Abstract
Complete Boltzmann sampling of reaction coordinates in biomolecular systems continues to be a challenge for unbiased molecular dynamics simulations. A growing number of methods have been developed for applying biases to biomolecular systems to enhance sampling while enabling recovery of the unbiased (Boltzmann) distribution of states. The Adaptive Biasing Force (ABF) algorithm is one such method, and works by canceling out the average force along the desired reaction coordinate(s) using an estimate of this force progressively accumulated during the simulation. Upon completion of the simulation, the potential of mean force, and therefore Boltzmann distribution of states, is obtained by integrating this average force. In an effort to characterize the expected performance in applications such as protein loop sampling, ABF was applied to the full ranges of the Ramachandran ϕ/ψ backbone dihedral reaction coordinates for dipeptides of the 20 amino acids using all-atom explicit-water molecular dynamics simulations. Approximately half of the dipeptides exhibited robust and rapid convergence of the potential of mean force as a function of ϕ/ψ in triplicate 50-ns simulations, while the remainder exhibited varying degrees of less complete convergence. The greatest difficulties in achieving converged ABF sampling were seen in the branched-sidechain amino acids threonine and valine, as well as the special case of proline. Proline dipeptide sampling was further complicated by trans-to-cis peptide bond isomerization not observed in unbiased control molecular dynamics simulations. Overall, the ABF method was found to be a robust means of sampling the entire ϕ/ψ reaction coordinate for the 20 amino acids, including high free-energy regions typically inaccessible in standard molecular dynamics simulations.
Keywords: biased Molecular Dynamics, Ramachandran maps, numerical integration, potential of mean force, dihedral bonds
Introduction
Molecular simulations are often used to compute a biomolecular system’s free energy landscape as a function of one or more particular reaction coordinates (progress variables), providing insight into conformational heterogeneity and the mechanism of structural transitions.1 Traditional, non-accelerated simulations suffer from poor or incomplete sampling of the full reaction coordinate in complex systems, particularly in high-energy regions.2-6 If a high-energy region is a transition state, its presence will also hamper converged sampling of the low-energy regions that it separates. On timescales presently achievable by molecular dynamics (MD) simulations, reaction coordinates typically of interest in biopolymers involve landscape ruggedness sufficient to hamper exhaustive equilibrium sampling. Therefore, an important goal is to overcome the sampling problem of unbiased MD simulations with computationally efficient methods.
A variety of methods exist to accelerate local barrier crossings,7-9 a subset of which make it possible to determine the Boltzmann distribution of conformational states. Examples of the latter include umbrella sampling,10 metadynamics,11,12 use of constraint forces,13-15 and the adaptive biasing force (ABF) method.3-6 ABF aims for uniform sampling of reaction coordinates using an “on the fly” estimate of the forces needed to overcome local free energy barriers encountered during the simulation.2-6,16,17 More precisely, the equilibrium force <F> is estimated by sampling the instantaneous force and then cancelled by application of the adaptive bias −<F>. Because the estimate relies on the previous history of sampling, the applied biasing force is “adaptive” and is continually updated to achieve the limit of uniform sampling along the reaction coordinate when the applied force is exactly the opposite of the true value of <F>. Per the reversible work theorem,18 the potential of mean force (PMF) is then obtained by integrating the adaptive biasing force along the reaction coordinate, similar to constrained simulations where the constraint force is integrated to yield the PMF.19 Modifications have been made upon the original method developed by Darve and Pohorille3 to expand its use to multiple coordinates.2 Analagous to metadynamics,11,12 applying the ABF methodology in multiple dimensions yields a multidimensional PMF.
The alanine dipeptide (Figure 1) is an oft-used and well-characterized test case for new computational techniques, where backbone dihedral angles ϕ and ψ are used as system descriptors. Alanine has previously been used to demonstrate the efficacy of ABF; using ranges of ϕ and ψ of [180°, 360°] and [−120°, 180°], respectively, Darve et al. rapidly achieved uniform sampling of the complete reaction coordinate, despite the presence of high free-energy regions that would be problematic for sampling in unbiased simulations.4 Alanine dipeptide has also served as an example to directly compare ABF and metadynamics,2 as both techniques are intended to improve sampling of rare events, and hence should sample even the high-energy conformations of alanine dipeptide to generate a full Ramachandran map (ϕ and ψ of [−180°, +180°] and [−180°, +180°]). In a comparison of these methods, ABF and metadynamics simulations yielded similar accuracy and results, but it was suggested that ABF might be a better approach in more ‘fragile’ systems.2 In a comparison of ABF and Umbrella Sampling,10 the reversible folding of deca-L-alanine was simulated under otherwise equivalent conditions; the ABF-generated PMF was comparable (in both qualitative and quantitative terms) to that of an umbrella sampling simulation forty times longer.17 This implies that ABF may indeed be more efficient in complicated systems.
