Concept and prototype of protein–ligand docking simulator with force ...

BIOINFORMATICS

Vol. 18 no. 1 2002 Pages 140–146

Concept and prototype of protein–ligand docking simulator with force feedback technology Hiroshi Nagata 1, Hiroshi Mizushima 1 and Hiroshi Tanaka 2 1 National

Cancer Center Research Institute, 5-1-1, Tsukiji, Chuo-ku, Tokyo 104-0045, Japan and 2 Tokyo Medical and Dental University, 1-5-45, Yushima, Bunkyo-ku, Tokyo 113-8549, Japan

Received on April 11, 2001; revised on July 28, 2001; accepted on September 5, 2001

ABSTRACT A novel concept for a protein–ligand docking simulator using Virtual Reality (VR) technologies, in particular the tactile sense technology, was designed and a prototype was developed. Most conventional docking simulators are based on numerical differential calculations of the total energy between a protein and a ligand. However, the basic concept of our method differs from that of conventional simulators. Our design utilizes the force between a ligand and a protein instead of the total energy. The most characteristic function of the system is its ability to enable the user to ‘touch’ and sense the electrostatic potential field of a protein molecule. The user can scan the surface of a protein using a globular probe, which is given an electrostatic charge, and is controlled by a force feedback device. The electrostatic force between the protein and the probe is calculated in real time and immediately fed back into the force feedback device. The user can easily search interactively for positions where the probe is strongly attracted to the force field. Such positions can be regarded as candidate sites where functional groups of ligands corresponding to the probe can bind to the target protein. Certain limitations remain; for example, only twenty protein atoms can be used to generate the electrostatic field. Furthermore, the system can only use globular probes, preventing drug molecules or small chemical groups from being simulated. These limitations are the result of our insufficient computer resources. However, our prototype system has the potential to become a novel application method as well as being applicable to conventional VR technologies, especially to force feedback technologies. Contact: [email protected]; [email protected]; [email protected]

1

INTRODUCTION

We adopted Virtual Reality (VR) technologies, especially tactile sense technology, to approach the protein–ligand docking problem and developed a new algorithm for docking simulations. We then designed a docking simulator 140

according to this algorithm and developed a prototype. The prototype enables users to tactually sense electrostatic forces between proteins and various probes, which represent functional groups or small molecules capable of becoming parts of ligand molecules, using a force feedback device known as ‘haptics.’ Through interactive scanning, the user can easily search for probe-docking points where the probes strongly adhere to the protein surface. Our concept and method are anticipated to be useful for simulating protein–ligand docking. Recent protein studies have identified many 3D structures of new therapeutically relevant target proteins through structural determination by x-ray crystallography, NMR spectroscopy, or homology modeling. As a result of this growing structural knowledge, the docking problem is becoming an essential concept in rational drug design. Kuntz et al. (1982) presented the DOCK program. DOCK is based on a sphere-matching procedure. Subsequently, many docking algorithms and systems have been developed (Goodford, 1985; Leach, 1994; Mizutani et al., 1994; Gschwend et al., 1996; Trosset and Scheraga, 1998; Verkhivker et al., 1999). In principle, however, the docking problem can also be tackled by applying energy minimization techniques. The main disadvantages of these algorithms are that the results depend on the initial placement of the ligands because calculations can be easily trapped at local minimum points. Goodsell and Olson (1990) applied simulated annealing to the docking problem. Although this global optimization technique in principle avoids trapping at local minima, simulations using this algorithm at different starting points produced different results. Automatically solving the local minima problem by computer seems to be very difficult. We thought that human decision and interactive intervention during docking simulation was necessary to overcome this problem. We decided to adopt VR technologies to solve the docking problem because VR provides an interactive and easy-to-use human interface. The incorporation of VR into molecular science has only just begun. Several types c Oxford University Press 2002


Protein–ligand docking simulator

of software for displaying 3D models of bio-molecules have been developed. Most of these programs use Virtual Reality Modeling Language (VRLM; Forster et al., 1999; Campagne et al., 1999; Reichert et al., 2000). However, drug design support using force feedback technology is still in its infancy. In this study, we attempted to use force feedback technology for protein–ligand docking and succeeded in the development of a prototype for a unique docking simulator. We consider force feedback technology to have enormous potential for improving the methodology of docking simulations and drug design.

