Novel phenolic inhibitors of the sarco/endoplasmic reticulum calcium ATPase: identification and characterization by quantitative structure–activity relationship modeling and virtual screening

Abstract

Inhibitors of the sarco/endoplasmic reticulum calcium ATPase (SERCA) are valuable research tools and hold promise as a new generation of anti-prostate cancer agents. Based on previously determined potencies of phenolic SERCA inhibitors, we created quantitative structure–activity relationship (QSAR) models using three independent development strategies. The obtained QSAR models facilitated virtual screens of several commercial compound collections for novel inhibitors. Sixteen compounds were subsequently evaluated in SERCA activity inhibition assays and 11 showed detectable potencies in the micro- to millimolar range. The experimental results were then incorporated into a comprehensive master QSAR model, whose physical interpretation by partial least squares analysis revealed that properly positioned substituents at the central phenyl ring capable of forming hydrogen bonds and of undergoing hydrophobic interactions were prerequisites for effective SERCA inhibition. The established SAR was in good agreement with findings from previous structural studies, even though it was obtained independently using standard QSAR methodologies.

• Calcium homeostasis
• calcium ion transport
• inhibitory potency
• in silico screening
• P-type ATPase
• partial least squares analysis

Introduction

The transmembrane enzyme sarco/endoplasmic reticulum calcium ATPase (SERCA) plays a central role in calcium homeostasis by utilizing the energy gained from ATP hydrolysis to transport calcium ions into intracellular storage compartments1^,2. SERCA can be inhibited by a variety of structurally diverse compounds, such as the natural product thapsigargin (TG), the fungal metabolite cyclopiazonic acid (CPA), the small molecule 2,5-di-tertiary-butylhydroquinone (BHQ), and several others3^,4. The ability to inhibit SERCA with high specificity and potency has been highly beneficial for studies of the enzyme’s physiological functions5. Moreover, SERCA inhibition is lethal for cells since it elevates intracellular calcium levels, a condition that ultimately triggers apoptosis. This effect has been elegantly exploited for the development of new TG-based anti-cancer pro-drugs that target SERCA in slowly proliferating prostate cancer cells6^,7. Specificity for cancerous cells was obtained by tethering a small deactivating peptide to the inhibitor, which can only be removed by a protease exclusively present on the surface of the prostate cancer cells. In cell culture studies and murine xenograft models, it was shown that a SERCA inhibitor prodrug could arrest the growth of prostate cancer cells while sparing healthy cells.

For the discovery of new SERCA inhibitors, high-resolution X-ray crystal structures of SERCA in complex with the abovementioned TG, CPA, and BHQ have been instrumental8–10. By providing a detailed account of the intermolecular interactions responsible for small-molecule inhibitor binding to SERCA, these studies have helped to define the structural requirements for small molecules to effectively interfere with SERCA function. In principle, this type of structural information can be the basis for virtual screens of large compound repositories, as demonstrated by one of our previous studies that used docking-based virtual screens to identify 19 hydroquinone derivatives capable of inhibiting SERCA at concentrations less than 50 µM11. Whereas docking-based virtual screens are undoubtedly of value, they also suffer from certain limitations, most of which relate to the proper prediction of changes in protein conformation upon binding of different ligands12. As a result, docking generally performs best in cases in which the screened molecules bear a significant resemblance to the original inhibitor present in the crystal structure, but may fail in the case of structurally more remote ligands. Moreover, the reliability of a particular docking algorithm and its scoring function may depend on the nature of the particular ligand/receptor system under investigation12.

In order to circumvent some of the limitations faced by structure-based screening methods, we have recently developed recursive partitioning protocols that utilize traditional two-dimensional molecular descriptors in combination with a sorting algorithm to classify molecules according to their predicted inhibitory potencies. Application of these protocols for the virtual filtering of a sizeable compound collection followed by experimental testing resulted in the discovery of seventeen new SERCA inhibitors with potencies below 100 µM13.

Another option for virtual screens is the use of classical quantitative structure–activity relationship (QSAR) modeling, which employs statistical methods to link bioactivities to chemical structure expressed in terms of molecular descriptors. QSAR models are powerful tools whose value has been proven in a vast number of drug design projects, both in industry and in academia14^,15. A particular strength of QSAR models is their independence from three-dimensional receptor structures, which are often unavailable or whose computational handling can be time consuming and highly demanding on resources. In contrast, QSAR models – particularly when developed with conformation-independent 2D descriptors – are fast and lend themselves to rapid screens of compound collections containing millions of entries.

