Scholarly article on topic 'Rapid, Accurate, Precise, and Reliable Relative Free Energy Prediction Using Ensemble Based Thermodynamic Integration'

Rapid, Accurate, Precise, and Reliable Relative Free Energy Prediction Using Ensemble Based Thermodynamic Integration Academic research paper on "Chemical sciences"

Share paper
Academic journal
J. Chem. Theory Comput.
OECD Field of science

Academic research paper on topic "Rapid, Accurate, Precise, and Reliable Relative Free Energy Prediction Using Ensemble Based Thermodynamic Integration"


[ournal of Chemical Theoiy and Computation

Subscriber access provided by NEW YORK UNIV

Rapid, accurate, precise and reliable relative free energy prediction using ensemble based thermodynamic integration

Agastya P. Bhati, Shunzhou Wan, David W. Wright, and Peter Vivian Coveney

J. Chem. Theory Comput., Just Accepted Manuscript • DOI: 10.1021/acs.jctc.6b00979 • Publication Date (Web): 08 Dec 2016

Downloaded from on December 12, 2016

Just Accepted

"Just Accepted" manuscripts have been peer-reviewed and accepted for publication. They are posted online prior to technical editing, formatting for publication and author proofing. The American Chemical Society provides "Just Accepted" as a free service to the research community to expedite the dissemination of scientific material as soon as possible after acceptance. "Just Accepted" manuscripts appear in full in PDF format accompanied by an HTML abstract. "Just Accepted" manuscripts have been fully peer reviewed, but should not be considered the official version of record. They are accessible to all readers and citable by the Digital Object Identifier (DOI®). "Just Accepted" is an optional service offered to authors. Therefore, the "Just Accepted" Web site may not include all articles that will be published in the journal. After a manuscript is technically edited and formatted, it will be removed from the "Just Accepted" Web site and published as an ASAP article. Note that technical editing may introduce minor changes to the manuscript text and/or graphics which could affect content, and all legal disclaimers and ethical guidelines that apply to the journal pertain. ACS cannot be held responsible for errors or consequences arising from the use of information contained in these "Just Accepted" manuscripts.

ACS Publications

Journal of Chemical Theory and Computation is published by the American Chemical Society. 1155 Sixteenth Street N.W., Washington, DC 20036 Published by American Chemical Society. Copyright © American Chemical Society. However, no copyright claim is made to original U.S. Government works, or works produced by employees of any Commonwealth realm Crown government in the course of their duties.

10 11 12

20 21 22

Rapid, accurate, precise and reliable relative free energy prediction using ensemble based thermodynamic integration

Agastya P. Bhati, Shunzhou Wan, David W. Wright, and Peter V. Coveney*

Centre for Computational Science, Department of Chemistry, University College London, 20 Gordon Street, London WC1H 0AJ, United Kingdom



The accurate prediction of the binding affinities of ligands to proteins is a major goal in drug discovery and personalised medicine. The time taken to make such predictions is of similar importance to their accuracy, precision and reliability. In the last few years, an ensemble based molecular dynamics approach has been proposed that provides a route to reliable predictions of free energies based on the molecular mechanics Poisson-Boltzmann surface area method which meets the requirements of accuracy, precision and reliability. Here, we describe an equivalent methodology based on thermodynamic integration to substantially improve the accuracy, precision and reliability of calculated relative binding free energies. We report the performance of the method when applied to a diverse set of protein targets and ligands. The results are in very good agreement with experimental data (90% of calculations agree to within 1 kcal/mol) while the method is reproducible by construction. Statistical uncertainties of the order of 0.5

*To whom correspondence should be addressed

a protein target, also referred to as the binding affinity, is a key property of interest in drug discovery as it correlates with the potency of possible drug candidates. The reliable prediction of binding affinities is important in drug discovery as well as in personalised medicine where it can be used to support clinical decision making. The free energy and its associated error need to be predicted within a span of a few hours in order to be useful in such real world applications.

4 kcal/mol or less are achieved. We present a systematic account of how the uncertainty

6 in the predictions may be estimated.

10 1 Introduction

13 The free energy change associated with the binding of a lead compound or drug with

20 21 22

29 Several in silico methods are available in the literature for calculating binding affinities.

31 All these methods are based on classical molecular dynamics which is used to determine

33 free energy and other macroscopic properties. Among them are so-called "exact" free energy

35 methods which, through the use of a thermodynamic cycle, allow the determination of the

37 relative binding affinity of two ligands. It is common in these methods to refer to a variable,

41 categories of such have been developed; A dynamics1,2 in which A is a dynamic variable and

47 -p/--\ T* r\ 1 -i n n ni i /"»I"» e\ n AnnTTfrr /-"\t> 4-1 i t* V\ e\ 4- -i s~\ v» ( L1 L'T) 3 4- V» /"M^w» s~\ /A t rv-» r\ -i tt Aifiril-i ati /"TT^ 4,5

A, which describes the (unphysical) path taken to transform one ligand into another. Two categories of such have been developed; A dynamics1,2 in which A is a dynamic variable and thermodynamic properties for multiple states are simultaneously evaluated in a single simulation and methods where separate simulations are run at fixed A values and analysed using formalisms such as free energy perturbation (FEP),3 thermodynamic integration (TI),4 Bennett acceptance ratio (BAR)6 or multistate Bennett acceptance ratio (MBAR).7 The double decoupling method8,9 can, in principle, be used to calculate absolute binding affinities. In this method the free energy changes of decoupling the ligand from the solvent in one case and the binding site of the solvated receptor in the other are computed with their difference providing the absolute binding affinity. Another set of methods are referred to as

10 11 12

20 21 22

"approximate" as they depend on underlying approximations. The empirically based linear interaction method (LIE)10-12 and the molecular mechanics Poisson-Boltzmann surface area (MMPBSA)13 methods fall in the latter category.

The first macromolecular free energy calculations were performed about three decades ago.14-16 Thereafter, with the increase in speed of computers, the methods have been increasingly applied for calculating protein-ligand binding affinities within the academic community. Unfortunately, however, as we have emphasised in recent years,17 "one-off" molecular dynamics simulations are not reproducible so most reported results lack scientific credibility. The lack of reproducibility of such an approach based on calculating free energy by performing just one MD simulation was further emphasized by Wright et al.18 who showed that two independent MD simulations of HIV-1 protease bound to inhibitors with identical initial structure and force field parameters can produce binding affinities varying by up to 12 kcal/mol. Such variations can be much larger for more flexible ligands.19

The free energy is a thermodynamic property. Statistical mechanics provides the prescription for calculating such macroscopic quantities as ensemble averages of microscopic states. A theoretical discussion has been provided by Coveney and Wan.20 Unfortunately, essentially all approaches described in textbooks and the research literature calculate these macroscopic properties from the time average of a single "long" duration trajectory. The key point is that, for systems which exhibit an equilibrium thermodynamic state, the microscopic dynamics must be at least mixing in the language of ergodic theory, hence chaotic.20 No single trajectory can describe the behaviour adequately. Various recent published works indeed demonstrate compellingly that multiple short simulations yield much more accurate binding affinities than a single long simulation.19,21-25 A new method, namely "enhanced sampling of molecular dynamics with approximation of continuum solvent (ESMACS)", based on ensemble averages, has been shown to produce precise and reproducible absolute binding affinity predictions.17-19,26-30

There has been very limited application of "exact" free energy simulation methods in industrially based drug discovery. Recently, the FEP/REST method of binding affinity calculation has been proposed by Wang et al.31 and applied to a large set of protein-ligand combinations, attracting interest within the pharmaceutical industry. Aldeghi et al.32 described a

12 method to calculate absolute binding affinities based on alchemical transformations, albeit

14 with scope restricted to a small number of molecules binding to a rigid active site. An in-

16 dependent trajectory TI (IT-TI) method25,33 has been reported where multiple independent

18 trajectories were found to improve the accuracy of free energy changes calculated using TI.

20 21 22

37 For successful uptake in drug design and discovery, reliable predictions of binding affinities

The purpose of the present paper is to describe a new method called thermodynamic integration with enhanced sampling (TIES) which yields reproducible, accurate and precise relative binding affinities. Based on the direct calculation of ensemble averages, it allows us to determine statistically meaningful results along with complete control of errors. It should be mentioned that the accuracy of the potential parameterisations used may also have substantial impact on the accuracy of results. 34 However, in the present paper, the systems studied are known to be well described by the potential parameterisations chosen.

need to be made on timescales which influence experimental programmes. For applications in personalised medicine, the selection of suitable drugs need to be made within a few hours to influence clinical decision making. 35 Therefore, speed is of the essence if we wish to use free energy based calculation methods in any of these areas. With TIES, it is possible to make predictions on timescales which meet these requirements. With sufficient compute resources available, TIES is currently able to calculate accurate, precise and reproducible binding affinities for a set of alchemical transformations within about 8 hours of wall clock time.

