J. Chem. Inf. Model. 2009, 49, 195–208
195
Analysis of Neighborhood Behavior in Lead Optimization and Array Design George Papadatos,† Anthony W. J. Cooper,‡ Visakan Kadirkamanathan,§ Simon J. F. Macdonald,‡ Iain M. McLay,‡ Stephen D. Pickett,‡ John M. Pritchard,‡ Peter Willett,† and Valerie J. Gillet*,† Krebs Institute for Biomolecular Research and Department of Information Studies, University of Sheffield, 211 Portobello Street, Sheffield S1 4DP, United Kingdom GlaxoSmithKline, Medicines Research Centre, Gunnels Wood Road, Stevenage SG1 2NY, United Kingdom, and Department of Automatic Control and Systems Engineering, University of Sheffield, Mappin Street, Sheffield S1 3JD, United Kingdom Received August 28, 2008
Neighborhood behavior describes the extent to which small structural changes defined by a molecular descriptor are likely to lead to small property changes. This study evaluates two methods for the quantification of neighborhood behavior: the optimal diagonal method of Patterson et al. and the optimality criterion method of Horvath and Jeandenans. The methods are evaluated using twelve different types of fingerprint (both 2D and 3D) with screening data derived from several lead optimization projects at GlaxoSmithKline. The principal focus of the work is the design of chemical arrays during lead optimization, and the study hence considers not only biological activity but also important drug properties such as metabolic stability, permeability, and lipophilicity. Evidence is provided to suggest that the optimality criterion method may provide a better quantitative description of neighborhood behavior than the optimal diagonal method. INTRODUCTION
The similar property principle1 states that structurally similar molecules tend to exhibit similar properties and underlies many techniques in chemoinformatics such as virtual screening and QSAR and ligand binding as well as the design of corporate screening collections.2,3 There are several exceptions to the Principle:4,5 even so, if it was not of general applicability, then it would be very difficult to attempt the development of systematic approaches to the identification of novel bioactive molecules, and there is a large, and increasing, body of evidence to support its use in drug discovery programs.6-9 Barbosa and Horvath10 note that the Principle has two major implications. First, if a molecule is known to be active against some biological target, then molecules that are structurally similar to the chosen molecule are likely to exhibit the same activity; second, it is possible to predict the properties of novel molecules, given a list of structurally similar molecules for which the requisite property data are already available. A further, closely related concept is the neighborhood principle,11,12 which states that molecules within the same local region, or neighborhood, of structural space tend to include more molecules with similar values of some desired property (usually biological activity) than would be expected in any other randomly selected region of the same size. Use of the word “similar” implies some way in which structural similarity can be measured, and there are two principal components of any similarity measure: a descriptor (or set of descriptors) by which molecules may be represented and a similarity coefficient to quantify the * Corresponding author e-mail:
[email protected]. † Krebs Institute for Biomolecular Research and Department of Information Studies, University of Sheffield. ‡ GlaxoSmithKline. § Department of Automatic Control and Systems Engineering, University of Sheffield.
Figure 1. Examples of good and bad neighborhood behavior. Adapted from ref 14.
degree of resemblance between two such representations. In this paper, we focus on molecular descriptors, specifically on the extent to which different types of descriptor are able to satisfy the neighborhood principle, an ability that is referred to as neighborhood behaVior.11 Informally, neighborhood behavior can be regarded as the extent to which a structural space can be mapped onto the property/biological activity space in such a way that neighboring points in the former are likely to correspond to neighboring points in the latter (as shown in Figure 1). Different descriptors will exhibit different levels of neighborhood behavior, and it is thus possible to compare the effectiveness of different types of structural representation by the extent to which they are able to map successfully between structural space and property space. We focus here on the design of arrays to support lead optimization, an application that seems to be very well suited to the neighborhood behavior concept. Lead optimization involves synthesizing and testing hundreds or thousands of structural analogues, in an effort to find the optimal combination of desired criteria such as potency, selectivity, lipophilicity, and solubility. Optimization is increasingly
10.1021/ci800302g CCC: $40.75 2009 American Chemical Society Published on Web 12/23/2008
196 J. Chem. Inf. Model., Vol. 49, No. 2, 2009
Figure 2. The Patterson or Neighborhood plot. Adapted from ref 14.
