Optimal strategies for virtual screening of induced-fit and flexible target in the 2015 D3R Grand Challenge

Induced fit or protein flexibility can make a given structure less useful for docking and/or scoring. The 2015 Drug Design Data Resource (D3R) Grand Challenge provided a unique opportunity to prospectively test optimal strategies for virtual screening in these type of targets: heat shock protein 90 (HSP90), a protein with multiple ligand-induced binding modes; and mitogen-activated protein kinase kinase kinase kinase 4 (MAP4K4), a kinase with a large flexible pocket. Using previously known co-crystal structures, we tested predictions from methods that keep the receptor structure fixed and used (a) multiple receptor/ligand co-crystals as binding templates for minimization or docking (“close”), (b) methods that align or dock to a single receptor (“cross”), and (c) a hybrid approach that chose from multiple bound ligands as initial templates for minimization to a single receptor (“min-cross”). Pose prediction using our “close” models resulted in average ligand RMSDs of 0.32 and 1.6 Å for HSP90 and MAP4K4, respectively, the most accurate models of the community-wide challenge. On the other hand, affinity ranking using our “cross” methods performed well overall despite the fact that a fixed receptor cannot model ligand-induced structural changes,. In addition, “close” methods that leverage the co-crystals of the different binding modes of HSP90 also predicted the best affinity ranking. Our studies suggest that analysis of changes on the receptor structure upon ligand binding can help select an optimal virtual screening strategy.     

Docking pose selection by interaction pattern graph similarity: application to the D3R grand challenge 2015

High affinity ligands for a given target tend to share key molecular interactions with important anchoring amino acids and therefore often present quite conserved interaction patterns. This simple concept was formalized in a topological knowledge-based scoring function (GRIM) for selecting the most appropriate docking poses from previously X-rayed interaction patterns. GRIM first converts protein–ligand atomic coordinates (docking poses) into a simple 3D graph describing the corresponding interaction pattern. In a second step, proposed graphs are compared to that found from template structures in the Protein Data Bank. Last, all docking poses are rescored according to an empirical score (GRIMscore) accounting for overlap of maximum common subgraphs. Taking the opportunity of the public D3R Grand Challenge 2015, GRIM was used to rescore docking poses for 36 ligands (6 HSP90α inhibitors, 30 MAP4K4 inhibitors) prior to the release of the corresponding protein–ligand X-ray structures. When applied to the HSP90α dataset, for which many protein–ligand X-ray structures are already available, GRIM provided very high quality solutions (mean rmsd = 1.06 Å, n = 6) as top-ranked poses, and significantly outperformed a state-of-the-art scoring function. In the case of MAP4K4 inhibitors, for which preexisting 3D knowledge is scarce and chemical diversity is much larger, the accuracy of GRIM poses decays (mean rmsd = 3.18 Å, n = 30) although GRIM still outperforms an energy-based scoring function. GRIM rescoring appears to be quite robust with comparison to the other approaches competing for the same challenge (42 submissions for the HSP90 dataset, 27 for the MAP4K4 dataset) as it ranked 3rd and 2nd respectively, for the two investigated datasets. The rescoring method is quite simple to implement, independent on a docking engine, and applicable to any target for which at least one holo X-ray structure is available.     

Molecular docking performance evaluated on the D3R Grand Challenge 2015 drug-like ligand datasets

The D3R Grand Challenge 2015 was focused on two protein targets: Heat Shock Protein 90 (HSP90) and Mitogen-Activated Protein Kinase Kinase Kinase Kinase 4 (MAP4K4). We used a protocol involving a preliminary analysis of the available data in PDB and PubChem BioAssay, and then a docking/scoring step using more computationally demanding parameters that were required to provide more reliable predictions. We could evidence that different docking software and scoring functions can behave differently on individual ligand datasets, and that the flexibility of specific binding site residues is a crucial element to provide good predictions.     

