New Research In
Physical Sciences
Social Sciences
Featured Portals
Articles by Topic
Biological Sciences
Featured Portals
Articles by Topic
 Agricultural Sciences
 Anthropology
 Applied Biological Sciences
 Biochemistry
 Biophysics and Computational Biology
 Cell Biology
 Developmental Biology
 Ecology
 Environmental Sciences
 Evolution
 Genetics
 Immunology and Inflammation
 Medical Sciences
 Microbiology
 Neuroscience
 Pharmacology
 Physiology
 Plant Biology
 Population Biology
 Psychological and Cognitive Sciences
 Sustainability Science
 Systems Biology
Free energy reconstruction from nonequilibrium singlemolecule pulling experiments

Communicated by David Chandler, University of California, Berkeley, CA (received for review November 28, 2000)
Abstract
Laser tweezers and atomic force microscopes are increasingly used to probe the interactions and mechanical properties of individual molecules. Unfortunately, using such timedependent perturbations to force rare molecular events also drives the system away from equilibrium. Nevertheless, we show how equilibrium free energy profiles can be extracted rigorously from repeated nonequilibrium force measurements on the basis of an extension of Jarzynski's remarkable identity between free energies and the irreversible work.
Recent advances in the micromanipulation of single molecules have led to new insights into the dynamics, interactions, structure, and mechanical properties of individual molecules (1–4). Singlemolecule manipulation with an atomic force microscope (AFM) (5–9), laser tweezer stretching (10), and analogous computer experiments pioneered by Schulten and coworkers (11–13) have revealed details about unfolding and unbinding events of individual proteins and their complexes. In an AFM experiment, a molecule is subjected to a timevarying external force, for instance by pulling on the end of a linear polymer (Fig. 1). The applied force is determined from the timedependent position of the cantilever tip with respect to the sample. Thus, one can drive rare molecular events (14), determine their force characteristics, and simultaneously monitor them with atomic resolution. However, both experiments and simulations actively perturb the system, leading to hysteresis and nonequilibrium effects.
How can one extract equilibrium properties from such measurements that drive the system away from equilibrium (15, 16)? From the second law of thermodynamics, we know that on average, the mechanical work of pulling will be larger than the free energy. Only if the experiment is performed reversibly, i.e., infinitely slowly, will the work equal the free energy. Thus, making rigorous thermodynamic measurements by pulling appears to require an extrapolation to zero pulling speed. However, Jarzynski (17, 18) recently discovered a remarkable identity between thermodynamic free energy differences and the irreversible work, thus extending the inequality of the second law of thermodynamics. This identity, although not directly applicable to atomic force measurements, suggests that in principle one should be able to extract free energy surfaces from repeated pulling experiments. In this paper, we show how this can be done in practice.
Theory.
We begin by showing that Jarzynski's identity follows almost immediately from the Feynman–Kac theorem for path integrals (19). This derivation leads directly to the appropriate extension, which forms the basis for the solution of the free energy reconstruction problem. This extension has been obtained by Crooks (20) as a special case of an even more general relation between forward and backward path averages. Consider a system whose phase–space density evolves according to a Liouvilletype equation: 1 ℒ_{t} is an explicitly timedependent evolution operator that has the Boltzmann distribution as a stationary solution, ℒ_{t}e^{−βℋ(x,t)} = 0, where ℋ(x, t) is a timedependent Hamiltonian, and β^{−1} = k_{B}T, with T the temperature and k_{B} Boltzmann's constant. For example, for diffusive dynamics on a potential V(x, t), the time evolution is governed by ℒ_{t = }D∇e^{−βV(x,t)}∇e^{βV(x,t)}, where D is the diffusion coefficient and ∇ = ∂/∂x. Other examples include systems that undergo Newtonian, Langevin, Nosé–Hoover thermostat, or Metropolis Monte Carlo dynamics. Now consider the unnormalized Boltzmann distribution at time t, 2 Because this distribution is stationary (ℒ_{t}p = 0), and because ∂p/∂t = −β(∂ℋ/∂t)p, it follows that the above p(x, t) is a solution of the sink equation, 3 as can be verified by direct substitution. However, the solution of this sink equation, starting from an equilibrium distribution at time t = 0, can also be expressed as a path integral by using the Feynman–Kac theorem (19). Equating these two different solutions immediately gives: 4 where δ(x) is Dirac's δ distribution. This identity between a weighted average of nonequilibrium trajectories (right) and the equilibrium Boltzmann distribution (left) is implicit in the work of Jarzynski (18), and is given explicitly by Crooks (equation 24, ref. 20). The average 〈…〉 is over an ensemble of trajectories starting from the equilibrium distribution at t = 0 and evolving according to Eq. 1. Each trajectory is weighted with the Boltzmann factor of the external work w_{t} done on the system, 5 Integrating both sides of Eq. 4 with respect to x, we obtain Jarzynski's identity (17, 18) 6 between the Boltzmannaveraged work w_{t} and the equilibrium free energy difference ΔG(t) between times t and 0.