The main objective of the present study was the identification of novel SERCA inhibitors via QSAR-based virtual screening of compound databases, followed by experimental testing. For this purpose, we developed several independent QSAR models using the potencies of phenol- and hydroquinone-based SERCA inhibitors previously published by our group11^,13^,16. Unlike analogs of the much larger TG and CPA, phenol or hydroquinone derivatives are abundant in many compound collections or can be obtained readily by synthesis17, which is a considerable advantage for a drug discovery project. In order to evaluate and potentially integrate different modeling approaches, we pursued three different QSAR model development strategies, each one performed independently by a different researcher. The programs employed for descriptor calculation and QSAR modeling encompassed the QuaSAR module of the modeling suite MOE, winMolconn/C-SAR, and Dragon/MobyDigs. Each model was subsequently used to screen a focused 2640 compound collection that had been obtained by filtering a 1.5 million compound repository previously assembled from several commercial vendor databases. Sixteen compounds were selected for experimental evaluation in SERCA inhibition assays. Incorporation of the experimental results into a comprehensive master QSAR model, followed by a physical interpretation of the model facilitated the identification of structural properties crucial for SERCA inhibition. At the present time, this study represents the first instance of a systematic application of QSAR modeling techniques for the discovery of novel SERCA inhibitors.

Methods

Molecular modeling of compound structures and assembly of the screening library

A set of training molecules for QSAR model development comprised of 54 phenol- or hydroquinone-based SERCA inhibitors was compiled (Supplementary materials, Figure S1). The potencies of all 54 inhibitors had been previously determined in the same type of activity assay that was used in the present study11^,13^,16. The structures of these inhibitors were modeled in Spartan (version 1.0.1; Wavefunction, Inc., Irvine, CA) and conformational energies were minimized in semi-empirical quantum mechanical calculations by the AM1 method18. All structures were stored in mol2 format for further use. For QSAR modeling, potencies originally reported as IC₅₀ values (units: µM) were converted into pIC₅₀ values (−log IC₅₀), implying that potencies of compounds with IC₅₀ values greater than 1 µM were negative.

Structural models for a total of 1 449 759 molecules were obtained by combining compound libraries from the vendors Maybridge (Tintagel, UK), Chembridge (San Diego, CA), ChemDiv (San Diego, CA), and InterBioScreen (Moscow, Russia). The size of this large collection was reduced drastically to 2640 entries by the elimination of entries that did not resemble the molecules in the training set and that were thus considered unlikely to inhibit SERCA by the same mechanism of action, if at all. The filtering was carried out by only including molecules that possessed a benzene-1,4-diol substructure, a feature present in almost all training set molecules. In addition, all molecules with more than 50% similarity according to the Tanimoto index to the most active compound in the training set (1) were admitted to the screening library. Conformational energies of all structures in this focused screening library were minimized in MOE (Molecular Operating Environment, version 2010.10; Chemical Computing Group, Montreal, Canada) via self-consistent field calculations with MOPAC and PM3 basis sets.

Descriptor calculation and QSAR modeling

QSAR model development was performed by three individual researchers working independently (D.S., J.B., and S.P.) and using a different software but the same set of training and test molecules along with their modeled structures and reported activities. During the first round of model development, the set of 54 compounds was split into a group of 43 training set and 11 external test set compounds (Supplementary materials, Table S1). Three types of initial models were generated and designated models type-1 (D.S.), type-2 (J.B.), and type-3 (S.P.). These model types, some of them after some additional development into “screening models” (see below), were the basis for subsequent virtual screening and selection of compounds for experimental evaluation. Finally, a comprehensive “master” model of type-1 was created based on a training set as large as possible. This was accomplished by considering all 65 compounds with known potencies, i.e. the 54 inhibitors from above plus 11 new active compounds evaluated in inhibition assays.

