Scaffold-Hopping And Optimization Of Small Molecule Soluble Adenyl Cyclase Inhibitors Led By Free-Energy Perturbation

Abstract

Free energy perturbation is a computational technique that can be used to predict how small changes to an inhibitor structure will affect the binding free energy to its target. In this paper we describe the utility of free energy perturbation with FEP+ in the hit-to-lead stage of a drug discovery project targeting soluble adenyl cyclase. The project was structurally enabled by X-ray crystallography throughout. We employed free energy perturbation to first scaffold hop to a preferable chemotype and then optimized the binding affinity to sub-nanomolar levels whilst retaining druglike properties. The results illustrate that effective use of free energy perturbation can enable a drug discovery campaign to progress rapidly from hit-to-lead, facilitating proof of concept studies that inform on target validation.

Introduction

Optimization of protein-ligand binding free energy is typically one of the key goals in the hit-to-lead process of drug discovery. For this reason, accurate prediction of protein-ligand binding free energy has long been a goal of computational chemistry in order to support structure-based drug design. Physics-based free energy calculations, such as free energy perturbation (FEP), are well-precedented and are generally considered to be the most rigorous approach to this problem ¹. Through time, there has been tremendous progress in sampling algorithms ², force-field development ³ and low-cost parallel computing. Accordingly, FEP simulation time and accuracy have improved significantly ^4–6. As a result, pharmaceutical companies have widely adopted relative binding free energy calculations, leading to successful prospective applications of FEP in industrial settings ⁷. In this study, we applied the FEP software FEP+ to the hit-to-lead stage of a drug discovery project targeting soluble adenylyl cyclase (sAC, ADCY10).

Since its initial discovery over 70 years ago, cyclic adenosine monophosphate (cAMP) has emerged as a key second messenger in a wide variety of cellular processes, such as apoptosis, differentiation, and proliferation ⁸. In a cell, cAMP signaling is compartmentalized into independently-regulated microdomains which prevent unintended interactions between these different cellular processes ⁹. Inside of these microdomains, levels of cAMP are controlled by adenylyl cyclases (ACs), which synthesize cAMP, and phosphodiesterase (PDEs), which degrade cAMP ^{10, 11}. Due to the relevance of cAMP signaling in numerous diseases, ACs and PDEs have been identified as promising drug targets. For example, PDE modulators have been established as an important class of therapeutics. In contrast, no AC regulators have been successfully advanced into clinical settings ¹².

In mammalian cells, the 10 known AC isoforms can be divided into two classes: the G-protein regulated transmembrane ACs (tmACs: ADCY1–9) and the bicarbonate-regulated soluble AC (sAC:ADCY10) ¹³. sAC, a non-membrane bound protein, is found within the cytosol and in several microdomains located inside organelles ^{11, 14}. In biological systems, CO₂ is maintained in a pH-dependent equilibrium with bicarbonate (HCO₃^-) and, as a result, sAC acts as Nature’s physiological CO₂/HCO₃^-/pH sensor ¹⁵. cAMP, produced by sAC in response to changes in CO₂/HCO₃^-/pH, has been linked to multiple biological functions, such as insulin-release from pancreatic β-cells ¹⁶, lysosomal acidification ¹⁷, regulation of intraocular eye pressure ¹⁸, in addition to controlling the motility and capacitation of sperm ¹⁹. In regulating these biological functions, sAC is a promising drug target, with sAC inhibitors having several potential therapeutic applications ^{20, 21}. This group has recently disclosed results describing biological application of a potent sAC inhibitor ^21–23.