Free Energy Surfaces from Cantilever Positions.
Eq. 4 allows us to reconstruct the underlying free energy surface from repeated singlemolecule AFM experiments. The setup in Fig. 1a can be described by a Hamiltonian ℋ(x, t) = ℋ_{0}(x) + k(z − vt)^{2}/2, where ℋ_{0}(x) is the Hamiltonian of the resting, unperturbed system, and z = z(x). Here, we have assumed for simplicity that the cantilever is vibrating in a harmonic potential with spring constant k. Substituting this Hamiltonian into Eq. 4, multiplying both sides by δ[z − z(x)], integrating over all x, and finally taking the logarithm, we have: 7 where G_{0}(z) is (to within an additive constant) the unperturbed free energy profile along the pulling coordinate z; and Δw_{t} is the external work minus the instantaneous biasing potential, Δw_{t} = w_{t} − k(z_{t} − vt)^{2}/2 = kv(vt^{2}/2 − ∫z_{t′} dt′) − k(z_{t} − vt)^{2}/2. At time t = 0, the trajectories are started from points z_{0} drawn from a Boltzmann distribution corresponding to Hamiltonian ℋ(x, 0) = ℋ_{0}(x) + kz^{2}/2.
Given the positions z_{t} of the cantilever at time t obtained from repeated pulling experiments, we can reconstruct the unperturbed free energy profile G_{0}(z) along z by using Eq. 7. At each time slice t, one can in principle obtain an estimate of the whole free energy surface. In practice, at any given time t, only a small window around the equilibrium position z = vt will be sampled adequately. Thus, an average over several time slices and repeated trajectories is required to obtain an optimal estimate of the free energy surface. At every time slice t, one obtains an ensemble of positions z_{t} and corresponding w_{t}s. The positions z_{t} are binned, and the corresponding histogram values are incremented by e^{−βwt}. The complete free energy surface G_{0}(z) can then be reconstructed by adapting the weighted histogram method (21): 8 where the sum is over time slices t and u(z, t) is the timedependent biasing potential [here: u(z, t) = k(z − vt)^{2}/2]. As in the weighted histogram method (21), this procedure can be refined by making Eq. 8 selfconsistent through replacement of 〈exp(−βw_{t})〉 with exp[−βΔG(t)] = ∫exp {−β[u(z, t) + G_{0}(z)]}dz/∫exp{−β[u(z, 0) + G_{0}(z)]}dz, thus requiring an iterative solution for ΔG(t).
To illustrate in detail how this formalism is implemented, we simulate a pulling experiment by conducting Browniandynamics simulations on a doublewell free energy surface. Ten trajectories were generated on the timedependent surface V(z) + k(z − vt)^{2}/2 with a pulling velocity v = 1 μm s^{−1}, a diffusion coefficient D = 10^{−7} cm^{2} s^{−1}, and a cantilever spring constant k = 20.6 pN nm^{−1} (=0.0206 mN m^{−1}). The pulling velocities and cantilever spring constants are typical of AFM experiments (2, 6). For this diffusion coefficient, the escape rate from the stable minimum is about 10^{−6} s^{−1} on the basis of Kramers' theory, in the range of unfolding rates for fairly large proteins under native conditions. The free energy difference between the two states separated by a barrier is ≈14 kcal/mol (at 298 K), again typical of protein unfolding. Nevertheless, this simple onedimensional model is only a crude approximation to real experiments where several other factors intrude, such as viscous damping of the cantilever motion or slow conformational relaxation orthogonal to the pulling direction (14).