This paper is organised as follows. Section 2 contains the underlying theory. In Section 3, the methodology, a detailed error analysis of the distribution of free energies produced by our ensemble simulations, and the required automation tools used are described. The results are reported in Section 4 followed by the discussion of a few special cases in Section

12 5; Section 6 presents our conclusions.

20 21 22

56 The thermodynamic cycle approach is employed to calculate the relative binding affinities

2 Theory

Thermodynamic integration (TI) is well known in the literature. 4,5,36,37 The relative binding affinity of two ligands L1 and L2 is calculated by considering an alchemical transformation between them connected through intermediate states defined by introducing a coupling parameter A, such that at A = 0 the system corresponds to ligand L1 (initial state) and at A =1 the system corresponds to ligand L2 (final state). The total energy of the system is taken to be its potential energy (V). The energy of the system can be defined as:

V(A,x) = (1 - A)Vi(A,x) + AV>(A,x) (1)

where V1 and V2 are the potential energies of ligands L1 and L2 calculated using a chosen molecular mechanics force field. The derivative of the energy with respect to A is used to compute the free energy difference as follows:

AGach =f dA

where (•••)* denotes an ensemble average in state A. Such an ensemble at each A window is generated using a single MD trajectory in most cases. By contrast, TIES uses an ensemble MD simulation approach as described in detail in the following section.

10 11 12

20 21 22

AAG between these two ligands associating with a protein using the following equation:

AAG = AGi - AG2 = AGlh - AG^

where AG1 and AG2 are the binding free energies of ligands L1 and L2 respectively. AG^

aq alch

and AGOX* are the free energy differences associated with the alchemical transformation of ligand L1 into L2 in free and bound states respectively.

Figure 1: Normalized frequency distribution of ensembles of dV/dA values for 4 different intermediate alchemical states (A=0.2, 0.4, 0.8, 0.9) for the transformation from ligand L1Q to ligand LI9 binding to CDK2 shown as open red circles. Gaussian distributions with the same mean and standard deviation are shown in blue.

2.1 Ensemble based thermodynamic integration

As noted, most published TI calculations have been based on performing a single MD simulation at each A-window. The derivative of the potential energy with respect to A, dV/dA,

lies at the heart of the binding affinity so computed by numerical quadrature (equation 2). However, to calculate a reliable free energy change the derivative at each A-window must be accurate, precise and reproducible. An intrinsic problem with approaches based on single MD trajectories is that they sample but one instance from a Gaussian random process and

12 are not as such reproducible; most authors have overlooked the need to compute macroscopic

14 averages by ensemble averaging (or they assume that their one-off MD trajectories provide

16 the equivalent via the ergodic theorem).20 Some authors have recognised the advantage of us-

18 ing ensembles over single trajectory calculations in such free energy calculations,21-25,33,38-42

20 but a systematic approach has been lacking.

38 each A-window. Examples of the frequency distribution of this ensemble of potential energy

40 derivatives are shown in Figure 1; these are all well described by a Gaussian distribution, just

42 as we have reported for all ensembles of MD simulations in recent publications.17 19,26 30 This

44 means that the integral in equation 2 should properly be interpreted as a stochastic integral.

46 The average of all the potential energy derivatives at each A-window is the value which is

48 used for calculating the integral of dV/dA with respect to A using a numerical method (eg.

50 the trapezoidal rule) to finally arrive at the alchemical free energy change AGach (equation

52 2).

ACS Paragon Plus Environment

At the same time, the problem of unreproducible results is removed by computing ensemble averages. Ensemble averaging provides us with the means to quantify uncertainty and hence to control errors. In TIES, multiple 'replica' MD simulations are performed at each Awindow, where all replicas have identical initial conditions other than their initial velocities, which are drawn randomly from a Maxwell-Boltzmann distribution. The set of multiple replica simulations constitute an ensemble simulation whose size is the number of replica simulations performed. Thus, we can compute an ensemble average of dV/dA values for

10 11 12

20 21 22

Figure 2: Structures of all the five target proteins (ribbon representation) studied in each case shown bound to a ligand (blue, in stick representation): (a) cyclin-dependent kinase 2 (CDK2); PDB:1H1Q, (b) protein tyrosine phosphatase 1B (PTP1B); PDB:2QBS, (c) myeloid cell leukemia 1 (MCL1); PDB:4HW3, (d) tyrosine kinase 2 (TYK2); PDB:4GIH and (e) thrombin; PDB:2ZFF. Structures of all ligands (which are all drawn from congeneric series) are provided in the Supporting Information.

3 Method

In this section, we describe the biomolecular systems studied and the standard TIES protocol. The ensemble size, intermediate À values and simulation length in this protocol have been arrived at after careful statistical analyses which are described in Section 3.4. The complete end-to-end execution of our protocol is facilitated using tools described in Section 3.5. As we shall see, the standard protocol is applicable to most of the ligand-protein systems studied here, although there are situations in which it needs to be modified.

10 11 12

20 21 22

3.1 Ligand-protein systems studied

In this work, we have applied TIES to calculate the relative free energies of a large set of ligands bound to five different target proteins (Figures S1 to S5 of the Supporting Information). They belong to different classes of protein including kinase and phosphatase with diverse functions in the human body and are important targets for a wide range of therapies. Myeloid cell leukemia 1 (MCL1) is a member of the Bcl-2 family of proteins. It is overexpressed and amplified in various cancers and promotes the aberrant survival of tumor cells that otherwise would undergo apoptosis.43 Protein tyrosine phosphatase 1B (PTP1B) is a negative regulator of the insulin and leptin receptor pathways and thus an attractive therapeutic target for diabetes and obesity.44 The serine protease thrombin is an established target for the prevention of cardiovascular diseases.45 Selective targeting of tyrosine kinase 2 (TYK2) is considered as a potential treatment of inflammatory diseases, such as psoriasis and inflammatory bowel diseases (IBD).46,47 Aberrant control of cyclin-dependent kinase 2 (CDK2) is a central feature of the molecular pathology of cancer. 48 The ligand transformations studied include chemical group modifications in which up to 10 heavy atoms are changed. In other recent work, we have applied TIES to studies of bromodomain and pan-TrkA inhibitors.29,30

3.2 Model building and simulation set-up

In the present paper, we study five different target systems (see Figure 2) using TIES. The starting structures for each of them were downloaded from the Protein Data Bank (PDB)49 with the corresponding PDB IDs as 1H1Q for CDK2, 2ZFF for thrombin, 2QBS for PTP1B, 4HW3 for MCL1 and 4GIH for TYK2. For thrombin, 2ZC9 PDB was chosen as the starting structure to study the transformation of unsubstituted benzylamine to meta-substituted benzylamide where a water molecule initially present in the S1 pocket of thrombin is displaced by the appearance of a meta-chlorine atom (see section 5 for details). For each of these biomolecular systems, a set of around 10 ligand pairs were chosen. The structures

of the ligands were downloaded from the Supporting Information of an open access article.

7 TIES currently uses a dual topology scheme.50 The disappearing part of the system (which

9 exclusively belongs to the initial state) and the appearing part of the system (which exclu-

11 sively belongs to the final state) are defined simultaneously. The hybrid potential energy

20 21 22

36 this criterion. These criteria ensure that the common region of the two ligands is indeed

38 chemically identical in both the ligands within the convergence limit of 0.1 e. It should be

40 noted that one or both of these criteria may need to be relaxed in the presence of highly

42 charged or polar groups in the ligands.