D3R grand challenge 2015: Evaluation of protein–ligand pose and affinity predictions

The Drug Design Data Resource (D3R) ran Grand Challenge 2015 between September 2015 and February 2016. Two targets served as the framework to test community docking and scoring methods: (1) HSP90, donated by AbbVie and the Community Structure Activity Resource (CSAR), and (2) MAP4K4, donated by Genentech. The challenges for both target datasets were conducted in two stages, with the first stage testing pose predictions and the capacity to rank compounds by affinity with minimal structural data; and the second stage testing methods for ranking compounds with knowledge of at least a subset of the ligand–protein poses. An additional sub-challenge provided small groups of chemically similar HSP90 compounds amenable to alchemical calculations of relative binding free energy. Unlike previous blinded Challenges, we did not provide cognate receptors or receptors prepared with hydrogens and likewise did not require a specified crystal structure to be used for pose or affinity prediction in Stage 1. Given the freedom to select from over 200 crystal structures of HSP90 in the PDB, participants employed workflows that tested not only core docking and scoring technologies, but also methods for addressing water-mediated ligand–protein interactions, binding pocket flexibility, and the optimal selection of protein structures for use in docking calculations. Nearly 40 participating groups submitted over 350 prediction sets for Grand Challenge 2015. This overview describes the datasets and the organization of the challenge components, summarizes the results across all submitted predictions, and considers broad conclusions that may be drawn from this collaborative community endeavor.     

Lessons learned from participating in D3R 2016 Grand Challenge 2: compounds targeting the farnesoid X receptor

D3R 2016 Grand Challenge 2 focused on predictions of binding modes and affinities for 102 compounds against the farnesoid X receptor (FXR). In this challenge, two distinct methods, a docking-based method and a template-based method, were employed by our team for the binding mode prediction. For the new template-based method, 3D ligand similarities were calculated for each query compound against the ligands in the co-crystal structures of FXR available in Protein Data Bank. The binding mode was predicted based on the co-crystal protein structure containing the ligand with the best ligand similarity score against the query compound. For the FXR dataset, the template-based method achieved a better performance than the docking-based method on the binding mode prediction. For the binding affinity prediction, an in-house knowledge-based scoring function ITScore2 and MM/PBSA approach were employed. Good performance was achieved for MM/PBSA, whereas the performance of ITScore2 was sensitive to ligand composition, e.g. the percentage of carbon atoms in the compounds. The sensitivity to ligand composition could be a clue for the further improvement of our knowledge-based scoring function.     

An iterative knowledge-based scoring function to predict protein-ligand interactions: I. Derivation of interaction potentials

Using a novel iterative method, we have developed a knowledge-based scoring function (ITScore) to predict protein-ligand interactions. The pair potentials for ITScore were derived from a training set of 786 protein-ligand complex structures in the Protein Data Bank. Twenty-six atom types were used based on the atom type category of the SYBYL software. The iterative method circumvents the long-standing reference state problem in the derivation of knowledge-based scoring functions. The basic idea is to improve pair potentials by iteration until they correctly discriminate experimentally determined binding modes from decoy ligand poses for the ligand-protein complexes in the training set. The iterative method is efficient and normally converges within 20 iterative steps. The scoring function based on the derived potentials was tested on a diverse set of 140 protein-ligand complexes for affinity prediction, yielding a high correlation coefficient of 0.74. Because ITScore uses SYBYL-defined atom types, this scoring function is easy to use for molecular files prepared by SYBYL or converted by software such as BABEL.     

An iterative knowledge-based scoring function to predict protein-ligand interactions: II. Validation of the scoring function

We have developed an iterative knowledge-based scoring function (ITScore) to describe protein-ligand interactions. Here, we assess ITScore through extensive tests on native structure identification, binding affinity prediction, and virtual database screening. Specifically, ITScore was first applied to a test set of 100 protein-ligand complexes constructed by Wang et al. (J Med Chem 2003, 46, 2287), and compared with 14 other scoring functions. The results show that ITScore yielded a high success rate of 82% on identifying native-like binding modes under the criterion of rmsd < or = 2 A for each top-ranked ligand conformation. The success rate increased to 98% if the top five conformations were considered for each ligand. In the case of binding affinity prediction, ITScore also obtained a good correlation for this test set (R = 0.65). Next, ITScore was used to predict binding affinities of a second diverse test set of 77 protein-ligand complexes prepared by Muegge and Martin (J Med Chem 1999, 42, 791), and compared with four other widely used knowledge-based scoring functions. ITScore yielded a high correlation of R2 = 0.65 (or R = 0.81) in the affinity prediction. Finally, enrichment tests were performed with ITScore against four target proteins using the compound databases constructed by Jacobsson et al. (J Med Chem 2003, 46, 5781). The results were compared with those of eight other scoring functions. ITScore yielded high enrichments in all four database screening tests. ITScore can be easily combined with the existing docking programs for the use of structure-based drug design.     

