End-to-end differentiable molecular mechanics force field construction

Yuanqing Wang, Josh Fass, and John D. Chodera
Chemical Science 13:12016, 2022 [DOI] [arXiv] [pytorch code] [JAX code]

Molecular mechanics force fields have been a workhorse for computational chemistry and drug discovery. Here, we propose a new approach to force field parameterization in which graph convolutional networks are used to perceive chemical environments and assign molecular mechanics (MM) force field parameters. The entire process of chemical perception and parameter assignment is differentiable end-to-end with respect to model parameters, allowing new force fields to be easily constructed from MM or QM force fields, extended, and applied to arbitrary biomolecules.

Teaching free energy calculations to learn from experimental data

Marcus Wieder, Josh Fass, and John Chodera
[bioRxiv] [code] [data]

We show, for the first time, how alchemical free energy calculations can be used to not only compute free energy differences between small molecules involving covalent bond rearrangements in systems treated entirely with quantum machine learning potentials, but that these calculations have the capacity to learn to efficiently generalize from conditioning on experimental free energy data.

Bayesian inference-driven model parameterization and model selection for 2CLJQ fluid models

Owen C Madin, Simon Boothroyd, Richard A Messerly, John D Chodera, Josh Fass, and Michael R Shirts
Preprint ahead of publication: [arXiv]

Here, we show how Bayesian inference can be used to automatically perform model selection and fit parameters for a molecular mechanics force field.

Fitting quantum machine learning potentials to experimental free energy data: Predicting tautomer ratios in solution

Marcus Wieder, Josh Fass, and John D. Chodera
Chemical Science, in press [bioRxiv] [code]

We demonstrate, for the first time, how alchemical free energy calculations can performed on systems simulated entirely with quantum machine learning potentials and how these potentials can be retrained on experimental free energies to generalize to new molecules from limited training data. We apply this approach to a difficult problem in small molecule drug discovery: Predicting accurate tautomer ratios in solution.

Capturing non-local through-bond effects when fragmenting molecules for quantum chemical torsion scans

Chaya D. Stern, Christopher I. Bayly, Daniel G. A. Smith, Josh Fass, Lee-Ping Wang, David L. Mobley, and John D. Chodera
Preprint ahead of submission
[bioRxiv] [GitHub code] [GitHub data] [Metadata] [OSF]

Towards chemical accuracy for alchemical free energy calculations with hybrid physics-based machine learning / molecular mechanics potentials

Dominic A. Rufa, Hannah E. Bruce Macdonald, Josh Fass, Marcus Wieder, Patrick B. Grinaway, Adrian E. Roitberg, Olexandr Isayev, and John D. Chodera.
Preprint ahead of submission.
[bioRxiv] [GitHub]

In this first use of hybrid machine learning / molecular mechanics (ML/MM) potentials for alchemical free energy calculations, we demonstrate how the improved modeling of intramolecular ligand energetics offered by the quantum machine learning potential ANI-2x can significantly improve the accuracy in predicting kinase inhibitor binding free energy by reducing the error from 0.97~kcal/mol to 0.47~kcal/mol, which could drastically reduce the number of compounds that must be synthesized in lead optimization campaigns for minimal additional computational cost.

Graph nets for partial charge prediction

Yuanqing Wang, Josh Fass, Chaya D. Stern, Kun Luo, and John D. Chodera.
Preprint ahead of publication. [arXiv] [GitHub]

Graph convolutional and message-passing networks can be a powerful tool for predicting physical properties of small molecules when coupled to a simple physical model that encodes the relevant invariances. Here, we show the ability of graph nets to predict partial atomic charges for use in molecular dynamics simulations and physical docking.

Toward learned chemical perception of force field typing rules

Camila Zanette, Caitlin C. Bannan, Christopher I. Bayly, Josh Fass, Michael K. Gilson, Michael R. Shirts, John Chodera, and David L. Mobley
Journal of Chemical Theory and Computation, 15:402, 2019. [DOI] [ChemRxiv] [GitHub]

We show how machine learning can learn typing rules for molecular mechanics force fields within a Bayesian statistical framework.

Quantifying configuration-sampling error in Langevin simulations of complex molecular systems

Josh Fass, David Sivak , Gavin E. Crooks, Kyle A. Beauchamp, Benedict Leimkuhler, and John Chodera.
Entropy 20:318, 2018. [DOI] [PDF] [GitHub] [bioRxiv preprint]

Molecular dynamics simulations necessarily use a finite timestep, which introduces error or bias in the sampled configuration space density that grows rapidly with increasing timestep. For the first time, we show how to compute a natural measure of this error---the KL divergence---in both phase and configuration space for a widely used family of Langevin integrators, and show that VRORV is generally superior for simulation of molecular systems.

Binding Modes of Ligands Using Enhanced Sampling (BLUES): Rapid Decorrelation of Ligand Binding Modes Using Nonequilibrium Candidate Monte Carlo

Samuel Gill, Nathan M. Lim, Patrick Grinaway, Ariën S. Rustenburg, Josh Fass, Gregory Ross, John D. Chodera, and David Mobley.
Journal of Physical Chemistry B 22:5579, 2018. [DOI] [ChemRxiv] [GitHub]

Nonequilibrium candidate Monte Carlo can be used to accelerate the sampling of ligand binding modes by orders of magnitude over instantaneous Monte Carlo.