In the pulling simulations, the “cantilever” is moved relative to the “sample” by 20 nm within a time of 20 ms. The position z_{t} and accumulated work w_{t} = kv(vt^{2}/2 − ∫z_{t′} dt′) are recorded for 10 runs and analyzed by using Eq. 7. For every pulling trajectory k (k = 1,… ,K; here: K = 10), we determine at discrete times t_{i} (i = 0,… ,N; here: N = 100 time points separated by 0.2 ms) the position of the cantilever z_{ik} with respect to the sample, and the corresponding deflection δz_{ik} = z_{ik} − vt_{i}. From the deflection, we obtain the instantaneous deflection energy u_{ik} = kδz/2. The external work is determined by numerical integration: w_{ik} = kv[vt/2 − Σ(t_{j} − t_{j−1})(z_{jk} + z_{j−1,k})/2] with w_{ik} = 0 for i = 0. The results of repeated trajectories are then combined. At every time slice t_{i}, we average 〈exp(−βw_{t})〉 ≈ η_{i} ≡ K^{−1}Σexp(−βw_{ik}). To approximate the average in the top numerator of Eq. 8, we also collect histograms h_{i}(l) at times t_{i} with zposition intervals of Δz: 9 where θ_{l}(z) is one if z is in the lth interval [(l − 1)Δz ≤ z < lΔz] and zero otherwise. We then estimate the free energy profile by averaging over all time slices t_{i}: 10 Fig. 2 compares the reconstructed and exact free energy surfaces. With the exception of the poorly sampled barrier region because of snapping motion (Fig. 2 Inset), the surface is accurately reproduced. We note that at high pulling speeds, the increasing variance of w_{t} will lead to a systematic bias of the estimator for −β^{−1}ln〈δ(z − z_{t})exp(−βw_{t})〉, which can be approximately corrected for by using a cumulant expansion.
Free Energy Surfaces from Pulling Forces.
So far, we made use of positionversustime curves. However, it is forceversusposition curves that are commonly reported from pulling experiments. As above, we assume a Hamiltonian ℋ(x, t) = ℋ_{0}(x) + u[z(x), t]. The unperturbed system described by ℋ_{0} is coupled to the measurement apparatus through u[z(x), t], which will typically be a harmonic potential, u[z(x), t] = k(z − vt)^{2}/2. The work entering the free energy expression Eq. 9 can be rewritten as 11 where the restoring force is F(z, t) = −∂u(z, t)/∂z, and the integral is along the positionversustime contour connecting z_{0} and z_{t}. This identity follows from du = (∂u/∂z)dz + (∂u/∂t)dt. Combining Eqs. 7 and 11 leads to the surprising result that the work entering into the free energy profile G_{0}(z) is not given simply by the forceversusdistance integral, but by an additional factor reflecting the biased choice of the initial state, 12 where the initial points are drawn from a Boltzmann distribution according to ℋ(x, 0) (i.e., for the system interacting with the cantilever, as above). Therefore, if force integrals of two or more trajectories are combined, individual trajectories should be weighted by the reciprocal Boltzmann factor exp[βu(z_{0}, 0)]; i.e., one should subtract from the integrated work the energy stored in the deflected cantilever at time t = 0.
In an AFM experiment, it is thus important to record the zeroforce position (δz = 0 in Fig. 1) with the cantilever sufficiently far from the sample. As the cantilever tip is brought into contact with the sample, a sufficiently long relaxation is important to draw initial states from an equilibrium Boltzmann distribution. We note that this new equilibrium will normally not be centered around δz = 0, but according to the combined Hamiltonian ℋ(x, 0) of the sample coupled to the cantilever spring. Nevertheless, the center of this initial distribution can be used to align the z positions of repeated pulling traces. As the cantilever is retracted, both z_{t} and δz_{t} are recorded. With known cantilever spring constant k and pulling speed v, the work can then be obtained from time or position integration.
As a final illustration, we analyze force spectra for extracting individual bacteriorhodopsin molecules from purple membrane patches of Halobacterium salinarum as measured by Oesterhelt et al. (8). In these experiments, a cantilever tip was attached to bacteriorhodopsin, and the protein was slowly pulled out of the membrane. This led to sequential unraveling of the transmembrane helices, resulting in unfolding of the protein. Specifically, we analyze the forceversusdistance curves for one of the experiments (G241C mutant) and estimate a free energy profile for extraction and unfolding of helices D and E, corresponding to the second peak at pulling distances of 20–30 nm in figure 5 of ref. 8. Because the force is fluctuating around the baseline before and after this force peak, we assume (i) that the additional bias u(z_{0}, 0) in Eq. 12 is of the order of k_{B}T and thus negligible on the scale of the integrated work; and that (ii) a quasiequilibrium was established before the force peak. We also assume (iii) that the force curves are aligned. We can then integrate the forces and use Eq. 12 to estimate the underlying free energy profile G_{0}(z), as shown in Fig. 3. Also shown is the mean restoring force, F(z) = dG_{0}(z)/dz, corresponding to reversible pulling. The noise in the force F(z) reflects uncertainties in the zposition alignment and the finite number of measurements. For extraction and stretching of the 57amino acid stretch with 56 amino acids contained in helices D and E, we estimate a free energy of about 64 kcal/mol, or about 1.1 kcal/mol per residue, and a meanforce peak of about 89 pN. Although these values seem overall reasonable, they should not be considered to be accurate before the above assumptions are removed by appropriate experimental design. In addition, one must account for instrument resolution, which leads to increased variation of the integrated work. Finally, the influence of the pulling speed needs to be investigated. In general, to extract accurate thermodynamic information from a small number of pulling experiments, it will be important to achieve small standard deviations of the work, possibly of a few k_{B}T, by pulling sufficiently slowly.