Figure 1.

Amino acid “dipeptide” representation of the protein backbone with dihedral angles ϕ/ψ. The −R group denotes the sidechain.
We investigated the free energy convergence of ABF on dipeptides of the 20 amino acids using backbone ϕ and ψ reaction coordinates. The effects of sidechain branching on backbone sampling convergence were examined, as well as the role of peptide bond isomerization. Such studies are relevant to molecular simulations of proteins, since changes in just one or two backbone dihedral angles can result in biologically-important structural changes such as flexible surface loop conformational switching.20,21
Methods
Preparation of dipeptide systems
Dipeptide systems were built in triplicate for the 20 amino acids, with three different ϕ/ψ conformations (180°/180°, “E” for extended; −60°/−60°, “HR” for right-handed helical; and 60°/60°, “HL” for left-handed helical) and having sidechain conformations based on the default internal coordinate values in the CHARMM force field topology files.22 Sidechains were protonated as appropriate for neutral pH. Histidine dipeptides were built with the sidechain imidazole singly protonated at either the delta (“Hsd”) or epsilon (“Hse”) nitrogen. Eachdipeptide was centered in a box of 12×12×12 = 1728 TIP3P23 water molecules placed at evenly-spaced grid points with a box size appropriate for the experimental density of water. Net system charges were balanced to 0 with the addition of either a single sodium or chloride ion as needed at the center of one of the faces of the box for dipeptides with charged sidechains. Systems were constructed using the CHARMM software.24,25
MD simulations and free energy calculations
Systems were represented using the CHARMM all-atom additive protein force field with the CMAP backbone potential22,26-29 and a modified TIP3P three-site rigid water molecule.23,30 Periodic boundary conditions31 were used for energy calculations with a cubic periodic unit. Bonds involving hydrogen were constrained to their equilibrium lengths using the SHAKE algorithm,32 and SETTLE,33 a faster, non-iterative algorithm was used to maintain rigid water geometries. A timestep of 0.002 ps was used with the BBK integrator34 in combination with Langevin thermostating35 and barostating.36 Lennard-Jones interactions were smoothly truncated using an 8 Å to 10 Å switching distance.37 Electrostatic interactions beyond the 10 Å cutoff were computed via the particle-mesh Ewald method using a ~1 Å grid-spacing,38 and the effect of long-range Lennard-Jones interactions on the system pressure was accounted for by an isotropic correction.31 Each system was energy minimized for 1000 steps using the conjugate-gradient algorithm39,40 and then heated to and equilibrated at 310 K and 1 atm for 40 ps by reassignment of velocities every 1000 steps, with harmonic positional restraints applied to dipeptide heavy atoms. The subsequent 50-ns production simulation replaced velocity reassignment with Langevin dynamics using a damping coefficient of 0.1 ps−1, and ABF was in effect for the full 50-ns duration.
The full two-dimensional domain defined by ABF degrees of freedom ϕ and ψ was sampled to generate a complete Ramachandran map from −180° to +180° in both dimensions.41 Periodicity was allowed, enabling a system to transition directly from e.g. = −179° to +179° without having to cross ϕ= 0°. Hydrogen atoms attached to non-hydrogen atoms defining the ϕ/ψ angles – namely the backbone amide hydrogens and the Cα hydrogen – were given increased masses of 12 amu. SHAKE bond constraints were not applied to bonds involving these hydrogen atoms in order to meet the ABF requirement that no constraints involve atoms used to define ABF degrees-of-freedom.17 Bin widths were 10° in each dimension. The adaptive biasing force was not applied fully to a given bin until nfull force samples had been collected for that bin. Specifically, until the number of force samples n equaled or exceed nfull, the adaptive biasing force for a bin was scaled by a factor f, where f = 0 for n < nmin and f = (n – nmin) / (nfull – nmin), with nfull = 100 and nmin = nfull / 2. ABF history files were saved every 100 ps during the first 2 ns of the simulation and every 1 ns afterwards for analysis. NAMD 2.842 was used for minimization, MD, and ABF free energy sampling. Gradient and histogram data from the history files were used to generate 2D ϕ/ψ PMFs using the “abf_integrate” utility program, distributed as part of the NAMD package. “abf_integrate” was run with default settings, where a Monte Carlo simulation is performed in gradient space, and sampling of ϕ/ψ bins is biased by a memory-dependent potential (i.e. computed PMF) in the same spirit as metadynamics11,12 so as to eventually achieve uniform sampling when the forces due to the computed PMF cancel out the ABF gradients. Per the default settings, the integration is deemed converged when the RMSD between the ABF simulation gradients and the gradients from the “abf_integrate” computed PMF does not improve with additional Monte Carlo sampling. For the dipeptide systems, this typically occurred after 5 to 10 million MC iterations, and increasing the number of iterations by 10-fold for several select dipeptide data sets yielded no additional improvement in integration convergence.
