Prediction of an MMP-1 inhibitor activity cliff using the SAR matrix approach and its experimental validation

A matrix metalloproteinase 1 (MMP-1) inhibitor activity cliff was predicted using the SAR Matrix method. Compound 4 was predicted as a highly potent activity cliff partner and found to possess 60 times higher inhibitory activity against MMP-1 than the structurally related compound 3. Furthermore, pharmacophore fitting of synthesized compounds indicated that the correctly predicted activity cliff was caused by interactions between the trifluoromethyl group at para position in compound 4 and residue ARG214 of MMP-1.

www.nature.com/scientificreports/ cliff prediction. To construct SARMs for MMP-1 inhibitors, we first obtained 644 compounds with available K i values from ChEMBL (data set ChEMBL332). The MMP-1 inhibitor set yielded 2,697 individual SARMs that were searched for regions of SAR discontinuity as described 7 , (i.e., regions existing analogues have large potency variations). In these regions, the potency of virtual analogues was predicted using local Free-Wilson models, as illustrated in Fig. 1a. Compound environments were inpected for predictions that would yield activity cliffs formed by weakly potent known inhibitors and virtual analogues predicted to be much more portent. Only a www.nature.com/scientificreports/ limited number of such putative activity cliff constellations were identified and we concentrated on a chemically attractive example where a phenyl ring in a weakly potent inhibitor was replaced by a trifluoromethyl group in a virtual analogue, shown in Fig. 1b. Compounds 1 15,16 , 2 17 , and 3 18 were known MMP-1 inhibitors included in the ChEMBL data sets. By contrast, virtual compound 4 originated from SARMs, representing a novel combination of a core and substituent extracted from structurally distinct inhibitors, as shown in Fig. 1b. It is emphasized that this compound 4 could not have been predicted using conventional QSAR methods on the basis of compound 1, which contains a distinct core structure and substituent, or compound 2, which contains a distinct core. Moreover, the activity cliff formed by compounds 3 and 4 could not possibly be predicted on the basis of compound 3 alone because the prediction fully depended on the local SARM environment of virtual analogue 4, as also illustrated in Fig. 1b. The potency of compound 4 was predicted to be at least one order of magnitude higher than of compound 3. This prediction was particularly attractive because the potency of compound 2 was improved ~ 10-fold by a corresponding replacement of the phenyl group in compound 1 with a trifluoromethyl group. Hence, the formation of an activity cliff by compound 3 and its virtual analogue compound 4 was predicted. We also emphasize that the prediction did not depend on prior SAR knowledge or subjective intervention. Instead, the SARM approach systematically generates all matrix neighborhoods containing existing and virtual compounds that are amenable to potency predictions and automatically prioritizes activity cliffs on the basis of potency differences between existing and virtual analogs, as illustrated in Fig. 1b. The formation of activity cliffs generally is a rare event in compound data sets 4,5 and in the case of MMP1 inhibitors, large numbers of known active compounds and new virtual analogues had to be systematically evaluated to predict the formation of an activity cliff formed by compounds 3 and 4.
On the basis of the prediction, we synthesized compound 4. In addition, compounds 3′ and 4′, which were the diastereomers of 3 and 4, respectively, were also synthesized in order to investigate the effect of stereochemical differences on the activity. Compound 5 18 , in which the phenyl group of 3 was replaced by hydrogen atom, and compound 6, in which the trifluoromethyl group of 4 was substituted at meta position, were also synthesized as control compounds for comparison (Fig. 1c).
Synthesis of compounds 3-6 is summarized in Scheme 1.18 Esters 7-10 were chosen as starting materials to be treated with lithium diisopropyl amide (LDA) in THF to introduce methyl and allyl groups stepwise at α position of each ester, and the resulting allylic esters were converted to the corresponding aldehydes 11-14 by ozonolysis. Reductive amination of the aldehydes 11-14 with D-alanine methyl ester followed by lactamization was carried out in the presence of zinc dust in acetic acid under reflux conditions in one pot to give the corresponding γ-lactams 15-18 in 29-49% yields with a 1:1 diastereomer ratio. After the diastereomers were separated by chromatography, γ-lactams 15-18 were converted into N-hydroxyamides 3-6 using NH 2 OH and KOH in 43-98% yields.