Improved pose and affinity predictions using different protocols tailored on the basis of data availability

The D3R 2015 grand drug design challenge provided a set of blinded challenges for evaluating the applicability of our protocols for pose and affinity prediction. In the present study, we report the application of two different strategies for the two D3R protein targets HSP90 and MAP4K4. HSP90 is a well-studied target system with numerous co-crystal structures and SAR data. Furthermore the D3R HSP90 test compounds showed high structural similarity to existing HSP90 inhibitors in BindingDB. Thus, we adopted an integrated docking and scoring approach involving a combination of both pharmacophoric and heavy atom similarity alignments, local minimization and quantitative structure activity relationships modeling, resulting in the reasonable prediction of pose [with the root mean square deviation (RMSD) values of 1.75 Å for mean pose 1, 1.417 Å for the mean best pose and 1.85 Å for the mean all poses] and affinity (ROC AUC = 0.702 at 7.5 pIC50 cut-off and R = 0.45 for 180 compounds). The second protein, MAP4K4, represents a novel system with limited SAR and co-crystal structure data and little structural similarity of the D3R MAP4K4 test compounds to known MAP4K4 ligands. For this system, we implemented an exhaustive pose and affinity prediction protocol involving docking and scoring using the PLANTS software which considers side chain flexibility together with protein–ligand fingerprints analysis assisting in pose prioritization. This protocol through fares poorly in pose prediction (with the RMSD values of 4.346 Å for mean pose 1, 4.69 Å for mean best pose and 4.75 Å for mean all poses) and produced reasonable affinity prediction (AUC = 0.728 at 7.5 pIC50 cut-off and R = 0.67 for 18 compounds, ranked 1st among 80 submissions).     

Scoring and lessons learned with the CSAR benchmark using an improved iterative knowledge-based scoring function

Based on a statistical mechanics-based iterative method, we have extracted a set of distance-dependent, all-atom pairwise potentials for protein-ligand interactions from the crystal structures of 1300 protein-ligand complexes. The iterative method circumvents the long-standing reference state problem in knowledge-based scoring functions. The resulted scoring function, referred to as ITScore 2.0, has been tested with the CSAR (Community Structure-Activity Resource, 2009 release) benchmark of 345 diverse protein-ligand complexes. ITScore 2.0 achieved a Pearson correlation of R(2) = 0.54 in binding affinity prediction. A comparative analysis has been done on the scoring performances of ITScore 2.0, the van der Waals (VDW) scoring function, the VDW with heavy atoms only, and the force field (FF) scoring function of DOCK which consists of a VDW term and an electrostatic term. The results reveal several important factors that affect the scoring performances, which could be helpful for the improvement of scoring functions.     

Significant enhancement of docking sensitivity using implicit ligand sampling

The efficient and accurate quantification of protein-ligand interactions using computational methods is still a challenging task. Two factors strongly contribute to the failure of docking methods to predict free energies of binding accurately: the insufficient incorporation of protein flexibility coupled to ligand binding and the neglected dynamics of the protein-ligand complex in current scoring schemes. We have developed a new methodology, named the 'ligand-model' concept, to sample protein conformations that are relevant for binding structurally diverse sets of ligands. In the ligand-model concept, molecular-dynamics (MD) simulations are performed with a virtual ligand, represented by a collection of functional groups that binds to the protein and dynamically changes its shape and properties during the simulation. The ligand model essentially represents a large ensemble of different chemical species binding to the same target protein. Representative protein structures were obtained from the MD simulation, and docking was performed into this ensemble of protein conformation. Similar binding poses were clustered, and the averaged score was utilized to rerank the poses. We demonstrate that the ligand-model approach yields significant improvements in predicting native-like binding poses and quantifying binding affinities compared to static docking and ensemble docking simulations into protein structures generated from an apo MD simulation.     