2

NEW CONCEPT FOR PROTEIN–LIGAND DOCKING SIMULATIONS A protein molecule is surrounded by the molecular potential field which consists of the electrostatic potential, van der Waals potential, hydrogen bond potential, etc. which together generate the potential energy between a protein and a ligand (Goodford, 1985). The total potential energy can be expressed by the following formula (1). E=

V  W 

(El j (v, w) + E el (v, w) + E hb (v, w)). (1)

v=1 w=1

In this formula, V and W are the numbers of atoms in the ligand and the protein, respectively. El j , E el , and E hb are the van der Waals (Lenard–Jones) potential energy, electrostatic potential energy, and hydrogen bond potential energy between the vth atom of the ligand and the wth atom of the protein. Each energy term is expressed as a function of the distance between the vth atom of the ligand and the wth atom of the protein. Most conventional docking algorithms calculate the numeric differentials of the total potential energy by using the atom distances to search local minimum points where the differential values become zero. The differential of the potential energy means the potential, i.e. the force, working between the protein and the ligand. Visually expressing the overall force potential affecting all ligands is very difficult. The strength and sign of the force change according to the ligand structure, the distance to the protein, and both the position and direction of the ligand. Moreover, many basins, valleys, mountains and peaks are present on the potential surface. However, force fields can be experienced by tactile sensations using force feedback technology. Figure 1 illustrates this concept. First, a protein molecule is set in a VR space and a ligand molecule is prepared in the same VR space. The position and coordinates of the ligand are controlled using a force feedback device. The total force affecting the ligand is calculated in real time and simultaneously fed back to the device. This method seems to be more effective and an easier way to search for ligand-

Fig. 1. Basic concept behind applying the force feedback technology to drug designing.

binding sites than conventional numerical differential methods. In this study, we used GRID potential energies (Goodford, 1985; Reynolds et al., 1989; Boobbyer et al., 1989). Each energy term is given by formulas (2)–(4). A B − 6 (2) 12 dvw d  vw  pv q w  1 (ξ − ε)/(ξ + ε)  +  (3) E el (v, w) = Kξ dvw 2 + 4s s dvw p q  C D − 4 (4) cosk θ cos2 ψ. E hb (v, w) = 6 dvw dvw El j (v, w) =

In these formulas, dvw is the distance between the vth and the wth atoms, pv and qw are the electrostatic charges of the atoms, K is the combination of a geometrical factor and natural constants, ξ is the dielectric constant of the protein surface, ε is the dielectric constant of the solvent; s p is the depth of the mth atom of the ligand molecule on the protein surface, sq is the depth of the nth atom of the protein molecule on its surface, and dmn is the distance between the mth atom of the ligand molecule and the nth atom of the protein molecule. The forces of each potential exerted on the vth atom by the protein are given by formulas (5)–(7) which represent the summations of differentiations according to the distance dvw of formulas (2)–(4). W   A B − 12 13 + 6 7 (5) Fl j (v) = dvw dvw w=1  W   dvw (ξ − ε)/(ξ + ε) pq 1 Fel (v) = + − (6) 2 2 + 4s s )3/2 K ξ dvw (dvw p q w=1 141

H.Nagata et al.

