International
Tables for
Crystallography
Volume F
Crystallography of biological macromolecules
Edited by E. Arnold, D. M. Himmel and M. G. Rossmann

International Tables for Crystallography (2012). Vol. F, ch. 18.10, pp. 534-538
https://doi.org/10.1107/97809553602060000864

Chapter 18.10. PrimeX and the Schrödinger computational chemistry suite of programs

J. A. Bell,a* Y. Cao,a J. R. Gunn,a T. Day,a E. Gallicchio,b Z. Zhou,a R. Levyb and R. Farida

aSchrödinger, 120 West 45th Street, 17th Floor, New York, NY 10036, USA, and bBioMaPS Institute for Quantitative Biology, Department of Chemistry and Chemical Biology, Rutgers, The State University of New Jersey, Piscataway, NJ 08854, USA
Correspondence e-mail:  jeff.bell@schrodinger.com

PrimeX is a new X-ray crystal structure refinement program for biological macromolecules from Schrödinger, Inc. It produces an all-atom model at moderate resolution that is in excellent agreement with the diffraction data and that is also suitable for immediate use in computational chemistry applications. The program features maximum-likelihood reciprocal-space minimization, simulated-annealing refinement, ligand placement, loop building and side-chain placement. PrimeX is integrated with the powerful molecular-graphics program Maestro, which provides an easy-to-use graphical interface. Command-line access to PrimeX tools provides for their scripting into complex workflows. Additional information about these features and the methods used therein are presented.

18.10.1. Introduction

| top | pdf |

PrimeX is a new crystal structure refinement program for large biological molecules. Among the tools included in the package are reciprocal- and real-space minimization, automated con­struction of polypeptides into electron density, side-chain placement, ligand placement and simulated-annealing refinement. PrimeX is unique among X-ray refinement packages for large biological molecules with respect to its origins: the popular docking program, Glide, and the protein structure prediction and refinement package, Prime.

18.10.1.1. Computational environment

| top | pdf |

PrimeX is fully integrated with Maestro, a general-purpose molecular graphics front end. It shares a consistently designed interface with a suite of other powerful programs in molecular modelling and computational chemistry. Some of these other programs and their functions are listed in Table 18.10.1.1[link].

Table 18.10.1.1| top | pdf |
Schrödinger software packages related to PrimeX

PackageBrief description
CombiGlide Lead compound design and optimization
Desmond High-performance molecular dynamics simulations
Glide Ligand–receptor docking and scoring
Induced-Fit Docking Modelling conformational changes induced by ligand binding
Liason Ligand–receptor binding free-energy prediction
MacroModel Full-featured program for molecular modelling
Prime Protein structure prediction and refinement
QSite Quantum mechanics/molecular mechanics calculations on biological systems
SiteMap Binding-site identification and analysis

18.10.1.2. Mission

| top | pdf |

The first goal of PrimeX is to produce protein models consistent with both the X-ray diffraction data and the needs of computational chemists. Each of the software packages listed in Table 18.10.1.1[link] depends, to some degree, on accurate protein structures as input for computations. However, the models of biological molecules produced by crystallographers typically cause problems for computational work in two ways. First, these models lack coordinates for hydrogen atoms. Because the experi­mental data rarely give any direct indication of the correct position of hydrogen atoms, crystallographers are reluctant to include them in their models. Secondly, these models frequently contain coordinates that place non-bonded atoms implausibly close together, leading to the calculation of extremely high internal energy in these molecules. Crystallographers historically have not restrained these interactions as strongly as they restrain other elements of molecular geometry such as bond lengths, bond angles and group planarity. A reason for placing atoms in these conflicting positions is to fit the atoms at the centre of the observed electron-density peak, even if, because of atomic motion, that peak extends over a large volume of space that intersects with the volume occupied by other atoms. The result is a model that may be representative of the time-averaged structure but which cannot be consistent with the structure at a given instance. PrimeX produces a model that is consistent with both timescale views.

The second goal of PrimeX is to minimize unnecessary experimenter interaction for repetitive tasks. At its current stage of development, PrimeX provides a comprehensive set of features for automation of the refinement of the X-ray crystal structures of large biological molecules. Future development will combine these tools in automated multi-step processes.

