1
Optimising the EVA descriptor for prediction of biological activity

2
EVA is a multivariate molecular descriptor for use in QSAR studies.

3
It is constructed from vibrational eigenvalues derived from either a quantum theoretical or molecular mechanical treatment of molecular structure.

4
This paper applies the method to biological-activity data using measures of the inotropic potential of a range of Calcium channel agonists.

5
The performance of the descriptor, as both an explanatory and a predictive tool, is analysed in relation to the way in which it is constructed using a rigorous statistical treatment.

6
Its capabilities are examined in relation to those of previously published methodology which used a composite descriptor.

7
It is shown to have improved performance and several procedural advantages, such as ease of calculation and operation.

8
It is a 3-D structural descriptor which does not require prior co-alignment of structures for a QSAR study.

Introduction

9
A Quantitative Structure-Activity Relationship (QSAR) is a statistical model relating a common property of a series of chemicals (such as their responses in an assay of a particular type of biological activity) to their molecular structure.

10
The major goal is to use this relationship to discover new chemicals with properties optimally suited to a defined application in an area such as pharmaceuticals, agrochemicals or catalysts.

11
The process entails relating structurally significant features of a set of chemicals to activity measurements (responses) that are the complex result of integration of many physical, chemical and metabolic processes.

12
The structural features are summarised numerically in a molecular descriptor, but it is often difficult to do so efficiently and effectively since these features may also be very complex.

13
Furthermore, many studies are performed on data sets comprising limited numbers of compounds and associated responses and this restricts the number of variables that can be used to describe structure in the QSAR model.

14
In order to maximise the value of this small number of variables, molecular descriptors that summarise efficiently a large amount of diverse chemical information are required.

15
One recent example is the theoretically-based multivariate descriptor, EVA.1–3

The EVA (EigenVAlue) descriptor

16
The premise on which the EVA approach is based is that a molecule's normal modes of vibration encode a suitable description of chemical structure, summarising the features necessary for the production of robust, predictive, QSARs.

17
The normal modes are readily derived, using computational techniques based on either Quantum Theoretical or Molecular Mechanical methods.

18
When suitably encoded they may be used as robust structural descriptors in QSAR lead-optimisation studies.

19
In the following section, an overview of the approach is provided; for a more complete discussion, the reader is referred to reference .1

20
QSAR studies demand as complete a structural description of a molecule as possible, in numerical terms.

21
Using Quantum Mechanics, the molecular wavefunction, Ψ, provides a thorough characterisation of the nuclear and electronic properties of a molecule (and in principle should also allow for the estimation of properties such as biological activity).

22
However, extraction of useful information for a QSAR study is difficult.

23
Following studies with experimentally-based Infra-red Spectra, theoretically-derived normal modes of vibration were identified as a suitable molecular descriptor encoding information on the identity of the constituent atoms, their bonding and spatial relationships (hence molecular shape and size), their vibrational modes and molecular electronic structure.

24
Intuitively, such a descriptor should constitute a robust, predictive, structural representation.

25
Calculation of a molecule's normal modes is achieved using the molecular electronic potential energy function, V, which is derived from the molecular wavefunction.

26
Normal Coordinate Analysis (NCA) is performed to characterise the molecular vibrational properties.

27
This approach involves the pointwise estimation of the potential function to determine the equations of motion, yielding both normal coordinate EigenVAlues (the normal mode frequencies of vibration) and their associated eigenvectors (the atom vibration vectors).

28
The EVA procedure makes use of the eigenvalues, the eigenvectors being retained for final back-transformation to the molecular structures.

29
The normal modes for a set of molecules cannot be used directly in a multivariate statistical analysis because current methods demand that the descriptor for each molecule in the training set is comprised of the same number of well-defined components, (i.e., the molecular descriptor must be a vector of fixed dimension).

30
Since the number of normal modes varies with the number of atoms N in a molecule (actually 3N-6 for a molecule without axial symmetry) each descriptor within a series will generally be of different dimension.

31
Data transformation is therefore required to convert, without loss of information, a varying number of vibrational eigenvalues into the descriptor EVA, which is of fixed dimension.

32
This is done by projecting the eigenvalues (vibrational frequencies) onto a bounded frequency scale (BFS) with individual vibrations represented by points along this axis.