Construction and test of ligand decoy sets using MDock: community structure-activity resource benchmarks for binding mode prediction

Two sets of ligand binding decoys have been constructed for the community structure-activity resource (CSAR) benchmark by using the MDock and DOCK programs for rigid- and flexible-ligand docking, respectively. The decoys generated for each complex in the benchmark thoroughly cover the binding site and also contain a certain number of near-native binding modes. A few scoring functions have been evaluated using the ligand binding decoy sets for their abilities of predicting near-native binding modes. Among them, ITScore achieved a success rate of 86.7% for the rigid-ligand decoys and 79.7% for the flexible-ligand decoys, under the common definition of a successful prediction as root-mean-square deviation <2.0 ? from the native structure if the top-scored binding mode was considered. The decoy sets may serve as benchmarks for binding mode prediction of a scoring function, which are available at the CSAR Web site ( http://www.csardock.org/).     

Large scale free energy calculations for blind predictions of protein–ligand binding: the D3R Grand Challenge 2015

We describe binding free energy calculations in the D3R Grand Challenge 2015 for blind prediction of the binding affinities of 180 ligands to Hsp90. The present D3R challenge was built around experimental datasets involving Heat shock protein (Hsp) 90, an ATP-dependent molecular chaperone which is an important anticancer drug target. The Hsp90 ATP binding site is known to be a challenging target for accurate calculations of ligand binding affinities because of the ligand-dependent conformational changes in the binding site, the presence of ordered waters and the broad chemical diversity of ligands that can bind at this site. Our primary focus here is to distinguish binders from nonbinders. Large scale absolute binding free energy calculations that cover over 3000 protein–ligand complexes were performed using the BEDAM method starting from docked structures generated by Glide docking. Although the ligand dataset in this study resembles an intermediate to late stage lead optimization project while the BEDAM method is mainly developed for early stage virtual screening of hit molecules, the BEDAM binding free energy scoring has resulted in a moderate enrichment of ligand screening against this challenging drug target. Results show that, using a statistical mechanics based free energy method like BEDAM starting from docked poses offers better enrichment than classical docking scoring functions and rescoring methods like Prime MM-GBSA for the Hsp90 data set in this blind challenge. Importantly, among the three methods tested here, only the mean value of the BEDAM binding free energy scores is able to separate the large group of binders from the small group of nonbinders with a gap of 2.4 kcal/mol. None of the three methods that we have tested provided accurate ranking of the affinities of the 147 active compounds. We discuss the possible sources of errors in the binding free energy calculations. The study suggests that BEDAM can be used strategically to discriminate binders from nonbinders in virtual screening and to more accurately predict the ligand binding modes prior to the more computationally expensive FEP calculations of binding affinity.     

Predicting binding poses and affinities for protein - ligand complexes in the 2015 D3R Grand Challenge using a physical model with a statistical parameter estimation

The 2015 D3R Grand Challenge provided an opportunity to test our new model for the binding free energy of small molecules, as well as to assess our protocol to predict binding poses for protein-ligand complexes. Our pose predictions were ranked 3–9 for the HSP90 dataset, depending on the assessment metric. For the MAP4K dataset the ranks are very dispersed and equal to 2–35, depending on the assessment metric, which does not provide any insight into the accuracy of the method. The main success of our pose prediction protocol was the re-scoring stage using the recently developed Convex-PL potential. We make a thorough analysis of our docking predictions made with AutoDock Vina and discuss the effect of the choice of rigid receptor templates, the number of flexible residues in the binding pocket, the binding pocket size, and the benefits of re-scoring. However, the main challenge was to predict experimentally determined binding affinities for two blind test sets. Our affinity prediction model consisted of two terms, a pairwise-additive enthalpy, and a non pairwise-additive entropy. We trained the free parameters of the model with a regularized regression using affinity and structural data from the PDBBind database. Our model performed very well on the training set, however, failed on the two test sets. We explain the drawback and pitfalls of our model, in particular in terms of relative coverage of the test set by the training set and missed dynamical properties from crystal structures, and discuss different routes to improve it.     

