|Home | About | Journals | Submit | Contact Us | Français|
Empirical scoring functions used in protein-ligand docking calculations are typically trained on a dataset of complexes with known affinities with the aim of generalizing across different docking applications. We report a novel method of scoring-function optimization that supports the use of additional information to constrain scoring function parameters, which can be used to focus a scoring function’s training towards a particular application, such as screening enrichment. The approach combines multiple instance learning, positive data in the form of ligands of protein binding sites of known and unknown affinity and binding geometry, and negative (decoy) data of ligands thought not to bind particular protein binding sites or known not to bind in particular geometries. Performance of the method for the Surflex-Dock scoring function is shown in cross-validation studies and in 8 blind test cases. Tuned functions optimized with a sufficient amount of data exhibited either improved or undiminished screening performance relative to the original function across all eight complexes. Analysis of the changes to the scoring function suggest that modifications can be learned that are related to protein-specific features such as active-site mobility.
The utility of molecular docking to drug discovery is well established, and has been highlighted in a number of recent reviews, benchmarking studies, and comparative evaluations [1; 2; 3]. There are a multitude of approaches, but they share the same underlying strategy: the marriage of a search strategy to a scoring function with the goal of identifying the optimal conformation and alignment (pose) of a ligand bound to a site within a protein of known structure. The earliest work in the area, pioneered by Blaney and Kuntz, used a physics-based formulation of scoring (essentially the non-bonded terms of a molecular mechanics force-field) and a method for rigid-body placement of small molecules . Later approaches introduced flexible search and empirically derived scoring functions [5; 6; 7; 8; 9; 10]. Surflex-Dock is a descendent of one of these earlier dockers, called Hammerhead [11; 12].
Our most recent work regarding Surflex’s scoring function focused on the idea of using negative training data to provide a sensible basis for optimizing the repulsive parameters of an empirical scoring function . These data took the form of computationally generated putative decoy ligands, which were produced by docking a decoy library to each of the protein structures from the complexes used in the original parameter estimation . By making use of such data, it was possible to estimate the value of repulsive terms such as protein-ligand interpenetration instead of relying on an ad hoc value. The difficulty with an approach relying only upon positive data (protein-ligand complexes of known affinity) is that the inductive bias of the most parsimonious estimation regime is to assume that if an example of an interaction does not exist (e.g. a ligand atom penetrating a protein atom) then nothing can be concluded, which leads to a value of zero on the associated term. This is in contrast with PMF-type approaches, where the normalization procedures lead to an inductive bias wherein the absence of an observation is indicative of low probability, which results in a preference against unobserved interactions [14; 15]. Our approach is related to one reported by Smith et al. , who used “noise” molecules in refining scoring functions for DOCK. However, the genesis of the work reported here and that which preceded it was our prior work that established the concept of multiple-instance learning in the area of 3D QSAR using both active and inactive ligands [17; 18].
When docking methods are evaluated, there are three criteria applied. First, docking accuracy measures the probability that a ligand will be docked in a pose that matches the experimental determination. Second, screening utility measures the ability of a docker to rank a list of known ligands of a protein above a set of decoys. Third, scoring accuracy measures the ability to rank a list of active ligands in order of binding affinity. In most work with scoring function development, the actual data for parameter estimation relates to scoring accuracy [7; 9; 19; 20]. Parameters are sought to minimize the difference between computed and experimental affinities for ligands with known bound geometries. In our recent report, we showed that it was possible to make use of negative data which related to screening utility. The approach sought parameters for a scoring function that would simultaneously minimize computed/experimental affinity differences and minimize the excursion of computed decoy affinities beyond a fixed threshold .
In this paper, we generalize this concept so that information from each of the three areas of docking application may be used to influence the refinement of Surflex’s scoring function. Data of the following form may be used to refine the scoring function:
The first three types of data bear on scoring, the fourth bears on screening utility, and the last bears on geometric docking accuracy. The optimization procedure implements a weighted objective function for parameter optimization based on simultaneous consideration of all types of data. In such an optimization problem, the issue of which pose of a ligand to consider becomes important. As with our previous work [9; 17; 21], we explicitly address this problem by making explicit choices of pose as the scoring function evolves. For example, given a protein/ligand complex with known affinity, it is appropriate to make use of the experimental ligand pose as the initial pose in parameterizing the scoring function. However, while the experimental pose may be a very good static approximation of the true interaction between the ligand and protein, small variations in the ligand position (within the accuracy of the crystallographic experiment) may yield different scores. Consider a computed pKd of 7.0 at the precise crystallographic pose of a ligand whose known pKd is 8.0. If a very close pose yields a maximum for the function of 8.0, one should use the 8.0 score, which entails no error for the scoring function. This issue is discussed in detail in our earliest work on scoring functions for docking , which was based on earlier work in 3D QSAR . The approach has been formalized within the machine-learning community as multiple-instance learning , and it has a substantial impact on the performance of systems where hidden variables (here the precise pose of a ligand) are present.