Model type-1

Models with eight variables were developed in C-SAR (CAChe C-SAR WorkSystem Pro 7.1, Build 3340; Fujitsu, Tokyo, Japan), using conformation-independent molecular descriptors computed with winMolconn (Hall Associates, Version 1.1.2.1, New York, NY). The full complement of 1650 descriptors was calculated for all the structures in the training and test set. The descriptors were then filtered to remove redundant and invariant ones, leaving a final set of 243. Model equations based on eight descriptors were obtained applying both simulated annealing and genetic algorithm methods, using degree of fit (r²) and leave-one-out cross-validation (q²) methods as the initial selection criteria. Partial least squares (PLS) regression and robust regression analysis were employed to evaluate candidate models for over-fitting and to detect influential observations. For database screening, the initial model was converted into a slightly more comprehensive screening model by the inclusion of all 54 inhibitors, except for six compounds that had relatively large discrepancies between predicted and observed potencies (6, 17, 29, 37, 38, and 48) (Tables S1 and S2). The resultant screening QSAR model whose equation still utilized eight descriptors was then used to predict the activities of the molecules in the screening database.

Model type-2

Models were generated using AutoQuaSAR in MOE, which implements the QuaSAR Evolution genetic algorithm. Both scripts were downloaded from the SVL Exchange (http://svl.chemcomp.com/). A total of 567 descriptors were calculated in MOE. At its default settings, the genetic algorithm was executed for the generation of a series of models based on eight descriptors, which were ranked by the lack-of-fit function (LOF) and cross-validated by the leave-one-out method. All models generated from five independent runs were pooled and ranked by LOF and the top 20 scoring models were written to a MOE database and used to evaluate the test set molecules. Only the model with the highest predictive value (as measured by r²) was retained and then expanded into a screening model by the inclusion of all 11 test set compounds (Tables S1 and S3). The descriptors that were part of this expanded model were the same as the ones used in the initial model. The screening model was then used to predict the activities of the molecules in the screening database.

Model type-3

For the calculation of molecular descriptors, the structures of the 54 inhibitors were imported into the program Dragon (version 5.4, Talete, Milan, Italy). All 1664 descriptors were calculated for each molecule. Descriptors that were constant (283), nearly constant (38), or correlated with each other (>95%, 891) across the data set were excluded, leaving 452 for further analysis. The descriptor values along with inhibitory potencies were then imported into MobyDigs (version 1.1, Talete, Milan, Italy). The genetic algorithm in MobyDigs was executed at its default settings, using RQK as the fitness function and allowing a maximum of eight variables (descriptors). The run was stopped after the fitness function ceased to improve further. For validation, Y-scrambling and boot-strapping were conducted at the default settings. Unlike models of type-1 and type-2, this model was not expanded further and used directly for the prediction of the activities of the compounds in the screening database (Table S1).

Experimental determination of inhibitory potencies

Rabbit hind leg tissue was obtained from Pel-Freez (Rogers, AR) and used in a standard procedure for the preparation of SERCA microsomes that are essentially free of other ATPases16^,19. The auxiliary enzymes pyruvate kinase (PK) and lactate dehydrogenase20 were received from Sigma-Aldrich (St. Louis, MO) whereas all other reagents were from Fisher Scientific (Pittsburgh, PA). Most potential inhibitors were obtained for testing from the original vendors ChemBridge, Maybridge, and InterBioScreen, unless they were available at lower cost from Sigma-Aldrich or Molport (Riga, Latvia). All compounds were used without further purification.

According to a previously described protocol, the SERCA-catalyzed rate of ATP hydrolysis coupled to the oxidation of NADH by the action of the auxiliary enzymes PK and LDH was measured spectroscopically with a plate reader (SpectraMax 190, Molecular Devices, Sunnyvale, CA) for 5 min at a wavelength of 340 nm in 96 well plastic plates11. ATP hydrolysis was initiated by the addition of 5 mM ATP to each sample (total volume: 225 µL), containing SERCA (6.7 µg protein/mL) in assay buffer (5 mM MgCl₂, 100 mM KCl, 0.7 mM CaCl₂, 0.5 mM EGTA, 3.3 µM calcimycin, 1.5 mM phosphoenolpyruvate, and 20 mM Tris; pH 7.3). Reaction rates were obtained by linear regression and measured at varying inhibitor concentrations for several minutes. Rates of NADH oxidation were fit to a three-parameter logistic equation21 and inhibitory potencies expressed in terms of IC₅₀ values, the inhibitor concentration required to reduce SERCA activity by 50%.

