The performance of ANI‐ML potentials for ligand‐n(H₂O) interaction energies and estimation of hydration free energies from end‐point MD simulations

Abstract

Here, we investigate the performance of "Accurate NeurAl networK engINe for Molecular Energies" (ANI), trained on small organic compounds, on bulk systems including non-covalent interactions and applicability to estimate solvation (hydration) free energies using the interaction between the ligand and explicit solvent (water) from single-step MD simulations. The method is adopted from ANI using the Atomic Simulation Environment (ASE) and predicts the non-covalent interaction energies at the accuracy of wb97x/6-31G(d) level by a simple linear scaling for the conformations sampled by molecular dynamics (MD) simulations of ligand-n(H 2 O) systems. For the first time, we test ANI potentials' abilities to reproduce solvation free energies using linear interaction energy (LIE) formulism by modifying the original LIE equation.Our results on $250 different complexes show that the method can be accurate and have a correlation of R 2 = 0.88-0.89 (MAE <1.0 kcal/mol) to the experimental solvation free energies, outperforming current end-state methods. Moreover, it is competitive to other conventional free energy methods such as FEP and BAR with 15-20 Â fold reduced computational cost.

Benchmarking ANI potentials as a rescoring function and screening FDA drugs for SARS-CoV-2 Mpro

“…A similar protocol to our previous works has been used in MD simulations [ 41 , 42 , 47 – 53 ]. Briefly, ligands were first optimized at B3LYP/6-31++G(d,p), and Merz–Kollman (MK) electrostatic potential (ESP) charges were computed at HF/6-31G* level using Gaussian 16 software.…”

“…[ 41 ]) with default and empirical parameters of 0.127 and −5.11, respectively. As reported by Akkus et al [ 41 ], these parameters were optimized SARS-CoV-2 M pro and have been successfully used in other studies [ 42 , 47 , 48 , 55 , 56 ].…”

Does a Machine-Learned Potential Perform Better Than an Optimally Tuned Traditional Force Field? A Case Study on Fluorohydrins

“…There are currently two models of the ANI family that may be applied to a broad spectrum of problems in chemical sciences. ANI-1ccx, which is trained to reproduce CCSD­(T*)/CBS, is the ANI NNP with the highest level of accuracy, ,,, although it can only simulate organic molecules containing elements H, C, N, and O. ANI-2x, on the other hand, is trained to reproduce ωB97X/6-31G* and has also shown promising results in some applications. ,, ANI-2x has the advantage of extending the chemical space covered by ANI-1ccx to organic molecules also containing elements F, S, and Cl, the addition of which is essential for day-to-day use in common applications. Because of this, ANI-2x covers a chemical space that encompasses 90% of druglike molecules and is the only ANI model that can be employed to simulate the γ-fluorohydrins considered in this study due to the presence of the fluorine atoms.…”

“…NNPs from the ANI family were designed to model neutral molecules in the singlet spin state, a limitation that has been recently circumvented by employing ML-based corrections to quantum methods. − ANI-2x in specific was trained using a data set comprising 8.9 million molecular equilibrium and nonequilibrium conformations of fragments, and it includes active learning refinements to torsional profiles, nonbonded interactions, and bulk water behavior. This NNP has been the focus of various benchmarks and applications in recent years. − Furthermore, as our traditional FF we used the General Amber FF (GAFF), which is commonly employed in the modeling of druglike molecules . Finally, our optimally tuned FF was an optimized version of the GAFF in which the bonded parameters were optimized to the same level of theory that ANI-2x was trained to reproduce (ωB97X/6-31G*).…”