50 electrostatically. The AMBER ff99SB-ILDN force field 51 was employed in all our simulations

function (equation 1) is set in such a way that the disappearing and appearing parts do not interact with each other. For an alchemical transformation from ligand L1 to ligand L2, the hybrid structure file is prepared by appending the coordinates of the appearing set of atoms from L2 to the structure file of L1 after making sure that their common parts are well aligned. The partial atomic charges for the hybrid ligand are derived from the partial atomic charges on the individual ligands such that the common atoms have identical charges, taken to be the average of their charges in the individual ligands. The charges on disappearing and appearing parts are adapted accordingly by reparameterising the ligands after constraining the charges on the common atoms to their new values. The set of atoms to be considered as part of the common region in the hybrid ligand is selected such that their overall charges in the individual ligands do not differ by more than 0.1 e and, in addition, each atom satisfies

The complexes of target protein and hybrid ligand were solvated in an orthorhombic water box with buffer width of 14 A. Counterions were added to neutralise the system

for protein parameters. Ligand parameters were produced using the general AMBER force field (GAFF).52 The restrained electrostatic potential (RESP) method was used to calculate partial atomic charges using Antechamber (AmberTools 12) after geometry optimization by Gaussian03 for all ligands. The package NAMD 2.953 was used for all the molecular

10 11 12

20 21 22

dynamics simulations with three-dimensional periodic boundary conditions. The systems were maintained at a temperature of 300 K and a pressure of 1 bar in an NPT ensemble using the standard NAMD protocol of Langevin dynamics (with a damping coefficient of 5 ps-1) and a Berendsen barostat (compressibility of 4.57 x 10-5 bar-1 and a relaxation time of 100 fs). A timestep of 2 fs was used. Van der Waals contributions were perturbed using linearly varying A across the full range (0 to 1). A soft core potential54'55 was used for the van der Waals interactions to avoid divergent potential energy due to the sudden appearance of atoms close to the end points of the alchemical transformation, often called "end point catastrophes". Moreover, the electrostatic interactions of the disappearing atoms were linearly decoupled from the simulations between A values of 0 and 0.55 and completely turned off beyond that, while those of the appearing atoms were linearly coupled to the simulations from A value 0.45 to 1, and completely extinguished otherwise.

Figure 3: TIES protocol requiring 5 replica simulations at each A window. For a single alchemical transmutation, 13 A-windows are used amounting to 65 molecular dynamics simulations in total. The number of cores and wall clock time employed on a Cray XC30 supercomputer are displayed on the right side of the figure.

A total of 13 A-windows was used for all TIES calculations. An ensemble simulation of size 5 was performed at each A-window. For each replica, energy minimization followed by 2 ns long equilibration was performed using the protocol defined in a previous publication.56

Production runs for each replica were 4 ns long. While the coordinates were recorded every 10 picoseconds, dV/dA values were recorded every 2 ps. The choice of 4 ns for the simulation length and 5 for the ensemble size is based on the uncertainty quantification and error analysis discussed in section 3.4.* Figure 3 exhibits the TIES workflow indicating the core counts

12 and wall clock time required for a typical TIES calculation. Given a sufficiently powerful

14 machine, it is possible to complete a TIES calculation using the protocol described in this

16 work (2 ns equilibration followed by 4 ns of production simulation) within 8 hours or less.

18 For any given system the turn around time will depend on both the number of atoms in the

20 system of interest and the available number of cores (in many cases GPUs can provide a

22 further increase in performance). The scalability of TIES calculations allows us to calculate

24 relative free energies for multiple alchemical transformations concurrently. These require

26 the same wall clock time as that for a single calculation simply by scaling up the required

28 resources accordingly. 57

32 3.3 Stochastic integration and error propagation

35 The relative binding affinity of two ligands, AAG, is given by the difference between free

37 energy changes associated with the alchemical transformation of one ligand into another in

39 free and bound states, AG<a<qch and AGaOCh^, respectively, as per equation 3. These two terms

41 are given by the integral of the ensemble average of dV/dA, equation 2. As earlier discussed,

43 the potential energy and hence its A-derivative behave like Gaussian random variables (Figure

45 1) and so each point in the integrand has an underlying Gaussian distribution. The integral

47 of a Gaussian random process is itself a Gaussian random process with variance given by the

49 convolution of the variance of all the points used to evaluate it.58 Therefore, we interpret

the integral in equation 2 in terms of stochastic calculus. We calculate the ensemble average of the potential derivative as the average of its values from all five replicas of our ensemble

55 *The protocol mentioned here is a general recommendation, but there may be cases where one may

56 wish or need to adjust it. The key point is that the ensemble size should be chosen such that the results are

57 converged, and hence, one may need to increase the ensemble size at different A-windows to handle particular

58 cases.

4 simulation, where the individual value for each replica is taken to be the average potential

6 derivative over the whole simulation length. The integral is calculated numerically using the

8 mean potential derivative at each A value. The error, a^, for a A window is taken to be the

10 bootstrapped standard error of the mean of the A-derivatives from all replicas. The variance

12 of alchemical free energy changes AG^C^ and AG^X^ and the variance of the final relative

14 binding affinity are calculated as:

17 2 2

(AA)2 (4a)

The sum of the AAG predictions for a set of ligand transformations forming a closed thermodynamic cycle is often referred to as the "hysteresis" of the cycle. Owing to the finite uncertainties associated with such AAG predictions, the hysteresis always deviates from its

18 ai/2 =

20 a2 = a2 + a2 (4b)

30 theoretical value of zero, and hence is one way of assessing the accuracy of the predictions.

32 It is noteworthy that on repeating all the calculations one may get a different value for the

34 hysteresis, and hence the value of such hysteresis has an accompanying uncertainty similar

36 to that of the AAG predictions. Wang et al.'s recently published FEP/REST method

38 attempts to artificially correct the AAG predictions made by shifting their values so as to

40 remove the associated hysteresis.31 However, the benefit of doing this is debatable given that

42 in their methodology there is an uncontrolled uncertainty associated with each prediction;

44 furthermore the approach may distribute a large error arising in one prediction over the entire

46 thermodynamic cycle, thereby distorting other correct predictions. TIES based predictions

48 of free energy changes do not exhibit significant hysteresis as adequate sampling at each A-

window in combination with stochastic integration minimise it. Table S6 of the Supporting Information lists the sets of ligands forming closed cycles in the transformations studied here along with the corresponding values of hysteresis. All but one of them have hysteresis values lying within the uncertainty range. The set of ligands L32,L42,L38 binding to MCL1

10 11 12

20 21 22

as listed in Table S6 of the Supporting Information is an example where large error in just one prediction (for the transformation L38-L42) leads to the large value of hysteresis in the closed cycle formed by them.

Figure 4: (a) Variation of error with ensemble size per A window and (b) the variation of AGb(OUh'd with simulation length for the transformation of ligand L1Q to ligand LI9 bound to CDK2. The above plots form a basis for our choice of simulation length as 4 ns and ensemble size as 5.

3.4 Uncertainty quantification and error analysis

The TIES protocol has three adjustable parameters: ensemble size, individual replica simulation time and the set of A values to be selected. In this work, we have chosen the ensemble size to be 5 and the simulation length to be 4 ns. Here we provide the justification for these choices which are made on the basis of an analysis of the variation of error with both of these parameters. Figure 4a shows the variation of TIES error with ensemble size in each A window keeping the simulation length fixed to 4 ns at each lambda window. We performed an ensemble of 105 replicas at each A-window for the transformation of ligand L1Q to ligand LI9 bound to CDK2. It is evident that the AAG converges fully after ensemble size 30 or so, where the error falls below 0.15 kcal/mol. We find that the error at ensemble size 5 is a little below 0.4 kcal/mol, reducing to about 0.3 kcal/mol on increasing the ensemble size to

10 and to about 0.2 kcal/mol on increasing it to 20. Thus, a four-fold increase in the computational cost yields marginal improvements. Therefore, here we take an ensemble size 5 as a good balance between computational cost and precision of results. However, this analysis clearly shows that a very high level of precision can be achieved given a sufficient amount of computation. It should be noted here that in ESMACS26 a similar analysis was used to

10 11 12

14 arrive at the appropriate ensemble size of 25, which is larger than that employed for each

16 A-window in the standard TIES protocol. This is because AAGties is evaluated from the

18 integral in equation 2, the error associated with which is in turn calculated as per equation

20 4. Thus, there is more smoothing of errors as compared to that in ESMACS where there is a

22 single ensemble computed at the end points. Figure 4b shows the variation of AGbOh^ with

24 simulation time for fixed ensemble size (in this case 5) per A window. It is evident that the

26 value of AGb°Und does not vary by more than 0.1 kcal/mol after 4 ns. A similar argument

28 based on minimising the cost-benefit ratio on increasing the simulation length accounts for

our choice of simulation length of 4 ns per replica simulation. Extensive studies of diverse ligand-protein systems all confirm that the cumulative averages of the derivatives dV/dA converge within 2 ns for all A windows.26,27,59 In the present case, we show the convergence in Figure S7 of the Supporting Information.