33
The frequency range is chosen to be 0 cm−1 to 4000 cm−1 to encompass all fundamental molecular vibrations.

34
Each of the 3N-6 vibrations is represented by an equivalent Gaussian curve, G{f(μ), σ2}, in which μ is the vibrational frequency; the area under each curve is assigned as unity.

35
Proximate or coincident Gaussian functions are permitted to overlap and the “intensity” is summed.

36
The choice of σ for each function defines the degree to which the vibrations overlap and is typically 10 cm−1 to 20 cm−1.

37
This smoothing operation introduces, deliberately, a “fuzziness” to the spectrum of vibrations and is a key step in the definition of EVA; a value that is too small will fail to detect similarities where they exist, and one that is too large will result in overlaps that obscure significant proportions of the variation in the descriptor.

38
The process results in a degree of serial correlation in the descriptor that causes inevitably some redundancy in the descriptor variables.

39
However, it also enables the significance of the presence or absence of peaks to be assessed in the subsequent analysis, together with the monitoring of changes in peak position.

40
Once a Gaussian smoothing function has been applied, the resultant spectrum is sampled across its whole width in fixed increments (L cm−1).

41
The choice of increment determines the number of variables in the EVA descriptor; thus, for example, a 2 cm−1 increment results in a descriptor string of 2000 variables.

42
Molecular vibrations are therefore depicted as “peaks” of intensity on a scale of frequency versus height.

43
This produces the molecular descriptor, EVA, which retains the integrity of its constituent frequency data.

44
The choice of the variables σ and L can have a marked effect on the quality of the subsequently derived QSAR/QSPR model, and this is discussed below (see section 3.2).

45
The EVA descriptor was originally formulated at Shell Research Ltd, Sittingbourne Research Centre, and was applied successfully in a series of agrochemical lead-optimisation studies.

46
While much of the experimental data used in the development remain proprietary and hence are not available for publication, its utility has been demonstrated by modelling experimentally observed logP values.1

47
The intention of the present study is to illustrate the application of the method to biological data that are within the public domain, incidentally providing a comparison of the performance of the EVA descriptor with that of other, more familiar, descriptors.

48
A convenient example lies in a set of calcium ion channel agonists and their associated responses in an assay of biological activity;4 this represents a particularly attractive set as it has been analysed previously by Davis et al5,6. with another theoretically-based 3-D descriptor, GRID.7

49
This paper compares the performance of EVA with the published results of this study.

50
In the previous work,5,6 the 3D structural descriptor, GRID,7 was used to analyse the molecular features which account for the observed activities.

51
The additional inclusion of selected physicochemical descriptors, and the consequences of variable scaling, was then considered in turn to assess whether the QSAR model could be improved upon.

52
All results were cited in terms of the squared correlation coefficient of the fitted PLS regression model, r2.

53
The intention of the present studies is to conduct a parallel investigation to examine the performance of the EVA descriptor.

54
A particular focus will be to examine the robustness of the predictive ability of the EVA based QSAR model.

55
Results will therefore be presented in terms of the squared correlation coefficients of the cross-validated equations, q2, rather than the fitted regression model.

56
The parameter q2 is commonly used as a measure of the predictive ability of the regression model, but in order to facilitate comparison between the two studies the values of r2 will also be included as appropriate.

57
This paper reports the results of an investigation to show how to optimise the EVA descriptor in order to increase the predictive power of a QSAR equation.

Experimental

The Calcium Channel Agonist data set

58
Table 1 shows the structures of the compounds used by Davis et al.5,6

59
The compounds were assayed in vitro for their inotropic potency, i.e., their ability to increase cardiac contractibility, using guinea pig atria paced at 1 Hz.

60
This was expressed as the concentration of the drug that increased the tension developed in the preparation to 50% of the isoprenaline maximum (EC50).

61
Results were expressed relative to the EC50 of the standard calcium channel agonist Bay K 8644 (Fig. 1) and these values are given in Table 1.

62
Also listed in Table 1 are values of the calculated log octanol/water partition coefficient (clogP) and the calculated molar refractivity (cMR) respectively, corresponding to the various structures.

63
These were calculated using MedChem software v.3.54 B (1989).

64
In the corresponding table in the paper by Davis et al. the quoted values for these parameters are different, having been modified by in-house fitting to a few experimentally determined values.

65
In the present paper we use the standard values in order to maintain external consistency.

GRID descriptor

66
For the 3-D QSAR analysis of the data set, Davis et al. made use of the GRID descriptor.7

67
The technique was developed originally as a method of probing proteins for potential binding sites and has been used successfully to predict sites of binding of small ligands in proteins,8 in addition to its more recent role as a structure descriptor.