18.10.1.3. Implementation path

| top | pdf |

In a field where most of the available software has been developed from academic sources, PrimeX is the only refinement program for biological crystallography that has been developed in a commercial environment. The advantage of commercial development includes high standards for consistency, professional software support and specific attention to the concerns of crystallographers in industry. All code is developed internally and intellectual property rights are monitored carefully. In an industrial setting, this assurance may be important to avoid future intellectual property entanglements.

18.10.2. Computational environment

| top | pdf |

Several aspects of the software packages listed in Table 18.10.1.1[link] are especially relevant to the application of PrimeX.

18.10.2.1. Molecular graphics

| top | pdf |

Maestro is the unified interface for all Schrödinger software. It has impressive rendering capabilities, a powerful selection of analysis tools and a set of fundamental tools for manipulating molecules. Hardware stereo is supported. Alternate conformations are displayed and can be manipulated in Maestro. Automation of complex tasks and development of custom applications can be accomplished through integrated Python scripting in Maestro, as well as through an internal macro facility. Function keys are also customizable. The results of each operation in a series of steps are organized in a project table (Fig. 18.10.2.1[link]). This feature of Maestro is well suited to documenting the multiple steps typical of crystal structure refinement.

[Figure 18.10.2.1]

Figure 18.10.2.1 | top | pdf |

The project table entries track the steps of this refinement in progress, using the model from PDB (Protein Data Bank; Berman et al., 2000[link]) entry 1mdq to solve the structure of entry 1mpd. The most important statistics for monitoring the course of the refinement are always easily available.

18.10.2.2. Supporting infrastructure

| top | pdf |

All Schrödinger programs are accessed through a clear, well organized graphical user interface (GUI). Nearly all computations are run through Schrödinger's job control facility. This facility provides a uniform mechanism for launching, monitoring and controlling calculations. Jobs can be submitted from any computer to any computer, cluster or grid, or to a batch queue, and the results returned to the computer from which the job was submitted. In addition, many of Schrödinger's tools have been integrated as KNIME extensions. KNIME (http://www.knime.org ) is an open-source pipelining program which provides a graphical interface for building workflows. PrimeX components will also be available for KNIME workflows in the future.

18.10.2.3. Molecular mechanics

| top | pdf |

OPLS (Jorgensen et al., 1996[link]; Kaminski et al., 2001[link]) is a general-purpose force field for modelling proteins, nucleic acids and small molecules. This accurate description of atomic level interactions provides consistency across all Schrödinger molecular modelling pack­ages, such as the core modelling functions of the MacroModel package or the molecular dynamics simulation tools of Desmond. PrimeX applies this consistent molecular description to the refinement of large biological crystal structures.

18.10.3. Achieving the mission of PrimeX

| top | pdf |

The main goal of any X-ray crystal structure refinement program is to produce an optimized fit to the X-ray diffraction data, consistent with known molecular stereochemistry. For PrimeX, the additional goal is to include hydrogen atoms in the model in a way that correctly reflects molecular stereochemistry, especially for non-bonded inter­actions.

18.10.3.1. Molecular models and computational chemistry

| top | pdf |

18.10.3.1.1. Hydrogen atoms

| top | pdf |

Although crystallographers have generally not included coordinates for hydrogen atoms in their models at moderate resolution, this situation has not always been so. During the early years of development of the very popular XPLOR package for refine­ment (Brünger, 1992[link]), which was later augmented to become the even more popular refinement package CNS (Brünger et al., 1998[link]), the geometric restraints were based on the CHARMm force field. The presence of all polar hydrogen atoms, i.e. hydrogen atoms attached to non-carbon atoms, was required by this force field. Quite a number of structures from that era remain in the Protein Data Bank and include coordinates for polar hydrogen atoms. This CHARMm-based restraint system was later modified to a unified-atom model that is used by CNS today.

Use of the all-atom force field OPLS in PrimeX requires the explicit definition of all hydrogen-atom positions for geometric calculations. Hydrogen atoms are added to the model automatically in a manner that satisfies local hydrogen-bonding opportunities with adjacent atoms, if they are not already present from previous cycles of refinement. An additional hydrogen-bond optimization tool in PrimeX (Fig. 18.10.3.1[link]) analyses donors and acceptors to define clusters of such sites that might be connected through hydrogen bonds, and that form networks. Within each cluster, hydrogen bonding is evaluated using a rule-based Monte Carlo method to find the optimal combinations of the variable components of these systems. The structural features adjusted during hydrogen-bond optimization are:

  • (i) alcoholic hydrogen-atom positions;

    [Figure 18.10.3.1]

    Figure 18.10.3.1 | top | pdf |

    The hydrogen-bond network optimizer clusters hydrogen-bond donors and acceptors to group together those that could possibly interact, so that these interactions can be efficiently optimized. Individual clusters may be examined and optimized by the user as shown, or all of the clusters may be optimized in a single operation.