Conclusions
We have shown how free energy surfaces can be reconstructed rigorously from repeated molecular pulling experiments. With this formalism, thermodynamic properties such as binding equilibria can be extracted from repeated measurements by using AFM (2, 5–9, 12), laser tweezers (1, 4), or magnetic beads (3). In addition, this work has implications in the area of computer simulations. For example, when a timedependent harmonic potential is added to the Hamiltonian as above, our procedure is a dynamic generalization of umbrella sampling (22). The weightedhistogram method permits the combination of dynamic umbrella sampling runs with different pulling speeds or directions. Moreover, this generalization has the immediate advantage that no equilibration is required in different windows, and thus the entire trajectory can be used to reconstruct the underlying free energy surface.
Footnotes
Abbreviation
 AFM,
 atomic force microscopy
 Received November 28, 2000.
 Accepted January 22, 2001.
 Copyright © 2001, The National Academy of Sciences
References
 ↵
 Perkins T T,
 Smith D E,
 Chu S
 ↵
 Florin E L,
 Moy V T,
 Gaub H E
 ↵
 ↵
 ↵
 ↵
 Rief M,
 Gautel M,
 Oesterhelt F,
 Fernandez J M,
 Gaub H E
 ↵
 Oesterhelt F,
 Oesterhelt D,
 Pfeiffer M,
 Engel A,
 Gaub H E,
 Müller D J
 ↵
 ↵
 Kellermayer M S Z,
 Smith S B,
 Granzier H L,
 Bustamante C
 ↵
 ↵
 ↵
 ↵
 ↵
 ↵
 ↵
 ↵
 ↵
 Schuss Z
 ↵
 ↵
 ↵
Citation Manager Formats
More Articles of This Classification
Physical Sciences
Chemistry
Biological Sciences
Biophysics
Related Content
Cited by...
 Enhanced stochastic fluctuations to measure steep adhesive energy landscapes
 Elasticity, friction, and pathway of {gamma}subunit rotation in FoF1ATP synthase
 Reconstructing folding energy landscapes from splitting probability analysis of singlemolecule trajectories
 Freeenergy inference from partial work measurements in small systems
 Mechanistic picture for conformational transition of a membrane transporter at atomic resolution
 Loop Motions Important to Product Expulsion in the Thermobifida fusca Glycoside Hydrolase Family 6 Cellobiohydrolase from Structural and Computational Studies
 Rate limit of protein elastic response is tether dependent
 Energy landscape analysis of native folding of the prion protein yields the diffusion constant, transition path time, and rates
 Product Binding Varies Dramatically between Processive and Nonprocessive Cellulase Enzymes
 Munc181 Tuning of Vesicle Merger and Fusion Pore Properties
 Probing Carbohydrate Product Expulsion from a Processive Cellulase with Multiple Absolute Binding Free Energy Methods
 Free energy profiles from singlemolecule pulling experiments
 Replica exchange with nonequilibrium switches
 Monitoring Early Fusion Dynamics of Human Immunodeficiency Virus Type 1 at SingleMolecule Resolution
 Axisdependent anisotropy in protein unfolding from integrated nonequilibrium singlemolecule experiments, analysis, and simulation
 Distinct kinetic and mechanical properties govern ALCAMmediated interactions as shown by singlemolecule force spectroscopy
 Singlemolecule analysis of cadherinmediated cellcell adhesion
 SingleMolecule Analysis of Human Immunodeficiency Virus Type 1 gp120Receptor Interactions in Living Cells
 Molecular dynamics and protein function
 Experimental test of Hatano and Sasa's nonequilibrium steadystate equality
 Using nonequilibrium measurements to determine macromolecule freeenergy differences
 Bias and error in estimates of equilibrium freeenergy differences from nonequilibrium measurements
 Beyond the conventional description of dynamic force spectroscopy of adhesion bonds
 Can energy landscape roughness of proteins and RNA be measured by using mechanical unfolding experiments?
 Developing an energy landscape for the novel function of a ({beta}/{alpha})8 barrel: Ammonia conduction through HisF
 A twostate kinetic model for the unfolding of single molecules by mechanical force
 Equilibrium Information from Nonequilibrium Measurements in an Experimental Test of Jarzynski's Equality
 Energetics of glycerol conduction through aquaglyceroporin GlpF
 How does a system respond when driven away from thermal equilibrium?