68
A detailed discussion of the derivation of the GRID descriptor can be found in reference .7

The inclusion of physicochemical descriptors

69
A limitation of GRID is that by generating the descriptors from the calculated interaction energy of a probe molecule, only the enthalpic component of the drug–receptor interaction can be modelled.

70
This neglects other factors which can be, potentially, highly influential, such as the entropic changes resulting from the interaction.

71
In recognition of this limitation, Davis et al5,6. included two physicochemical descriptors, clogP and cMR.

72
ClogP represents a particularly useful descriptor since by accounting for lipophilicity and solvation, features that are influenced heavily by hydrogen bonding, it can give information on the entropic changes associated with a molecule's coming out of solution and binding at a receptor site.

EVA descriptor generation

73
EVA descriptors were generated for the 36 compounds in the calcium channel agonist set.5,6

74
A typing error in the published table5,6 cites incorrectly the substitution about the benzoylpyrrole backbone of compounds 8, 12, 19, 33, 35 and 36 as 2,4 rather than as .2,59

75
The correct structures were used in the present study and were created using Sybyl 6.2 from Tripos Associates10 on a Silicon Graphics Indy R4000 platform.

76
In the previous studies, Davis et al. constrained the energy minimisation of the structures in an attempt to minimise noise in the GRID descriptors and hence optimise the subsequent QSAR investigations.5,6

77
In order to perform the calculation of the normal modes of vibration required for EVA, it is a prerequisite to ensure that the potential energy of a molecular conformation is at a stationary point.

78
Slight deviations away from this energy minimum can lead to the introduction of rotational and translational modes during the determination of the equations of motion in the NCA.

79
This in effect ‘contaminates’ the vibrational modes, resulting in the determination of vibrations that are not purely harmonic in nature and degrading significantly the quality of the calculated vibrational spectrum.

80
A more detailed discussion of this is provided in reference .11

81
No attempt was made therefore to reproduce directly the conformational co-ordinates used by Davis et al.5,6

82
Molecular structures were optimised using the MOPAC 6.0 AM1 Hamiltonian12 (keywords: SCFCRT = 1.D-12 GNORM = 0.05).

83
Successfully minimised structures were then used to generate the NMs (additional keyword: FORCE).

84
These were then converted into EVA descriptors in the manner described.

85
It is acknowledged that the calculation of NMs using a semi-empirical route is less accurate than for example, by the use of Density Functional Theory methods.13

86
However, the use of the MOPAC AM1 Hamiltonian has previously been shown to produce qualitatively good results.14

87
Furthermore, EVA descriptors generated from the MOPAC AM1 Hamiltonian have been applied successfully in previously reported studies.14

88
Consequently, it is believed that the EVA descriptors generated in this way are acceptable for the present investigation.

Statistical analysis

89
To generate useful QSARs between the activity, EC50, and the EVA descriptor, the Partial Least Squares regression method (PLS) was adopted for the investigations; this is in line with the GRID-based studies of Davis et al.5,6

90
PLS is a supervised approach that provides a reduced solution of new ‘latent’ variables that are linear combinations of the original variables and are well correlated with Y. In essence, it is a projection method that defines a hyperplane through the x-descriptor space.

91
The latent variables represent the projected ‘summaries’ of the original variables onto this plane and each PLS component represents a co-ordinate dimension of the hyperplane.

92
A more complete discussion of the PLS approach can be obtained from references 15 and .16

93
The PLS regression approach is an accepted technique for handling situations where over-square matrices are encountered and has some advantages over unsupervised techniques such as, for example, Principal Components analysis.

94
By constructing the components in a supervised fashion, it is possible to produce fewer numbers of components correlated with Y, thereby summarising the data more efficiently.

95
With matrices such as EVA and GRID this is a clear advantage.

96
As has been mentioned earlier in section 1.1, by having the ability to vary the form and dimension of the EVA descriptor (i.e., σ and L), it is possible to change the manner in which the structural information is presented.

97
In effect, this can be thought of as a form of data scaling, in line with log and variance scaling methods.

98
This affects ultimately the regression analysis, particularly with respect to the co-variance with the activity data and the way in which the information is loaded onto the components.

99
Therefore an essential stage in the application of EVA is the generation of descriptors with different σ and L values and their subsequent systematic validation to identify the optimum predictive model.