  • (ii) sulfhydril hydrogen-atom positions;

  • (iii) phenolic hydrogen-atom positions;

  • (iv) charge and tautomeric states of Asp and Glu side chains;

  • (v) charge, tautomeric states and orientation (flip) of His side chains;

  • (vi) orientation (flip) of Asn and Gln side chains; and

  • (vii) positions for hydrogen atoms in water molecules.

18.10.3.1.2. Electrostatics and implicit solvation

| top | pdf |

For hydrogen atoms bonded to more electronegative atoms, electrostatic forces are a major determinant of non-bonded interactions, and they must be evaluated during the calculation of geometric gradients in the refinement. Thus, PrimeX employs the complete molecular mechanics description of atomic interactions embodied in OPLS, including electrostatic terms. A solvation energy term (Zhu, Shirts & Friesner, 2007[link]) can account for the effects of solvation on electrostatic interactions without the use of explicit coordinates for all solvent atoms. This term may be optionally included in the molecular-mechanics description used by PrimeX. Refinement incorporating a complete electrostatic description, including an implicit solvation term, has been shown to lead to lower Rfree values compared with refinement excluding those interactions (Knight et al., 2008[link]). For cases where coordinates for bound water molecules were explicitly included in the refinement, the effects of an implicit solvation term on Rfree were favourable but small (J. A. Bell, unpublished work). A more complete analysis of the effects of the implicit solvation term in PrimeX is in progress.

18.10.3.1.3. van der Waals interactions

| top | pdf |

The full OPLS force field includes terms for repulsive and attractive non-bonded interactions. Most significant for res­traints during crystallographic refinement is the utilization of the full repulsive term for non-bonded interactions, unlike other leading biological crystal structure refinement pack­ages where these terms are weighted relatively lightly.

18.10.3.2. Molecular models consistent with X-ray data

| top | pdf |

All current programs for refinement of biological crystal structures employ geometric restraints that either originate with or are consistent with a survey of small-molecule structures by Engh & Huber (1991[link]). Other geometric restraints for group planarity and non-bonded interactions have been added to each refinement package independently from different sources. PrimeX employs the internally consistent molecular description embodied in the OPLS force field to achieve the same goal of restraining stereochemistry to values typical of small-molecule structures.

During refinement, all crystallographic calculations in PrimeX are made without any contribution from hydrogen atoms. This strategy avoids exacerbation of the problem of overfitting that would occur if hydrogen coordinates were also considered to be adjustable parameters in refinement. Rather, the hydrogen-atom coordinates in the model can be thought of as a means of applying the more accurate molecular description found in the OPLS force field, while the positions of non-hydrogen atoms are influenced by the crystallographic gradients. This distinct hybrid approach for the calculation of the two types of terms during refinement works well, despite the seeming asymmetry of the treatment (J. A. Bell, unpublished work).

18.10.3.3. Automation of refinement tasks

| top | pdf |

As well as executing through the graphical interface, most PrimeX tasks can be run through command-line execution, so these tasks can be easily integrated into workflows. Future devel­opment will be focused on combining multiple tasks and diagnostic information into larger automated tasks that will reduce the amount of routine work required from users during refinement.

18.10.4. PrimeX implementation and theory

| top | pdf |

18.10.4.1. Force-field model and automatic atom typing

| top | pdf |

An important feature of PrimeX is that all molecules are auto­matically treated with atom typing and parameters are generated without the intervention of the crystallographer. This capability saves a large amount of time that might otherwise be spent collecting this information from disparate sources. The use of OPLS in PrimeX thus ensures that all ligand and protein parameters are similarly scaled and are consistent between all parts of the structure.