In what follows, we demonstrate that this generalized multiple-constraint optimization procedure is able to improve the screening performance of Surflex-Dock in a protein-target specific manner. Given that operational use of docking programs typically involves a user with large amounts of non-public data relating to the particular target under study, we expect that the ability to specifically tune docking parameters based on such data will lead to substantial practical benefits in all three areas of docking performance.
The optimization procedure has been implemented as a standalone Surflex program (Surflex-Dock-Optimize, version 1.0). The scoring function parameter files can be used by the released version of Surflex-Dock which has been updated to allow for loading parameter files (version 2.11-lp). A future release of the Surflex-Dock software will incorporate the optimization feature directly. The software that implements the algorithms described here is available free of charge to academic researchers for non-commercial use (contact the corresponding author for details on obtaining the software). Molecular data sets presented herein are also available.
The optimization procedure described herein is general enough for use with any parameterized scoring function. For the purposes of this paper, results are reported for the scoring function used in Surflex-Dock. A relatively brief review of this scoring function and its parameters will be given as other work offers a more detailed account [3; 9]. This will be followed by a description of the data used for training and testing optimized scoring functions. The last section will describe the optimization procedure itself. All training and testing data sets used in this study have been taken from published docking benchmarks that are freely available. They may be obtained by contacting the corresponding author of this paper.
The scoring function employed by Surflex-Dock was originally trained on 34 protein-ligand complexes representing a variety of functional classes whose dissociation constants ranged from 10−3 to 10−14. This function was optimized to predict the experimental binding affinities of each complex, resulting in an effective means for modeling the non-covalent interactions between small organic molecules and proteins. The function is continuous and piece-wise differentiable with respect to pose. Listed in order of import, the terms of the scoring function are hydrophobic complementarity, polar complementarity, and entropy. Parameters are listed in Table 1. The following four equations define the scoring function:
The hydrophobic and polar terms (Eqs. 1 and 2) dominate the scoring function. These terms operate on the pair-wise van der Waals surface distance r between atoms, coupled with information such as element type, formal charge, and atom status as a hydrogen bond donor or acceptor. The distance dependence of the hydrophobic and polar interactions are composed of a Gaussian, sigmoid, and quadratic penetration term. The polar term is further scaled by directionality and formal charge. The directionality term between atoms I and J is computed based on three vectors (normalized to unit length): the vector from between I and J (bij in Eq. 2), the preferred direction of interaction of I (vi), and the preferred direction of interaction of J (vj). If multiple directional preferences are present (as for a carbonyl moiety), the preference that yields the maximal polar interaction is used. Additional details can be found in the original paper describing the scoring function . Figure 1 plots the relative hydrophobic and polar scores for an ideal contact. Due to the large number of hydrophobic contacts typically seen between a protein and a ligand, on average the hydrophobic term tends to dictate scoring despite a smaller peak value per ideal contact. An ideal hydrogen bond for the scoring function exists, for example, when the center of the O in C=O is 1.97Å away from the center of the H in an N-H and the four atoms are co-linear. This results in a contribution of 1.25 pKd units to the interaction score.
The polar repulsion term (Eq. 3) measures the penalty for placing atoms of similar polarity in close proximity and is scaled by direction. The remaining entropic term (Eq. 4) captures the degrees of rotational and translational freedom lost to the ligand upon binding. This ligand-centric penalty scales linearly in the number of rotatable bonds and linearly with the log of its molecular weight.
As described in the Introduction, our previous work refined the original scoring function, determining weights for penalty terms that govern steric interpenetration and noncomplementary polar contacts . This new function (Surflex-Dock v1.31 and all succeeding versions) was shown to be an improvement over the original and is the default scoring function used by the program. The set of parameters that define the default function are the starting point for further optimization in this work.
The protocol used for generating training data, performing scoring function optimization, and testing protein-specific scoring function optimization was implemented as a standalone Surflex module (Surflex-Dock-Optimize v1.0). Test set validation performance in this work was calculated using the standalone optimization suite. Scoring function parameter files generated by this process can be loaded into the latest released Surflex-Dock program  (v2.11-lp, which has an added –lparam command-line switch). The results presented here are statistically indistinguishable from those generated by employing the derived parameters from optimization to dock the test ligands using Surflex-Dock v2.11-lp with the –lparam option.
This work employed several publicly available molecular datasets for training the scoring function. The original 34 complex set used to train the original scoring function was used to provide an “anchor” for the scoring function during further optimization . Each of the 34 complexes was annotated by an experimentally derived Kd. This set served as a control during optimization to ensure that the parameters of the tuned function do not stray exceedingly far from the values of the default function.