100
Previous studies have shown that a σ value of between 10 and 20 cm−1 is a reasonable starting point for such investigations.1

Model validation and significance testing

101
A key step in regression analysis is the validation to determine the correct dimensionality of the model, i.e., the number of significant PLS components to be included.

102
Without this validation there is a risk of ‘over-training’, leading to apparently well-fitting models with effectively little predictive capacity.

103
As the intention of the present study is to assess the predictive capability of the EVA descriptor, two forms of model validation have been incorporated.

104
In line with the study by Davis et al., Cross-Validation (CV) has been used to test each successive PLS component for significance.

105
In addition, randomisation tests have been included to establish confidence limits on whether or not random correlations are being selected in place of real relationships.

106
CV represents one of the most favoured methods of testing the PLS components for significance.

107
In essence, CV randomly divides the data set into groups of compounds.

108
Each group is then omitted from the data set in turn and the remaining set used to construct a new regression model, which is then used to calculate the responses for the members of the omitted set.

109
From the predicted y-values for each of the omitted ‘test’ sets, the sum of squares of the differences from actual values can be calculated.

110
This yields the PRESS score (Predictive Residual Sum of Squares) and by weighting this against both the inclusion of successive components and the number of cases in the set, the standard error of cross-validation, Scv, can be calculated.

111
This provides a good test of the significance of the inclusion of each component upon the overall regression model and has some advantages over other procedures such as, for example, the F-ratio test.

112
By simulating how well the model predicts new data, it provides an estimate of the robustness of the model and potentially can highlight models that are influenced by data points with high leverage (e.g., outliers).

113
As the underlying intention of a QSAR is to make predictions, then this is useful.

114
Initial analysis of the EVA descriptor models was performed using Leave-One-Out Cross-Validation (LOO CV).

115
To determine the optimum regression model, the Scv scores for the first 10 PLS components were calculated and the minimum Scv score identified.15

116
Previous experience has shown that the relevant information linking the response and descriptor variables is summarised in the first few components (typically ≤ 6).1

117
The squared correlation coefficient of the cross-validated equation, q2, was then cited for the inclusion of each successive component.

118
In a previous publication,1 it was shown that the LOO CV q2 obtained on a training set of 135 structures, using EVA as the descriptor and logPow as the measured response, was a very good indicator of subsequent predictive power for logPow in an unrelated test set of 76 compounds.

119
However, work by Shao17 has challenged the suitability of LOO CV for testing the predictive robustness of a QSAR model and suggests that Leave-n-Out Cross-Validation, where n is greater than 1, represents a more robust alternative.

120
Consequently, in the present investigations we have compared LOO CV results with those obtained using Leave-Group-Out Cross-Validation (LGO CV) with n = 7 (see section 3.2); this is also in line with the validation approach applied by Davis et al.5,6

121
However, while LOO CV represents an easily reproducible test, i.e., each run will yield the same q2 result, in LGO CV the cases assigned to each group to be left out are selected at random.

122
This means that LGO CV must be repeated a large number of times (e.g., 1000) in order to obtain a realistic estimate of the internal predictivity and is hence more computationally demanding than LOO CV.

123
Our work shows that with this data set and analytical approach, LOO CV gives very similar results to LGO CV and so we have adopted the simpler procedure.

124
It must be acknowledged that neither LOO CV nor LGO CV can be regarded as a sufficient indicator of model validity; in addition, some measure of the probability that the result may be a chance occurrence is needed.

125
This is particularly important in regression studies where there are a smaller number of cases than the dimensionality of the x-descriptor, i.e., where an over-square matrix is present.

126
In such instances, there is an increased possibility of the extraction of a random correlation.

127
Consequently, in order to address this with regard to the EVA descriptor, a randomised permutation test was performed to examine the model for chance correlations; this is in line with previous studies that have applied the EVA descriptor.2,18

128
The process involves:

129
1.

130
Randomly assigning observations to structures,

131
2.

132
Performing a regression analysis using EVA and then evaluating q2 (or r2),

133
3.

134
Repeating steps 1 and 2 a large number of times (e.g., 1000 times) to ensure thorough sampling,

135
4.

136
Determining the frequency distribution of q2 (or r2),

137
5.