Results

QSAR models for virtual screens of compound libraries

Fifty-four hydroquinone derivatives with previously reported inhibitory potencies for SERCA were the basis for the independent development of three types of QSAR models (types 1–3, see “Methods” section) using the software packages MOE (QuaSAR module), winMolconn/C-SAR, and Dragon/MobyDigs. With the exception of a central phenol moiety present in each molecule, these compounds varied considerably in nature and size of functional groups (Supplementary materials, Figure S1). All inhibitory potencies had been obtained by our group using the same method, a coupled ATPase activity assay, which ensured consistency within the dataset. This aspect is important because different types of assays can generate variations in the reported activity for a compound, particularly if potencies are reported as IC₅₀ values, which, unlike inhibitor dissociation constants, are relative quantities.

Three different model development strategies were pursued starting with the same set of 54 inhibitors that was split into two pools of 43 training and 11 external test set molecules. The latter were unknown to the initial models and therefore served as a benchmark for their ability to predict potencies of external molecules. As shown in Figure 1, all three model types were capable of properly accounting for the variability of potency within the training set, which encompassed slightly more than two orders of magnitude (pIC₅₀ range: 2.35). Comparisons of statistical parameters indicative of model quality did not reveal any drastic differences between the model types (Tables 1–3). Moreover, the ability of the three models to predict the potencies of the 11 external test compounds was comparable (Figure 1).

Figure 1. Correlation between experimentally determined and predicted inhibitory potencies (pIC₅₀) for initial models type-1 (left), type-2 (middle), and type-3 (right). External test set compounds are shown as empty circles.

Table 1. Descriptors and regression parameters for initial QSAR model type-1.

Variable	Regression coefficient	Error regression coefficient
y-Intercept	−0.4857	0.1587
SHBint3	−0.02555	0.00575
xp6	−1.3081	0.1406
xp7	0.6846	0.1463
xp8	1.1357	0.2911
dxp5	1.2460	0.1614
n3pag23	−0.14825	0.01676
n4pag23	−0.08370	0.01715
n4pag33	0.09731	0.02737

n = 43, r²= 0.815, q²= 0.719, s = 0.265, LOF = 0.103, PRESS = 3.61, F-value = 18.68.

Table 2. Descriptors and regression parameters for initial QSAR model type-2.

Variable	Regression coefficient	Error regression coefficient
y-Intercept	4.241	1.470
PM3_HOMO	0.792	0.167
vsurf_DW12	0.027	0.009
vsurf_HB6	−0.032	0.011
vsurf_Wp7	7.373	1.152
Smallest_Pos_Hardness	8.187	2.473
RDF060v	−0.069	0.014
TRDF010p	−0.025	0.016
TRDF050m	0.052	0.007

n = 43, r²= 0.827, q²= 0.743, s = 0.255, LOF = 0.131, PRESS = not determined, F-value = 20.3.

Table 3. Descriptors and regression parameters for initial QSAR model type-3.

Variable	Regression coefficient	Error regression coefficient
y-Intercept	−4.36	0.31
MATS5e	0.801	0.204
Mor21m	−0.807	0.214
Mor30v	1.17	0.33
Gu	3.94	0.38
Dm	3.74	0.56
(RDF030m)^2	−0.00246	0.0068
(G1u2)^2	−1.46	0.22
(Du)^2	3.46	0.99

n = 43, r²= 0.863, q²= 0.808, s = 0.227, LOF = 0.103, PRESS = 2.456, F-value = 26.9.