carried out using arrays of molecules. An initial array, based on the original lead molecule, is synthesized, purified, and tested, and then the most promising member of the array (i.e., the one that best fulfills the chosen criteria) is used as a starting point, or seed, for the design of the next array. This seed will be structurally modified, typically by altering one or two R-groups, with the hope that at least one member of the new array will be able to serve as the seed for a subsequent array. This is a simple application of the similar property principle, with small variations in structure being carried out to enable a systematic exploration of the structural space surrounding the current seed molecule. The exploration will be effective only if the descriptor used to determine the coverage of structural space does exhibit neighborhood behavior: otherwise the systematic exploration will rapidly degenerate into a random search. The paper is organized as follows. The next section discusses previous studies of neighborhood behavior, focusing on two methodssthe optimal diagonal method of Patterson et al.11 and the optimality criterion method of Horvath and Jeandenans13sthat provide a quantitative basis for assessing the extent to which a descriptor does indeed exhibit neighborhood behavior. We then report the results of an extended set of experiments that use array data generated in lead optimization and array design programs at GlaxoSmithKline. The neighborhood behavior of several widely used 2D and 3D fingerprints is evaluated using the two chosen methods, and we provide evidence to suggest that the optimality criterion method may provide a better quantitative description of neighborhood behavior than the optimal diagonal method. PREVIOUS STUDIES OF NEIGHBORHOOD BEHAVIOR
The neighborhood behavior concept was first presented in the much-cited paper by Patterson et al.11 Here, 20 small QSAR data sets were characterized by different molecular descriptors, including 2D fingerprints, ClogP, connectivity indices, topomeric steric field-based descriptors, and random numbers (as a blind test). The dissimilarities were computed between each pair of molecules in a data set, in terms of both structure (as characterized by the Tanimoto coefficient or Euclidean distance for the chosen descriptor) and activity (as characterized by the absolute difference in logIC50). The resulting Patterson or neighborhood plot reflects the neighborhood behavior of the descriptor, as shown in Figure 2. An effective descriptor, i.e., one that exhibits neighborhood
PAPADATOS
ET AL.
behavior as illustrated in the upper part of Figure 1, will yield a plot in which pairs of molecules with small structural dissimilarities (i.e., similar as defined by the chosen descriptor) have small differences in activity; an ineffective descriptor will have many data points in the forbidden area, corresponding to pairs of molecules that have large activity differences but small structural dissimilarities. Patterson et al. defined the neighborhood enhancement (or NBE) of a descriptor, this quantifying the extent of the neighborhood behavior exhibited by that descriptor, and described a procedure based on χ2 to test the statistical significance of the resulting NBE (as described in detail in the section Experimental Methods). A modification of this test was used in a study of ACE inhibitors characterized by UNITY fingerprints, molecular steric fields, WHIM indices, pharmacophore atom-pairs, spatial autocorrelation functions, and, for comparison, molecular weight and random numbers.15 The study concluded that UNITY fingerprints and the pharmacophore atom-pairs exhibited the best neighborhood behavior. The studies of Patterson et al. and Matter suggested that a good descriptor would be one that had most of the data points in the lower-right part of a Patterson plot, i.e., below the diagonal in Figure 2. Subsequently, Dixon and Merz suggested that neighborhood behavior should be associated with a distinct pattern of progressively wider ranges in the biological activity differences at increasing dissimilarity scores, i.e., a “gradual fanning out” of points as one moves to the right of a Patterson plot.16 They also showed that the χ2 test used by Patterson et al. tends to overestimate any genuine neighborhood behavior, markedly in some cases, and proposed an adjusted χ2 test based on repeated randomization of the activity values in the data set being analyzed. Finally, Dixon and Merz also advocated the use of the correlation coefficient r as an alternative indicator of the strength of the relationship between intermolecular activity differences and structural