138
Determining where in this frequency distribution the q2 (or r2) of the real assignment lies.

139
An actual result that is extreme in the upper confidence tail of the distribution is regarded as having a low probability of occurring by chance

Results and discussion

140
In the following section, the results are reported for the QSAR investigation of a set of 36 calcium channel agonists (Table 1), using EVA as the structural descriptor and PLS regression as the statistical approach.

141
Except where stated otherwise, all results will be reported in terms of the squared correlation coefficient (q2) of the Leave-One-Out Cross-Validated (LOO CV) equation.

142
To facilitate comparisons, Table 2 is taken directly from the work of Davis et al., detailing the PLS regression models reported for the full 36 compound data set.5,6

Determination of optimum PLS regression model

Optimum EVA descriptor model

143
Initial identification of the optimum EVA model was performed by comparing the results obtained with a series of descriptors using σ values of 10, 20, 30 and 40 cm−1, respectively.

144
The results established a narrowed focus of attention around σ = 20 cm−1 and a further series of EVA descriptors was then generated (16 ≤ σ ≤ 26 cm−1).

145
Comparison of the predictive performance of the models identified the optimum EVA model as having σ = 24 cm−1.

146
Throughout, the BFS start point, S, was fixed at 1 cm−1 and L was defined as σ/2.

147
This has been assumed previously1 to represent a reasonable sampling frequency for a given value of σ and is justified below (section 3.2).

148
The results for the full range of EVA models considered are included in Table 3.

Model transformations and variance scaling effects

149
As was demonstrated earlier,5,6 PLS analysis can be highly sensitive to the relative scales of variance between the x-block and y-variable (this is especially significant when different x-descriptors, each with its own unit and scale, are included in the regression equation, although this situation does not apply to EVA).

150
A series of transformations and scaling procedures was applied to both the activity and EVA variables, to investigate the impact on performance.

151
Scaling and transformation tests were conducted in two parts.

152
In the case of the activity variable, EC50, both log scaling and variable standardisation were considered, as in the study of Davis et al.5,6

153
For the EVA descriptor, two forms of variance scaling were applied, Autoscaling and Blockscaling.

154
Autoscaling standardises each individual variable column to zero mean and unit variance, while Blockscaling scales the entire descriptor block to zero mean and unit variance.

155
The impact of these two scaling methods is of particular interest, because differences in column variance across the descriptor are considered to contain structurally significant information.

156
The full results are reported in Table 4.

157
Variance scaling with the EC50 variable has no impact on the PLS regression model (4c model, q2 = 0.498).

158
Similarly, blockscaling the EVA descriptor does not affect the performance.

159
However, log10 transformation of the activity values does result in a significant degradation of the PLS model (1c model, q2 = 0.194).

160
Notably, use of the autoscaled EVA descriptor yields the worst model of all.

EVA QSAR model testing and stability

161
As discussed above (section 2.4.1), with over-square x-matrices the PLS regression approach, potentially, can extract chance correlations with the y-variable.

162
In the case of the EVA descriptor, an extreme example of an over-square matrix, then the likelihood of such an occurrence may be increased.

163
Therefore, a series of rigorous validation exercises was applied to establish confidence in the QSAR model obtained.

164
These include sensitivity tests on the effects of changes in the EVA L variable, LGO CV (in line with references 5 and 6) and randomised activity permutation tests.

165
Fundamentally, two variables, namely σ (Gaussian standard deviation) and L (the BFS sampling interval), determine the form and dimension of the EVA descriptor when derived from the normal coordinate spectrum.

166
Potentially, the value chosen for L can impact on the information content retained in the final descriptor model and an extreme illustration of this would occur if the sampling interval were so large that not all of the intensities under the Gaussian curves were to be sampled.

167
As outlined earlier in the initial determination of the optimum EVA model, L was defined as σ/2.

168
This ensures that all structural information is retained, while keeping the associated computational time within reason.

169
Validation of this procedure was carried out by comparing the effects of using a range of sampling intervals (σ/4 ≤ L ≤ 4σ).

170
The full set of results is shown in Table 5 and illustrated in Fig. 2.

171
For sampling intervals where Lσ, the information content remains intrinsically stable, with no significant impact on the statistical performance of the EVA models.

172
However when L > σ, decreasing stability in the performance in the EVA model is observed.

173
The results confirm that L = σ/2 is a reasonable rule.

174
To test rigorously the predictive performance of the optimum EVA model (EVA{24,12}), LGO CV was used and compared with the results of LOO CV.