Screening enrichment data was gathered from two primary sources: the PDBBind database  and the DUD database . The former was the basis for our previous work  that is used again here (called the Pham Benchmark). The Pham benchmark consisted of 27 protein structures with 256 known active ligands from the PDBBind database along with two decoy databases. Those targets from the Pham benchmark that overlapped with the DUD database  were chosen as case studies for training our scoring function. Table 2 lists the targets along with information about the number and composition of active ligands: acetylcholinesterase (AChE), estrogen receptor (ER), coagulation factor Xa (FXa), HIV-1 protease (HIVPR), poly(ADP-ribose) polymerase (PARP), thrombin, trypsin, and thymidine kinase (TK). The protein structures for these eight targets served as the docking targets for both the training and test sets. The 107 cognate ligands of these eight targets from the Pham set became the training data from which the scoring function would learn to distinguish active from non-active. Active ligands may be referred to as positive examples in what follows.
Two decoy libraries taken from the Pham benchmark were used as the decoy training background in optimizing for screening enrichment. Two sources were used to test the potential of training bias towards a particular set of decoys. One library was derived from the work of Bissantz et al.  which contained 990 randomly selected nonreactive organic molecules with 0 to 41 rotatable bonds from the Available Chemicals Directory (ACD). This benchmark (hereon referred to as the Rognan set) was culled to a more drug-like set of 861 molecules with a maximum of 15 rotatable bonds . The ZINC database (version 07.26.2004)  was the source for the second decoy set. This database was compiled from the catalogs of numerous small molecule vendors and represents a collection of purchasable compounds suitable for virtual screening. A random subselection of 1000 molecules was taken from the drug-like subset (1,847,466 total) to generate the ZINC1 decoy benchmark . We will refer to compounds from a background library interchangeably as either decoys or negative ligands. Recent work by Irwin and Shoichet  considered multiple decoys sets and compared their physical properties as well as the degree of challenge they posed in screening. The DUD set itself, which was designed with knowledge of the specific active ligands for the targets under construction, was the most challenging in their experiments. Among the “agnostic” decoy sets (constructed with no specific knowledge of the targets under consideration), the ZINC1 set was the most challenging, and the Rognan set was the least. Consequently, in what follows, we focus most of our attention on the ZINC1 results.
Data was uniformly prepared by an automated procedure. All protein structures were converted from PDB to Sybyl mol2 format and protonated at physiological pH. Active site rotamers such as hydroxyls and thiols, as well as imidazole tautomers, were sampled and selected for interaction with the co-crystallized cognate ligand. Ligands were minimized using a DREIDING-like force-field as implemented within Surflex [25; 12]. The active site models (called protomols) necessary for docking with Surflex were generated using the crystallized ligand (surflex-dock – proto_bloat 1.0 proto xtal-lig.mol2 protein.mol2 p1). Initial ligand poses used as input to the scoring function refinement algorithm were generated using Surflex-Dock-Optimize, which yields equivalent poses to Surflex-Dock version 2.11 with default screening parameters (surflex-dock –pscreen dock_list mol-archive.mol2 p1-protomol.mol2 protein.mol2 log).
To cleanly assess the performance of our tuned function, we conducted screening enrichment experiments on positive and decoy ligands that were never encountered in training. As described above, each of the 8 test targets were shared between the Pham and DUD benchmarks. We made use of the Pham actives for training, which contained fewer examples of known ligands than present in the DUD set. We used the DUD actives that did not include any from the Pham set as ligands to test scoring functions that had been optimized with knowledge of the Pham actives. A fair test required a new decoy background. Another 1,000 unique molecules were randomly selected from the drug-like subset of ZINC, this time from version 2007. The process of generating the new decoy set made use of 2D molecular similarity to eliminate the overlap between the testing and training decoy libraries. The test decoy set will be referred to as ZINC2.
This work introduces a constraint based optimization scheme that allows the use of several different sources of data in customizing a scoring function. We will begin by defining the available constraints and how they might be utilized to create scoring functions optimized for a particular task. We will then cover the optimization protocol in detail, along with the options that govern its use.
During any parameter optimization regime, the goal is to extremize the value of an objective function as we explore the parameter space. Our objective function is described by user-defined constraints on training data. Constraints come in three flavors: scoring, screening, and geometric. Together these constraints combine to form the objective function.
Score constraints relate a particular protein and a single ligand or set of ligands to a target score. The user can specify whether the predicted score should be exactly/above/below the target score. Moving in an undesired direction from the target score incurs a squared penalty (see Table 3). This is, in fact, the original training regime where the scoring function was tuned to fit experimental binding affinities . In the current formulation, we would create 34 individual score constraints of equal weight, one for each of the 34 protein-ligand complexes, indicating success as an exact match to the experimental Kd. Using additional such constraints, a user could potentially tune the performance of a scoring function for more accurate rank-order prediction of novel ligands. By focusing, for example, on training data that was dominated by the lead series of interest, better predictions of potency for new ligands in the series could result.
