Introduction

Cytosolic phospholipase A2α (cPLA2α) is an esterase that selectively catalyzes the hydrolysis of the sn-2 ester of arachidonate-containing membrane phospholipids to generate free arachidonic acid and lysophospholipids (Ghosh et al., 2006; Kita et al., 2006). The freed arachidonic acid is rapidly oxidized via cyclooxygenase (COX) and lipoxygenase (LO) pathways to eicosanoids such as prostaglandins, which play a major part in the inflammatory response, and leukotrienes, which play a main role in the pathogenesis of asthma. Remaining lysophospholipids with an alkyl ether moiety at the sn-1 position can be acetylated to platelet-activating factor (PAF), another mediator of inflammation. Although several other phospholipases A2 are present in the mammalian organism, the predominance of cPLA2α for lipid mediator generation was demonstrated especially by studies with cPLA2α-deficient mice (Bonventre et al., 1997; Hegen et al., 2003; Miyaura et al., 2003; Nagase et al., 2000; Sapirstein and Bonventre, 2000; Uozumi et al., 1997). These animals, which show a reduced eicosanoid production, are immune to disease in a variety of models of inflammation, including collagen-induced arthritis. Therefore, this enzyme (cPLA2α) can be regarded as a target for inflammatory diseases (Bonventre, 2004).

Despite the fact that there have been intense efforts for developing inhibitors of cPLA2α (Clark and Tam, 2004; Connolly and Robinson, 1995; Lehr, 2006; Magrioti and Kokotos, 2010), only a few substances with high in vitro potency have been found so far, such as the thiazolidinedione (Seno et al., 2000, 2001), the benzhydrylindole (efipladib) (Lehr, 2006; Lee et al., 2007, 2008; McKew et al., 2008), and the 1,3-diaryloxypropan-2-one (AR-C70484XX) (Chen et al., 2009). A common deficiency of these inhibitors is their high lipophilicity, which leads to low aqueous solubility and as a result of this to poor bioavailability. For this reason, efforts have been started to reduce the lipophilicity of such cPLA2α inhibitors. Recently, a series of structurally related indole-5-carboxylic acids with low lipophilicity were reported to be potent inhibitors of human cPLA2α (Drews et al., 2010). Structure–activity relationship studies on these inhibitors were investigated by the modification of the electrophilic ketone group in the middle part of the molecule, only to find that all derivatives were less active than the lead compound (Kaptur et al., 2011). A 3D-QSAR pharmacophore model consisting of hydrogen bond acceptors, negative, and aromatic rings was developed, giving some ideas about possible interactions (Jain et al., 2013). However, these SAR and quantative structure–activity relationship (QSAR) studies failed to synthesize or design more potent cPLA2α inhibitors based on the 1-(5-carboxyindol-1-yl)propan-2-one scaffold. The necessity of developing more potent QSARs with good predictive insight into the structure requirement for cPLA2α-binding affinity is urgently needed, which can facilitate the discovery of cPLA2α inhibitors with improved biological activity. Due to its effectiveness, fastness, and high predictive power, holographic QSAR is becoming increasingly an important drug design tool that encodes fragment-based information of molecular structures. Therefore, as part of this research program aimed at discovering more potent cPLA2α inhibitors with appropriate lipophilicity, we have employed the hologram QSAR (HQSAR) method to generate predictive 2D QSAR models, which has rarely been reported before. On the basis of the established HQSAR model, we attempted to elucidate a structure–activity relationship to provide useful guidelines for the design of more potent cPLA2α inhibitors.

Experiment and computation

Datasets and molecular modeling

In vitro inhibitory activity of the 1-(5-carboxyindol-1-yl)propan-2-one inhibitors of human cPLA2α, which has been reported (Drews et al., 2010), was taken for the study (Table 1). The biological data taken from the literature as IC50 value of cPLA2α inhibition was converted to the corresponding pIC50 (−log IC50) and used as dependent variables in HQSAR analysis. The pIC50 values span a range of 3 log units, providing a broad and homogenous dataset for the HQSAR study. Taking the structural diversities and wide range of activity into account, the compounds were divided randomly into training and test set. Meanwhile, a little care was taken in the selection of test sets, so that representatives of all compounds were included for prediction. 23 out of total 28 compounds were included in the training set to derive the HQSAR model while the remaining five were used as test set to validate the external predictability of model. Molecular modeling studies were performed using the SYBYL 8.1.1 software package (Tripos, L.P., St. Louis, MO, USA) running on a HP Z600 workstation.

Table 1 Chemical structures, experimental and predicted activities, and residuals of compounds included in the training set and test set
Full size table

HQSAR analysis

HQSAR is a modern QSAR technique developed from unity hashed fingerprint concept, which employs specialized fragment fingerprints as predictive variables of biological activity (Heritage Trevor and Lowis David, 1999). Compared with other existing QSAR methods, HQSAR not only avoids the need for 3D structure, putative binding conformations, and molecular alignment in CoMFA (Cramer et al., 1988) and CoMSIA (Klebe et al., 1994), but also averts the selection and calculation or measurement of the physicochemical descriptors required by classical QSAR. HQSAR analysis involves three main steps: the generation of substructural fragments for each of the molecules in the training set; the encoding of these fragments in holograms; and correlation of the latter with the available biological data.