In this study all systems under investigation contain a relatively rigid ligand bound to a small globular protein, justifying the use of the the protocol established for CDK2 in all cases. Should the system of interest differ, for example in containing a flexible ligand, it is trivial to increase replica number and/or simulation length as necessary to improve sampling and reduce statistical uncertainty.

3.5 Automation and high performance computing

The TIES protocol would be very lengthy, tedious and error-prone to perform manually. Its execution is much faster and more error-proof when performed in an automated fashion. We automated the implementation of TIES by a judicious combination of our software tools

4 known as the Binding Affinity Calculator (BAC)56 and FabSim60 as we have also done for

6 ESMACS. BAC incorporates the entire sequence of steps to be performed to produce the

8 final results as shown in Figure 3. Inter alia it automatically builds the input files required

10 for the dual topology scheme. The final step of statistical analysis of the data from the

12 ensemble of replicas can be performed on a desktop (or remote machine) yielding the final

14 results including error estimation and uncertainty quantification. A user friendly version

16 of BAC, namely uf-BAC, has been developed to extend its accessibility to non-technical

18 users. 26

20 21 22

40 4 Results

58 error (RMSE) is 0.9 kcal/mol (Figure 5a). The calculated relative free energies strongly cor

FabSim is a python-based toolkit developed to simplify a range of computational tasks for researchers in diverse disciplines. It comes into play in the TIES workflow during the production phase. Since TIES involves performing ensemble simulations, it is highly desirable to have a well-defined scheme for data management and curation. The arrangement of input and output files in a standard way makes portability between different users and environments very convenient. FabSim is configured to automatically transfer data and stage the equilibration and production jobs on to the desired supercomputer using a set of simple one-line commands.60

In this study we applied our TIES methodology to carry out relative free energy calculations for 55 alchemical transformations between ligands binding to five different target proteins (Table 1). We were interested in evaluating the accuracy and precision of our predictions when applied to diverse sets of protein-ligand systems encapsulating different types of physico-chemical interactions. Table 1 summarises the results obtained from this study. The majority of calculations agree extremely well with the experimentally determined values (Figure 5). The overall mean absolute error (MAE) is 0.7 kcal/mol and the root mean square

10 11 12

20 21 22

relate with the experimental results with a Pearson's r of 0.84 and also rank the ligands well based on their relative binding affinities with a Spearman's p of 0.85 as shown in Figure 5a. Our calculations have a high level of precision with a range of uncertainty from 0.1 to 0.5 kcal/mol (Table 1).

Table 1: Summary of TIES results for the five different target proteins studied. The number of alchemical transformations, crystal structures used, original publications reporting the experimental binding affinities and the experimental method used to determine the binding affinities are provided. Isothermal titration calorimetry is abbreviated as ITC, /C50 stands for half maximal inhibitory concentration and Ki denotes inhibition constant. The range of uncertainty obtained is mentioned such that all the predictions have uncertainties lying within it. Values of several statistical parameters - root mean squared error (RMSE) and mean absolute error (MAE) for all TIES predictions as well as Pearson's r and Spearman's p between AAGTIES and experimental results - are also reported in order to assess the quality of TIES results.

CDK2 Thrombin TYK2 MCL1 PTP1B

No. of transformations 7 11 11 16 10


Exp metrics 1C50 ITC Ki Ki Ki

Reference 48 45 46, 47 43 44

Range of uncertainty (kcal/mol) 0.2-0.3 0.2-0.4 0.1-0.2 0.2-0.5* 0.2-0.5

RMSE (kcal/mol) 0.8 0.8 0.5 1.4 0.5

MAE (kcal/mol) 0.7 0.7 0.4 1.2 0.4

Pearson's r 0.87 0.90 0.88 0.80 0.84

Spearman's p 0.86 0.91 0.88 0.80 0.82

t 5 of the 16 transformations have uncertainties between 0.7-0.8 kcal/mol due to the presence of a charged carboxylate group in the transforming part of the ligands. See details in Section 5.

Table 1 provides the evaluation of the performance of the TIES methodology for individual target proteins. For each of the target proteins, TIES predictions have the same level of accuracy and precision as for the overall results. The RMSE and MAE are below 0.8 kcal/mol and 0.7 kcal/mol respectively for all targets except MCL1, for which the respective values are 1.4 kcal/mol and 1.2 kcal/mol. The Pearson coefficient is no less than 0.80 for any of the targets and reaches 0.90 for thrombin. Similarly, the Spearman coefficient is always higher than 0.80, reaching 0.91 for thrombin. The range of uncertainties for each target protein, as mentioned in Table 1, shows the high level of precision achieved with our methodology for the

10 11 12

20 21 22

individual biomolecular systems studied. It should be noted that 5 out of 16 transformations of MCL1 ligands have uncertainties in the range of 0.7-0.8 kcal/mol. This is due to the presence of a charged carboxylate group in the mutating part of the ligands. More details are provided in section 5.

-2 0 2 AAG-experimental (kcal/mo!)

-3 -2 -1 0 AAG-experimental (kcal/mol)

Figure 5: (a) Correlation between TIES-predicted relative binding affinities and experimental data for all five protein targets studied. The black line is the perfect correlation line. Blue and pink dotted lines show the ±1 kcal/mol and ±2 kcal/mol ranges respectively. The majority of points lie within the ±1 kcal/mol band, a few points lie in the ±1 kcal/mol to ±2 kcal/mol band and only two points lie outside the ±2 kcal/mol range. (b) An alternative representation of the same data such that all the experimental values are negative. Blue squares are the directionally agreeing predictions while red stars are the directionally disagreeing ones. Red and blue dotted lines show the boundary of experimental values equal to -0.6 kcal/mol and -0.9 kcal/mol. All the predictions with AAG < -0.9 kcal/mol are in directional agreement, while all directionally disagreeing predictions lie on the right side of the red dotted line except one.

Figure 5a is a scatter plot of predicted versus experimental relative binding affinities for all the transformations. The blue and pink dotted lines show the ±1 kcal/mol and ±2 kcal/mol bands respectively from the perfect correlation black line. All but two points lie within the pink lines, while a few points lie between the blue and pink lines. The majority

of points are within the blue lines, serving to emphasise the high level of accuracy of our predictions. 45 out of 55 predictions (81.82 %) deviate from the experimental values by less than 1.0 kcal/mol, while 53 out of 55 predictions (96.36 %) do so by less than 2 kcal/mol (Table 2 (left)). More than half of our predictions have an error less than 0.6 kcal/mol.

bound by significant uncertainties. 61,62 However, for the systems studied here the error bars

12 It is important to mention here that experimental binding affinities contain errors and are

16 on experimental results are unavailable. Figure S6 in the Supporting Information contains

18 correlation plots for each biomolecular system separately along with the uncertainties in the

20 TIES predictions shown as error bars.

24 Wang et al.31 recently reported binding affinity predictions for the same protein targets

38 and 96.36 % of our predictions are accurate within the absolute error of 1.0 kcal/mol and

40 2.0 kcal/mol respectively, while the same numbers from the FEP/REST study are 63.3 %

42 and 92.4 % respectively. A visual comparison of Figure 5a with Figure 3 of Wang et al.31

44 shows that the number of points in the ±1 kcal/mol to ±2 kcal/mol band is much lower

46 in the former as compared to the latter.* Furthermore, these author's results are effectively

48 obtained from running a single replica, meaning that they are liable to suffer from a lack of

50 reproducibility.

as here using the FEP/REST method and hence a direct comparison between their results and ours can be made. The RMSE and MAE for each protein target reported here (Table 1) are smaller than those given by FEP/REST, except for MCL1, for which they are the same in both studies. The correlation coefficients for each system reported here (Table 1) when compared to those given by the FEP/REST method are better for CDK2 (0.87 versus 0.48) and thrombin (0.90 versus 0.71), while being almost the same for the others. 81.82 %

52 ^Out of the 55 ligand transformations studied here, 18 are in common with the perturbations studied

53 by Wang et al. using their FEP/REST methodology.31 Figure S10 in the Supporting Information shows a

54 direct comparison of these 18 predictions from both the methods. TIES exhibits marginally better accuracy

55 with slightly smaller RMSE and MAE and slightly larger Pearson's r and Spearman's p. Only one of the

56 18 TIES predictions lie outside the 1 kcal/mol window from the experimental data and one directionally

57 disagrees with the experimental value. On the other hand, the corresponding numbers for the FEP/REST

58 predictions are three and two respectively.