Fhb (v) =

W  

−6

w=1

C D +4 5 7 dvw dvw



cosk θ cos2 ψ . (7)

Table 1. The specifications of the prototype system

Computers

1. SGI workstation (OCTAINE MXI; memory 256 MB) 2. PC (Pentium 500 MHz; memory 196 MB; OS Windows NT 4.0)

Force feedback device

PHANToMTM DESKTOP

Force feedback

0 to 1.5 N, Prevent sticking of the protein and ligand

Potential field

Electrostatic potential field generated by maximum 20 atoms

The total force exerted on the ligand is given by formula (8). FTotal =

V 

(Fl j (v) + Fel (v) + Fhb (v)).

(8)

v=1

Ligand molecules approach the specific binding sites of target proteins guided by the FTotal . In the FTotal , the most important terms are the electrostatic forces. The van der Waals and hydrogen bond potentials, which operate only over short distances, work mainly to assure the final binding of the ligand and protein. In contrast, the electrostatic potential is thought to play a major role in attracting drug molecules to nearby binding sites, because it can work over relatively long distances. Several experiments and simulations support this hypothesis (Tan et al., 1993; Kozack et al., 1995; Wade et al., 1998; Wlodek et al., 2000). Our ultimate goals are to calculate and interactively feedback the FTotal to a force feedback device in real time. However, achieving these goals simultaneously appears to be very difficult. Therefore, we decided to only utilize the electrostatic potential. We then designed and implemented the simple prototype described below.

3

PROTOTYPE REQUIREMENTS, SPECIFICATIONS AND DESIGN We designed and developed a prototype system in which variations in electrostatic forces can be felt while scanning the surface of the protein using the ligand molecule as a probe. Our prototype system has the following requirements: (a) a VR space must be created and protein and ligand molecules, which are displayed as 3D computer graphics, are placed within this space; (b) the coordinates and positions of the ligand molecule are input using a force feedback device; (c) the user can scan the electrostatic potential field of the protein surface using the ligand as a probe; (d) the electrostatic force, which the probe receives from the potential, is calculated and fed back to the force feedback device in real time; (e) the system recognizes collisions between the surfaces of the ligand and protein, and prevents them from sticking to each other. The specifications for the prototype system are shown in Table 1. 142

Molecular graphics

Space filling model, Connolly model

Probe

Globular. The charge and the radius can be changed freely

Operation of protein

Rotatable to the X –Y –Z axis

Calculation of atomic charges

MOPAC

We used two computers to develop the prototype. A graphic workstation was used to create a VR space and a graphical user interface. In addition, a Windows NT Personal Computer (PC) was used to control the PHANToMTM (SensAble Technologies Inc.) and calculate the electrostatic force. The PHANToM series is the most popular force feedback device in the world. Its control software can be easily developed using GHOST, the exclusive programming library for the PHANToM. Initially, we planned to implement all of the software on a PC, but this was difficult because the CPU power was insufficient. Therefore, we divided the software between the two computers. The normal space filling model and the Connolly surface model were adopted to create the protein molecule graphics (Langridge et al., 1981; Connolly, 1983). The position of the probe is input by the PHANToM interactively. The PC immediately calculates the electrostatic force working on the probe and feeds this information back to the PHANToM. The force exerted on the probe is then fed back to the PHANToM according to a linear relationship based on 1 atomic unit (au) = 1 newton (N). However, we limited the maximum output of the PHANToM to 1.5 N, to avoid system crashes. The electrostatic force on the probe must be calculated using all of the atoms in the protein, but we had to limit the number of atoms to 20 to calculate the force in real time. As mentioned above, the probe is a spherical object, and an arbitrary radius and charge can be assigned to it. Positional control of the probe was not attempted at this time. The placement of a protein and a probe in a VR space is an important problem. The probe is operated by the PHANToM, and the operation range is limited to the

Protein–ligand docking simulator

Fig. 2. The VR space coordinates.