Virtual screens of compound libraries, compound selection, and experimental testing

Before predicting the potencies of the 2640 compounds contained in the focused screening library, initial models of types 1 and 2 were slightly expanded at the discretion of the individual modeler into “screening” models by including some (model type-1) or all (model type-2) of the 11 test set compounds into the training set. These expansions were done without altering the general nature of the models by retaining the protocol for model development as well as the number and type of descriptors (see “Methods” section and Tables S2 and S3). The initial type-3 model was used directly for virtual screening without any alterations. The model-based predictions were then factored in the selection of compounds for experimental evaluation (T01–T16), which was also influenced by other considerations. Since the major objective was the identification of novel SERCA inhibitors with good potencies, a compound’s potency needed to be predicted by at least one of the three screening models in the low micromolar concentration range (pIC₅₀ > −0.50). Furthermore, structural diversity was a criterion because we wished to avoid focusing on a particular subset of structurally similar molecules. Therefore, the test pool encompassed molecules of varying structural complexity, ranging from the rather small compounds T04–T07 to sizeable molecules, such as T14–16. Bisphenols T01 and T02 were included because previous studies have identified several representatives from this compound class as potent SERCA inhibitors22^,23. Finally, cost and availability were also factors that influenced the selection.

The 16 compounds shown in Figure 2 were obtained for testing in coupled ATPase activity assays. Figure 3 illustrates the good quality of the performed assays by displaying three representative curves for a high (T01), medium (T08), and low potency (T03) inhibitor. Overall, 11 of the 16 tested compounds showed measurable inhibitory effects, with average IC₅₀ values ranging from 1 to 500 µM. All test results are presented in Table 4 and analyzed in detail in the “Discussion” section.

Figure 2. Chemical structure of the 16 compounds selected for testing.

Figure 3. Representative inhibition curves obtained for a high (•, T01), medium (▪, T08), and low (♦, T03) potency inhibitor.

Table 4. Experimentally determined and screening model-predicted inhibitory potencies of compounds selected for testing.

Compound no.	IC50 (µM)	pIC50	pIC50 (type-1)	pIC50 (type-2)	pIC50 (type-3)
T01	0.94 ± 0.62	0.03	−0.01	−0.14	−1.83
T02	1.01 ± 0.50	0.00	−0.15	0.19	−1.90
T03	6.94 ± 1.63	−0.84	−0.43	−0.81	−0.94
T04	17.9 ± 4.0	−1.25	−0.51	0.38	0.23
T05	33.3 ± 2.5	−1.52	0.10	0.79	0.32
T06	38.8 ± 4.9	−1.59	0.03	0.32	0.59
T07	53.0 ± 7.0	−1.72	0.03	−1.47	−1.46
T08	112 ± 9	−2.05	−0.45	−0.08	−0.40
T09	185 ± 46	−2.27	−0.84	−0.72	0.28
T10	409 ± 52	−2.61	0.14	−0.34	0.99
T11	419 ± 107	−2.62	−0.73	−0.52	0.21
T12	Inactive	–	−1.03	−1.45	1.54
T13	Inactive	–	0.12	−0.79	−1.32
T14	Inactive	–	1.56	−0.05	−0.68
T15	Inactive	–	0.08	0.58	−1.24
T16	Inactive	–	0.23	1.07	−1.67

Development of a comprehensive master QSAR model

In addition to being a crucial component of the virtual screening procedure, the descriptors that are part of a QSAR model can help to identify structural features of compounds that make them bioactive. For this purpose, a master model containing a maximum amount of information was developed that considered all 65 active SERCA inhibitors in its training set. For the sake of model quality, however, 13 compounds were eliminated that either had poorly predicted activities or represented overly influential data points that hindered the derivation of a robust model for the remainder of the training set. The master model was of type-1 since this model type offered the distinct advantage of only requiring conformation-independent descriptors. In comparison to all other models, the quality of the master model was somewhat superior as evident from the fact that it was able to account for the variation in activity of a larger pool of compounds (52 versus 43 in the initial models) using a smaller number of descriptors (seven versus eight) without a decline in parameters indicative of statistical quality, such as r² or q². Predicted versus observed potencies for the final model are shown in Figure 4 and Table S1 and statistical parameters are listed in Table 5.

Figure 4. Correlation between experimentally determined and model-predicted inhibitory potencies (pIC₅₀) for the final “master” model.

Table 5. Descriptors and regression parameters for the master QSAR model.

Variable	Regression coefficient	Error regression coefficient
y-Intercept	−1.85	0.128
SHCsatu	−0.242	0.0490
SHHBd	0.225	0.0348
dxp6	−1.02	0.112
dxvp4	−0.387	0.0939
dxvp7	1.94	0.348
np7	0.0218	0.00345
n3pag33	0.142	0.0371

n = 52, r²= 0.802, q²= 0.743, s = 0.272, LOF > 0.1, PRESS = 4.14, F-value: 25.52.