RMSD values referenced in the text refer to the root mean square deviation (in kcal/mol) between pairs of PMFs, each generated from one of the three independent ABF simulations for a dipeptide. Pairs of PMFs were first least-squares aligned in the energy coordinate prior to RMSD calculation. For graphical display, PMFs were aligned along the energy coordinate such that the lowest free energy in each PMF is 0 kcal/mol.
Results and Discussion
The PMFs obtained from three independent simulations of each dipeptide were used to gauge convergence as a function of simulation time. Two metrics were used for comparison: (1) RMSD values between the three PMFs, and (2) consistency of the free energy differences ΔG between different free-energy minima in ϕ/ψ space. Because RMSD measures convergence across the entire reaction coordinate space, it can be skewed by poor estimates for high free-energy (low probability) regions. Therefore, the ΔG metric is complementary in that it measures convergence in the low free-energy (high-probablity) regions. For a given dipeptide, as convergence is achieved the RMSD should approach zero and the ΔG values computed from each PMF should converge to a single consensus value; in practice, due to the stochastic nature of the “abf_integrate” program used to integrate the ABF gradients to generate a ϕ/ψ PMF, pairwise RMSDs for PMFs generated using the exact same alanine dipeptide ABF data are 0.1 – 0.2 kcal/mol, and such RMSD values thus represent a practical minimum in the limit of full convergence. As discussed below in detail, quantitative convergence was observed for some dipeptides, qualitative convergence for others, and poor convergence for only a remaining few.
The dipeptides were clustered into three groups with increasing PMF RMSD cutoffs for more detailed analysis: RMSD < 0.5 kcal/mol (“Group 0.5”), < 1.0 kcal/mol (“Group 1.0”), and < 1.5 kcal/mol (“Group 1.5”), where a given dipeptide was assigned to a particular group if all three pairwise inter-simulation RMSDs met the cutoff at 50 ns. Group 0.5 and Group 1.0 contained nine members each, while Group 1.5 had two members (Table 1). Group 0.5 contains the two amino acids with the simplest sidechains: Gly and Ala. Gly and Ala had both the quickest-converging and lowest final RMSDs, with values of 0.4 kcal/mol achieved at 5 ns and values of 0.2–0.3 kcal/mol at 50 ns. The other two amino acids that demonstrated quick RMSD convergence were Ser and Cys, with values of 0.4 and 0.5 kcal/mol achieved at 5 ns, respectively. Following Ala, Ser and Cys have the next simplest sidechains, composed of two non-hydrogen atoms. At the other end of the spectrum was Pro, which fell outside of all three of the groups, as its inter-simulation RMSDs were 3.1, 1.3, and 2.6 kcal/mol. Though not as problematic as Pro, Thr and Val, with two heavy atoms bound to their beta carbon, showed relatively poor convergence, placing them in Group 1.5. Surprisingly, Ile, which like Thr and Val is also branched at the beta carbon, exhibited good convergence, with final inter-simulation RMSDs ranging from 0.5 to 0.6 kcal/mol. Differences in sidechain conformational behavior are discussed further below, and time courses for the inter-simulation RMSDs are shown in Figure 2 and in Supporting Information Figure S1.
Table 1.
Dipeptide groupings based on final (50 ns) inter-simulation PMF RMSDs.
| Group 0.5a | Group 1.0a | Group 1.5a |
|---|---|---|
| Ala, Asp, Cys, Glu, Gly, Hse, Ser, Trp, Tyr |
Arg, Asn, Gln, Hsd, Ile, Leu, Lys, Met, Phe |
Thr, Val |
All three inter-simulation PMF RMSDs values for a given dipeptide must fall below the indicated value in kcal/mol. Proline values were 1.3, 2.6 and 3.1 kcal/mol.
Figure 2.
PMF heterogeneity in independent dipeptide simulations. The three inter-simulation PMF RMSDs (kcal/mol) are shown as a function of simulation time. Note that the RMSD scale for Pro is 2x that of the other amino acids. Data for all 21 systems are presented in Supporting Information Figure S1.