In HQSAR, the input molecule is broken into a series of unique structural fragments (linear, branched, and overlapping) containing user-defined minimum and maximum number of atoms. According to a predefined set of rules that encodes the frequency of occurrence of various molecular fragment types, the hashed fingerprint is obtained. Then, this hashed fingerprint is divided into strings at a fixed interval as determined by a hologram length (HL) parameter. The strings are then aligned and the sum of each column constitutes the individual component of the molecular hologram of a particular length.

A number of parameters concerning hologram generation, such as HL, fragment size, and fragment distinction, prevailingly affect the HQSAR model quality (Heritage Trevor and Lowis David, 1999). In order to derive the best HQSAR model, it is necessary to discuss the influence of various combinations of parameters on the HQSAR model. All models generated in these studies were evaluated using full cross-validated q 2, partial least squares (PLS), and leave-one-out (LOO) method.

Predictive correlation coefficient (\( r_{\text{pred}}^{2} \))

The predictive ability of the HQSAR models was expressed with predictive correlation coefficient (\( r_{\text{pred}}^{2} \)), defined as:

$$ r_{\text{pred}}^{2} = \left( {{\text{SD}} - {\text{PRESS}}} \right)/{\text{SD}} $$

where SD is the sum of squared deviations between the biological activity of the test set and the mean activity of the training set molecules and the PRESS is the sum of squared deviations between predicted and observed activity values for every molecule in the test set.

HQSAR analysis for various fragment distinction combinations

For the sake of reducing the chances of bad collisions, the defaults of the HLs are set automatically by software as several prime numbers, such as 53, 59, 61, 71, 83, 97, 151, 199, 257, 307, 353, and 401. Employing these prime numbers as HLs, several combinations of these parameters were considered using the fragment size default (4–7) as follows: A/B, A/B/C, A/B/C/H, A/B/H, A/B/DA, A/B/C/DA, A/B/H/DA, and A/B/C/H/DA. The fragment distinction parameters are described as follows: A, atoms; B, bonds; C, connections; H, hydrogen atoms; DA, donor and acceptor. Due to the lack of chiral carbon atom of all the 28 molecules, the fragment distinction of chirality was not discussed in Table 2.

Table 2 HQSAR analysis for various fragment distinction combinations on the key statistical parameters using default fragment size (4–7)
Full size table

From what has been demonstrated in Table 2, we can obviously see that the best statistical model was derived using atoms, bonds, connections, and donor and acceptor (DA) as fragment distinction with 6 being the optimum number of PLS components showing cross-validated q 2 value of 0.773 and conventional r 2 value of 0.978. As can be seen from the comparison between model 2 and model 6, the DA played an important role in ameliorating the model quality. The important role of hydrogen bond acceptor was also reflected in the 3D-QSAR pharmacophore model (Jain et al., 2013). In our study, the model 6 was indicative of the possibility that the hydrogen bond donor exerted positive impact on the inhibitory activity, which is a new discovery but needs to be confirmed in further investigation. In particular, based on the model 6 in which the DA flag was enabled, the additional selection of hydrogen flag in model 8 cannot make the model better, which is ascribed to the drastic increase in the number of fragments generated when both of these options are considered (Heritage Trevor and Lowis David, 1999).

HQSAR analysis for the influence of various fragment size

Based on the best HQSAR model generated above (model 6, Table 2), the influence of different fragment sizes on statistical parameters was further investigated and summarized in Table 3. As can be seen from Table 3, the r 2 values of all models were greater than 0.94, and the q 2 values are also satisfactory. The result shown in bold fonts in Table 3 indicated that the fragment size (3–6) led to better statistical results in comparison with other fragment sizes. Therefore, the best final HQSAR model obtained from training set with 23 compounds was established using atoms, bonds, connections, DA as fragment distinction, and 1–7 as fragment size with 5 being the optimum number of PLS components showing cross-validated q 2 value of 0.790 and conventional r 2 value of 0.961.

Table 3 HQSAR analysis for the influence of various fragment size using the best fragment distinction (A/B/C/DA)
Full size table

The evaluation of HQSAR model quality

Since the structure encoded within a 2D fingerprint is directly related to biological activity of molecules, the HQSAR model is able to predict the activity of new related molecules according to its fingerprint. In virtue of the finally optimal QSAR model showing non-cross-validated (r 2 = 0.961) and cross-validated (q 2 = 0.790) correlation coefficients, which manifested a good internally predictive power, the predicted pIC50 values of both test set and training set compounds are listed in Table 1. Furthermore, the graphic results for the experimental versus predicted activities of both training set and test set are displayed in Fig. 1. The constructed HQSAR model showed good agreement between experimental and predicted values for the test set compounds with the higher predictive correlation coefficient (\( r_{\text{pred}}^{2} = 0. 60 5 \)), which signified a high external predictability of model. As far as the satisfactory performance of this holographic QSAR is considered, the model can be used to predict the biological activity of novel compounds within this structural class.