Discussion

Qualitative establishment of SARs

Even without the aid of QSAR models, several SAR trends can be inferred from a qualitative inspection of chemical structures and potencies of the 16 experimentally tested compounds (Figure 2 and Table 4). First, there appears to be a size limit for inhibitors, most likely dictated by the finite volume of the inhibitor binding pocket. This conception stems from on the observation that most of the larger compounds are inactive (T13–T16) or have low potencies (T08–T11), presumably as the direct result of inability or difficulty to fit into the binding site. In contrast, some of the smaller inhibitors (T04–T07) show good potencies (IC₅₀: 15–60 µM), with the exception of T12 which might be too small to effectively interact with the enzyme.

Many inhibitors with potencies below 100 µM carry at their phenol scaffold at least one bulky non-polar substituent such as alkyl groups, halides, or both (T03–T07), suggesting an important role of hydrophobic contacts for binding. The two most active inhibitors, however, are T01 and T02, both of which are bisphenols whose two methylated phenyl rings are connected by a methylene bridge carrying either a bulky isobutyl (T02) or a cyclohexyl substituent (T01). The high potency of these molecules is consistent with a recent study on the ability of certain bisphenols to inhibit SERCA with remarkable potencies22.

A medium-sized compound with a somewhat unique structure and a relatively high potency (IC₅₀ = 6.9 µM) is the naphthoquinone-based T03. Previously, certain naphthoquinones have been described as SERCA inhibitors24–26. In comparison, the somewhat larger anthracene T09 whose functional groups are identical to those of T03 has a significantly reduced potency, which is most likely due to T09 reaching the size limit of the binding site.

Physical interpretation of the descriptors employed by the comprehensive master QSAR model

As noted under “Results” section, 13 compounds from the pool of 65 were excluded during the development of the master model, most of them because of significant disagreements between predicted and observed potencies (Table S1). According to solubility data obtained from Chemical Abstracts SciFinder (computed by ACD, version 11.02, Columbus, OH), many of these outliers were only sparingly soluble in water, which may offer a plausible explanation for the molecules’ behavior. In addition to causing solubility issues in the activity assay at higher inhibitor concentrations, a significant fraction of these non-polar compounds could be absorbed by the hydrophobic interior of microsomal membranes present in the assay buffer, thereby lowering the effective concentration available to the enzyme and ultimately impairing the accuracy of the potency measurements.

For a rigorous physical interpretation of the master QSAR model, a previously published method based on PLS analysis was employed27^,28. PLS analysis shows that the first four components account for 96.5% of the SAR information contained in the model. Therefore, the descriptors dominating these four components were inspected further in order to link structural features captured by these descriptors to inhibitory potency.

The first component accounts for 59.6% of the SAR and is dominated by three descriptors (Table 6). The most important one is SHHBd (sum of hydrogen-atom electrotopological state indices for hydrogen-bond donor atoms) which provides 46.5% of the information contained in component-1, followed by dxp6 (sixth-order path difference molecular connectivity index, 23.6% contribution to component-1) and SHCsatu (sum of electrotopological indices for hydrogen atoms on a methyl group attached to an aromatic carbon atom, 15.1% contribution to component-1). SHHBd carries a positive weight and reflects the fact that compounds with two hydrogen-bond donors – such as the hydroquinones with two hydroxyl groups – frequently have higher potencies than those with a single hydrogen-bond donor. The two other descriptors carry a negative weight and correct potencies overestimated by SHHBd. They accomplish this by favoring compounds that have a high degree of branching close to the phenyl ring (reducing the numerical values for both dxp6 and SHCsatu). Molecules illustrating the features captured by component-1 are the active 3, 4, and 49 and the less active 8, 12, 13, and 16. Compound 34 illustrates an interesting departure from the benzene-1,4-diol (hydroquinone) class. In this case, the hydroxyl groups are on separate rings. However, conformational analysis and geometry optimization by semi-empirical quantum mechanics (AM1) provided evidence that the hydroxyl groups for 34 occupy nearly the same relative positions in space as the hydroxyl groups on compound 3 (Figure S2), suggesting that molecules that mimic the geometry of the hydroquinones can exert significant potency.

