MD potentials from ML are more effective when protein-specific

Summary

Potentials for MD simulations that are derived from machine learning are more effective when they are specific to each protein, rather than general-purpose (1). These force fields were calculated using Graph neural networks.

Figures

Protein	FF-NNP Min RMSD (Å)	FF-NNP Mean RMSD (Å)	FF-NNP Macro prob. (%)	G-NNP Min RMSD (Å)	G-NNP Mean RMSD (Å)	G-NNP Macro prob. (%)
Chignolin	0.3	0.9 ± 0.7	42.5 ± 0.2	0.2	1.5 ± 0.4	16.3 ± 0.2
Trp-cage	1.1	3.2 ± 1.4	5.5 ± 0.1	4.0	5.5 ± 0.5	11.0 ± 0.1
BBA	0.4	1.3 ± 0.7	30.0 ± 0.1	1.6	2.5 ± 0.4	42.9 ± 0.1
WW-domain	0.4	1.0 ± 0.6	1.8 ± 0.2	2.7	4.9 ± 0.9	4.4 ± 0.1
Villin	0.4	1.0 ± 0.5	18.6 ± 0.1	2.2	6.9 ± 1.4	6.9 ± 0.1
NTL9	0.5	0.9 ± 0.3	16.0 ± 0.1	2.7	4.7 ± 0.4	2.9 ± 0.1
BBL	0.5	1.8 ± 0.6	4.6 ± 0.2	3.3	6.1 ± 1.3	17.0 ± 0.1
Protein B	4.0	8.6 ± 2.5	9.9 ± 0.1	4.5	6.6 ± 0.7	4.8 ± 0.1
Homeodomain	0.5	3.6 ± 3.9	35.0 ± 0.1	4.1	7.2 ± 0.9	7.1 ± 0.1
Protein G	0.5	1.0 ± 0.4	1.5 ± 0.1	1.7	2.8 ± 0.4	6.1 ± 0.1
a3d	3.4	8.5 ± 2.3	5.8 ± 0.1	4.1	7.0 ± 2.3	5.7 ± 0.1
λ-repressor	4.9	6.8 ± 1.5	0.4 ± 0.2	4.6	7.5 ± 1.1	4.9 ± 0.1
Ref (1)

See also

• MD simulations

1.

Navarro C, Majewski M, De Fabritiis G. Top-Down Machine Learning of Coarse-Grained Protein Force Fields. Journal of Chemical Theory and Computation. 2023;19(21):7518–26. Available from: https://doi.org/10.1021/acs.jctc.3c00638