Fig. 3. Design of prototype system.

operation radius of the PHANToM. The VR space must be defined based on the operation radius of the PHANToM. Additionally, the radius of the VR space can be used as the operation radius of the PHANToM. A protein molecule can be placed in an arbitrary position within this VR space. However, for our purpose, we decided that the center of the protein should correspond to the center of the VR space. Also, the space must be secured in order to scan the protein surface with the probe. To facilitate scanning, the protein should be rotatable on the X , Y and Z axes of the VR space. Therefore, we defined the VR space as shown in Figure 2. The design for the prototype system is shown in Figure 3. The PC and the workstation are linked by a TCP/IP LAN, which translates data on the coordinates of the probe and the protein.

Figure 4 shows the software design for the PC, and Figure 5 illustrates the software design for the workstation. The workstation software provides a graphical user interface, which consists of functions for drawing and rotating a protein molecule, transmitting data between the PC, and drawing and moving the probe in the VR space according to the coordinate data transmitted from the PC. It also provides the user for the necessary interface for changing the properties of the probe. Several kinds of probes can be registered, and the user can select and change them interactively. The points where the probe is strongly attracted can also be marked and stored. Protein 3D data can be obtained from the Brookhaven Protein Data Bank over the Internet. The user selects twenty atoms to calculate the electrostatic force. The electrostatic charges of the selected atoms must then be calculated. We decided to calculate the charges using MOPAC (Stewart, 1990), which is one of the most popular types of software in the field of computer chemistry suitable for such calculations. The properties of the probe, the protein structure data, and information on the rotation of the protein and the 20 atoms selected by the user are stored in shared memory and sent to the PC. The PC calculates the electrostatic force and feeds the information back to the PHANToM. The PC also calculates the coordinates of the probe in the VR space according to the movements of the PHANToM and sends the results to the workstation.

4 IMPLEMENTATION AND RESULTS Figure 6 shows the user interface of the workstation software. The graphical user interface was implemented using Open GL, and the internal processing functions were implemented using the C++ programming language. The left half of the user interface shows the VR space, in which the protein is displayed. The prominently displayed points are the atoms that have been selected to generate the force field. Several control buttons are also implemented in this area. All of the buttons are implemented in VR and can be selected using the PHANTOM. The right half shows the system status and the communication situation with the PC. The PC software was implemented using Visual C++ and GHOST, the control function library for the PHANToM. It has two windows as does the workstation software. One is for displaying the same VR space as that of the workstation, with a limitation that it can display only the atoms selected to generate the force field. Another window monitors the communication status. The quality of the molecular graphics is much lower than that of the workstation, such that the VR space of the PC is mainly for debugging. However, it may be possible to generate and display high-quality molecular graphics on the PC if its CPU power is greatly improved. In such a situation, the workstation would become unnecessary, and 143

H.Nagata et al.

Fig. 4. Design of the workstation software.

Fig. 5. Design of the PC software.

the system design would become much simpler. We carried out various tests to confirm the operation of the system. For example, a virtual atom with a radius 144

of 1 xxx (approximately equal to that of an oxygen atom) was given a charge of −1 and placed in the VR space. We then scanned it using a probe with a

Protein–ligand docking simulator

Fig. 7. Illustration of the scanning of a molecule using the PHANToMTM . Fig. 6. User interface of the prototype.

radius of 0.5 xxx (approximately equal to that of a hydrogen atom) and a charge of +1. The probe was strongly attracted to the virtual atom. The electrostatic force was inversely proportional to the second power of the distance. The attractive force increased rapidly as the probe approached the virtual atom, while the force was rapidly attenuated as the probe was moved away. We also increased the number of virtual atoms and repeated the tests, ultimately concluding that the system had been implemented correctly. Next, we evaluated the system employing actual protein data. We used several enzymes that are known to be targets of anti-cancer drugs. For example, dihydrofolate reductase (Matthews et al., 1977; Olliaro and Yuthavong, 1999; Burgen, 2000), one of the most important enzymes for cancer chemotherapy, was prepared. The binding sites of these drugs have been identified. Twenty atoms on the exposed surface of the protein binding site were selected, and the force field was generated. The charges of these atoms were calculated using MOPAC. We scanned the force field using the PHANToM (Figure 7) and succeeded in ‘feeling’ the complex structure of the electrostatic force field. Because the probe was subjected to the attractive and repulsive forces from twenty atoms simultaneously, the direction and strength of the force changed even when the probe was moved only slightly. Moreover, several deep basins surrounded by high barriers of repulsive forces were identified. The probe naturally dropped into these basins, which were local minimum points in the total energy. Conventional docking algorithms are usually easily trapped at these local minimums and have difficulty in escaping. In our system, on the other hand, the probe was able to escape local minimums smoothly by the