18.10.4.2. Reciprocal-space calculations

| top | pdf |

18.10.4.2.1. Restrained maximum-likelihood minimization

| top | pdf |

For reciprocal-space refinement with a maximum-likelihood target, the formulation of Pannu & Read (1996[link]) was implemented following the general concepts developed in the publications of Brünger and co-workers (Brünger, 1989[link]; Brünger et al., 1998[link]). A maximum-likelihood target has been shown to improve the convergence of refinement and to reduce the effects of model bias (Murshudov et al., 1997[link]). Optionally, a more traditional least-squares target may be employed. Three different minimizers are provided: conjugate gradient, quasi-Newton and truncated Newton (Zhu, Shirts, Friesner & Jacobson, 2007[link]). Parts of the model related by noncrystallographic symmetry may be restrained to be similar with respect to both coordinates and displacement parameters during refinement. All or parts of the structure can be easily grouped for rigid-body coordinate refinement. Refinement will automatically account for alternate conformations. Alternate conformations may vary in size and may encompass from just one atom to many residues. Restrained isotropic B-factor refinement is also provided. B factors may optionally be grouped by main-chain and side-chain atoms for each residue. Grouped occupancy refinement is also available.

18.10.4.2.2. Simulated annealing

| top | pdf |

Simulated-annealing refinement was implemented within the general-purpose molecular modelling package IMPACT (Banks et al., 2005[link]), employing concepts for simulated-annealing refinement validated in the program CNS (Adams et al., 1997[link]). PrimeX simulated annealing provides two alternative energy models for dynamic simulation refinement. In the complete method, all molecular-mechanics terms are evaluated during the simulation. In the approximate method, the electrostatic and implicit solvation terms are not evaluated, a method similar to that employed in CNS (Adams et al., 1997[link]).

18.10.4.2.3. Map generation

| top | pdf |

Map calculations in PrimeX are based on the sigmaA weighting scheme of Read (1986[link]), a data treatment that has been shown to decrease the bias in electron-density maps. Optionally, an unweighted map may be calculated.

18.10.4.2.4. Omit maps

| top | pdf |

Omit maps (Bhat & Cohen, 1984[link]; Bhat, 1988[link]) have been implemented because of their demonstrated ability to reduce model-derived bias in maps (Vellieux & Dijkstra, 1997[link]). The kicked omit maps described by Guncar et al. (2000[link]) have also been included in PrimeX. After all the atoms to be omitted have been removed from the coordinates for the calculation, the model is modified by adding a small displacement of random size and direction to each coordinate before the phases are calculated. This process is repeated multiple times and the multiple maps are averaged. Either of these methods can be applied in a composite manner where small portions of the structure are omitted systematically over the entire map volume. Finally, simulated-annealing omit maps can be calculated by deleting the atoms to be omitted from the model, followed by simulated-annealing refinement and automatic map generation.

18.10.4.3. Real-space tools

| top | pdf |

18.10.4.3.1. Loop refinement

| top | pdf |

Polypeptide segments, both internal (loops) and external (tails), can be built from the target sequence as stored in the program. The program builds polypeptides according to an algorithm derived from the protein-modelling program Prime (Jacobson et al., 2004[link]; Zhu, Shirts & Friesner, 2007[link]). Using the end of the existing structure as an anchor point, a residue is added and sampled in various conformations. All conformations that demonstrate even a modest fit to the electron density are kept. During the next addition, all stored conformations are used as starting points for sampling the conformation of the next residue. The total number of conformations may reach many thousands as additional residues are added. When all amino acids have been added, the conformations are clustered to reduce the redundancy of the set, treated with real-space minimization, scored by conformation (OPLS energy) and fitted to the electron density. When building loops, the structures are built from both ends of the structure gap. Only those pairs of built polypeptides that meet to form a contiguous loop are clustered, refined and evaluated. Typically, one to three of the highest scoring structures are returned by the program, depending on the quality of the electron density.

18.10.4.3.2. Side-chain placement

| top | pdf |

Also based on Prime technology, one or multiple side chains may be simultaneously evaluated and placed into electron density. Even if two side chains have overlapping conformations at the start of the process, they will both be placed into appropriate electron density. As an alternative, various common rotamers of side chains may be explored interactively in Maestro and the best fit to the electron density selected visually or automatically.

18.10.4.3.3. Minimization

| top | pdf |

Real-space minimization provides an additional method of optimizing model coordinates according to the electron density and force field, as well as playing an integral part in the building of polypeptide chains, the placement of amino-acid side chains and the placement of ligands.