We next examined the inhibitory activity of the synthesized compounds 3-6 against MMP-1 using a colorimetric assay. This assay was performed using the MMP-1 Inhibitor Screening Assay Kit (ab139443) according to the manufacturer's protocols. The reaction was started by the addition of the diluted MMP-1 substrate. The continuous absorbance of the wells was measured at A 412nm using a microplate reader. The results are summarized in Table 1. The IC 50 value of compound 4 (IC 50 = 0.18 ± 0.03 µM) was 60-fold lower than that of compound 3 (IC 50 = 11.5 ± 1.3 µM), hence confirming the formation of the predicted activity cliff. In contrast, both diastereomers 3′ and 4′ did not display significant inhibitory activity even at a 100 µM concentration. On the other hand, the inhibitory activity of compound 5, which had no substituent at the phenyl ring, was moderate (IC 50 = 1.54 ± 0.08 µM), indicating that the trifluoromethyl substituent is more favorable than the phenyl group of compound 3, probably due to improved steric/hydrophobic compatibility. Compound 6, which had trifluoromethyl group at meta position, exhibited similar potency (IC 50 = 11.1 ± 0.5 µM) to compound 3.
To evaluate possible binding interactions between compound 4 and MMP-1, a pharmacophore model was constructed 19 from the crystal structure of compound SC44463 in complex with MMP-1 (PDB entry 1FBL). SC44463 is a substrate-based inhibitor with a hydroxamic acid moiety, which chelates the active site zinc cation in MMP-1. Here, three pharmacophore features of SC44463 were used including a hydrogen bond acceptor (HA), hydrophobic moiety (Hy), and zinc binding location features (ZL) (Fig. 2a). The zinc-chelating and hydrophobic interactions with the S1′ pocket, which is sequence-variable within the MMP family, are related to potency and selectivity of MMP inhibitors 19,20 . Although compounds synthesized in Scheme 1 were racemic, the (R,S)enantiomer of compound 4 was superimposed on SC44463 pharmacophore model and putative interactions were refined (Fig. 2b). Table 1. MMP-1 inhibitory activity of synthesized compounds. a The compound concentration required for 50% inhibition (IC 50 ) was determined from semi-logarithmic dose-response plots, and the results represent the mean ± standard deviation of triplicated samples. www.nature.com/scientificreports/ As a result, the trifluoromethyl phenyl group of compound 4 was placed into the hydrophobic S1′ pocket where the isobutyl group of SC44463 was located. Considering this model, the biphenyl group of compound 3 was thought to be too bulky to optimally match the S1′ pocket (see Figure S1 of the supporting information). Also, compound 4 might form a halogen bonding interaction involving the trifluoromethyl group and residue ARG214 of MMP-1 (with a calculated distance of ~ 2.5 Å; see Figure S2a www.nature.com/scientificreports/ toward MMP-1 (IC 50 > 100 µM) than (R,S)-isomers 3 and 4. In fact, the carbonyl group of the γ-lactam is expected to play an important role for the MMP-3 inhibitory activity through hydrogen bond formation with the backbone of residues LEU164 and ALA165 18 . Consistent with this notion, our model suggested hydrogen bond formation between the γ-lactam of compound 4 with LEU181 and ALA182 in MMP-1 (with calculated length of ~ 2.9 and ~ 2.8 Å, respectively; see Figure S2b of the supporting information for details), whereas this hydrogen bond formation could not be observed for isomer 4′ ( Figure S1).