175
In line with the LGO CV in references 5 and 6, the 36 compounds were divided randomly into 7 groups.

176
Repeated runs of the LGO CV (1000 iterations) were carried out with different randomisations to ensure that the range of possibilities was thoroughly sampled.

177
The results for the LGO CV study are presented in Table 5.

178
These show that the 4-component model occurs most frequently (560 iterations; Fig. 3a), in agreement with LOO CV.

179
The range of q2 scores for the LGO CV 4 component model is approximately normal in distribution (Fig. 3b and 3c).

180
Determination of the mean q2 reveals that q2LGO compares closely with q2LOO (Fig. 2c: q2LGO = 0.474, q2LOO = 0.498), which gives confidence in the use of LOO CV to test the range of EVA models.

181
Figs. 4 and 5 show the results of the randomised activity permutation tests on the same data set, obtained by using 1000 permutations of the response data with respect to structure.

182
Fig. 3 shows the frequency distribution and normalised plot of q2, obtained from the LOO CVs; the actual experimental result of 0.498 lies well outside the upper 95% confidence limit (>99%).

183
A similar result is shown in Fig. 4 for the fitted r2 distribution.

184
Table 6 summarises these results.

Augmentation of EVA {24,12} with the physicochemical descriptors clogP and cMR

185
Davis et al5,6. showed that the 3D molecular descriptor GRID did not by itself encode sufficient information to give a high-quality QSAR regression model of −log EC50 (r2 = 0.42).

186
However, inclusion of clogP and cMR, in linear combination with GRID, gives a significant improvement (r2 = 0.86).

187
It is therefore of interest to examine the influence of including clogP and cMR with EVA.

188
The results are given in Table 6 and show that inclusion of these parameters (either singly or in combination) degrades the quality of the fit (r2) and also the predictive power (q2).

EVA and analysis of the data of Davis et al5,6.

189
QSAR methods that relate biological response to molecular structure are traditionally based on linear free energy relationships, as illustrated in the pioneering work of Hansch et al.19

190
Typically, these methods attempt to draw correlations between empirical descriptors which supposedly represent generalised forms of enthalpic and entropic effects associated with a drug molecule's transport to, and interaction with, a receptor or enzyme active site.

191
The current research is aimed at further developing a new molecular descriptor, EVA, which encodes a large amount of diverse chemical information and can be used more effectively than the previous empirical forms.

192
Analysis of the predictive performance of a range of EVA PLS models (Table 3) identifies the optimum descriptor as having a Gaussian sigma; = 24 cm−1 and a sampling interval L = 12 cm−1, EVA {24,12}.

193
Cross-Validation (using both LOO CV and LGO CV methods) suggests that the first 4 components are significant, yielding a QSAR that is able to account for circa. 80% of the variation in the EC50 activity (fitted r2 = 0.801).

194
Rigorous testing of the internal predictivity of the EVA QSAR model (Tables 3 and 5) shows that for structures outside the training set circa. 50% of the activity variation may be predicted with confidence, an experimentally useful result (4-component models, q2LGO = 0.474; q2LOO = 0.498).

195
The close agreement between the two validation methods, illustrated in Fig. 2, gives considerable confidence in the use of LOO CV to validate the chosen EVA model and indicates that EVA is robust as a predictive QSAR tool, i.e., the model is stable.

196
Since Davis et al5,6. analysed their data with log-transformation of the EC50 values (i.e., the y-variate was −log(EC50), Table 1) and their original data are not available to us, an exact comparison of the two sets of results is difficult.

197
Moreover, cross-validation tests were not reported so it is only possible to compare values of r2, obtained by fitting the full-regression equation, rather than q2.

198
The best model of Davis et al5,6. uses a linear combination of GRID, clogP and cMR.

199
(GRID alone gives a 1-component model, r2 = 0.42, clogP alone a 1-component model, r2 = 0.69, and GRID in combination with clogP and cMR gives a 4-component model, r2 = 0.86).

200
The values of −log(EC50) predicted by this model are not quoted, but a comparison with experimentally observed values is given graphically in Fig. 5 of reference 5; by careful measurement, it is possible to obtain reliable estimates for the predicted −log(EC50) values.

201
Taking antilogarithms yields values for predicted EC50 and the regression statistics may be calculated and compared with those obtained for the EVA model.