Screening constraints allow a user to denote that one set of positive ligands (e.g. a set of cognate ligands) should score measurably higher than a set of negative ligands (e.g. a set of decoys). Performance is assessed by ROC AUC. A function that could flawlessly determine whether a ligand is positive or negative would have an AUC of 1.0. Conversely, a classifier which randomly assigned ligands a positive or negative label would achieve an AUC of 0.5 in the average case. The impact of a screening constraint on the objective function is formulated as the square of its ROC area’s deviation from 1.0 (see Table 3), scaled by 100 to ensure that its value shares the same effective range as the other constraint types. Using such data, a user can tune a scoring function to perform well in finding new leads for a particular protein of interest in a screening experiment. This particular scenario will be presented in detail in the results that follow, owing to the existence of a large publicly available database for testing.
Geometric constraints offer a method for addressing what are termed “hard failures” in docking. Given an incorrect prediction of a ligand’s pose, it may stem from either a failure of the search method (the best pose was not found, but it would have scored best, termed a soft failure). Or, it may stem from a problem in the scoring function: the best-scoring pose may actually score higher than the correct one (a hard failure). A geometric constraint enforces the rule that no incorrect pose may score higher than the best correct pose. Any deviation results in a squared penalty (see Table 3). In focused medicinal chemistry efforts that are guided in part by docking, the geometric predictions can be very important. By providing a method to learn from hard docking failures, a user can take advantage of structures where docking predictions were wrong to improve future performance.
Constraints can be organized further into weighted groups. This feature allows one to arbitrate the influence of certain constraints over the objective function. Consider the following scenario: one has 34 protein-ligand complexes whose scores the function should predict exactly (34 score constraints). One also has a set of known actives and inactives for a given protein, necessitating the need for a single screening constraint. It is important to explicitly be able to control the relative importance of these two types of constraints in modifying the scoring function. To ensure that a single constraint is not overwhelmed by the presence of numerous competing constraints, we can place the 34 score constraints in one group and the single screening constraint in a second group. The optimization procedure is implemented such that each constraint group has an equal bearing on the objective function. In this example, the objective function essentially will see first the 34 individual scoring constraints and the single screening constraint as having equal relative importance. Users may additionally specify a weight be given to a group, providing more control of influence of different data on the objective function.
Figure 2 depicts a high level view of the optimization procedure. Our method can be organized concisely into three components: Input → Optimize → Output. The input consists of constraint information and an initial set of parameters from which the optimization will begin. The constraint information is simply a set of proteins and ligands coupled with metadata informing the objective function as to how it should interpret its training data. The initial values used in all experiments were the default Surflex-Dock parameters reported previously [9; 13].
Each epoch of optimization proceeds as follows:
The individual steps are described in more detail below.
We use the scoring function with the current set of parameters to score each ligand pose. As discussed in the Introduction, one complication that arises from the optimization exercise is that as the scoring function changes, so too does the optimal pose which extremizes the value of the function. Initially, we begin with poses provided as input by the user with the underlying assumption that the provided pose is also the highest scoring pose. However, as parameters change, the original pose may no longer lie at the extremum of the scoring function. The solution is to interleave local pose optimization along with parameter optimization. Pose optimization occurs on a schedule during the overall procedure when a certain number of successful function parameter modification steps have been taken. Following a local gradient-based optimization of the current ligand pose, the new pose is added to a “pose cache” for that ligand. Each time a ligand is scored, all cached poses are scored with the highest score returned as the representative score for this ligand. The results reported here used a pose cache that stored five of the most recent high scoring poses.
Note that the most general approach would require re-docking of ligands whose true pose was unknown. However, due to computational complexity concerns, this was not implemented. The effect may be approximated by interleaving re-docking between separate invocations of the optimization procedure.
The objective function is defined as the mean squared error (MSE) over all constraints n:
Refer to Table 3 for the error forms of each constraint type. Since the best possible MSE is zero, the procedure seeks to minimize MSE. A good step in the course of optimization is defined as one in which the current epoch MSE is lower than the previous epoch.
All three stopping conditions (maximum number of epochs, MSE goal, and maximum number of epochs with no MSE improvement) are user definable options. In this work, we used values 100,000, 0.0001, and 200, respectively.
The most important output is the newly optimized parameter set, which is a text file containing scoring function parameter values (e.g. “new.param”). These can be used immediately by Surflex-Dock to perform a task of interest (scoring function parameters are loaded with -lparam new.param as an argument to Surflex-Dock v2.11 or later).
This scheme interleaves two ways of sampling the parameter space: random walking and line optimization. A random walk is used to ensure broad parameter space exploration and to overcome local minima. Line optimization yields precisely optimized local minima from any given starting point. Each search method is used for a number of iterations, then the search method is switched. Of course, many more complex search strategies exist. However, this procedure yielded robust results and required little time for optimization. On a typical example requiring both scoring and screening constraints, the parameter optimization process took under an hour on typical desktop hardware.