Mammalian sAC features an ~470 residues Class III AC catalytic core (sAC-cat) at its N-terminus, connected to ~1100 residues with putative functions in regulation and localization ²⁴. Human sAC-cat structures confirmed that it features the generic Class III overall core structure and catalytic residues despite its unique regulation ^{24, 25}. The core comprises two structurally similar catalytic domains, C1 and C2, in a pseudo-heterodimeric head-to-tail association (Figure 1A). C1 and C2 contribute complementary catalytic residues to the active site at the heterodimer interface, and another set of residues to the pseudo-symmetric site, which is catalytically incompetent and serves a regulatory function. In the active site, two conserved Asp residues (47 and 99) coordinate two Mg²⁺ ions, ion A and B, which bind the ATP phosphates to support substrate turnover and binding, respectively ²⁴. sAC displays an unusually low apparent ATP affinity (K_M ~10 mM), which can be increased by Ca²⁺ acting as ion B (K_M ~1 mM) ^{26, 27}.

Figure 1 –

Overlaid crystal structures of sAC bound to bicarbonate from PDBID 4CLL (pink) and 1 from PDBID 5IV4 (grey). Structures of the (a) global protein structure and (b) bicarbonate binding site are shown. Proteins are displayed as cartoon ribbons with wire atoms and ligands are shown as balls and sticks.

The unique feature of sAC enzymes is the activation by bicarbonate, which accelerates substrate turnover ^{24, 27}. Structures of hsAC-cat/ligand complexes revealed that the degenerate, pseudo-symmetric active site serves as bicarbonate binding site (BBS) and identified local rearrangements contributing to activation ^{24, 25}. The BBS residues Lys95 and Arg176 bind bicarbonate like tweezers (Figure 1B). For this binding mode, Arg176 leaves an inhibitory interaction with the conserved active site Asp99 and flips into the BBS, enabling rearrangements yielding a catalytically competent active site ^{24, 25}. Importantly, the pseudo-symmetric site differs largely from that of tmACs, which accommodates the pharmacological tmAC activator forskolin and is speculated to serve as binding site for a so far unknown physiological tmAC regulator ²⁴. The unique features of the sAC BBS render it an attractive site for specific modulation and indeed contributes to accommodation of several sAC inhibitors ²⁴. While the bicarbonate transporter blocker DIDS (4,4′-diisothiocyanatostilbene-2,2′-disulfonic acid) was surprisingly found to block the active site entrance rather than the BBS ²⁵, the sAC inhibitors ASI-8 and bithionol indeed exploit the BBS ^{28, 29}. Bithionol occupies the mostly hydrophobic BBS access channel for a mixed-type inhibition and positions a chlorine in the bicarbonate pocket ²⁸. It acts as an inhibitor (K_i 2.3 mM) and is inactive against tmACs, but shows effects on many other targets. ASI-8, which was developed through a structure-based fragment screen ²⁹ inhibitor (IC₅₀ inhibitor (IC ~0.4 mM) at low ATP concentrations in vitro, but no data on compound specificity or cellular effects have been reported. ~0.4 mM) at low ATP concentrations in vitro, but no data on compound specificity or cellular effects have been reported. In contrast, plenty of biological data is available for KH7, (IC₅₀ ~3 mM). It is remotely related to ASI-8 and might have a similar binding mode. However, KH7 has off-target effects which make it less attractive for drug development ³⁰. In a more recent high-throughput screen, the chemically distinct compound LRE1 was identified as a sAC-specific inhibitor (IC₅₀ ~1 mM) ³⁰ inhibitor in various assays and biological systems, and crystal structures of sAC/LRE1 complexes revealed that the compound’s 2-amino-6-chloropyrimidine occupies the BBS (Figure 1B/LRE1 complex revealed that the compound’s 2-amino-6-chloropyrimidine occupies the BBS). LRE1’s small cyclopropane moiety reaches into the channel connecting BBS and active site but does not overlap with ATP. Consistently, LRE1 is competitive with bicarbonate but non-competitive with substrate, rendering it a fully allosteric BBS-targeting. LRE1 binding widens the BBS through rearrangements of Lys95 and Arg176, and the pocket accommodating the thiophene moiety is created mostly through rearrangements of Arg412 and Phe338 (Figure 1B). These reorientations lead to ATP binding in a catalytically incompetent conformation and illustrate the plasticity of this site.