10 11 12

20 21 22

Table 2: A summary of the level of accuracy obtained for the total set of TIES predictions. The number of predictions found to be accurate for a specified absolute error range (left) and the number of predictions found to be in directional agreement with the increasing absolute values of experimental results (right). The lack of experimental errors means that the entire discrepancy is assigned to the theoretical predictions. In practice, the agreement is likely to be even better than shown here.

MAE < No. of predictions

jAAGexp! >

0.4 19

0.6 30

0.8 36

1.0 45

1.2 49

1.4 50

No. of directionally disagreeing predictions

0.0 0.3 0.4 0.5 0.6 0.9

9 8 6 3 1 0

Figure 5b provides a different representation of our results. A prediction is said to be directionally agreeing with experimental observations if AAGties has the same sign as that of AAGexp, otherwise it is deemed to be directionally disagreeing. In Figure 5b we have flipped the signs of both AAGexp and AAGTIES for all the points which originally had positive AAGexp. In other words, we rearranged equation 3 such that ligand L1 always has a more negative binding free energy than ligand L2. Such a representation is useful in order to show if the calculations exhibit the same trend as that of the experimental results. Thus, we are left with all the points on the left hand side of the y-axis. Therefore, all the points which lie above the x-axis directionally disagree (red stars in Figure 5b) while those below the x-axis directionally agree (blue squares in Figure 5b). The red and blue dotted lines here are the boundaries of AAGexp equal to -0.6 kcal/mol and -0.9 kcal/mol respectively. As shown in the figure, all the points in directional disagreement lie to the right of the red line except one, which lies between the red and the blue lines. This means that all our results predict the direction of the change in the binding affinities correctly if the absolute change in the corresponding experimental values is greater than or equal to 0.9 kcal/mol, with the proviso that there are no error bars on the experimental values that are available in this study, as mentioned earlier. Wanget al. suggest that for high quality measurements,

10 11 12

20 21 22

the uncertainties on the experimental relative binding affinities AAGexp are of the order of 0.4-0.7 kcal/mol.31 More generally, Chodera and Mobley state that published affinities may have errors of around 24%, reported errors near universally underestimating the error obtained from inter-laboratory variation by one to two orders of magnitude.62 All but one TIES predictions which directionally disagree lie within this range of experimental noise as shown in Table 2 (right). Therefore, when the overlap between the error bars of both the experimental and the theoretical AAG predictions is taken into account, even better agreement can be achieved than that reported here.

Figure 6: Reproducibility of TIES: The relative binding affinities of the CDK2 ligands (black circles) were calculated using 10 values resampled from an ensemble of 20 replica simulations. Error bars are represented as standard deviations of a and 2a. For each calculated AAG value, results are also shown for two randomly chosen non-overlapping 10-replica samples (blue and red dots). The data demonstrate that a 10-replica prediction will lie within ±a and ±2a of the averaged relative binding affinities (open black circles) with confidence intervals of 0.68 and 0.95, respectively.

4.1 Reproducibility of the predictions

Another strength of TIES is its ability to yield reproducible predictions. By reproducibility, we mean that if a TIES calculation is repeated for a particular transformation, the new

prediction of the relative binding affinity for that transformation would lie within ±a and ±2a range of the original prediction with a probability of 0.68 and 0.95 respectively. The typical uncertainty on our predictions is 0.4 kcal/mol. Thus, in principle, any repeated TIES calculation for a typical transformation should yield a new prediction within ±0.8 kcal/mol

12 of the original prediction 95% of the time.

20 21 22

38 using extended simulations

41 We compared the predictions from the use of the "standard" TIES protocol with the case

43 when only one "long" simulation is performed at each A-window. For this purpose, we

45 chose the transformation from ligands L1Q to LI9 binding to CDK2. Five independent TI

47 calculations were performed for this transformation based on a 20 ns molecular dynamics

49 simulation at each A window to check its reproducibility. The AGbOUnd from these five TI

51 calculations vary by as much as 0.8 kcal/mol. On the other hand, the AGb(OUhd of the same

53 transformation from TIES has the 1a uncertainty range as 0.38 kcal/mol. Therefore, at

55 least in this case, the results from TIES employing the standard protocol are only slightly

57 better than those from TI employing a single longer simulation at each A-window for the

ACS Paragon Plus Environment

To illustrate this property of our methodology, we chose six of our relative free energy calculations between ligands binding to the CDK2 target protein. An ensemble of twenty MD simulations at each A-window was performed and the final relative free energies were calculated (represented as open black circles along with their ±a and ±2a ranges in Figure 6). Thereafter, two random non-overlapping sub-samples of ten MD simulations at each A-window was chosen in order to obtain two independent TIES predictions (denoted by red and blue filled circles in Figure 6) for the 6 transformations. Both the predictions (red and blue circles) for four of the transformations lie within the ±a range, while for the remaining two they lie within the ±2a range.

4.2 Comparison of results from ensemble based TIES with those

same amount of computer time (equivalent to 20 ns per lambda window in each case). This observation may be compared with other studies,17-19,26-28 where it has been found that the prediction from an ensemble of 50 short (4 ns) simulations is better than using a single 1 ps simulation. This may be related to our earlier observation (section 3.4) that the TIES/TI

5 Discussion

Mobley and Klimovitch63 quantified the impact of reliable binding free energy predictions on

12 algorithm involves a lot of internal cancellation of "errors", unlike in end point MD. However,

14 for cases with larger and more flexible ligands and/or proteins and when the size of the

16 alchemically changing part of the ligand is large, TIES yields much more precise predictions

18 than can be realised from single "long" simulations. 27 Moreover, the use of multiple replicas

20 within TIES has the important advantage of reducing the wall clock time required to get

22 results since these can all be performed concurrently on a supercomputer.

32 the drug discovery process. According to them, during lead optimisation, a computational

34 method screening 10-100 molecules per week with an error of 0.5 kcal/mol would reduce by

36 a factor of 8 the number of compounds requiring synthesis in order to achieve a ten-fold

38 improvement in binding free energy. A similar conclusion can be drawn from the analysis

40 of biological assay variability.61 In our study, TIES has been shown to achieve this level of

42 precision with a short turn around time and hence has considerable potential to influence

44 drug design. Moreover, in this study we have employed TIES on a broad range of target

46 proteins and for a large set of alchemical transformations including many types of chemical

48 interactions and processes such as hydrophobic, hydrogen-bonding and electrostatic inter-

actions and solvent effects including displacement of water molecules from binding pockets. All these interactions have been accurately predicted with TIES within the limits of the accuracy of the classical force-field and ligand parameters employed. Below we draw attention to several cases in which specific interactions impact results.

10 11 12

20 21 22

Figure 7: Cross sections of the SI pocket of thrombin for the two end A-windows of an alchemical transformation involving mutation of m,-chloro benzylamide to beuzamidine. Experimentally, raeia-substituents of benzylamides displace the water molecule from the SI pocket which is present there in the case of an unsubstituted benzylamine ring. Red wireframes show the regions of water occupancy averaged over all the conformations from the 5 replica simulations aligned to the corresponding initial structures, (a) No water molecule bound in the SI pocket in the presence of chlorine (green) at A = 0; (b) Water molecule enters the SI pocket through channels CI and C2 on fully transforming CI to II at A = 1. The protein surface is shown in grey and ligand atoms are colored by element; hydrogen in white, carbon in cyan, oxygen in red and nitrogen in blue.

Figure 7 shows the SI pocket of the active site of the human thrombin protein which contains a water molecule in case of the ligand containing unsubstituted benzylamine ring. This water molecule is locked inside the pocket and mediates hydrogen bonding from the ligand amidino group to the protein. However, the water molecule is displaced by the mife-substituted benzylamide ring of the ligand. We studied the transformation from meta-substituted benzylamide to benzylamine (A) and the reverse transformation (B) using TIES; our results are in excellent agreement with the experimental values. A A Gties for the transformations A and B are —0.6 ± 0.2 kcal/mol and 1.0 ± 0.2 kcal/mol respectively while the corresponding experimental values are -0.9 kcal/mol and 0.9 kcal/mol respectively (no error bars reported). It is noteworthy here that the A A G prediction of the transformation B from FEP/REST31 is 1.5 kcal/mol. In order to quantify the contribution arising from the

presence of the water molecule we calculated the probability of occurrence of a water molecule in a cubic box of volume 1000 Â3 centered around the pocket over all conformations taken from the 5 replica simulations aligned to the initial structure. In the case of transformation A, the probability increases from 0 at A = 0 to 0.7 at A = 1, while in the case of transformation

12 B, it decreases from 0.6 at A = 0 to 0.3 at A =1. Figure 7 shows the two end points (A = 0,1)

14 of transformation A where the red frames denote the space with the number density of water

16 oxygen centers greater than or equal to that in bulk water. Figure 7a shows that the presence

18 of water is confined to the two distinct channels C1 and C2 in the presence of m-Cl while

20 Figure 7b shows that when the chlorine atom is transformed fully to a hydrogen atom the

22 water enters into the S1 pocket as well as through C1 and C2.