In order for the tests on the eight protein targets described to be fair and appropriately “blind” we needed to determine the preferred optimization regime using other data. The goal is to combine protein-specific screening constraints with the scoring constraints that gave rise to the original Surflex-Dock scoring function. The critical issue has to do with the relative weighting of the two types of constraints. To understand how these constraints interact, we selected two DUD proteins not in our prediction set: P38 MAP kinase (P38) and dihydrofolate reductase (DHFR). These were chosen because of their large number of known actives (256 and 201 actives, respectively). We performed 10-fold cross-validation using several group weight combinations. For each training fold iteration, actives were randomly partitioned 30%–70% into training and testing sets. So, while the 34 score constraints provided an anchor for the current scoring function, the protein-specific screening constraint provided pressure to learn to score the active training molecules above the ZINC1 decoy set. The optimizer, which is stochastic, was run beginning with default scoring function parameters three times. The best scoring-function parameter set by MSE for each fold was chosen to run a screening enrichment test on the remaining 70% of active compounds against the ZINC1 background. We tested multiple constraint group weight combinations, and we computed the mean ROC AUC over the ten cross-validation folds for each weight combination (Table 4 summarizes the results). Note that in testing a particular scoring function, a full docking was carried out.
When using default parameters, the scoring function is just better than random on P38 with a mean ROC AUC = 0.549 over the ten folds screened against ZINC1. The scoring combination that gives the screening data zero weight gives nearly the same results, as it should (see row 6 in Table 4). Utilizing this weight mixture is equivalent to freeing all scoring function parameters as we re-optimize using only the binding affinities for the 34 complex set. The original parameters appear to be stable under re-optimization, despite employing a more exhaustive search procedure in the present work. Note, the same effect was seen in the DHFR case.
Conversely, zeroing the scoring constraint weight (row 7 of Table 4), improved ROC area to almost 0.70 for P38 (from 0.549) and to 0.941 for DHFR (from 0.750). However, by ignoring the scoring constraint, occasionally pathological behavior resulted in terms of the magnitude of the scores computed using the optimized scoring functions. Since the internal docking search strategy makes use of some thresholds on scores, it was important to retain a similar scale. As we increased the relative weight of the screening constraint, we observed both the improvements in screening performance under cross-validation while maintaining a sensible scale where the scores could be interpreted as pKd. Note that increasing weights on the scoring constraints yielded the expected regression toward the use of only the scoring constraints.
Given the evidence from the cross validation study on P38 and DHFR, we chose to test the optimization scheme on the blind data for the eight targets with two constraint groups: a scoring constraint group defined by the scoring constraints within the 34 complex set; and a screening constraint group comprised of that complex’s training actives and a set of decoys. The scoring and screening constraint groups were assigned weights of 1 and 5, respectively.
The primary test of the scoring function optimization method is in a screening enrichment assessment against eight different protein targets (see Table 2). We have been careful to avoid any contamination of the test by either the active ligands used for scoring function tuning or by the decoys used. The test data for each of the eight targets includes novel active ligands and employs a different set of decoy molecules (the ZINC2 set). We also uniformly applied the procedure that was developed in our preliminary work (which included cross-validation on two other targets). The overall numerical results are presented in Table 5, with plots of the relevant ROC curves presented in Figure 3 and Figure 4. As has become standard practice, we have characterized screening performance in terms of ROC AUC, and we have also computed 95% confidence intervals to bracket the performance of the tuned function in each of the eight test cases. The results are broken into three groups, based on the performance changes.
In the six cases where 10 or more active ligands were available in the Pham set, we observed increased or unchanged performance in all cases, with significant improvements in two cases. These two cases (PARP and HIVPR) will be discussed in detail here.
Poly-(ADP-ribose)-polymerase is involved in the response to genomic damage that results in strand breaks. For specific proteins, PARP can add up to 200 residues of ADP-ribose to form branched polymers, which act as binding sites for repair proteins that play a central role in DNA metabolism . The majority of inhibitors used to tune the scoring function for PARP were small and had relatively weak binding, typically in the micromolar range (see Figure 5 for example structures). The first ROC plot of Figure 3 corresponds to the test of the PARP-focused tuned scoring function on the blind test data. The improvement in screening enrichment for the blind test molecules in this case was pronounced, with an improvement in ROC AUC of 0.10, corresponding to an increase in true-positive rate from approximately 20% to 90% at a false positive rate of less than 5%.
HIV-1 protease is an aspartic protease with a large, solvent-accessible active site with several charged polar moieties both interior and proximally exterior to the pocket. Crystallo-graphic studies have shown that interaction with the interior catalytic triad Asp25-Thr26-Gly27 as well as surface residues, Asp29 and Asp30, is important for enzyme inhibition [27; 28]. The majority of inhibitors used in training bind in the nanomolar range (example structures are shown in Figure 5). The second plot of Figure 3 shows the ROC curves for HIVPR. The tuned function shows a substantial increase in true positive rates at a false positive rate of 5% relative to the default function from roughly 60% to roughly 85%, corresponding to enhanced early enrichment.