interactive operations. We also confirmed that the force changes when the radius and charge of the probe are changed. These results confirm the ability to scan and ‘feel’ the electrostatic potential field of a protein using tactile sense technologies.

5 DISCUSSION We have developed a new concept for a protein–ligand docking simulator based on force feedback VR technology. We have also implemented a prototype system that enables users to ‘feel’ the electrostatic potential field surrounding a protein molecule and search for docking sites of probes representing chemical groups of ligand molecules. The prototype still has several limitations, mainly as a result of our insufficient computer resources. Several improvements mentioned below are necessary to realize a practical docking simulator. First, the computer power must be improved by 10to 100-fold. Our prototype can use only twenty atoms to generate the electrostatic field, but the ligand-binding site of a protein is usually composed of several hundred atoms. Parallel processing is suitable to this problem because every term in the formula (7) has the same form and only the parameter values are different. There are three methods of parallel processing: (a) use a super computer that supports parallel computing, (b) distributed parallel processing over a network, and (c) integration of a special computer using Application Specific Integrated Circuits (ASICs) to calculate the electrostatic forces. A dedicated super computer for docking simulation is preferable, but such a system may be fairly difficult to obtain. Network parallel processing seems to be feasible because it allocates a small amount of calculations to many PCs and workstations over the network that are relatively under utilized. However, real time processing may be 145

H.Nagata et al.

difficult because this method involves large overheads of exchanging packets among computers. We think that the third solution is the most feasible. At this time, we have only included calculations for the electrostatic potential. However, several other potentials, including van der Waals potential and hydrogen bond potential, are more important than the electrostatic potential in the final stage of creating chemical bonds between a drug and a protein. These potentials must be considered in accurate searches for probe binding sites. To achieve this, ASICs seem to be the best solution. We are presently preparing designs for the implementations of ASICs. It is also necessary to improve the probe. The probe that was used in this study was globular and had a single charge. However, our goal is to use a ligand molecule as the probe. A ligand molecule generally consists of 10– 100 atoms, which have individual charges and various other chemical properties. Additionally, ligand molecules have unique shapes and structures. Therefore, the position of the prove must be controllable if it is to be used to represent a ligand. Ligands are affected by various repulsive and attractive forces from various directions, depending on the molecular potential of the protein. These forces produce torques on the ligands around the holding points of the PHANToM. Automatic positional control utilizing these torques may be possible. In this study, we did not confirm whether or not the deepest force basin was the global energy minimum. We are planning to challenge this problem in future studies. With these improvements and additional investigations, we believe that our system will be useful for molecular docking and drug design in the future.

REFERENCES Boobbyer,D.N., Goodford,P.J., McWhinnie,P.M. and Wade,R.C. (1989) New hydrogen-bond potentials for use in determining energetically favorable binding sites on molecules of known structure. J. Med. Chem., 32, 1083–1094. Burgen,A.S. (2000) Targets of drug action. Ann. Rev. Pharmacol. Toxicol., 40, 1–16. Campagne,F., Jestin,R., Reversat,J.L., Bernassau,J.M. and Maigret,B. (1999) Visualisation and integration of G protein-coupled receptor related information help the modelling: description and applications of the Viseur program. J. Comput. Aided Mol. Des., 13, 625–643. Connolly,M.L. (1983) Solvent-accessible surfaces of proteins and nucleic acids. Science, 221, 709–713. Forster,M., Heath,A. and Afzal,M. (1999) Application of distance geometry to 3D visualization of sequence relationships. Bioinformatics, 15, 89–90. Goodford,P.J. (1985) A computational procedure fordetermining