36 the standard TIES protocol as mentioned in section 3.2. In this study, 5 out of the 16

38 transformations of MCL1 ligands have a charged carboxylate group in the perturbing region,

40 with the uncertainties of their AAG predictions lying in the range of 0.7-0.8 kcal/mol, while

42 the uncertainties of the remaining 11 predictions for MCL1 ligands lie within the range of

44 0.2-0.4 kcal/mol (see Table 1). In such cases, one can modify the standard protocol described

46 here by, for example, increasing the ensemble size at various A-windows and/or excluding

48 the charged group from the alchemically mutating part of the ligands to further emphasise

50 this point. A comparison of results with and without charged groups in PTP1B ligands is

52 examined in detail in SI.

56 It is evident from Table 1 and Figure 5 that the TIES predictions for MCL1 have larger

58 deviations from the experimental results, with the largest RMSE and MAE. In this case

In the intermediate A-windows of the TIES calculation the electrostatic interactions are weakly scaled. Therefore, when the alchemically changing groups contain charges, it is sometimes possible to encounter larger fluctuations in dV/dA due to instabilities caused by the loss or increase of strong electrostatic interactions. This results in larger variation in the stochastic integral (equation 2), and hence, less precise prediction of AAG using

4 interactions with R263 becomes too weak to hold it in a stable position at intermediate

6 A-windows due to the scaling down of the electrostatic interactions. Such behavior can be

8 attributed to the highly flexible nature of the ligand (detail shown in SI).

11 The free energy methods based on alchemical transformation found in the literature have

13 been applied to cases where the size of the change in the chemical group is small, usually

20 21 22

40 6 Conclusions

56 gives predictions in very good agreement with the experimental values for a large set of

58 ligands bound to five different target proteins. Out of the total 55 predictions made here,

limited to one or two heavy atoms. However, the recently reported FEP/REST method handles perturbations up to 10 heavy atoms.31 In this study, we have achieved the same level by successfully applying TIES to a diverse range of chemical modifications with "perturbations" including up to 10 heavy atoms. For example, the range of chemical transformations with corresponding absolute errors (AE) include benzamide to p-fluoro benzenesulfonamide (AE=0.6 kcal/mol), benzyloxy to ethanamide (AE=0.5 kcal/mol), hydrogen to methyl cy-clohexane (AE=0.1 kcal/mol), cyclohexyl to phenyl (AE=0.5 kcal/mol), cyclohexyl methyl to 2,2,4,4-tetra methyl cyclohexyl (AE=0.2 kcal/mol), cyclopentyl to cycloheptyl (AE=0.0 kcal/mol) and indole to indane (AE=2.1 kcal/mol) to mention a few. TIES correctly predicts that small changes in binding affinity are associated with these large perturbations (up to 10 heavy atoms).30

In this article, a new approach based on thermodynamic integration is described to predict relative binding free energies which has a high level of accuracy and precision. The methodology described here is robust and reliable. Automation of the entire workflow yields results within a few hours. It provides users with the freedom to set desired statistical uncertainties for purposes of accuracy and precision, based on both the length of the MD simulations and the number of replicas per A-window, which can all be adjusted separately. The method

30 deviate from experimental values by less than 0.6 kcal/mol, with an overall mean absolute error of 0.7 kcal/mol and root mean square error of 0.9 kcal/mol. A rich diversity of chemical modifications has been successfully captured with this approach, ranging from single atom perturbation to large functional group modifications (up to 10 heavy atoms).

12 Accurate predictions for flexible ligands with large numbers of rotatable bonds were made.

14 Most importantly, the calculated relative binding affinities are shown to be reproducible

16 within carefully controlled statistical uncertainties. Such a high level of accuracy and preci-

18 sion make the predictions reliable and, hence, this approach may prove of substantial value

20 in the drug discovery and design process.

37 In the present study, the potential parametrisations for both proteins and ligands are all

47 64,65

A direct comparison of our results with those from FEP/REST31 demonstrates enhanced accuracy of predictions for the five biomolecular systems studied in terms of reduced mean absolute error and root mean squared error and improved correlation coefficients (section 4). Unlike TIES where it is a built-in feature, there is no mention of the reproducibility of the FEP/REST calculations. Compounding the issue, the FEP/REST methodology is proprietary and not accessible to open evaluation.

known to be reliable. However, this cannot be guaranteed to be the case in new situations. Care must always be taken in selecting such parametrisations, particularly for new ligands. Problems pertaining to alchemical transformations between congeners involving a change in the net charge using free energy methods are well known and TIES is no exception to them. 64,65 The current version of the TIES protocol cannot reliably handle such situations. Moreover, when alchemically mutating parts of the ligands contain charges it may be necessary to modify the standard TIES protocol to ensure adequate sampling in the calculation of ensemble averages.

TIES represents one of a new class of ensemble-based free energy methods which is rapid, accurate, precise and reproducible. The approach has the potential to make an impact in drug design and personalised medicine.

12 Supporting Information

Tables of all the experimental and predicted AAG values for all five protein targets bound to pairs of ligands are provided, along with the structures of all the ligands studied as well as discussion of some notable ligand-protein interactions studied with TIES. Additionally, hysteresis values for selected ligands forming closed cycles and plots exhibiting the convergence of (dV/dA) are provided.

20 21 22

28 Acknowledgements

50 the UK's national High Performance Computing Service, funded by the Office of Science

52 and Technology through EPSRC's High-End Computing Programme. Access to ARCHER

54 was provided through the 2020 Science programme. This research was supported in part by

56 PLGrid Infrastructure (specifically the Prometheus supercomputer run by ACK Cyfronet

58 AGH in Krakow). A.P.B. is supported by an Overseas Research Scholarship from UCL and

We thank Dr. Frank Lovering of Pfizer for helpful discussions pertaining to the comparison of experimental and theoretical free energy predictions. The authors would like to acknowledge the support of EPSRC via the 2020 Science programme (, EP/I017909/1), the Qatar National Research Fund (7-1083-1-191), the MRC Medical Bioin-formatics project (MR/L016311/1), the EU H2020 projects ComPat (, Grant No. 671564) and CompBioMed (, Grant No. 675451), and funding from the UCL Provost. We acknowledge the Leibniz Supercomputing Centre for providing access to SuperMUC ( and the very able assistance of its scientific support staff. We also made use of ARCHER,

an Inlaks Scholarship from the Inlaks Shivdasani Foundation.

'1) Kong, X.; Brooks, C. L. A-dynamics: A new approach to free energy calculations. J. Chem. Phys. 1996, 105, 2414.

(4) Straatsma, T. P.; Berendsen, H. J. C.; Postma, J. P. M. Free energy of hydrophobic

8 Bibliography

16 (2) Knight, J. L.; Brooks, C. L. A-dynamics free energy simulation methods. J. Comput.

18 Chem. 2009, 30, 1692.

21 (3) Zwanzing, R. W. High-temperature equation of state by a perturbation method in non

23 polar gases. J. Chem. Phys. 1954, 22, 1420.

28 hydration: A molecular dynamics study of nobel gases in water. J. Chem. Phys. 1986,

30 85, 6720.

49 (8) Hamelberg, D.; McCammon, J. A. Standard free energy of releasing a localized water

56 (9) Gilson, M.; Given, J.; Bush, B.; McCammon, J. The statistical-thermodynamic basis

58 for computation of binding affinities: a critical review. Biophys. J. 1997, 72, 1047.

(5) Straatsma, T. P.; Berendsen, H. J. C. Free energy of ionic hydration: Analysis of a thermodynamic integration technique to evaluate free energy differences by molecular dynamics simulation. J. Chem. Phys. 1988 , 89, 5876.

(6) Bennett, C. H. Efficient estimation of free energy differences from Monte Carlo data. J. Comput. Phys. 1976, 22, 245.

(7) Shirts, M. R.; Chodera, J. D. Statistically optimal analysis of samples from multiple equilibrium states. J. Chem. Phys. 2008, 129, 124105.

molecule from the binding pockets of proteins: Double-decoupling method. J. Am. Chem. Soc. 2004, 126, 7683.

