Efficient Sampling of High-Dimensional …
131
comparable (Fig. 3c, d), the LJ energy dominates the binding to the hydrophobic
surface.
Example 2 Elucidating reaction networks with SPRINT coordinates
Fu and Pfaendtner [82] have demonstrated an approach to study chemical reaction
networks by using PBMetaD to explore the connections of atoms in a molecule. The
CVs that they utilized to describe these connections are called Social PeRmutation
INvarianT (SPRINT) coordinates [83]. A SPRINT coordinate is a type of generic
CV that describes the relative connectivity of atoms in a system and is derived from
the contact matrix of all relevant atoms. These generalized CVs are unique in that
they require no chemical intuition about a system a priori, and the only “chemical
information required” is the equilibrium bond lengths of the atoms involved. In
practice, the number of SPRINT CVs scales massively with the size of the system,
due to the nature of creating a contact matrix for every atom in the system. Fu and
Pfaendtner have shown that chemical reaction networks can be explored by using
the PBMetaD scheme to bias all of the atomic SPRINT coordinates of a system.
Particularly, they employed this PBMetaD + SPRINT framework to explore the
decomposition of γ-ketohydroperoxide (KHP), also known as the Korcek reaction
mechanism.
From their study, they concluded that the recovered reaction networks and pathway
selectivities were largely unaffected by the choice of conservative or aggressive bias
parameters (i.e., Gaussian hill height and width). It was noted, however, that the use
of aggressive bias parameters may increase the number of reactive events that are
sampled, but also carries the potential risk of missing reverse reactions.
They also investigated how the choice of starting structure affected the sampled
pathways. For KHP decomposition, the lowest energy pathway proceeds via a ring
formation step to form 1,2-dixolan-3-ol (CYCP), shown in Fig. 4, from which the
network diverges into four pathways. They found that the choice of starting structure
(i.e., KHP or CYCP) did not impact the sampled pathways but affected the recovered
selectivity. Since KHP and CYCP exhibit different bonding topologies, the atomic
SPRINT coordinates are distributed differently between the two structures, and thus,
the history-dependent bias potential accumulates at different points in phase space.
Lastly, one important feature of PBMetaD + SPRINT that separates it from other
automated reaction prediction methods is the ability to distinguish between pathways
that produce enantiomers, whereas other methods lose chirality in their predictions.
Example 3 Effect of surfactant on the surface induced denaturation of proteins
It is known that surfactants prevent surface induced denaturation of proteins, and to
study this mechanism, Arsiccio et al. employed atomistic MD enhanced by PBMetaD
in order to exhaustively sample peptide configurations at three interfaces, namely air
water, ice water, and water silica [84]. This study uses GB1, a 16-residue peptide that
folds into a stable beta hairpin, and Tween 80, a polysorbate as the model protein and
surfactant, respectively. By employing PBMetaD, the authors were able to include
131
comparable (Fig. 3c, d), the LJ energy dominates the binding to the hydrophobic
surface.
Example 2 Elucidating reaction networks with SPRINT coordinates
Fu and Pfaendtner [82] have demonstrated an approach to study chemical reaction
networks by using PBMetaD to explore the connections of atoms in a molecule. The
CVs that they utilized to describe these connections are called Social PeRmutation
INvarianT (SPRINT) coordinates [83]. A SPRINT coordinate is a type of generic
CV that describes the relative connectivity of atoms in a system and is derived from
the contact matrix of all relevant atoms. These generalized CVs are unique in that
they require no chemical intuition about a system a priori, and the only “chemical
information required” is the equilibrium bond lengths of the atoms involved. In
practice, the number of SPRINT CVs scales massively with the size of the system,
due to the nature of creating a contact matrix for every atom in the system. Fu and
Pfaendtner have shown that chemical reaction networks can be explored by using
the PBMetaD scheme to bias all of the atomic SPRINT coordinates of a system.
Particularly, they employed this PBMetaD + SPRINT framework to explore the
decomposition of γ-ketohydroperoxide (KHP), also known as the Korcek reaction
mechanism.
From their study, they concluded that the recovered reaction networks and pathway
selectivities were largely unaffected by the choice of conservative or aggressive bias
parameters (i.e., Gaussian hill height and width). It was noted, however, that the use
of aggressive bias parameters may increase the number of reactive events that are
sampled, but also carries the potential risk of missing reverse reactions.
They also investigated how the choice of starting structure affected the sampled
pathways. For KHP decomposition, the lowest energy pathway proceeds via a ring
formation step to form 1,2-dixolan-3-ol (CYCP), shown in Fig. 4, from which the
network diverges into four pathways. They found that the choice of starting structure
(i.e., KHP or CYCP) did not impact the sampled pathways but affected the recovered
selectivity. Since KHP and CYCP exhibit different bonding topologies, the atomic
SPRINT coordinates are distributed differently between the two structures, and thus,
the history-dependent bias potential accumulates at different points in phase space.
Lastly, one important feature of PBMetaD + SPRINT that separates it from other
automated reaction prediction methods is the ability to distinguish between pathways
that produce enantiomers, whereas other methods lose chirality in their predictions.
Example 3 Effect of surfactant on the surface induced denaturation of proteins
It is known that surfactants prevent surface induced denaturation of proteins, and to
study this mechanism, Arsiccio et al. employed atomistic MD enhanced by PBMetaD
in order to exhaustively sample peptide configurations at three interfaces, namely air
water, ice water, and water silica [84]. This study uses GB1, a 16-residue peptide that
folds into a stable beta hairpin, and Tween 80, a polysorbate as the model protein and
surfactant, respectively. By employing PBMetaD, the authors were able to include