dissimilarity. More recently, Horvath and Jeandenans have discussed neighborhood behavior in the context of actiVity profiles, where molecules have multiple activity values associated with them.13,17 Specifically, an activity profile is a consistent set of experimental activity values for a molecule when measured against a panel of different biological assays. A profile is hence a vector that describes the location of a molecule in activity space, and the dissimilarity between pairs of these vectors can be measured using Euclidean or Dice distances (see also work by Fliri et al.18 and by Schuffenhauer et al.19). Horvath and Jeandenans described two parameters that quantify the neighborhood behavior of a descriptor. The consistency criterion (χ) is the propensity of a similarity measure to selectively rank pairs of activityrelated molecules among the structurally most similar pairs; the oVerall optimality criterion (Ω) is the fraction of activityrelated pairs that rank among the structurally most similar pairs. These two parameters, consistency and optimality, are analogous to the precision and recall parameters that are used to evaluate text-retrieval systems and that have been studied for chemoinformatics applications by Edgar et al.20 The performance of different similarity measures can be assessed and compared by generating Ω-χ plots, where good neighborhood behavior is associated with high optimality values at high consistency values, and this approach was
NEIGHBORHOOD BEHAVIOR
IN
LEAD OPTIMIZATION
J. Chem. Inf. Model., Vol. 49, No. 2, 2009 197
followed in a study of 584 commercially available drugs and reference compounds screened against a panel of 42 assays.17 A large number of 2D and 3D topological and structural descriptors were generated, with the Euclidean and Dice distance coefficients being used to measure structural dissimilarity. FBPA (Fuzzy Bipolar Pharmacophore Autocorrelogram) descriptors were found to exhibit the best neighborhood behavior, closely followed by global 2D and 3D descriptors, with the Dice coefficient performing noticeably better than the Euclidean distance at measuring intermolecular dissimilarity. A subsequent study of FBPA descriptors compared the extent of the relationship between neighborhood behavior and clustering behavior.21 Most neighborhood behavior studies have used relatively small, and often homogeneous, data sets. However, Perekhodtsev used four large, structurally diverse sets of serine protease inhibitors in a neighborhood behavior study to validate the use of standard 2D similarities (Accelrys Accord fingerprints and the Tanimoto coefficient) for the prediction of binding free energies (both measured and computed).22 Analyses based on Patterson plots and on the correlation coefficient r (following Dixon and Merz16) indicated that all four data sets demonstrate neighborhood behavior, i.e., that structurally similar compounds tend to exhibit more similar binding affinities than do those that are less structurally similar. Perekhodtsev’s focus on the data set, rather than on the descriptors used to characterize molecules within it, mirrors recent work that seeks to quantify and classify the nature of the underlying structure-activity relationships. Methods such as the SAR Index (SARI)23 and the Structure-Activity Landscape Index (SALI)24 are based on simple formulas that involve the direct calculation of activity distances ∆A and structural dissimilarities ∆S among molecule-pairs and are thus directly related to neighborhood behavior; for example, SALI (eq 1) is the slope of the diagonal between a given point (i,j) and the origin of a Patterson plot (Figure 2). SALIij )
|∆Aij| ∆Sij
(1)
EXPERIMENTAL METHODS
Data Sets and Descriptors. Data from three GSK lead optimization projects (Projects I-III) were used for our experiments. The Project I data set contained data for 2331 molecules, with IC50 values against two protein targets (Target 1 and Target 2). The Project II data set contained data for 2971 molecules, including IC50 values against three protein targets (Targets 3-5) as well as measurements of lipophilicity, membrane permeability, and metabolic stability. Finally, the Project III data set contained 3286 molecules with IC50 values against one target (Target 6). An extensive series of preprocessing routines was applied to the raw project data. Duplicate entries were merged, and the associated property values averaged. Entries with obsolete or obviously mistaken values or values with modifiers (i.e., ‘’) were excluded, as were singleton compounds that had not been synthesized via an array and salts, mixtures, and stereoisomers. Low-potency compounds (pIC50