146

energetically favorable binding sites on biologically important macromolecules. J. Med. Chem., 28, 849–857. Goodsell,D.S. and Olson,A.J. (1990) Automated docking of substrates to proteins by simulated annealing. Proteins, 8, 195–202. Gschwend,D.A., Good,A.C. and Kuntz,I.D. (1996) Molecular docking towards drug discovery. J. Mol. Recognit., 9, 175–186. Kozack,R.E., d’Mello,M.J. and Subramaniam,S. (1995) Computer modeling of electrostatic steering and orientational effects in antibody–antigen association. Biophys. J., 68, 807–814. Kuntz,I.D., Blaney,J.M., Oatley,S.J., Langridge,R. and Ferrin,T.E. (1982) A geometric approach to macromolecule–ligand interactions. J. Mol. Biol., 161, 269–288. Langridge,R., Ferrin,TE., Kuntz,I.D. and Connolly,M.L. (1981) Real-time color graphics in studies of molecular interactions. Science, 211, 661–666. Leach,A.R. (1994) Ligand docking to proteins with discrete sidechain flexibility. J. Mol. Biol., 235, 345–356. Matthews,D.A., Alden,R.A., Bolin,J.T., Freer,S.T., Hamlin,R., Xuong,N., Kraut,J., Poe,M., Williams,M. and Hoogsteen,K. (1977) Dihydrofolate reductase: x-ray structure of the binary complex with methotrexate. Science, 197, 452–455. Mizutani,M.Y., Tomioka,N. and Itai,A. (1994) Rational automatic search method for stable docking models of protein and ligand. J. Mol. Biol., 243, 310–326. Olliaro,P.L. and Yuthavong,Y. (1999) An overview of chemotherapeutic targets for antimalarial drug discovery. Pharmacol. Ther., 81, 91–110. Reichert,J., Jabs,A., Slickers,P. and Suhnel,J. (2000) The IMB Jena Image Library of biological macromolecules. Nucleic Acids Res., 28, 246–249. Reynolds,C.A., Wade,R.C. and Goodford,P.J. (1989) Identifying targets for bioreductive agents: using GRID to predict selective binding regions of proteins. J. Mol. Graph., 7, 103–108, 100. Stewart,J.J. (1990) MOPAC: a semiempirical molecular orbital program. J. Comput. Aided Mol. Des., 4, 1–105. Tan,R.C., Truong,T.N., McCammon,J.A. and Sussman,J.L. (1993) Acetylcholinesterase: electrostatic steering increases the rate of ligand binding. Biochemistry, 32, 401–403. Trosset,J.Y. and Scheraga,H.A. (1998) Reaching the global minimum in docking simulations: a Monte Carlo energy minimization approach using Bezier splines. Proc. Natl Acad. Sci. USA, 95, 8011–8015. Verkhivker,G.M., Rejto,P.A., Bouzida,D., Arthurs,S., Colson,A.B., Freer,S.T., Gehlhaar,D.K., Larson,V., Luty,B.A., Marrone,T. and Rose,P.W. (1999) Towards understanding the mechanisms of molecular recognition by computer simulations of ligand–protein interactions. J. Mol. Recognit., 12, 371–389. Wade,R.C., Gabdoulline,R.R., Ludemann,S.K. and Lounnas,V. (1998) Electrostatic steering and ionic tethering in enzymeligand binding: insights from simulations. Proc. Natl Acad. Sci. USA, 95, 5942–5949. Wlodek,S.T., Shen,T. and McCammon,J.A. (2000) Electrostatic steering of substrate to acetylcholinesterase: analysis of field fluctuations. Biopolymers, 53, 265–271.