The ROC plots are sensitive to the relative separation of active from decoy ligands. Increases in the scores for active ligands, decreases for decoys, or a combination of both can lead to improvements in recovery of active ligands and increase in ROC AUC. The cumulative distributions of positive and negative scores for the default and tuned functions (Figure 6) reveal the underlying impetus for enrichment improvement.
In the case of PARP, the increase in ROC AUC from 0.89 to 0.99 for the tuned function stemmed from a decrease in the scores of the decoys relative to the untuned function with a simultaneous increase in the scores of the active ligands. The bulk of the actives, when docked with the tuned scoring function, had scores approximately 1 log unit higher than when docked using the default scoring function (this corresponds to the rightward shift from the solid red curve to the solid green curve in the top plot of Figure 6). Conversely, the inactives exhibited decreases of roughly 0.5 log units. In the case of HIVPR, performance increased from a ROC AUC of 0.913 to 0.964. However, in this case, the distribution of decoy scores changed only slightly and did so in the wrong direction. The improvement in enrichment came from a significant upward shift of the lowest scoring active ligands by about 1.0 log units. With the default function, 40% of actives had pKd < 7.5, but only 20% of actives scored by the tuned functions had pKd < 7.5.
The underlying reasons for the performance increases observed with PARP and HIVPR stemmed from different sources. In the former case, we observed increased ability to recognize actives and reject decoys. In the latter case, both sets of scores increased, but with a specific advantage to the actives. Inspection of the individual terms of the scoring function before and after the optimization procedure (Figure 7) lends insight into the reasons for these differences. Three plots are given for each case, showing the default and tuned functions for the hydrophobic, polar, and polar mismatch terms. The axes are the same as for Figure 1, with the Y axis being the interaction score in pKd, and with the X axis being the inter-atomic surface distance in Angstroms. Negative distances indicate nominal interpenetration of van der Waals radii; note that radii for polar atoms are not scaled, so ideal polar contacts exhibit numerical interpenetration.
The hydrophobic terms show markedly different modifications in response to tuning for PARP and HIVPR. In the former case (top left plot), the penalty for atomic surface interpenetration is decreased somewhat, and the area of positive hydrophobic interaction (from the Gaussian in Eq. 1) is both more narrow and has lower amplitude. In the latter case, the softening of the overlap penalty is more significant, and the area of positive interaction increases. The tuned scoring function parameters are given in Table 6. The decrease in sensitivity to inter-atomic clashes is reflected in the value of the hrd parameter, which changed from −0.95 (default function) to −0.16 (tuned function). For HIVPR, we also considered the effect of generating the training poses (and testing the resulting tuned function) without the use of Surflex’s ligand pre-minimization and post-docking all-atom optimization. These procedures are part of the default screening protocol of Surflex, since they help decrease dependence on input ligand preparation and allow access to Cartesian movements that can ameliorate clashes between the protein and ligand . This can be especially important for large ligands. The blue curve in Figure 7 shows that the tuned clash penalty is even softer when the docking process is restricted to a ligand’s internal coordinates. In order to obtain reasonably high scores for the large HIVPR ligands, it is necessary to relax the clashing penalty, and this effect is larger when the ligands are unable to bend outside of torsional and alignment space. Differences in docking protocol can yield marked differences in the resulting tuned functions, so particular attention must be given to replicating the protocol used for generating training data as will be used for operational application of the resulting tuned function.
The differences in clashing penalty between the PARP and HIVPR cases can be seen in the polar terms (middle plots of Figure 7), since the hrd parameter also controls excessive interpenetration between polar atoms. Apart from that, the positive aspect of polar interactions exhibited similar behavior in the two tuned function, with both increases in the maximal value of a single polar contact (controlled by the poz parameter) and a slight increase in the distance from which complementary polar contacts obtain positive scores (controlled by the pom parameter, and corresponding to a change from 1.97 Å to 2.09 Å in inter-atomic center distances for N-H and C=O).
In contrast to the decrease in the repulsive effect of inter-atomic clashes, we see a marked increase in the repulsive effect of proximal same-charge moieties for both PARP and HIVPR. The rightmost plots of Figure 7 show increases both in the overall magnitude of same-charge repulsion penalty (controlled by the pr2 parameter) as well as an increase in the distance at which the effect becomes important (controlled by the prm parameter). In the case of HIVPR, the magnitude of the same-charge repulsion penalty increased 75% over the default function, and for PARP, it increased approximately 45%.
The changes in the tuned scoring functions are evident in the behavior of specific test ligands. Figure 8 (top panel) shows the experimentally determined pose of a cyclic urea HIVPR inhibitor bound to the protease (PDB code: 1BVE). Note the position of the hydroxyl groups of the central 7-member ring relative to the catalytic aspartic acids ASP-A25 and ASP-B25 in dark blue. The test ligand (ZINC03833842) has a very similar structure and is shown in its docked pose using the tuned scoring function in the middle panel. Despite a poor ring geometry that was present in the input structure (ring search within Surflex was not employed), the docked pose with the tuned scoring function is reasonable, with sensible interactions between the hydroxyls on the central ring system to the aspartic acid residues as well as good placement of the “arms” of the ligand. The bottom panel shows the same ligand docked using the default scoring function. In this case, the inhibitor was clearly docked poorly.