Computing the free energy changes for transition between the β,
α, and αL (left-handed α) regions of
ϕ/ψ space showed trends
consistent with the RMSD data. These three different conformational states were, for
the purposes of this study, defined as different 10° × 10°
regions of the ϕ/ψ space. For the
β conformation, this region was centered at
ϕ/ψ =
−135°/135°, for α at
ϕ/ψ =
−65°/−45°, and for αL at
ϕ/ψ = 65°/45°. The
β to α transition ΔG(β
α) was computed as G(α) –
G(β) and the β to αL transition
ΔG(β
αL) as G(αL) –
G(β). Following these intra-PMF
ΔG values as a function of time (Figure 3 and Supporting Information Figure S2), a similar trend is observed as in the
inter-PMF RMSD data: most of the change occurs within the first 5 ns of the 50-ns
ABF simulations. In the case of Val, which showed poor RMSD convergence, estimates
for ΔG(β
αL) reach stable values of 0, 3, and 4 kcal/mol by 5 ns and
remain unchanged (and unconverged) for the next 45 ns. This is in contrast to
quickly converging systems like Ser, for which nearly identical estimates of
ΔG are reached within 5 ns. In aggregate, 95% confidence
intervals at the 50-ns timepoint for
ΔG(β
α) and ΔG(β
αL) for dipeptides in Group 0.5 are 0.49 and 0.86 kcal/mol,
respectively (Table 2), showing the ability
of ABF to achieve near-quantitative convergence. Error estimates were 1.19 and 1.33
kcal/mol for Group 1.0, consistent with qualitative convergence, and 2.10 and 3.51
for Group 1.5, consistent with poor convergence (Table 2). As mentioned previously, while RMSD measures convergence
across the entire ϕ/ψ reaction
coordinate including free-energy maxima, ΔG values measure
convergence in the lowest free-energy, and hence thermodynamically most important,
regions of the reaction coordinate. However, these separate measures of convergence
are correlated, as demonstrated by increasing errors for ΔG
estimates with increasing Group RMSD values (Table
2).
Figure 3.
Intra-PMF ΔG’s for the β
α (solid) and β
αL (dash) transitions in independent dipeptide
simulations. Note that the ΔG scale for Pro is 4x that
of the other amino acids. Data for all 21 systems are presented in Supporting Information Figure
S2.
Table 2.
Confidence intervals for β
α and β
αL transition free energies. All data are from the final
50-ns timepoints.
<ΔG(β α)> |
95% CI | <ΔG(β αL)> |
95% CI | |
|---|---|---|---|---|
| Group 0.5 | ||||
| Ala | −1.22 | 0.20 | −0.29 | 0.66 |
| Asp | −1.78 | 0.13 | −0.55 | 1.01 |
| Cys | −0.76 | 0.11 | −0.87 | 1.72 |
| Glu | −0.73 | 1.27 | 0.79 | 1.13 |
| Gly | −2.41 | 0.14 | −2.28 | 0.63 |
| Hse | −0.38 | 0.55 | −0.80 | 0.59 |
| Ser | −0.75 | 0.56 | −0.20 | 0.34 |
| Trp | −0.72 | 0.74 | −0.31 | 1.20 |
| Tyr | −0.55 | 0.69 | 0.32 | 0.46 |
| average | 0.49 | 0.86 | ||
| Group 1.0 | ||||
| Arg | −0.48 | 0.73 | 0.50 | 0.43 |
| Asn | −0.76 | 0.47 | −0.46 | 0.59 |
| Gln | −0.49 | 1.18 | 0.69 | 1.20 |
| Hsd | −0.63 | 0.91 | −0.33 | 1.39 |
| Ile | −0.47 | 1.39 | 2.81 | 0.28 |
| Leu | −1.23 | 1.21 | 0.19 | 2.19 |
| Lys | −0.38 | 0.89 | 1.36 | 1.82 |
| Met | −1.06 | 2.14 | 0.33 | 1.80 |
| Phe | −0.72 | 1.84 | 0.21 | 2.25 |
| average | 1.19 | 1.33 | ||
| Group 1.5 | ||||
| Thr | −1.71 | 1.97 | 0.09 | 2.08 |
| Val | 0.29 | 2.22 | 2.60 | 4.94 |
| average | 2.10 | 3.51 | ||
| Pro | −11.83 | 3.57 | −4.90 | 7.78 |
Despite the differences in convergence properties and level of accuracy in
free-energy estimates, comparing ϕ/ψ
PMFs from each Group demonstrates the ability of ABF to consistently sample the
entire ϕ/ψ space and locate minima,
maxima, and transition states. Figure 4 shows
the 50-ns ϕ/ψ PMFs for Ser, Gln,
Val, as well as Pro, which had the poorest ABF convergence based on RMSD and
ΔG values. In the case of Ser dipeptide, there is
quantitative agreement in the relative heights of the three minima in the β,
α, and αL regions, as well as the heights and locations of
maxima (near 0°/0°, −15°/−165°,
135°/−115°, and 145°/0°) and saddle point regions
between the minima. For Gln, there is qualitative agreement between the three
simulations, but differences are observed in the αL minimum and
0°/0° maximum. In contrast, for Val, which demonstrated poor
convergence in RMSD and ΔG, agreement is lacking across the
three simulations with regard to minima, maxima, and transition state free energies.
