Application of the FMO prediction models to 3D protein structures predicted by AlphaFold2 and MD simulations.
Tomohiro SATO *1, Chiduru WATANABE1, Yoshio OKIYAMA2
1Center for Integrative Medical Sciences, RIKEN
2Graduate School of Human Development and Environment, Kobe University
Along with the experimentally determined structures, recent advances of structural prediction algorithms and molecular dynamics (MD) simulations have been rapidly increasing the amount of theoretically predicted 3D structures of biomolecules. Since these methods evaluate the physical stability of the predicted structures by conventional molecular mechanics, application of quantum mechanics (QM) methods to access the structures could improve the quality of the predicted structures. Currently, application of QM calculation to biomolecules is getting popular due to the development of methods to accelerate QM calculation for large molecular systems such as QM/MM or fragment molecular orbital (FMO) method. However, the computation cost is still high to routinely apply QM calculation to numerous protein structures generated by structural prediction methods and molecular dynamics. To enable the fast estimation of FMO calculation, we presented machine learning models to predict Inter-Fragment Interaction Energy (IFIE) between amino acid residues based on FMO results registered in FMODB in 2024. The FMO prediction models achieved high prediction performance of R2=0.914 on average and fast calculation speed which is roughly 500 times faster than that of actual FMO calculation. In this study, we apply the FMO prediction models to protein structures predicted by AlphaFold2 and MD snapshots to assess the applicability of the prediction model as the tool to evaluate theoretically generated structures. Validation using 500 AlphaFold2 structures with high confidence (pLDDT>98.5), 22 AlphaFold2 structures with low prediction confidence (pLDDT<30) and 991 MD snapshots of Trp-Cage protein showed that the FMO prediction models generally show excellent prediction performances (R2=0.96 in all three datasets), leaving some outliers in AlphaFold2 structures which FMO prediction models clearly underestimate their IFIE values. The details of prediction performance and characteristics of the outlier structures are to be discussed.
Acknowledgment
This work was partially supported by JSPS KAKENHI Grant Number 23K18192 and AMED BINDS Grant Number JP25ama121030. FMO calculations were performed using the supercomputer Fugaku (HPCI ID: hp250154).