This ligand was ranked 44th out of 1038 molecules (38 actives + 1000 ZINC2 decoys) by the default scoring function. However, when re-docked using the tuned function, it was ranked 1st. The pose resulting from application of the tuned function is very different, owing to the differences in the penalty terms. If we rescore the tuned function pose using the default function, the steric clashing term alone generates more than 10 pKd units in additional penalty. The large difference in penalty terms between the two functions leads to widely different poses among the active ligands when using the different functions. Among all of the active test ligands, the typical deviation in top-scoring pose between the application of the two functions was quite high (mean rmsd of 5.5Å), reflecting both the flexibility of the ligands as well as the substantial change in the scoring function, especially the parameters that controlled steric clashing. With HIV protease, it is known that ligand binding causes substantial conformational changes to the enzyme . Treatment of the protein structure as rigid has obvious computational benefits in terms of search complexity, but in cases where this treatment is especially inaccurate (e.g. with large ligands), lowering the steric penalty serves as a surrogate for modeling induced fit. The scoring function optimization scheme provides a systematic method to exploit such protein-specific features.
In the case of PARP, the active site is much smaller, and it appears to undergo a smaller degree of movement on binding inhibitors. This is evidenced by the stronger penalty for steric clashes as compared with HIVPR, and it also shows in the degree to which docking with the tuned scoring function yields different top scoring poses compared with docking with the default function. For PARP, 17/31 test ligands dock within 0.5Å rmsd between tuned and default functions, with only 4 ligands above 1.0Å rmsd, including 2 above 2.0Å rmsd. Figure 9 shows the ligand with the largest geometric deviation between docking with the two different scoring functions. The left panel shows the pose generated using the tuned function, and the middle panel shows the pose from the default function. While the pose is not grossly different, as in the HIVPR case shown above, the pose from the tuned function is clearly closer to correct, making the appropriate contacts common to PARP inhibitors. Note that this case was the exception. Most of the top-scoring poses changed very little, but the tuned function yielded systematically higher scores for the actives. In the rightmost panel, a relatively high-ranking decoy is shown as docked using the default scoring function (it ranked 138/1031 molecules). The tuned function, when used to rescore the poses produced by docking using the default function ranked the decoy at 736/1031. In the full docking that gave rise to the ROC performance shown in Figure 3, this decoy ranked 961/1031. So, while the changes in the scoring function had relatively subtle effects on the active ligands, the effects on the decoys were more substantial.
Optimization yielded small, but not statistically significant improvements in three of four cases (ER, Thrombin, and TK), and produced an insignificant decrease in performance in the other case (Trypsin). Perusal of the training results revealed that there was little information to be extracted from the input data. The default function yielded a mean ROC AUC in the training data of 0.95 (minimum 0.93, maximum 1.0). Following optimization, the average training performance was 0.98 (minimum 0.97, maximum 1.0). While the training procedure yielded the desired effect on the training data, given that there was very little room for improvement, the net result was that little improvement was seen in the test data. Lacking a significant number of examples that are poorly ranked, there should be no expectation of a significant change after training. However, the fact that the function parameters are stable in this situation is a useful characteristic. This is aided by inclusion of the original set of 34 complexes as part of the weighted training regime.
In the two cases (FXa and AChE) where just six molecules were available as active ligands from which to tune the scoring function, overall performance on the test libraries was reduced in a statistically significant fashion (see Table 5). To test whether the lack of active ligand examples was the source of the reduction in performance, we added 20 randomly selected actives from the test sets for FXa and AChE to their training sets. After retuning the scoring functions using the same procedure as before, the addition of active training ligands was shown to reverse the degradation. Figure 4 shows the ROC plots corresponding to these additional experiments. In the case of FXa, the tuned function yielded a significant increase, though just marginally in a statistical sense. In the case of AChE, while performance was improved (instead of substantially decreased), the screening utility was still low.
Training on the very small number of actives pushed the optimization protocol toward specific parameter changes for a skewed population of actives. Increasing the number of training examples avoided this skew, but still did not address the problem of weak screening performance. AChE contains a long, narrow binding pocket formed by the aromatic rings of 14 conserved residues . Two active sites are known to exist: a main site located at the bottom of the aromatic gorge, and a secondary site 14Å away near the opening of the binding cavity . This target represents a difficult case in that inhibitors may occupy just the main site or interact with both sites. The active site used in our study (pdb code: 1E66) was taken from the structure of AChE in complex with huprine X , a small molecule with only 40 atoms that binds the primary active site at the bottom of the long pocket. We repeated the optimization under identical conditions with the extended training set, but we used a different structure for AChE (pdb code: 1H23). In this structure, the bound ligand was much larger, huperzine A, which occupies both the primary and secondary sites . We then executed the screening enrichment test making use of both structures, keeping the highest score from either run as the representative ligand score. The results from this experiment were encouraging. Under this treatment, the ROC AUC of the tuned function improved to 0.753, which was significantly better than the default function performed using only a single structure. In this case, scoring function tuning alone was not sufficient to overcome serious limitations imposed by the structure that was used for docking.