Nevertheless, the locations of these regions are consistent across the Val
simulations, and the entire reaction coordinate space is indeed sampled,
demonstrating the utility of the ABF method. Finally, while the RMSD and
ΔG time series for Pro suggest it to be the most
challenging dipeptide with regard to ABF sampling, there is some notable consistency
between the three independent PMFs. Specifically, there are two major minima on the
surface, one in the α region and the other near
−35°/145° in the poly-proline region. The
ΔG between these two regions is quantitatively
consistent across the three simulations, and estimates of the transition state
connecting them are within 1 kcal/mol. In the case of proline, it is the high
free-energy regions of the reaction coordinate, located at
ϕ< −120° and
ϕ> 60° (Figure
4, “Pro”), which are inconsistently estimated, leading to
large RMSD values (Figure 2). Additionally,
because the β and αL conformations are in these high energy
regions, the 50-ns estimates for ΔG(β
α) and ΔG(β
αL) are inconsistent (Table
2).
Figure 4.
ϕ/ψ PMFs from three independent 50-ns ABF simulations of Ser (from Group 0.5), Gln (Group 1.0), Val (Group 1.5) and Pro. “E” simulations were seeded with dipeptide conformations having ϕ/ψ =180°/180° (extended), “HR” with −60°/−60° (right-handed helical), and “HL” with 60°/60° (left-handed helical). Contours are drawn every kcal/mol. Note difference in z-axis for Pro. Data for all 21 systems are presented in Supporting Information Figure S3.
We end the discussion on proline dipeptide by noting that, because the N and Cα atoms are connected in a 5-membered ring, the ϕ/ψ surface is in fact not periodic, unlike the other dipeptides. Rather, in the ϕ coordinate, rotation of the backbone dihedral angle will lead to build-up of ring strain due to twisting of the ring, and this strain may be released by trans-cis isomerization of the N-terminal peptide bond, leading to a new stable peptide backbone conformation with a different ϕ/ψ PMF.43 This occurred in one of the three present simulations, namely the “HR” simulation, and may be a contributing factor to the PMF from this simulation being conspicuously different than those from the “E” and “HL” simulations (Figure 4, “Pro HR”). The three Pro simulations were therefore repeated with half-harmonic restraining potentials in place for ϕ < −150° and 150° < ϕ; however, this still did not prevent isomerization, which was observed in both the “HR” and “HL” simulations. In contrast, all of the other amino acid dipeptide ABF simulations retained trans peptide bonds for their entirety. And, repeating the triplicate Pro simulations without ABF resulted in no isomerization, though naturally in such unbiased simulations a very limited region of the ϕ component of the reaction coordinate is sampled. Such results suggest special care in applying ABF sampling to the peptide backbone of proline residues, for example, by limiting the extent of ϕ sampling to only the low free-energy regions (Figure 4, “Pro”).
One possible explanation for the dramatic sidechain dependence of ABF convergence involves differences in the characteristic relaxation times associated with sidechain conformational sampling. If sidechain conformation affects ϕ/ψ energetics and rotamer sampling times are on the order of or greater than the simulation length, then incomplete rotamer sampling will hinder PMF convergence. Time autocorrelation functions C(Δt)18 were analyzed for the χ1 dihedral angle, defined as N-Cα-Cβ-Xγ, where Xγ is a non-hydrogen atom whose identity depends on the amino acid sidechain. This dihedral is relevant as it determines the location of Xγ relative to the backbone and its degree of interaction with the dipeptide. In contrast, Cβ is fixed relative to the backbone and the Xδ position only exists in a subset of amino acids.
Relaxation times were compared for Group 0.5 and Group 1.5 dipeptides Cys and Ser, and Thr and Val (Figure 5). C(Δt) was computed by defining three χ1 states: trans: χ < −120° or χ ≥ 120°; gauche minus: −120° ≤ χ < 0°; and gauche plus: 0° ≤ χ < 120°. C(Δt) was computed for the memory of the three χ1 states in each independent simulation. For example, for a trans C(Δt) calculation, raw dihedral values were converted to 1’s and 0’s, with a raw value converted to 1 if it met the criteria for being in the trans state and 0 otherwise. C(Δt) was then computed from the time series of 1’s and 0’s using the CORREL facility in CHARMM. Not surprisingly, C(Δt) values for Cys and Ser decay rapidly to ~0 by Δt < 3 ns. Consistent with the hypothesis that sidechain sampling affects ABF convergence, C(Δt) for Thr shows slow decay. The C(Δt) decay behavior for Val was intermediate between these two extremes.
Figure 5.
Time autocorrelation functions C(Δt) for sidechain dihedral χ1 sampling in ABF dipeptide simulations. Data are shown for trans (solid), gauche minus (dash) and gauche plus (dot) states. Results for each state are shown only if its sampling exceeds a 15% threshold in two of the three simulations.
