P02-05
NNP-based Force Field Optimization to Improve RBFEP Performance
Junya YAMAGISHI *, Yunoshin TAMURA
Preferred Networks
( * E-mail: jyamagishi@preferred.jp )
Accurate binding affinity prediction techniques, such as free energy perturbation (FEP), have become very effective tools for enhancing small molecule drug discovery. However, since FEP is based on molecular dynamics simulations, it is known that the accuracy of binding affinity prediction depends on the accuracy of the molecular force field (FF). Our validation studies have shown that the accuracy of the existing FFs for complicated drug-like compounds is not sufficient. It is necessary to improve the accuracy of FF for more accurate prediction of binding affinity.
There are several challenges to optimizing FF parameters. First, each FF parameter, including bonded terms and partial charges, interferes with each other. This means that there are no universal parameters that can be applied to multiple molecules: ideally, tailor-made FF parameters should be made for each molecule.
The second is related to accurate energy and force calculations used as a reference for optimizing FF parameters, which are typically performed by quantum mechanical (QM) calculations. QM calculations of drug-like compounds (Mw < 500) with thousands of conformations are very time-consuming and impractical when tailoring FF parameters.
To overcome these challenges, we applied a new protocol to tailor-make FF parameters for each drug-like molecule using neural network potential (NNP) instead of QM calculations. Preferred Potential (PFP) [1] was used as the NNP, whose accuracy has been verified for drug-like molecules. Because PFP can predict accurate energies and forces thousands of times faster than QM calculations, we were able to generate tailor-made FF parameters for each full-sized drug-like molecule. In this presentation, we will show the accuracy of PFP for drug-like molecules and the performances of FF parameter optimization using PFP. We will also show comparisons of the results of relative binding free energy perturbation (RBFEP) performed with our service, called P-FEP[2], using existing and optimized FF parameters.
[1] Takamoto, S., Shinagawa, C., Motoki, D. et al. Towards universal neural network potential for material discovery applicable to arbitrary combination of 45 elements. Nat Commun 13, 2991 (2022).
[2] https://tech.preferred.jp/ja/blog/pfep-launch/