Our results show very little effect of changing the decoy set used in training. Using either the Rognan decoy set or the ZINC1 set yielded nearly identical performance (see Table 5). This makes sense, since the effect of a decoy set in the optimization exercise is based upon the small proportion of difficult cases that show up as nominal false positives when using the default scoring function. As long as a decoy set contains some reasonable candidates to be such false positives, it will serve adequately. Note, however, that there are limits to this. A decoy set containing only a large collection of different Fullerenes probably would be of no utility in refining scoring functions for the proteins under study. With respect to the effect of different decoy sets on testing the performance of docking systems, experience is somewhat mixed. While our results  agree with those of Irwin and Shoichet  that the ZINC1 set (called the “Jain set” in ) is more challenging than the Rognan set, the difference we observed was much smaller in magnitude .
In this work, we have chosen to continue to use decoy sets that have been constructed with no specific knowledge of active ligand structures. We have done so for three reasons. First, it provides a direct comparison to our previous studies, which employed the same (or similarly constructed) decoy sets as well as overlapping protein structures [13; 12; 11]. Second, while the statistical likelihood of finding true ligands among a random collection of screening compounds is known to be low (1/1000 to 1/10,000), it is not at all clear what the likelihood might be if one selects a set of decoys that have similar size, charge, and hydrophobicity characteristics, though it is almost certainly higher. Third, even decoy sets that have been shown to have relatively non-drug-like properties are sufficient to distinguish the performance of many docking protocols .
One might expect that having close to correct poses for active ligands used in training would have a beneficial impact on the tuned scoring functions. This is a difficult effect to measure, in part because one typically employs a single protein structure in screening, so we have used single structures in our experiments. While all of the active ligands in the Pham set (by construction) had known bound poses, since protein conformations change, not all of those poses would serve as appropriate starting points using a single protein structure. Rather than using those directly, we re-docked the active ligands using more aggressive search parameters. In cases where a pose existed within 2Å rmsd of correct, and whose score was within 80% of the highest score for any pose, we replaced the highest scoring pose with this pose for purposes of training. This is related to an approach reported by Smith et al. , where the closest-docked-pose was used in scoring function refinement and compared with making use of crystallographic poses. After this filtering method was applied, 76% our training poses were within 2Å rmsd of correct, vs. 46% without filtering. We repeated the optimization experiment summarized in Table 5. Virtually no difference in test performance was detected across all eight complexes. To a degree, this parallels what was found by Warren et al. , where they observed little relationship between docking accuracy and screening utility. However, this is not an intuitive result and requires more investigation.
The results reported here clearly demonstrate that the parameters governing a scoring function for protein-ligand interactions can be optimized to improve performance for a particular task. Moreover, the multiple constraint approach for constructing an objective function for optimization of scoring functions introduces an extensible framework for making use of many types of data. In this work, we have optimized the Surflex-Dock scoring function to enhance screening enrichment for particular targets. Significant screening improvement was possible when training on as few as 15 known actives, with substantial increases in early enrichment for HIV protease and PARP. In all cases with 10 or more actives, screening performance was improved or stayed the same. For those complexes with less than 10 training ligands, use of the very small data sets was problematic but was reversed by including additional data.
As a practical matter, many practitioners of docking spend a great deal of effort on very small numbers of targets. Frequently, such situations involve access to large quantities of proprietary crystal structures as well as structure-activity data. While refinement of scoring functions for docking will continue toward addressing the general case of application to any target, focused refinement may prove to be of great utility to those whose interests lie in studying a particular target as opposed to caring about the generality of the methodology. By providing the tools for rapid optimization of scoring function parameters to users, we hope that the subtle parameter refinements seen here to yield large changes in performance will be demonstrated on targets “in the wild.”
As a theoretical matter, a rigorous treatment of the multiple instance problem (which pose do we listen to?) coupled with creative use of objective functions (can we enforce a constraint that this ligand or pose is supposed to score better than these others?) may prove to be of use beyond scoring functions in docking or methods in 3D QSAR. The place where such an approach has obvious applicability, but has not yet been tried to our knowledge, is in the development and refinement of empirical scoring functions for use in protein folding.
The authors gratefully acknowledge NIH for partial funding of the work (grant GM070481). Dr. Jain has a financial interest in BioPharmics LLC, a biotechnology company whose main focus is in the development of methods for computational modeling in drug discovery. Tripos Inc., has exclusive commercial distribution rights for Surflex-Dock, licensed from BioPharmics LLC.