ConfGen: A Conformational Search Method for Efficient Generation of Bioactive Conformers K. Shawn Watts, † Pranav Dalal, ‡ Robert B. Murphy, § Woody Sherman, § Rich A. Friesner, | and John C. Shelley* ,† Schro ¨dinger, LLC, 101 SW Main Street, Suite 1300, Portland, Oregon 97204 and 120 West 45th Street, 17th Floor, New York, New York 10036, D. E. Shaw India Software Private Limited, Sanali Infopark, 8-2-120/113, Road No. 2, Banjara Hills, Hyderabad 500 034, Andhra Pradesh, India, and Department of Chemistry, Columbia University, 3000 Broadway, New York, New York 10027 Received January 9, 2010 We describe the methodology, parametrization, and application of a conformational search method, called ConfGen, designed to efficiently generate bioactive conformers. We define efficiency as the ability to generate a bioactive conformation within a small total number of conformations using a reasonable amount of computer time. The method combines physics-based force field calculations with empirically derived heuristics designed to achieve efficient searching and prioritization of the ligand’s conformational space. While many parameter settings are supported, four modes spanning a range of speed and quality trades-offs are defined and characterized. The validation set used to test the method is composed of ligands from 667 crystal structures covering a broad array of target and ligand classes. With the fastest mode, ConfGen uses an average of 0.5 s per ligand and generates only 14.3 conformers per ligand, at least one of which lies within 2.0 Å root-mean-squared deviation of the crystal structure for 96% of the ligands. The most computationally intensive mode raises this recovery rate to 99%, while taking 8 s per ligand. Combining multiple search modes to “fill-in” holes in the conformation space or energy minimizing using an all-atom force field each lead to improvements in the recovery rates at higher resolutions. Overall, ConfGen is at least as good as competing programs at high resolution and demonstrates higher efficiency at resolutions sufficient for many downstream applications, such as pharmacophore modeling. INTRODUCTION The ability to accurately and robustly generate bioactive conformations of small molecules is a key step in both ligand- and structure-based drug design. Many widely used techniques, such as docking, 1-4 pharmacophore searching, 5 and shape-based screening, 6,7 depend heavily on the ability to generate conformers that are close to the structure that the ligand assumes in the protein-ligand complex of interest. In benchmarking studies, closeness is usually measured by calculating the root-mean-squared displacement (rmsd) of the heavy atoms of the ligand relative to a conformation of the ligand from a crystal structure of a protein-ligand complex. Existing conformational search algorithms sample the ligands using a variety of techniques, including random torsional angle changes, 8-10 random coordinate changes, 11 distance geometry, 12,13 and rule-based methods, 14,15 or locate minima using normal modes. 16,17 A number of recent studies have been published that attempt to assess and/or optimize the performance of various tools for predicting bioactive conformations. 18-27 One challenge in conformational searching comes from the fact that bound-state ligand conformations are often not in the global minimum energy state. 22 Ligand conformations in protein-ligand complexes do not exactly correspond to conformations at local potential energy minima for unbound ligands because the binding process induces some degree of strain in the ligand. Therefore, one would not expect that the lowest energy conformation for the unbound state, even with a perfect energy function and complete sampling, to always be a bioactive conformer. However, recent results suggest that bioactive conformers are often closer to energy minima than previously reported. 28 Another important ques- tion to consider when analyzing conformational search methods is whether there can be multiple bioactive confor- mations for a given ligand, and if there is something special about a bioactive conformation relative to other energetically accessible unbound-state conformations. It has been shown that very similar ligands can adopt different conformations upon binding 29 or even that the same ligand can adopt multiple binding modes with different conformational states. 30-33 Clearly, to reliably reproduce bioactive confor- mations for ligands, a conformational search algorithm must produce multiple conformations. Regarding whether there is something special about bioactive conformers relative to other low-energy conforma- tions, thermodynamics tells us that any state can be a bioactive conformer as long as the energy cost associated with adopting that conformation can be compensated by favorable interactions in the receptor-ligand complex. While this is true in theory, general observations as well as previous work have shown that ligands typically bind in relatively * Corresponding author. E-mail: John.Shelley@schrodinger.com. † Schro ¨dinger, LLC, Portland, Oregon. ‡ D. E. Shaw India Software, Andhra Pradesh, India. § Schro ¨dinger, LLC, New York, New York. | Columbia University, New York, New York. J. Chem. Inf. Model. 2010, 50, 534–546 534 10.1021/ci100015j  2010 American Chemical Society Published on Web 04/07/2010