Pharmaceutical firms are demanding better tools for computational drug design, and presentations during last week's Structure-Based Drug Design conference indicate that software developers in the academic and private sectors are eager to deliver.
As BioInform reported last week [BioInform 06-16-06], pharmaceutical researchers are seeing positive results from virtual screening, molecular modeling, and other computational methods, but are outspoken in describing the current shortcomings of the field. In particular, pharma is seeking improved methods for sampling the conformational space of ligands, better ways of accounting for protein and ligand flexibility, and more accurate scoring functions.
In response, a number of software vendors and academic developers discussed ways in which they are addressing those issues. One such effort is a collaboration between Schrödinger and William Jorgensen's computational chemistry team at Yale University and involves an improved version of the OPLS (Optimized Potential for Liquid Simulations) force field for molecular modeling.
Woody Sherman, application scientist at Schrödinger, said that the current version of the force field, OPLS 2005, includes 1,780 torsion types â€" many more than some other popular force fields, such as MMFF (Merck Molecular Force Field), which includes 918 torsion types. However, Sherman noted, in a study of commercially available compounds, the company found that there are approximately 7,000 additional torsion types that could be added that would greatly improve the performance of the force field.
Sherman said that Schrödinger has purchased a compute cluster in order to run "quantum mechanical torsional scans" on these compounds as it works on developing the next version of the force field, OPLS 2007.
Jorgensen, the original developer of OPLS, said that Schrödinger's work in expanding torsion types is "very important" for the field. His lab, meanwhile, is developing software for designing inhibitors for a given target binding site by "growing" molecules. The program, called BOMB (biochemical and organic model builder), begins with a "core" fragment in the binding site, and then builds a combinatorial library of compounds by adding substituents from a list of more than 600 drug fragments.
Jorgensen said that there are 100 possible cores, and up to four positions where substituents can be added, resulting in a virtual library of up to 10 trillion possible molecules.
Hans Briem, senior scientist in the compound design and computational chemistry group at Schering, noted that these large combinatorial libraries require extremely fast docking algorithms, but said that drugmakers also require methods that consider pharmacophore restraints while docking. This is a "difficult problem," he said.
Briem said that Schering has been working with BioSolveIT to combine the software company's FlexXc program for docking combinatorial libraries with its FlexX-Pharm program for docking with pharmacophore constraints.
The new module, called FlexXc-Pharm, has shown promising results so far, Briem said. In a study involving a combinatorial library of 22.4 million compounds built upon the core scaffold for Gleevec, FlexXc-Pharm docked around 70,000 in 3 hours on a 20-node cluster. The closest Gleevec analog was ranked No. 124, placing it in the top 0.001 percent of all the compounds, Briem said.
A BioSolveIT representative at the meeting said that the FlexXc-Pharm module will be available in FlexX 2.1, which is due out in the late fall.
Another company that is addressing the challenge of pharmacophore modeling is Tripos. Bob Clark, senior director of software research at the company, discussed the module, called GALAHAD (Genetic Algorithm with Linear Assignment for Hypermolecule Alignment of Datasets), which Tripos added to its Sybyl molecular modeling environment last year [BioInform 08-15-06].
Clark said that GALAHAD also helps address the challenge of binding site flexibility, which varies greatly between proteins. The method first builds a set of low-energy conformations, which are run through a "multi-objective" genetic algorithm that looks for conformers with minimal energy, maximum steric overlap, and maximum pharmacophore concordance, Clark said.
While the genetic algorithm did a good job of finding the best conformers, Clark said it fell short when it came to alignment, so Tripos has incorporated a version of the LAMDA (Linear Assignment Method for Database Alignment) algorithm developed at the University of Sheffield into the method.
â€" Bernadette Toner ([email protected])