Fig. 1
figure 1

Plot of experimental versus predicted pIC50 values of the training set and test set molecules. The training set and test set molecules are shown in black (squares) and red (triangle) spots, respectively

Full size image

Interpretation of HQSAR contribution map

A significant role of a QSAR model is not only to predict the activities of untested molecules, but also to throw light on what molecular fragments play key roles to the contribution of biological activity. The results of the HQSAR analysis can be graphically displayed as color-coded structure diagrams in which the color of each atom reflects its contribution to the molecule’s overall activity. The colors at the red end of the spectrum (red and orange) represent poor contributions, while colors at the green end (yellow, blue, and green) indicate favorable contributions. HQSAR offers a good way of accounting for the variance of molecular activity by condensing information on the structural fragment.

Using the best HQSAR model, which factored atoms, bonds, connections, and DA into fragment distinction parameters, the atomic contribution maps of 23 compounds included in the training set were generated. The individual atomic contributions maps of the most (compound 26) and least (compound 5) potent cPLA2α inhibitors, resulting from the best HQSAR model, are displayed in Fig. 2. From Fig. 2, it can be seen that the structural fragment containing the carboxylic acid moiety in 5 position of the indole ring (compound 26) was colored yellow indicating its positive contribution to its inhibitory activity, while the amide group in the same place (compound 5) was colored heavily red signifying its negative effect on the activity. This was a possible reason why compounds 2, 6, and 7 with carboxylic acid moiety on the 5 position of indole ring have higher potency than compounds 1, 4, and 5. This result was consistent with previous studies, which reinforced the importance of the carboxylic acid moiety in establishing the pharmacophore of these inhibitors (Jain et al., 2013). In consideration of the preeminence of the carboxylic acid moiety and the newly discovered role of hydrogen bond donor in our study, we deduced that the carboxylic acid group may function as hydrogen bond donor in the interaction between the inhibitors and the active site of target enzyme, which also supported the above-mentioned hypothesis in HQSAR analysis for various fragment distinction combinations.

Fig. 2
figure 2

Atomic contribution maps for compound 26 and compound 5

Full size image

In particular, it is found that one fragment of the compound 26, represented by oxygen-(4-CF3-phenyl) moiety, was indicated to be strongly related to its biological activity. Compared with the molecular structure of compound 25 and compound 27, the introduction of trifluoromethyl played an important role in improving the activity of compound 26, which was probably attributed to the electron-withdrawing effect of the trifluoromethyl. The electron-withdrawing group attached to the oxygen-(4-phenyl) may be able to strengthen inhibitory activity, which was the basis for our follow-up molecular design.

Compounds designed and activity predicted

In view of the information derived from these contribution maps, we modified the structure of compound 26 by substituting the R4 and R5 fragment with other groups such as propionyl and amide group, respectively. Nevertheless, the predicted activities resulting from the best HQSAR model herein established were far from satisfaction. Taking the dominant role of the carboxylic acid moiety and the electron-withdrawing effect into consideration, we further modified the structure of 1-(5-carboxyindol-1-yl)propan-2-one cPLA2α inhibitors. The structures of new compounds with potentially improved biological activity are displayed in Fig. 3. In terms of the best holographic QSAR model established above, the activity of the new compounds thus designed were predicted, as shown in Table 4. From the prediction results, the biological activities (pIC50) of new compounds were all greater than 2.0. These new compounds are likely to possess higher inhibitory activity but still remain to be experimentally verified.

Fig. 3
figure 3

Structures of new compounds and predicted pIC50

Full size image
Table 4 Chemical structures of designed molecules and predicted activities
Full size table

Conclusions

In summary, we successfully generated a hologram QSAR model for 1-(5-carboxyindol-1-yl)propan-2-one cPLA2α inhibitors with good statistical results. The model (N = 5) displayed significant cross-validated (q 2 = 0.790) and non-cross-validated (r 2 = 0.961) correlation coefficients. The high agreement between the experimental and predicted values for the test set compounds verified the reliability and robustness of the constructed HQSAR model, indicating a high external predictability of model. The importance of the structural fragment to the overall activity of this series was interpreted by the HQSAR contribution maps. Contribution maps showed that the carboxylic acid moiety in position 5 of the indole scaffold and the electron-withdrawing effects increased the inhibitory activity. Moreover, we have designed novel inhibitors of cPLA2α possessing better inhibitory activity. Therefore, the HQSAR model can provide guidelines for future efforts in the design of new more active cPLA2α inhibitors that are structurally related with the training set compounds.