Protein–protein docking with multiple residueconformations and residue substitutionsDAVID M. LORBER,1MARIA K. UDO,1,2AND BRIAN K. SHOICHET11Northwestern University, Department of Molecular Pharmacology and Biological Chemistry,Chicago, Illinois 60611, USA2Loyola University Chicago, Department of Physics, Chicago, Illinois 60626, USA(RECEIVED July 18, 2001; FINAL REVISION March 4, 2002; ACCEPTED MARCH 5, 2002)AbstractThe protein docking problem has two major aspects: sampling conformations and orientations, and scoringthem for fit. To investigate the extent to which the protein docking problem may be attributed to thesampling of ligand side-chain conformations, multiple conformations of multiple residues were calculatedfor the uncomplexed (unbound) structures of protein ligands. These ligand conformations were docked intoboth the complexed (bound) and unbound conformations of the cognate receptors, and their energies wereevaluated using an atomistic potential function. The following questions were considered: (1) does theensemble of precalculated ligand conformations contain a structure similar to the bound form of the ligand?(2) Can the large number of conformations that are calculated be efficiently docked into the receptors? (3)Can near-native complexes be distinguished from non-native complexes? Results from seven test systemssuggest that the precalculated ensembles do include side-chain conformations similar to those adopted in theexperimental complexes. By assuming additivity among the side chains, the ensemble can be docked in lessthan 12 h on a desktop computer. These multiconformer dockings produce near-native complexes and alsonon-native complexes. When docked against the bound conformations of the receptors, the near-nativecomplexes of the unbound ligand were always distinguishable from the non-native complexes. Whendocked against the unbound conformations of the receptors, the near-native dockings could usually, but notalways, be distinguished from the non-native complexes. In every case, docking the unbound ligands withflexible side chains led to better energies and a better distinction between near-native and non-native fits.An extension of this algorithm allowed for docking multiple residue substitutions (mutants) in addition tomultiple conformations. The rankings of the docked mutant proteins correlated with experimental bindingaffinities. These results suggest that sampling multiple residue conformations and residue substitutions ofthe unbound ligand contributes to, but does not fully provide, a solution to the protein docking problem.Conformational sampling allows a classical atomistic scoring function to be used; such a function maycontribute to better selectivity between near-native and non-native complexes. Allowing for receptor flex-ibility may further extend these results.Keywords: Protein–protein docking; mutant; combinatorial; flexibilitySupplemental material: See www.proteinscience.org.Interactions between proteins are critical in biology, andhave been widely studied. With the advent of genome andproteome projects, there is much interest in predicting thestructures of protein–protein complexes. This, however,turns out to be difficult, and is often referred to as the“Protein Docking Problem” (Richmond 1984; Connolly1986). This problem has two aspects: enumeration of pos-sible states, and evaluation of their complementarity.Since the early 1990s, docking programs have been ableto regenerate near-native structures of protein–protein com-plexes using the complexed (bound) conformations of thetwo proteins (Cherfils et al. 1991; Shoichet and Kuntz 1991;Reprint requests to: Brian K. Shoichet, Northwestern University, De-partment of Molecular Pharmacology and Biological Chemistry, 303 E.Chicago Ave., Chicago, Illinois 60611, USA; e-mail: [email protected]; fax: (312) 503-5349.Article and publication are at http://www.proteinscience.org/cgi/doi/10.1110/ps.2830102.Protein Science (2002), 11:1393–1408. Published by Cold Spring Harbor Laboratory Press. Copyright © 2002 The Protein Society1393Hart and Read 1992; Vakser 1995). The protein dockingproblem only becomes acute when docking the uncom-plexed (unbound) conformations of the two proteins; theseare the relevant states for true prediction (Totrov and Aba-gyan 1994; Vakser 1995; Jackson et al. 1998; Norel et al.1999; Vakser et al. 1999; Camacho et al. 2000; Kimura et al.2001). Although unbound proteins often adopt main-chainconformations similar to their bound counterparts, their sol-vent-exposed side chains commonly adopt conformationsthat are not complementary to their binding partner (Conteet al. 1999). Thus, when docking algorithms generate near-native complexes from the unbound conformations of thepartners, atoms in the interface clash. Such near-native fitsscore poorly in classical, atomistic energy potentials be-cause of these clashing atoms. Even a single noncomple-mentary atom can lead to very unfavorable energies becauseof the steepness of the steric repulsion term of the van derWaals energy (Weiner et al. 1984).The problem of docking unbound proteins can be ad-dressed either by explicitly sampling many conformationsor by modifying the scoring function to accommodateclashes. Several methods have been published that usemodified scoring functions. Soft docking (Jiang and Kim1991; Palma et al. 2000), Fourier correlation (Gabb et al.1997; Ritchie and Kemp 2000), and shape fitting methods(Shoichet and Kuntz 1991; Norel et al. 1994, 1995, 1999),smooth the details of protein–protein interactions, therebyallowing clashes to occur, at least before minimization. Em-pirically derived, or trained, scoring functions (Weng et al.1996; Moont et al. 1999; Palma et al. 2000) have also beenused to address the protein docking problem. Althoughthese methods allow near-native structures to be identified,they cannot reliably distinguish the near-native complexesfrom non-native complexes when docking unbound pro-teins. Vajda and coworkers have explored applying exten-sive minimization to a series of predocked complexes (Ca-macho et al. 2000). Although this has shown some promisein discriminating near-native complexes from non-nativecomplexes, the discrimination is not always as convincingas one might like, and the high computational expense of theprocedure makes it prohibitive for on the fly docking.An alternative to avoiding the van der Waals violationproblem with a trained or attenuated scoring