Valine was further studied by repeating the ABF triplicate simulations with a biasing potential on χ1 to restrain it to the trans state. The heating/equilibration protocol was unchanged, and the production ABF protocol only differed by the addition of a harmonic restraint to the N-Cα-Cβ-Cγ1 dihedral of the form k(χ – χ0)2 / 2 where χ0 = 180° and k = 0.050 kcal mol−1 degree−2. With this modification, Val switches from Group 1.5 to Group 0.5 and from inconsistent to consistent ΔG estimates (Table 3), providing additional evidence that limitations in Val convergence are due to rotamer sampling. Therefore, if computational resources allow, it may be advisable to repeat a simulation several times with different starting conditions and compare inter-simulation PMFs. Unfortunately, this is not always possible given that many biomolecular simulations tax the limits of available computational resources.
Table 3.
Final (50-ns) ΔG and inter-simulation PMF RMSD values (kcal/mol) for Val in unrestrained and χ1-restrained ABF simulations.
| Unrestrained | Restrained | |
|---|---|---|
ΔG(β α) |
0.36, −0.65, 1.14 | −1.39, −1.21, −1.66 |
ΔG(β αL) |
3.11, 0.40, 4.28 | 2.09, 2.50, 2.03 |
| RMSD | 1.03, 1.14, 1.10 | 0.42, 0.49, 0.48 |
As a final measure of the efficacy of ABF for peptide ϕ/ψ sampling, the ABF data for alanine were compared with standard MD simulations performed in triplicate. PMFs from these latter simulations were generated by computing G = −RT ln p(ϕ/ψ) + C, where p(ϕ/ψ) is the number of MD samples in a 10° × 10° bin centered at ϕ/ψ and C is selected such that G = 0 at the global minimum. Figure 6 reveals the differences in sampling of minima, transition state regions, and maxima early in (5 ns) and at the end of (50 ns) the simulations. Two key differences between standard MD and ABF sampling are apparent: only ABF samples high free-energy regions of ϕ/ψ space, and only ABF demonstrates convergent sampling early in the simulations (Figure 6). Both standard MD and ABF of alanine give similar estimates at 50 ns for low-lying basins and their minimum energy transitions. Standard MD, however, fails to sample the less accessible, high-energy regions that separate conformational basins.
Figure 6.
ϕ/ψ PMFs after 5- and 50-ns sampling in independent Ala simulations using ABF (Ala) and standard MD (Alanobias). Contours are drawn every 1 kcal/mol. Bright yellow areas in the Alanobias landscapes indicate unsampled regions.
Conclusions
While ABF sampling in the 50-ns time regime investigated here was not sufficient to generate quantitative ϕ/ψ free energy estimates for all 20 amino acids, it was sufficient to generate such estimates for nearly half of them. And, ABF did sample the minima, maxima, and transition states across the entire span of ϕ/ψ values for all dipeptides. Backbone-rotamer coupling appears to explain the poor convergence of certain dipeptides. Rotational sampling of the Thr sidechain is slow relative to the 50-ns timescale. Although sampling of the Val sidechain is comparatively fast, removing sidechain rotation converted the Val dipeptide system from poorly- to well-converged. Finally, for the trivial alanine dipeptide, ABF sampling is clearly superior to unbiased MD with regard to generating converged estimates of ΔG(ϕ/ψ) across the full ϕ/ψ reaction coordinate.
With regard to generalizing the results to MD studies on proteins, our own work has shown that in certain instances the overall conformation of a protein loop region is dependent upon the backbone conformation of a single key residue in the loop. In one instance, the peptide-binding cleft of an SH2 domain, formed by two loop regions, could be induced to go from a closed conformation to an open one capable of peptide binding solely by transition of a Tyr residue backbone in one of the loops from the extended region of ϕ/ψ space to the left-handed helical region.20 In another instance, the transition from the extended region of ϕ/ψ space to the left-handed helical region for a Tyr residue in a loop forming the carbohydrate-binding cleft of the CD44 protein induced a loop conformational change allowing for formation of a direct contact between the sidechain of an Arg residue in the loop and the carbohydrate ligand.21 This latter conformational change is thought to be a molecular basis for affinity-switching in CD44 binding to its carbohydrate ligand hyaluronan.44 The ABF methodology can therefore be a very attractive solution for determining the thermodynamics associated with loop sampling in such instances. From the dipeptide data here, it is expected that ABF studies of loop conformation as a function of Ala, Asp, Cys, Glu, Gly, His, Ser, Trp, or Tyr backbone conformation ought to yield quantitative results on the 50 ns timescale. Results will be more qualitative for the other 11 amino acids, with Pro, Thr, and Val being especially challenging. Special attention should be paid to trans-cis peptide bond isomerization in proline, a phenomenon observed in the present ABF studies that does not occur in the absence of ABF sampling on the same timescale. Of course, in the case of loop sampling, in addition to sidechain effects on backbone dynamics, other interactions involving additional degrees of freedom may also affect convergence times. Based on the present results, the most reliable way to assess convergence is using multiple simulations with different starting conditions. The present results caution against over-confidence in measures of time-dependent behavior in single simulations as yardsticks for PMF convergence. The fact that ΔG between two states reaches a stable, seemingly time-invariant value is not a guarantee that it has converged. In conclusion, when combined with appropriate measures of convergence, the ABF methodology provides a useful means of free energy measurement for conformations of the protein backbone.