Table 6. Results of PLS analysis of the master QSAR model and descriptors making significant contributions to the first four PLS components.

Descriptor	x-Weight comp-1	x-Weight comp-2	x-Weight comp-3	x-Weight comp-4
SHCsatu	−0.388	−0.370	0.275	0.240
SHHBd	0.682	−0.353	−0.096	−0.330
dxp6	−0.486	−0.313	0.208	−0.595
dxvp4	−0.136	0.336	−0.196	−0.670
dxvp7	0.007	0.707	0.188	0.000
np7	0.093	0.145	0.734	−0.022
n3pag33	0.348	−0.079	0.510	−0.174

The most influential descriptors are shown in boldface.

Component-2 codes for an additional 14.8% of the SAR and is dominated by the descriptors dxvp7 (seventh-order valence-corrected path difference molecular connectivity index), and – as in component-1 – by a combination of SHCsatu and SHHBd. The descriptor dxvp7 has a negative weight and reduces the predicted activities of compounds that are large and structurally complex (22 and 40, for instance). In component-3, which contributes another 10.2% to the overall SAR, the descriptors np7 (count of occurrence of a seventh-order path) and n3pag33 (count of occurrence of tertiary carbon atoms separated by three bonds) are weighted heavily (53.8% and 26.0% of component-3, respectively). Both descriptors carry positive weights and compounds 19 and 53 are examples for active molecules that take high values for these descriptors. Finally, component-4 adds 12.0% to the SAR, bringing the cumulative total for the first four components to 96.5%. This component is dominated by two negatively weighted descriptors, dxvp4 (fourth-order valence-corrected path difference molecular connectivity index, 44.9% of component-4) and dxp6 (see component-1, 35.4% of component-4). Component-4 corrects the activities of compounds that are overestimated otherwise. For example, 3 (high potency) and 9 (lower potency) have very similar groups attached to their central phenyl ring, but possess a different substitution pattern. This difference is reflected in the values for dxvp4 computed for the two compounds and therefore taken into account during the activity prediction.

In summary, the PLS analysis reveals that hydrogen bonding and hydrophobic interactions are crucial factors of inhibitory potency against SERCA. More specifically, both structural elements need to be properly positioned at the phenyl scaffold for maximum potency. Two hydrogen-bond donors in para position and the presence of hydrophobic substituents on opposite sides of the phenyl ring that are moderately large, space-filling, but not too rigid appear ideal. It is remarkable that this SAR solely derived from PLS analysis of a QSAR model is in excellent agreement with the findings of structure-based studies8^,16^,17, thereby illustrating the value of QSAR modeling in enzyme inhibitor development projects. In particular, the QSAR-derived SAR information was obtained readily, relying exclusively on standard bioassays and well-established modeling and statistical methods without requiring high resolution X-ray crystal structures of the inhibitor/receptor complexes, which – unlike experimentally measured inhibitory potencies – can be difficult to obtain.

Conclusion

In conclusion, the present study is the first instance of the application of rigorous QSAR methodology in the search for new SERCA inhibitors. We were able to discover 11 active, not previously tested compounds, some of which had good potencies in the low micromolar concentration range. Furthermore, PLS analysis of our master model showed that the most important structural requirements for effective SERCA inhibition were hydrogen bond donor groups and compact hydrophobic substituents, both located at opposite sides of the central phenyl ring. Based on these findings, future efforts will be directed at the study of compounds that possess these structural features. As stated above, SERCA inhibitors hold promise as a novel class of pro-drugs against prostate cancer, as already shown for TG-derivatives tethered to a short peptide. In principle, the established strategy of conveying specificity for prostate cancer cells by peptide attachment should be applicable to any type of SERCA inhibitor, including hydroquinone-based compounds.

Supplementary material available online

Supplementary Tables S1–S3 Supplementary Figures S1 and S2

Declaration of interest

The authors report no declaration of interest.

This work was supported by grants from the Kentucky Biomedical Research Infrastructure Network (P20GM103436-12) and the National Institutes of Health (1R15GM084431-01 and 2R15GM084431-02) to S.P. Moreover, J.B. is grateful to the School of Natural Science at Spalding University for the purchase of some of the test compounds and for covering software license fees.