A nonredundant structure dataset for benchmarking protein‐RNA computational docking

Protein–RNA interactions play an important role in many biological processes. The ability to predict the molecular structures of protein–RNA complexes from docking would be valuable for understanding the underlying chemical mechanisms. We have developed a novel nonredundant benchmark dataset for protein–RNA docking and scoring. The diverse dataset of 72 targets consists of 52 unbound–unbound test complexes, and 20 unbound–bound test complexes. Here, unbound–unbound complexes refer to cases in which both binding partners of the cocrystallized complex are either in apo form or in a conformation taken from a different protein–RNA complex, whereas unbound–bound complexes are cases in which only one of the two binding partners has another experimentally determined conformation. The dataset is classified into three categories according to the interface root mean square deviation and the percentage of native contacts in the unbound structures: 49 easy, 16 medium, and 7 difficult targets. The bound and unbound cases of the benchmark dataset are expected to benefit the development and improvement of docking and scoring algorithms for the docking community. All the easy‐to‐view structures are freely available to the public at http://zoulab.dalton.missouri.edu/RNAbenchmark/ . © 2012 Wiley Periodicals, Inc.     

Ligand-biased ensemble receptor docking (LigBEnD): a hybrid ligand/receptor structure-based approach

Ligand docking to flexible protein molecules can be efficiently carried out through ensemble docking to multiple protein conformations, either from experimental X-ray structures or from in silico simulations. The success of ensemble docking often requires the careful selection of complementary protein conformations, through docking and scoring of known co-crystallized ligands. False positives, in which a ligand in a wrong pose achieves a better docking score than that of native pose, arise as additional protein conformations are added. In the current study, we developed a new ligand-biased ensemble receptor docking method and composite scoring function which combine the use of ligand-based atomic property field (APF) method with receptor structure-based docking. This method helps us to correctly dock 30 out of 36 ligands presented by the D3R docking challenge. For the six mis-docked ligands, the cognate receptor structures prove to be too different from the 40 available experimental Pocketome conformations used for docking and could be identified only by receptor sampling beyond experimentally explored conformational subspace.     

Multiple ligand simultaneous docking: Orchestrated dancing of ligands in binding sites of protein

Present docking methodologies simulate only one single ligand at a time during docking process. In reality, the molecular recognition process always involves multiple molecular species. Typical protein–ligand interactions are, for example, substrate and cofactor in catalytic cycle; metal ion coordination together with ligand(s); and ligand binding with water molecules. To simulate the real molecular binding processes, we propose a novel multiple ligand simultaneous docking (MLSD) strategy, which can deal with all the above processes, vastly improving docking sampling and binding free energy scoring. The work also compares two search strategies: Lamarckian genetic algorithm and particle swarm optimization, which have respective advantages depending on the specific systems. The methodology proves robust through systematic testing against several diverse model systems: E. coli purine nucleoside phosphorylase (PNP) complex with two substrates, SHP2NSH2 complex with two peptides and Bcl‐xL complex with ABT‐737 fragments. In all cases, the final correct docking poses and relative binding free energies were obtained. In PNP case, the simulations also capture the binding intermediates and reveal the binding dynamics during the recognition processes, which are consistent with the proposed enzymatic mechanism. In the other two cases, conventional single‐ligand docking fails due to energetic and dynamic coupling among ligands, whereas MLSD results in the correct binding modes. These three cases also represent potential applications in the areas of exploring enzymatic mechanism, interpreting noisy X‐ray crystallographic maps, and aiding fragment‐based drug design, respectively. © 2010 Wiley Periodicals, Inc. J Comput Chem, 2010     

Robust optimization of scoring functions for a target class

Target-specific optimization of scoring functions for protein–ligand docking is an effective method for significantly improving the discrimination of active and inactive molecules in virtual screening applications. Its applicability, however, is limited due to the narrow focus on, e.g., single protein structures. Using an ensemble of protein kinase structures, the publically available directory of useful decoys ligand dataset, and a novel multi-factorial optimization procedure, it is shown here that scoring functions can be tuned to multiple targets of a target class simultaneously. This leads to an improved robustness of the resulting scoring function parameters. Extensive validation experiments clearly demonstrate that (1) virtual screening performance for kinases improves significantly; (2) variations in database content affect this kind of machine-learning strategy to a lesser extent than binary QSAR models, and (3) the reweighting of interaction types is of particular importance for improved screening performance. Electronic supplementary material  The online version of this article (doi:) contains supplementary material, which is available to authorized users.     