Supplementary Material
Acknowledgments
O.G. thanks University of New England for startup funding and Garlic Jones for helpful discussion. This work was supported by National Institutes of Health grant R15 GM099022 (O.G. and R.D.H) and by the National Science Foundation through XSEDE resources provided by the National Institute for Computational Sciences (grant TG-MCB120007; O.G.).
Footnotes
Supporting Information Available: Figures of PMF heterogeneity plots, intra-PMF ΔGs of transition for all systems, ϕ/ψ PMFs from three independent 50-ns ABF simulations for all dipeptide systems are available. This material is available free of charge via the Internet at http://pubs.acs.org.
References
- (1).Christ CD, Mark AE, van Gunsteren WF. J Comput Chem. 2010;31:1569. doi: 10.1002/jcc.21450. [DOI] [PubMed] [Google Scholar]
- (2).Hénin J, Fiorin G, Chipot C, Klein ML. Journal of Chemical Theory and Computation. 2010;6:35. doi: 10.1021/ct9004432. [DOI] [PMC free article] [PubMed] [Google Scholar]
- (3).Darve E, Pohorille A. J Chem Phys. 2001;115:9169. [Google Scholar]
- (4).Darve E, Rodriguez-Gomez D, Pohorille A. J Chem Phys. 2008;128:144120. doi: 10.1063/1.2829861. [DOI] [PubMed] [Google Scholar]
- (5).Darve E, Wilson MA, Pohorille A. Molecular Simulation. 2002;28:113. [Google Scholar]
- (6).Rodriguez-Gomez D, Darve E, Pohorille A. J Chem Phys. 2004;120:3563. doi: 10.1063/1.1642607. [DOI] [PubMed] [Google Scholar]
- (7).Bhatt D, Zuckerman DM. Journal of Chemical Theory and Computation. 2011;7:2520. doi: 10.1021/ct200086k. [DOI] [PMC free article] [PubMed] [Google Scholar]
- (8).Isralewitz B, Gao M, Schulten K. Current Opinion in Structural Biology. 2001;11:224. doi: 10.1016/s0959-440x(00)00194-9. [DOI] [PubMed] [Google Scholar]
- (9).Schlitter J, Engels M, Krüger P. Journal of Molecular Graphics. 1994;12:84. doi: 10.1016/0263-7855(94)80072-3. [DOI] [PubMed] [Google Scholar]
- (10).Torrie GM, Valleau JP. Journal of Computational Physics. 1977;23:187. [Google Scholar]
- (11).Laio A, Parrinello M. Proc Natl Acad Sci U S A. 2002;99:12562. doi: 10.1073/pnas.202427399. [DOI] [PMC free article] [PubMed] [Google Scholar]
- (12).Leone V, Marinelli F, Carloni P, Parrinello M. Current Opinion in Structural Biology. 2010;20:148. doi: 10.1016/j.sbi.2010.01.011. [DOI] [PubMed] [Google Scholar]
- (13).Carter EA, Ciccotti G, Hynes JT, Kapral R. Chemical Physics Letters. 1989;156:472. [Google Scholar]
- (14).den Otter WK. Journal of Chemical Physics. 2000;112:7283. [Google Scholar]
- (15).den Otter WK, Briels WJ. Journal of Chemical Physics. 1998;109:4139. [Google Scholar]
- (16).Chipot C, Henin J. J Chem Phys. 2005;123:244906. doi: 10.1063/1.2138694. [DOI] [PubMed] [Google Scholar]
- (17).Henin J, Chipot C. J Chem Phys. 2004;121:2904. doi: 10.1063/1.1773132. [DOI] [PubMed] [Google Scholar]
- (18).Chandler D. Intorduction to Modern Statistical Mechanics. Oxford University Press; New York: 1987. p. 201. [Google Scholar]
- (19).Darve E. In: Numerical Methods for Calculating the Potential of Mean Force New Algorithms for Macromolecular Simulation. Leimkuhler B, Chipot C, Elber R, Laaksonen A, Mark A, Schlick T, Schütte C, Skeel R, editors. Vol. 49. Springer; Berlin Heidelberg: 2006. p. 213. [Google Scholar]
- (20).Guvench O, Qu C-K, MacKerell AD., Jr. BMC Structural Biology. 2007;7:14. doi: 10.1186/1472-6807-7-14. [DOI] [PMC free article] [PubMed] [Google Scholar]