10 11 12

20 21 22

'10) Hansson, T.; Äqvist, J. Estimation of binding free energies for HIV proprotein inhibitors by molecular dynamics simulations. Protein Eng. 1995, 8, 1137.

'11) Äqvist, J.; Medina, C.; Samuelsson, J. E. A new method for predicting binding affinity in computer-aided drug design. Protein Eng. 1994, 7, 385.

'12) Äqvist, J. Calculation of absolute binding free energies for charged ligands and eeffect of long-range electrostatic interactions. J. Comput. Chem. 1996, 17, 1587.

'13) Kollman, P. A.; Massova, I.; Reyes, C.; Kuhn, B.; Huo, S.; Chong, L.; Lee, M.; Lee, T.; Duan, Y.; Wang, W.; Donini, O.; Cieplak, P.; Srinivasan, J.; Case, D. A.; Cheatham, T. E. Calculating structures and free energies of complex molecules: Combining molecular mechanics and continuum models. Acc. Chem. Res. 2000, 33, 889, PMID: 11123888.

14) Postma, J. P. M.; Berendsen, H. J. C.; Haak, J. R. Thermodynamics of cavity formation in water. Faraday Symp. Chem. Soc. 1982, 17, 55.

'15) Tembe, B. L.; McCammon, J. A. Ligand-receptor interactions. Comput. Chem. 1984, 8, 281.

'16) Jorgensen, W. L.; Ravimohan, C. Monte Carlo simulation of differences in free energies of hydration. J. Chem. Phys. 1985, 83, 3050.

'17) Sadiq, S. K.; Wright, D. W.; Kenway, O. A.; Coveney, P. V. Accurate ensemble molecular dynamics binding free energy ranking of multidrug-resistance HIV-1 protease. J. Chem. Inf. Model. 2010, 50, 890.

'18) Wright, D. W.; Hall, B. A.; Kenway, O. A.; Jha, S.; Coveney, P. V. Computing clinically relevant binding free energies of HIV-1 protease inhibitors. J. Chem. Theory Comput. 2014, 10, 1228.

10 11 12

20 21 22

(19) Wan, S.; Knapp, B.; Wright, D. W.; Deane, C. M.; Coveney, P. V. Rapid, precise and reproducible prediction of peptide-MHC binding affinities from molecular dynamics that correlate well with experiment. J. Chem. Theory Comput. 2015, 11, 3346.

(20) Coveney, P. V.; Wan, S. On the calculation of equilibrium thermodynamic properties from molecular dynamics. Phys. Chem. Chem. Phys. 2016, 18, 30236, 10.1039/c6cp02349e.

(21) Genheden, S.; Ryde, U. How to obtain statistically converged MM/GBSA results. J. Comput. Chem. 2010, 31, 837.

(22) Auffinger, P.; Louise-May, S.; Westhof, E. Multiple molecular dynamics simulations of the anticodon loop of tRNAAsp in aqueous solution with counterions. J. Am. Chem. Soc. 1995, 117, 6720.

(23) Zagrovic, B.; van Gunsteren, W. F. Computational analysis of the mechanism and thermodynamics of inhibition of phosphodiesterase 5A by synthetic ligands. J. Chem. Theory Comput. 2007, 3, 301, PMID: 26627173.

(24) Genheden, S.; Ryde, U. A comparison of different initialization protocols to obtain statistically independent molecular dynamics simulations. J. Comput. Chem. 2011, 32, 187.

(25) Lawrenz, M.; Baron, R.; McCammon, J. A. Independent-trajectories thermodynamic-integration free-energy changes for biomolecular systems: Determinants of H5N1 avian influenza virus neuraminidase inhibition by peramivir. J. Chem. Theory Comput. 2009, 5, 1106, PMID: 19461872.

(26) Wan, S.; Bhati, A. P.; Zasada, S.; Coveney, P. V. Rapid, accurate, precise and reproducible ligand-protein binding free energy prediction. 2016; preprint.

10 11 12

17 2011, 8, 1114.

33 ACS National Meeting. Philadelphia, USA.

38 son, S.; Dahlgren, M. K.; Greenwood, J.; Romero, D. L.; Masse, C.; Knight, J. L.;

40 Steinbrecher, T.; Beuming, T.; Damm, W.; Harder, E.; Sherman, W.; Brewer, M.;

42 Wester, R.; Murcko, M.; Frye, L.; Farid, R.; Lin, T.; Mobley, D. L.; Jorgensen, W. L.;

44 Berne, B. J.; Friesner, R. A.; R., A. Accurate and reliable prediction of relative lig-

46 and binding potency in prospective drug discovery by way of a modern free-energy

48 calculation protocol and force field. J. Am. Chem. Soc. 2015, 137, 2695.

51 (32) Aldeghi, M.; Heifetz, A.; Bodkin, M. J.; Knapp, S.; Biggin, P. C. Accurate calculation

53 of the absolute free energy of binding for drug molecules. Chem. Sci. 2016, 7, 207.

56 (33) Lawrenz, M.; Baron, R.; Wang, Y.; McCammon, J. A. In Computational Drug Discovery

58 and Design; Baron, R., Ed.; Springer New York, 2012; Vol. 819; Chapter Independent-

27) Bunney, T. D.; Wan, S.; Thiyagarajan, N.; Sutto, L.; Williams, S. V.; Ashford, P.; Koss, H.; Knowles, M. A.; Gervasio, M. A.; Coveney, P. V.; Katan, M. The effect of mutations on drug sensitivity and kinase activity of fibroblast growth factor receptors: A combined experimental and theoretical study. EBioMedicine 2015, 2, 194.

28) Wan, S.; Coveney, P. V. Rapid and accurate ranking of binding affinities of epidermal growth factor receptor sequences with selected lung cancer drugs. J. R. Soc. Interface

(29) Wan, S.; Bhati, A. P.; Zasada, S. J.; Wall, I.; Green, D.; Bamborough, P.; Coveney, P. V. Rapid and reliable binding affinity prediction for analysis of bromodomain inhibitors: a computational study. 2016; preprint.

(30) Wan, S.; Bhati, A.; Coveney, P.; Skerratt, S. E.; Gore, K.; Bagal, S. K.; Shanmugasun-daram, V.; Omoto, K. Rapid, accurate and reproducible binding affinity calculation for drug discovery: A retrospective analysis of the Pfizer Pan-Trk program. 2016; 252nd

(31) Wang, L.; Wu, Y.; Deng, Y.; Kim, B.; Pierce, L.; Krilov, G.; Lupyan, D.; Robin-

10 11 12

20 21 22

42 (40) Genheden, S.; Ryde, U. Will molecular dynamics simulations of proteins ever reach

44 equilibrium? Phys. Chem. Chem. Phys. 2012, 14, 8662.

47 (41) Genheden, S.; Mikulskis, P.; Hu, L.; Kongsted, J.; Söderhjelm,; Ryde, U. Accurate

49 prediction of nonpolar solvation free energies require explicit consideration of binding-

51 site hydration. J. Am. Chem. Soc. 2011, 133, 13081.

54 (42) Fujitani, H.; Tanida, Y.; Ito, M.; Jayachandran, G.; Snow, C. D.; Shirts, M. R.;

56 Sorin, E. J.; Pande, V. S. Direct calculation of the binding free energies of FKBP

58 ligands. J. Chem. Phys. 2005, 123, 084108.

ACS Paragon Plus Environment

Trajectory Thermodynamic Integration: A Practical Guide to Protein-Drug Binding Free Energy Calculations Using Distributed Computing, pp 469-486.

(34) Lindorff-Larsen, K.; Maragakis, P.; Piana, S.; Eastwood, M. P.; Dror, R. O.; Shaw, D. E. Systematic validation of protein force fields against experimental data. PLoS One 2012, 7, 1.

(35) Sadiq, S. K.; Mazzea, M. D.; Zasada, S.; Manos, S.; Stoica, I.; Gale, C. V.; Watson, S. J.; Kellam, P.; Brew, S.; Coveney, P. V. Patient-specific simulation as a basis for clinical decision-making. Phil. Trans. R. Soc. A 2008, 366, 3199.

(36) Straatsma, T. P.; McCammon, J. A. Multiconfigurational thermodynamic integration. J. Chem. Phys. 1991, 95, 1175.

(37) Straatsma, T. P.; McCammon, J. A. Computational alchemy. Annu. Rev. Phys. Chem. 1992, 43, 407.