Conclusion
We have investigated activity cliff formation based on SARM analysis targeting MMP-1 inhibitors. Compound 4 was selected as a virtual candidate for activity cliff formation on the basis of a thorough search using SARMs. Subsequently, compound 4 and derivatives were synthesized and examined for MMP-1 inhibitory activity. The predicted compound 4 was found to exhibit 60-fold higher potency than its analogue compound 3, thereby confirming the predicted activity cliff. Retrospective pharmacophore analysis was consistent with the prediction and experimental observations, indicating a prominent interaction between the trifluoromethyl group at para position of compound 4 and ARG214 of MMP-1. Our case study and proof-of-concept investigation suggests that SARM-based analysis of compounds and associated SAR data extends the spectrum of compound design methods and enables the prediction of potent compounds and activity cliffs, representing scientifically stimulating and practically relevant applications. Further studies will aim to investigate the generalization potential and compound design perspectives suggested by this work. Specifically, our proof-of-concept study indicates that SARM-based virtual analogue populations of compound data sets can be systematically screened for predicted activity cliffs, without subjective intervention, which would not be possible using contemporary QSAR methods. Thus, prediction of activity cliffs can be attempted on a large scale for compounds with any biological activity.
To a solution of the crude material in THF (7.5 mL), x LDA 1.0 M solution in hexane/THF (1 : 2) (3.30 mL, 3.30 mmol) at − 78 °C was slowly added. After the mixture was stirred at 0 °C under argon atmosphere for 1 h, allyl bromide (433 µL, 5.12 mmol) was added. The resulting mixture was stirred at room temperature for 1 h and the reaction mixture was concentrated under reduced pressure. The reaction was quenched with brine and the mixture was extracted with EtOAc, washed with hexane, dried over sodium sulfate, and concentrated under vacuum. The crude product was used in the next step without further purification.
Ozone was pumped into a -78 °C solution of the above crude material in CH 2 Cl 2 (5 mL) until the starting material disappeared, as monitored by TLC analysis. The mixture was purged with argon. Triphenylphosphine (806 mg, 3.07 mmol) was added. After 1 h at room temperature, the mixture was concentrated under vacuum. Purification by short column chromatography on silica gel (20% EtOAc in Hexane) gave the crude product 12 which was used to the next step without further purification.
To a solution of the above product (12) and D-alanine methyl ester hydrochloride (275 mg, 1.97 mmol) in acetic acid (10.7 mL) was added portion-wise zinc powder (1.17 g, 17.9 mmol). The mixture was heated to reflux for 12 h, and then cooled to room temperature. Following addition of CH 2 Cl 2 , the mixture was filtered and the filter cake washed with methanol/ CH 2 Cl 2 . The filtrate was concentrated at 45 °C in vacuo to remove acetic acid. The residue was treated with ethyl acetate and filtered to remove insoluble materials. The filtrate was concentrated and purified by column chromatography on silica gel (40% EtOAc in hexane) to afford 179 mg of fast eluting isomer (16′), 135 mg of slow eluting isomer (16), and 80 mg of mixture containing a mixture of both isomers (total 394 mg, 1.20 mmol, 47% yield for 4 steps) as a colorless oil.  N-hydroxy-2-((S)-3-methyl-2-oxo-3-(4-(trifluoromethyl)phenyl)pyrrolidin-1-yl)

Synthesis of (R)-
Pharmacophore fitting. To predict binding interaction between compounds and MMP-1, a pharmacophore model was constructed from the crystal structure of SC44463/MMP-1 (PDB: 1FBL) using LigandScout 4.4 (InteLigand GmbH). Then, three pharmacophore features of SC44463 were used including a hydrogen bond acceptor (HA), hydrophobic (Hy) moiety, and zinc binding location feature (ZL). For pharmacophore evaluation, the scoring function was set to 'Relative Pharmacophore-Fit' . For all other parameters, default values were used. Compounds were fit to the SC44463 pharmacophore model followed by interaction energy minimization with MMP-1.