Statistical potential for modeling and ranking of protein-ligand interactions

Applications in structural biology and medicinal chemistry require protein-ligand scoring functions for two distinct tasks: (i) ranking different poses of a small molecule in a protein binding site and (ii) ranking different small molecules by their complementarity to a protein site. Using probability theory, we developed two atomic distance-dependent statistical scoring functions: PoseScore was optimized for recognizing native binding geometries of ligands from other poses and RankScore was optimized for distinguishing ligands from nonbinding molecules. Both scores are based on a set of 8,885 crystallographic structures of protein-ligand complexes but differ in the values of three key parameters. Factors influencing the accuracy of scoring were investigated, including the maximal atomic distance and non-native ligand geometries used for scoring, as well as the use of protein models instead of crystallographic structures for training and testing the scoring function. For the test set of 19 targets, RankScore improved the ligand enrichment (logAUC) and early enrichment (EF(1)) scores computed by DOCK 3.6 for 13 and 14 targets, respectively. In addition, RankScore performed better at rescoring than each of seven other scoring functions tested. Accepting both the crystal structure and decoy geometries with all-atom root-mean-square errors of up to 2 ? from the crystal structure as correct binding poses, PoseScore gave the best score to a correct binding pose among 100 decoys for 88% of all cases in a benchmark set containing 100 protein-ligand complexes. PoseScore accuracy is comparable to that of DrugScore(CSD) and ITScore/SE and superior to 12 other tested scoring functions. Therefore, RankScore can facilitate ligand discovery, by ranking complexes of the target with different small molecules; PoseScore can be used for protein-ligand complex structure prediction, by ranking different conformations of a given protein-ligand pair. The statistical potentials are available through the Integrative Modeling Platform (IMP) software package (http://salilab.org/imp) and the LigScore Web server (http://salilab.org/ligscore/).     

The ensemble performance index: an improved measure for assessing ensemble pose prediction performance

We present a theoretical study on the performance of ensemble docking methodologies considering multiple protein structures. We perform a theoretical analysis of pose prediction experiments which is completely unbiased, as we make no assumptions about specific scoring functions, search paradigms, protein structures, or ligand data sets. We introduce a novel interpretable measure, the ensemble performance index (EPI), for the assessment of scoring performance in ensemble docking, which will be applied to simulated and real data sets.     

VoteDock: Consensus docking method for prediction of protein–ligand interactions

Molecular recognition plays a fundamental role in all biological processes, and that is why great efforts have been made to understand and predict protein–ligand interactions. Finding a molecule that can potentially bind to a target protein is particularly essential in drug discovery and still remains an expensive and time‐consuming task. In silico, tools are frequently used to screen molecular libraries to identify new lead compounds, and if protein structure is known, various protein–ligand docking programs can be used. The aim of docking procedure is to predict correct poses of ligand in the binding site of the protein as well as to score them according to the strength of interaction in a reasonable time frame. The purpose of our studies was to present the novel consensus approach to predict both protein–ligand complex structure and its corresponding binding affinity. Our method used as the input the results from seven docking programs (Surflex, LigandFit, Glide, GOLD, FlexX, eHiTS, and AutoDock) that are widely used for docking of ligands. We evaluated it on the extensive benchmark dataset of 1300 protein–ligands pairs from refined PDBbind database for which the structural and affinity data was available. We compared independently its ability of proper scoring and posing to the previously proposed methods. In most cases, our method is able to dock properly approximately 20% of pairs more than docking methods on average, and over 10% of pairs more than the best single program. The RMSD value of the predicted complex conformation versus its native one is reduced by a factor of 0.5 Å. Finally, we were able to increase the Pearson correlation of the predicted binding affinity in comparison with the experimental value up to 0.5. © 2010 Wiley Periodicals, Inc. J Comput Chem 32: 568–581, 2011