Comparison of Machine Learning and Classical Force Field in Peptide Simulation

Compared with the classical force field, the new machine learning (ML) force field does not describe the energy of the molecule in a parameterized way, but uses the AI framework to learn the energy of the atom in the corresponding chemical environment, and compare the energy of each atom in the molecular structure. The local chemical environment is transformed into a descriptor, including information about the radial and angular distribution of neighboring atoms. These descriptors use neural networks to learn and predict the energy of each atom in the system based on a large number of quantum mechanical calculations. Compared with the classical force field, the ML force field does not depend on the allocation of atom types with preset parameters, which makes the ML model more transferable between different molecules. In April 2021, David Rosenberger's group from Los Alamos National Laboratory published "Modeling of Peptides with Classical and Novel Machine Learning Force Fields: A Comparison." on JPCB, introducing the existing dynamic simulation based on machine learning. And compared with the classical force field in water and peptides.

Background

Result

Although the ML force field has achieved some success, we know very little about its performance in systems such as biomolecules. The author of this paper compares the ML force field ANI-2x and two classical force fields CHARMM27FF and GROMOS96 43a1FF in water and two model peptides (Ala3/Aib9) systems. The author first compared the bulk water system modeled by ANI-2x, TIP3PCHARMM and SPC, and calculated the radial distribution between different atoms of various models. It can be seen that ANI-2x is more structured than the classical force field. Excessive structuring is due to the stronger hydrogen bonding caused by the lack of dispersion interaction in the DFT calculation as the training set (Figure 1).

Figure 1. Radial distribution between different atoms in water system. (Justin S. Smith et al, 2021)

Furthermore, the author calculated the probability distribution of the distance (PDA) and angle between the hydrogen bond donor-acceptor (D-A). Compared with the two classic waters, SPC and TIP3PCHARMM, the distance and angle distributions obtained by the ANI-2x model are narrower. This means that for ANI-2x, hydrogen bonds are more likely to occur at shorter distances and smaller angles, which also confirms the stronger hydrogen bonds. (Figure 2).

Figure 2. Statistics of hydrogen bond properties in different water systems. (Justin S. Smith et al.2021)

After that, the author studied the three force fields under the Ala3 system. Table 1 shows the end-to-end distance, the radius of gyration, and the average distance between the α carbon and nitrogen atoms along the skeleton. It can be seen that ANI-2x can model the distance between Cα and the radius of gyration, and it is in good agreement with the classical force field.

Statistics

Conclusions

In this article, the author uses the latest ML force field ANI-2x to calculate the energy distribution of two small peptides and compares it with the energy distribution generated by two mature classical force fields CHARMM27 and GROMOS96 43a1. For the Ala3 system studied in vacuum and water, the performance of ANI-2x is similar to the classical force field. Research on Aib9 shows that even though the information about Aib9 is not used as part of the training model, ANI-2x can still model the two helical conformations, which shows the transferability of the machine learning force field. In addition to the short-distance nature of the current ML force field, the lack of dispersion correction in the DFT calculation may also have an adverse effect.

Reference:

Justin S. Smith et al. Modeling of Peptides with Classical and Novel Machine Learning Force Fields: A Comparison. J. Phys. Chem. B 2021, 125, 14, 3598–3612.

* For Research Use Only.

Related Services