18.10.4.4. Water placement

| top | pdf |

Waters may be added by PrimeX to difference electron-density maps for positive peaks exceeding a selected level of significance. Prospective water sites in the electron-density map are screened by three different criteria related to distance from other atoms. Water molecules that refine to a B factor above a selectable cutoff are deleted.

18.10.4.5. Ligand placement

| top | pdf |

Ligand placement can be accessed through a convenient GUI (Fig. 18.10.4.1[link]). The algorithm for ligand placement was derived from the docking program Glide (Friesner et al., 2004[link]). In brief, the program abstracts the protein model as an energy grid in which the properties of the residues are encoded. Multiple conformations are generated for the ligand, and then a search is made for the best possible position and orientation in the vicinity of the electron density that the user has identified. The many alternative solutions are reduced in number with the application of a dual scoring system: (i) the GlideScore is an evaluation of the chemical complementarity between the ligand and the protein; (ii) the DensScore is a measure of the fit of the ligand to the observed electron density. Both of these scores are used to rank the fitting of each ligand conformation in each orientation. The weighting of the DensScore versus the GlideScore is under user control and it strongly favours the DensScore by default. Surviving candidates are subjected to rigid-body and torsion-angle refinement against an energy function that includes the OPLS energy and the DensScore. The weighting of the DensScore versus the OPLS energy is also under user control, and it strongly favours the DensScore by default. Finally, the top few dozen candidates are treated with real-space refinement and are ranked by real-space R factor. The top five poses for the ligand are returned with the protein structure for evaluation by the user.

[Figure 18.10.4.1]

Figure 18.10.4.1 | top | pdf |

The ligand/solvent placement GUI provides a list of electron-density blobs from a map with coefficients Fo − Fc, together with their volume, and several ways to specify the molecules to be placed into the electron density. Multiple molecules may be included in a single placement operation at a single site. Ligand-placement options include the generation of charge and tautomer states of the ligand with the program LigPrep and the inclusion of symmetry-related molecules in the calculation at the ligand site.

18.10.4.6. Validation

| top | pdf |

PrimeX provides three main validation tools. `Protein reports' lists geometric variations from expected values in table form. This table can be sorted by residue or by the magnitude of the deviation, to highlight parts of the structure most in need of attention. A density fit table (Fig. 18.10.4.2[link]) lists the real-space R factors for each main chain and side chain and for the entire residue. This table can also be sorted by any column. Clicking on any row of either table centres the molecular display on that residue. An interactive Ramachandan plot provides additional analysis of main-chain structure.

[Figure 18.10.4.2]

Figure 18.10.4.2 | top | pdf |

The density fit table provides a convenient method for finding residues where the main chain or side chain in the model poorly fits the electron density. Clicking on the line for Lys118 in chain A centres that residue in the workspace display, allowing the user to bring to bear the appropriate tool to resolve the problem.

18.10.5. Conclusion

| top | pdf |

PrimeX is a combination of well established methodology with innovations focused on producing the most useful model from the refinement of biological molecules. Future directions to be explored will include TLS scaling (Winn et al., 2003[link]), increasing levels of automation, more reliable water placement, and refinement of ensembles of structures as a means of representing the flexibility of the molecule and the uncertainty of the atomic positions (Knight et al., 2008[link]).