(38) Caves, L. S. D.; Evanseck, J. D.; Karplus, M. Locally accessible conformations of proteins: Multiple molecular dynamics simulations of crambin. Protein Sci. 1998, 7, 649.

(39) Elofsson, A.; Nilsson, L. How consistent are molecular dynamics simulations? J. Mol. Biol. 1993, 233, 766.

(43) Friberg, A.; Vigil, D.; Zhao, B.; Daniels, R. N.; Burke, J. P.; Garcia-Barrantes, P. M.; Camper, D.; Chauder, B. A.; Lee, T.; Olejniczak, E. T.; Fesik, S. W. Discovery of potent myeloid cell leukemia 1 (Mcl-1) inhibitors using fragment-based methods and structure-based design. J. Med. Chem. 2013, 56, 15, PMID: 23244564.

(44) Wilson, D. P.; Wan, Z.-K.; Xu, W.-X.; Kirincich, S. J.; Follows, B. C.; JosephMcCarthy, D.; Foreman, K.; Moretto, A.; Wu, J.; Zhu, M.; Binnun, E.; Zhang, Y.-L.; Tam, M.; Erbe, D. V.; Tobin, J.; Xu, X.; Leung, L.; Shilling, A.; Tam, S. Y.; Man-

(45) Baum, B.; Mohamed, M.; Zayed, M.; Gerlach, C.; Heine, A.; Hangauer, D.; Klebe, G.

10 11 12

19 sour, T. S.; Lee, J. Structure-based optimization of protein tyrosine phosphatase 1B

21 inhibitors: From the active site to the second phosphotyrosine binding site. J. Med.

23 Chem. 2007, 50, 4681, PMID: 17705360.

28 More than a simple lipophilic contact: A detailed thermodynamic analysis of nonbasic

30 residues in the S1 pocket of thrombin. J. Mol. Biol. 2009, 390, 56.

35 Blair, W. S.; Chang, C.; Driscoll, J.; Eigenbrot, C.; Ghilardi, N.; Gibbons, P.; Halla-

37 day, J.; Johnson, A.; Kohli, P. B.; Lai, Y.; Liimatta, M.; Mantik, P.; Menghrajani, K.;

39 Murray, J.; Sambrone, A.; Xiao, Y.; Shia, S.; Shin, Y.; Smith, J.; Sohn, S.; Stan-

41 ley, M.; Ultsch, M.; Zhang, B.; Wu, L. C.; Magnuson, S. Lead identification of novel

43 and selective TYK2 inhibitors. Eur. J. Med. Chem. 2013, 67, 175.

46 (47) Liang, J.; van Abbema, A.; Balazs, M.; Barrett, K.; Berezhkovsky, L.; Blair, W.;

48 Chang, C.; Delarosa, D.; DeVoss, J.; Driscoll, J.; Eigenbrot, C.; Ghilardi, N.; Gib-

50 bons, P.; Halladay, J.; Johnson, A.; Kohli, P. B.; Lai, Y.; Liu, Y.; Lyssikatos, J.;

52 Mantik, P.; Menghrajani, K.; Murray, J.; Peng, I.; Sambrone, A.; Shia, S.; Shin, Y.;

54 Smith, J.; Sohn, S.; Tsui, V.; Ultsch, M.; Wu, L. C.; Xiao, Y.; Yang, W.; Young, J.;

56 Zhang, B.; Zhu, B.-y.; Magnuson, S. Lead optimization of a 4-aminopyridine benzamide

(46) Liang, J.; Tsui, V.; Abbema, A. V.; Bao, L.; Barrett, K.; Beresini, M.; Berezhkovskiy, L.;

10 11 12

20 21 22

24 28, 235.

scaffold to identify potent, selective, and orally bioavailable TYK2 Inhibitors. J. Med. Chem. 2013, 56, 4521, PMID: 23668484.

48) Hardcastle, I. R.; Arris, C. E.; Bentley, J.; Boyle, F. T.; Chen, Y.; Curtin, N. J.; En-dicott, J. A.; Gibson, A. E.; Golding, B. T.; Griffin, R. J.; Jewsbury, P.; Menyerol, J.; Mesguiche, V.; Newell, D. R.; Noble, M. E. M.; Pratt, D. J.; Wang, L.-Z.; ; Whit-field, H. J. N2-substituted O6-cyclohexylmethylguanine derivatives: Potent inhibitors of cyclin-dependent kinases 1 and 2. J. Med. Chem. 2004, 47, 3710, PMID: 15239650.

49) Berman, H. M.; Westbrook, J.; Feng, Z.; Gilliland, G.; Bhat, T. N.; Weissig, H.; Shindyalov, I. N.; Bourne, P. E. The Protein Data Bank. Nucleic Acids Res. 2000,

(50) Pearlman, D. A. A comparison of alternative approaches to free energy calculation. J. Phys. Chem. 1994, 98, 1487.

(51) Wang, B.; Li, L.; Hurley, T. D.; Meroueh, S. O. Molecular Recognition in a Diverse Set of Protein-Ligand Interactions Studied with Molecular Dynamics Simulations and End-Point Free Energy Calculations. J. Chem. Inf. Model. 2013, 53, 2659, PMID: 24032517.

(52) Wang, J.; Wolf, R. M.; Caldwell, J. W.; Case, D. A. Development and testing of a general amber force field. J. Comput. Chem. 2004, 25, 1157.

(53) Phillips, J. C.; Braun, R.; Wang, W.; Gumbart, J.; Tajkhorshid, E.; Villa, E.; Chipot, C.; Skeel, R. D.; Kal, L.; Schulten, K. Scalable molecular dynamics with NAMD. J. Comput. Chem. 20 05 , 26, 1781.

(54) Zacharias, M.; Straatsma, T. P.; McCammon, J. A. Separation-shifted scaling, a new scaling method for Lennard-Jones interactions in thermodynamic integration. J. Chem. Phys. 1994, 100, 9025.

10 11 12

20 21 22

(55) Beutler, T. C.; Mark, A. E.; Schaik, R. C. v.; Gerber, P. R.; Gunsteren, W. F. v. Avoiding singularities and numerical instabilities in free energy calculations based on molecular simulations. Chem. Phys. Lett. 1994, 222, 529.

(56) Sadiq, S. K.; Wright, D. W.; Watson, S. J.; Zasada, S.; Stoica, I.; Coveney, P. V. Automated molecular simulation based binding affinity calculator for ligand-bound HIV-1 protease. J. Chem. Inf. Model. 2008, 48, 1909.

(57) Highfield, R. Science Museum Blog. supercomputer-bid-to-create-the-first-truly-personalised-medicine/,

(accessed on November 2, 2016).

(58) 0ksendal, B. K. Stochastic differential equations : an introduction with applications; Universitext; Springer: Berlin, Heidelberg, New York, 1998; Corrected second printing 2000.

(59) Wan, S.; Coveney, P. V.; Flower, D. R. Molecular Basis of Peptide Recognition by the TCR: Affinity Differences Calculated Using Large Scale Computing. J. Immunol. 2005, 175, 1715-1723.

(60) Groen, D.; Bhati, A. P.; Suter, J.; Hetherington, J.; Zasada, S. J.; Coveney, P. V. FabSim: Facilitating computational research through automation on large-scale and distributed e-infrastructures. Comput. Phys. Commun. 2016, 207, 375.

(61) Kramer, C.; Dahl, G.; Tyrchan, C.; Ulander, J. A comprehensive company database analysis of biological assay variability. Drug Discov Today 2016, 21, 1213.

(62) Chodera, J. D.; Mobley, D. L. Entropy-enthalpy compensation: role and ramifications in biomolecular ligand recognition and design. Annu. Rev. Biophys. 2013, 42, 121.

(63) Mobley, D. L.; Klimovich, P. V. Perspective: Alchemical free energy calculations for drug discovery. J. Chem. Phys. 2012, 137, 230901.

10 11 12

20 21 22

(64) Lin, Y.-L.; Aleksandrov, A.; Simonson, T.; Roux, B. An overview of electrostatic free energy computations for solutions and proteins. J. Chem. Theory Com,put. 2014, 10, 2690, PMID: 26586504.

(65) Rocklin, G. J.; Mobley, D. L.; Dill, K. A.; Hunenberger, P. H. Calculating the binding free energies of charged species based on explicit-solvent simulations employing lattice-sum methods: An accurate correction scheme for electrostatic finite-size effects. J. Chem,. Phys. 2013, 139, 184103.

TOC graphic