- (21).Jamison FW, 2nd, Foster TJ, Barker JA, Hills RD, Jr, Guvench O. J Mol Biol. 2011;406:631. doi: 10.1016/j.jmb.2010.12.040. [DOI] [PubMed] [Google Scholar]
- (22).MacKerell AD, Jr., Bashford D, Bellott, Dunbrack RL, Evanseck JD, Field MJ, Fischer S, Gao J, Guo H, Ha S, Joseph-McCarthy D, et al. The Journal of Physical Chemistry B. 1998;102:3586. doi: 10.1021/jp973084f. [DOI] [PubMed] [Google Scholar]
- (23).Jorgensen WL, Chandrasekhar J, Madura JD, Impey RW, Klein ML. J Chem Phys. 1983;79:926. [Google Scholar]
- (24).Brooks BR, Brooks CL, 3rd, Mackerell AD, Jr., Nilsson L, Petrella RJ, Roux B, Won Y, Archontis G, Bartels C, Boresch S, et al. J Comput Chem. 2009;30:1545. doi: 10.1002/jcc.21287. [DOI] [PMC free article] [PubMed] [Google Scholar]
- (25).Brooks BR, Bruccoleri RE, Olafson BD, States DJ, Swaminathan S, Karplus M. J Comput Chem. 1983;4:187. [Google Scholar]
- (26).Guvench O, MacKerell AD., Jr. Methods Mol Biol. 2008;443:63. doi: 10.1007/978-1-59745-177-2_4. [DOI] [PubMed] [Google Scholar]
- (27).MacKerell AD, Jr., Brooks B, Brooks CL, 3rd, Nilsson L, Roux B, Won Y, Karplus M. Encyclopedia of Computational Chemistry. John Wiley & Sons, Ltd; Chichester: 1998. CHARMM: The Energy Function and Its Parameterization. [Google Scholar]
- (28).Mackerell AD, Jr., Feig M, Brooks CL., 3rd. J Comput Chem. 2004;25:1400. doi: 10.1002/jcc.20065. [DOI] [PubMed] [Google Scholar]
- (29).MacKerell AD, Jr., Feig M, Brooks CL., 3rd. J Am Chem Soc. 2004;126:698. doi: 10.1021/ja036959e. [DOI] [PubMed] [Google Scholar]
- (30).Durell SR, Brooks BR, Ben-Naim A. The Journal of Physical Chemistry. 1994;98:2198. [Google Scholar]
- (31).Allen MP, Tildesley DJ. Computer Simulation of Liquids. Oxford University Press; Oxford, U.K.: 1987. [Google Scholar]
- (32).Ryckaert JP, Ciccotti G, Brenderson HJC. Journal of Computational Physics. 1977;23:327. [Google Scholar]
- (33).Miyamoto S, Kollman PA. J Comput Chem. 1992;13:952. [Google Scholar]
- (34).Brünger A, Brooks CL, 3rd, Karplus M. Chemical Physics Letters. 1984;105:495. [Google Scholar]
- (35).Kubo R, Toda M, Hashitsume N. Statistical Physics II: Nonequilibrium Statistical Mechanics. 2 ed Springer; New York: 1991. [Google Scholar]
- (36).Feller SE, Zhang Y, Pastor RW, Brooks BR. J Chem Phys. 1995;103:4613. [Google Scholar]
- (37).Steinbach PJ, Brooks BR. J Comput Chem. 1994;15:667. [Google Scholar]
- (38).Darden T, York D, Pedersen L. J. Chem. Phys. 1993;98:10089. [Google Scholar]
- (39).Hestenes MR, Stiefel E. J. Res. N. B. S. 1952;49:409. [Google Scholar]
- (40).Fletcher R, Reeves CM. The Computer Journal. 1964;7:149. [Google Scholar]
- (41).Ramachandran GN, Ramakrishnan C, Sasisekharan V. J Mol Biol. 1963;7:95. doi: 10.1016/s0022-2836(63)80023-6. [DOI] [PubMed] [Google Scholar]
- (42).Phillips JC, Braun R, Wang W, Gumbart J, Tajkhorshid E, Villa E, Chipot C, Skeel RD, Kalé L, Schulten K. J Comput Chem. 2005;26:1781. doi: 10.1002/jcc.20289. [DOI] [PMC free article] [PubMed] [Google Scholar]
- (43).Fischer S, Dunbrack RL, Karplus M. J Am Chem Soc. 1994;116:11931. [Google Scholar]
- (44).Banerji S, Wright AJ, Noble M, Mahoney DJ, Campbell ID, Day AJ, Jackson DG. Nat Struct Mol Biol. 2007;14:234. doi: 10.1038/nsmb1201. [DOI] [PubMed] [Google Scholar]
Associated Data
This section collects any data citations, data availability statements, or supplementary materials included in this article.



α)>
αL)>

α)
αL)