methods:

• Developed two novel methods for optimizing FM parameters within the BICePs framework.
• Method 1: Treats FM parameters as nuisance parameters, integrating over them in the full posterior distribution.
• Method 2: Employs variational minimization of the BICePs score, representing the free energy of experimental restraints.
• Incorporated improved likelihood functions for handling experimental outliers.

Main Results:

• Successfully refined parameters modulating the Karplus relation, essential for predicting nuclear spin-spin coupling constants (J-couplings).
• Validated the approach using a toy model and human ubiquitin, predicting six sets of Karplus parameters for various couplings.
• Demonstrated the framework's generalizability to any differentiable FM, including neural network-based models.

Conclusions:

• The enhanced BICePs algorithm effectively refines FM parameters, improving the agreement between theoretical predictions and experimental measurements.
• This approach offers a robust method for force field validation and optimization.
• The generalized framework provides a promising direction for training and validating advanced neural network-based FMs.