202
Overall, the full EVA model (r2 = 0.80, standard error = 4.1) gives a considerably better fit to the untransformed experimental data than does the GRID + clogP + cMR model (r2 = 0.42, standard error = 7.0).

203
The results are shown in Table 7 and are further illustrated in Fig. 6 which shows plots of residuals for both treatments.

204
For small values of EC50 (typically less than 4.1, the standard error of the EVA treatment) the two models (not surprisingly) cannot be distinguished; however, for higher values of EC50 EVA consistently outperforms the other model by an impressive margin.

205
A model-validation test using randomised activity permutations establishes confidence intervals for both the predictive (Fig. 4) and fitted (Fig. 5) regression models and demonstrates that, although the EVA descriptor presents an over square x-matrix in the PLS regression model, random correlations with the activity data are highly unlikely (>99% probability that the correlations have not arisen by chance).

206
EVA contains structurally significant features that account well for the observed biological activity.

207
It should also be emphasised that EVA is a robust descriptor of 3D molecular structure, which does not require prior alignment of structures in order for a QSAR study to be performed.

208
This is especially useful when dealing with non congeneric series, where alignment about a common skeletal feature is not feasible.1,3

209
Although the relevance, multicolinearity and redundancy of the EVA descriptors have not been addressed in this study, they probably provide the most likely explanation for the relatively small values obtained for the estimates of LOO q2.

210
These aspects will be reported in a future publication.

211
Clearly, a model that yielded a q2 of >50% would be even more useful.

212
Plots of the predicted vs. actual activity for both the fitted and cross-validated regression models (Figs. 7 and 8) highlight problems with outliers, particularly those compounds with EC50 values which are essentially zero; these structures can only contribute noise to the analysis, and the consequences are addressed in detail in section 3.7.

Data scaling

213
In studies where the x-variate function is a vector consisting of a number of variables such as molecular properties with different dimensionalities and variances, careful tailoring of the variable variances is critical in extracting an optimum QSAR model.

214
This is true, for example, of the work by Davis et al.5,6

215
In the present study however, Table 4 shows that blockscaling and autoscaling of the x-variate function does not improve the EVA QSAR model.

216
In particular, autoscaling (which scales columns independently to zero mean and unit variance), performs significantly worse.

217
In the EVA descriptor, all variable columns are presented on the same scale with the differences in variance between them intrinsically encoding structurally relevant information.

218
Consequently, individual columns should not be standardised since this would result in giving columns with little variation and therefore little information, undue weighting.

219
Indeed, in-house development with proprietary data sets (Shell Research Ltd., the former Sittingbourne Research Centre) has in the past demonstrated that scaling typically does not improve the performance of EVA-based QSAR models.

220
Log transformation of the y-variate, EC50, leads to a substantially worse EVA model.

221
This contrasts markedly with the results of a similar transformation in the work of Davis et al., which gave the best result.

222
This is investigated further in section 3.7, below.

Inclusion of the physicochemical descriptors

223
It is reasonable to ask the question as to whether or not the inclusion of the physicochemical parameters clogP and cMR would improve the EVA model.

224
The results of doing so are shown in Table 8 and provide an insight into the information content of EVA.

225
It is clear that the inclusion of either clogP or cMR, or both, does not lead to an improved QSAR model, even with scaling of the variables in this more complex x-variate vector.

226
The implication may be that the information contained within the physicochemical parameters is already encoded intrinsically within the EVA descriptor.

227
In order to test this hypothesis, namely that EVA encodes physicochemical information, regression models were constructed for clogP and cMR, respectively.

228
The results (Table 9) show that while EVA accounts for the variation in cMR extremely well (r2 = 0.983, q2 = 0.963), it does not do so at all well for clogP (r2 = 0.415, q2 = 0.281).

229
The result with clogP is surprising, since in previous work it has been shown that EVA QSAR can explain the variation in experimentally measured logP values across a wide variety of structurally-diverse, non-related compounds and be usefully predictive (135 compounds, r2 = 0.96, q2 = 0.68).1

230
Why should the approach be less successful with the congeneric set of structures described in Table 1?

231
The most likely reason for this anomaly is in the variance in clogP; the range of values is relatively small, i.e., only 3 log units in comparison to 6 log units in the logP study.

Investigation of outliers

232
It was noted that eight of the structures in the data set have corresponding values of EC50 that were approaching zero, ranging from 0.01 to 0.09 (see Table 1).

233
Given the nature of the bioassay from which they have been derived, within experimental error, they probably may not be significantly different from zero.

