Journal: bioRxiv

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

doi: 10.1101/2020.07.29.227959

Figure Lengend Snippet: Top: We can construct a hybrid machine learning / molecular mechanics (ML/MM) potential that treats ligand intramolecular interactions with higher accuracy than achievable by MM potentials by subtracting the MM energy of the ligand in vacuum and adding the more accurate ML energy of the ligand in vacuum. Here, the MM model uses the Open Force Field Initiative [ http://openforcefield.org ] OpenFF 1.0.0 (“Parsley”) small molecule force field , AMBER14SB , and TIP3P while the ML model uses the ANI-2x neural network potential parameterized using DFT ω B97X/6-31G* QM calculations. Bottom: The ANI-2x ML potential first computes radial and angular features for each atom and then sums energetic contributions by atom using deep learning models specific to each element-element pair.

Article Snippet: Further study will indicate whether other MM force fields—including the GAFF force field [ , ] and more recent iterations of the OpenFF force fields—generally provide sufficient phase-space overlap with ML models for ML/MM corrections to remain computationally convenient and accurate with respect to experiment.

Techniques: Construct

Journal: bioRxiv

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

doi: 10.1101/2020.07.29.227959

Figure Lengend Snippet: (A) Absolute binding free energies for the MM small molecule OpenFF 1.0.0 (“Parsley”) force field used with AMBER14SB and TIP3P water computed from relative free energy calculations estimated using perses 0.7.1 [ http://github.com/choderalab/perses ] and the maximum-likelihood estimator to integrate estimates from redundant transformations in the relative alchemical transformation network. The same redundant network of relative alchemical transformations used in was used here. (B) Absolute free energies (ΔG) corrected to ML/MM (using ANI-2x for the ML model) using the nonequilibrium correction scheme depicted in . (C) Relative MM binding free energies (ΔΔG) for computed relative free energy transformation edges, with correction using MLE. (D) Relative ML/MM binding free energies obtained from differences in the corrected absolute binding free energy estimates (top right). Blue scatter points are MM results, and orange are ML/MM results. Dark and light grey shaded regions indicate the region of ±0.5 and ±1.0 kcal mol −1 error respectively. Vertical error bars (which appear smaller than the symbols) show one standard deviation in the free energy, calculated by MBAR, while the experimental error bar of 0.18 kcal mol −1 is used . Statistical analysis was performed using the Arsenic package [ http://github.com/openforcefield/arsenic ], with 95% confidence intervals calculated by bootstrap analysis. For all plots, an additive constant was added to all computed values, such that the mean computed value is equal to the mean experimental value, such as to minimise the RMSE as in .

Article Snippet: Further study will indicate whether other MM force fields—including the GAFF force field [ , ] and more recent iterations of the OpenFF force fields—generally provide sufficient phase-space overlap with ML models for ML/MM corrections to remain computationally convenient and accurate with respect to experiment.

Techniques: Binding Assay, Transformation Assay, Standard Deviation