References

Adams, P. D., Pannu, N. S., Read, R. J. & Brünger, A. T. (1997). Cross-validated maximum likelihood enhances crystallographic simulated annealing refinement. Proc. Natl Acad. Sci. USA, 94, 5018–5023.
Banks, J. L., Beard, H. S., Cao, Y., Cho, A. E., Damm, W., Farid, R., Felts, A. K., Halgren, T. A., Mainz, D. T., Maple, J. R., Murphy, R., Philipp, D. M., Repasky, M. P., Zhang, L. Y., Berne, B. J., Friesner, R. A., Gallicchio, E. & Levy, R. M. (2005). Integrated Modeling Program, Applied Chemical Theory (IMPACT). J. Comput. Chem. 26, 1752–1780.
Berman, H. M., Westbrook, J., Feng, Z., Gilliland, G., Bhat, T. N., Weissig, H., Shindyalov, I. N. & Bourne, P. E. (2000). The Protein Data Bank. Nucleic Acids Res. 28, 235–242.
Bhat, T. N. (1988). Calculation of an OMIT map. J. Appl. Cryst. 21, 279–281.
Bhat, T. N. & Cohen, G. H. (1984). OMITMAP: an electron density map suitable for the examination of errors in a macromolecular model. J. Appl. Cryst. 17, 244–248.
Brünger, A. T. (1989). A memory-efficient fast Fourier transformation algorithm for crystallographic refinement on supercomputers. Acta Cryst. A45, 42–50.
Brünger, A. T. (1992). X-PLOR. Version 3.1. A system for X-ray crystallography and NMR. Yale University, Connecticut, USA.
Brünger, A. T., Adams, P. D., Clore, G. M., DeLano, W. L., Gros, P., Grosse-Kunstleve, R. W., Jiang, J.-S., Kuszewski, J., Nilges, M., Pannu, N. S., Read, R. J., Rice, L. M., Simonson, T. & Warren, G. L. (1998). Crystallography and NMR system: a new software suite for macromolecular structure determination. Acta Cryst. D54, 905–921.
Engh, R. A. & Huber, R. (1991). Accurate bond and angle parameters for X-ray protein structure refinement. Acta Cryst. A47, 392–400.
Friesner, R. A., Banks, J. L., Murphy, R. B., Halgren, T. A., Klicic, J. J., Mainz, D. T., Repasky, M. P., Knoll, E. H., Shelley, M., Perry, J. K., Shaw, D. E., Francis, P. & Shenkin, P. S. (2004). Glide: a new approach for rapid, accurate docking and scoring. 1. Method and assessment of docking accuracy. J. Med. Chem. 47, 1739–1749.
Guncar, G., Klemencic, I., Turk, B., Turk, V., Karaoglanovic-Carmona, A., Juliano, L. & Turk, D. (2000). Crystal structure of cathepsin X: a flip-flop of the ring of His23 allows carboxy-monopeptidase and carboxy-dipeptidase activity of the protease. Structure, 8, 305–313.
Jacobson, M. P., Pincus, D. L., Rapp, C. S., Day, T. J., Honig, B., Shaw, D. E. & Friesner, R. A. (2004). A hierarchical approach to all-atom protein loop prediction. Proteins, 55, 351–367.
Jorgensen, W. L., Maxwell, D. S. & Tirado-Rives, J. (1996). Development and testing of the OPLS all-atom force field on conformational energetics and properties of organic liquids. J. Am. Chem. Soc. 118, 11225–11236.
Kaminski, G., Friesner, R. A., Tirado-Rives, J. & Jorgensen, W. L. (2001). Evaluation and reparametrization of the OPLS-AA force field for proteins via comparison with accurate quantum chemical calculations on peptides. J. Phys. Chem. B, 105, 6474–6487.
Knight, J. L., Zhou, Z., Gallicchio, E., Himmel, D. M., Friesner, R. A., Arnold, E. & Levy, R. M. (2008). Exploring structural variability in X-ray crystallographic models using protein local optimization by torsion-angle sampling. Acta Cryst. D64, 383–396.
Murshudov, G. N., Vagin, A. A. & Dodson, E. J. (1997). Refinement of macromolecular structures by the maximum-likelihood method. Acta Cryst. D53, 240–255.
Pannu, N. S. & Read, R. J. (1996). Improved structure refinement through maximum likelihood. Acta Cryst. A52, 659–668.
Read, R. J. (1986). Improved Fourier coefficients for maps using phases from partial structures with errors. Acta Cryst. A42, 140–149.
Vellieux, F. M. D. & Dijkstra, B. W. (1997). Computation of Bhat's OMIT maps with different coefficients. J. Appl. Cryst. 30, 396–399.
Winn, M. D., Murshudov, G. N. & Papiz, M. Z. (2003). Macromolecular TLS refinement in REFMAC at moderate resolutions. Methods Enzymol. 374, 300–321.
Zhu, K., Shirts, M. R. & Friesner, R. A. (2007). Improved methods for side chain and loop predictions via the protein local optimization program: variable dielectric model for implicitly improving the treatment of polarization effects. J. Chem. Theory Comput. 3, 2108–2119.
Zhu, K., Shirts, M. R., Friesner, R. A. & Jacobson, M. P. (2007). Multiscale optimization of a truncated Newton minimization algorithm and application to proteins and protein–ligand complexes. J. Chem. Theory Comput. 3, 640–648.








































to end of page
to top of page