234
Inclusion of these data has the consequence of biasing the activity data set (y-variate) heavily, and introducing noise in the form of irrelevant descriptor information (x-variate).

235
The consequence of removing these data to leave a set of 28 structures and activity scores was investigated.

236
Table 10 compares the regression results with those obtained for the original set of 36 structures.

237
For the untransformed EC50 data, there was a slight improvement in both the fit of the full regression equation and in the predictive power, albeit with the addition of a further 2 components in the regression model (6 component model, r2 = 0.939, q2 = 0.527, cf. a 4-component model, r2 = 0.802, q2 = 0.498).

238
In contrast, there was a highly significant change in the performance of the regression results for the −log(EC50) scores, Figs. 9 and 10 (5-component model, r2 = 0.963, q2 = 0.721, cf. a 1-component model, r2 = 0.332, q2 = 0.194).

239
This compares favourably with the results of a similar transformation in the work of Davis et al. (4-component model, r2 = 0.86).

240
More importantly, the q2 result would provide significant confidence in support of the application of this model in a predictive role.

241
However, it should be recognised that chemical interpretability will be non-trivial, given the Gaussian transformation and subsequent sampling of the normal modes of vibration.

242
The degree of improvement in the log transformed EC50 was marked, however, it was not unexpected.

243
As was highlighted earlier, PLS analysis can be highly sensitive to the relative scales of variance between the x-block and y-variable.

244
Certainly, for the log-transformed y-variable set, this variance would have been significant.

245
Furthermore, given the likely experimental error in the bioassay, it is reasonable to postulate that relative differences in the log values for these compounds would have been subject to significant error.

246
Consequently, in attempting to generate a model, these data would have had a greater bias in the log-transformed set.

247
It remains interesting to speculate on whether the sensitivity in model formation was also in part a product of the choice of an essentially linear regression model.

248
While outside the scope of this study, it remains an appealing point for future studies to test the performance of the EVA descriptor across a series of non-linear statistical methods.

Chemical similarity searching

249
The formulation of the EVA descriptor as a bit string opens up the possibility of using the encoded vibrational mode information for chemical similarity searching.

250
There is a substantial body of literature that describes the procedures and software available for this purpose and which points out the strengths and pitfalls of the approach.20–22

251
Current procedures are based principally on the use of chemical fragments represented as topological maps, atom counts and bond types.

252
Use of EVA for similarity searches would provide useful additional information to complement and extend this structural data.

253
Applications might include selective compound acquisition20 and identification of compounds with similar biological activities21.

Conclusions

254
Within the context of the present study, EVA is well suited to the production of usefully predictive QSARs with both biological and physicochemical data.

255
The absence of a requirement for additional descriptors demonstrates that EVA has a very high information content encompassing, intrinsically, physicochemical descriptors such as clogP and cMR.

256
In direct comparison with a composite descriptor (GRID + clogP + cMR), EVA has been shown to have superior predictive performance.

257
As a general tool for QSAR studies, EVA advantageously may be applied without the prior alignment of structures required for other 3-D descriptors such as COMFA and GRID.

258
However, optimisation of the EVA descriptor by adjusting how it is constructed can influence the predictive power of a QSAR equation.

259
Further studies with other sets of biological data will be necessary in order to demonstrate the robustness of these conclusions.

Abbreviations

260
QSAR, Quantitative Structure-Activity Relationship; EVA {σ, L, S}, Eigen VAlue molecular spectral descriptor model where: σ (sigma) = standard deviation of the EVA Gaussian curve, L = sampling interval of summed intensities and S = sampling interval start point on the bounded frequency scale; BFS, Bounded Frequency Scale (0 to 4000 cm−1); NCA, Normal Co-ordinate Analysis; NMs, Normal Modes (of Vibration); EC50, relative force of contraction; clogP, calculated log octanol/water partition coefficient; cMR, calculated molar refractivity; PLS, Partial Least Squares regression analysis; r2, squared correlation coefficient; CV, cross-validation; LOO, ‘Leave One Out’ cross-validation; LGO, ‘Leave One Group Out’ cross-validation; q2, squared correlation coefficient of the cross-validation equation; PRESS, Predictive REsidual Sum of Squares; Scv, cross-validated standard error, where ; n, number of compounds in data set; c, number of PLS components included in QSAR model; F-ratio, weighted ratio of r2 to 1.0 − r2, where