LRE1 has multiple deficiencies, including modest intrinsic potency, poor pharmacokinetic characteristics, and the presence of a thiophene moiety that raises safety concerns ³¹. Therefore, a hit-to-lead campaign was undertaken to improve intrinsic potency, engineer in “drug-like” properties to enhance cellular activity and yield a preclinical proof-of-concept sAC inhibitor as described previously ^{22, 32}. Here we illustrate how this process was driven by the extensive use of FEP calculations. To assess the effectiveness, FEP prediction accuracy is compared with standard Glide SP docking and MM-GBSA approaches.

Materials and Methods

Protein Structure Preparation

A previously published protein structure was taken from the protein databank entry with PDB ID 5IV4 ³⁰. In all cases, selenomethionines were changed to methionines and missing side-chains were added using Schrödinger’s Preparation Wizard, which was also used to evaluate the orientations of the asparagine, glutamine, and histidine residues, as well as the protonation state of all ionizable residues ³³. All heteroatomic species such as buffer solvents and ions were removed. Water molecules outside the LRE1 binding site were retained. The structure was then minimized whilst converging the heavy-atom RMSD to 0.3 Å.

Free-Energy Perturbation Calculations

FEP calculations were performed using the FEP+ tool from Schrödinger Suite ⁵. The starting pose of each compound was generated by aligning the new ligand structures with the co-crystalized ligand using the Schrödinger ligand alignment tool. The perturbation map was either manually edited or automatically generated using the Mapper tool. In the automated set-up, ligand pairs with high similarity scores were connected by edges. The system was solvated in a water box with buffer width of 5Å for the complex simulations and 10Å for the solvent simulations. The systems were relaxed and equilibrated using the default relaxation protocol. Molecular dynamics simulations were carried out used OPLS3e force field ³⁴ for the protein and the ligands along with the SPC water model ³⁵. Force Field Builder was employed to generate custom torsional parameters for ligand torsions that were not include in the default force field. FEP+ jobs were run for between 5ns to 20ns until converged (as noted). After the simulations were completed, the hysteresis along closed thermodynamic cycles was calculated. In some cases, ligands were modelled in different protonation states. The effect of potentially having multiple different protonation states was incorporated using a pKa state correction approach ³⁶. For the purposes of prediction, experimental pIC₅₀ values were used in place of experimental pKd values.

Glide Ensemble Docking Calculations

Molecular docking studies were performed using the Schrödinger software suite (2020–2). For each of the structures, a receptor grid that can accommodate ligands with length up to 20 Å, was created at the centroid of the co-crystalized ligand. All ligands underwent preparation by LigPrep (Schrödinger, LLC, New York, NY, 2020) and were then docked into all 19 receptors with using Glide standard precision (SP) docking using the default settings ³⁷. The top scoring poses were output for MMGBSA analysis.

MM-GBSA Calculations

Compounds that were successfully docked were subjected to a rescoring process using the Molecular Mechanics with generalized Born Surface Area (MM-GBSA) method ³⁸ implemented in Prime (Schrödinger, LLC, New York, NY, 2020). The MMGBSA binding free energy was evaluated by minimization of the protein-ligand complex to address the errors generated by using rigid protein structure in docking procedures. Protein flexibility was allowed for residues up to 5 Å from the ligand.

Kendall’s Tau Calculation

Kendall’s rank correlation coefficient (τ) is calculated by using the software Kendall tau Rank Correlation (v1.0.13) in Free Statistics Software (v1.2.1) ³⁹.

Chemical Synthesis

Chemical synthesis of pyrazole compounds was completed as described previously ^{22, 32}.

Crystallography

hsAC-cat protein was expressed and purified as described previously ⁴⁰. hsAC apo crystals (space group P63) were grown at 278 K in hanging drops of 1.5 μL protein stock (10 mg/ml), 1 μL reservoir solution (0.1 M sodium acetate pH 5.0, 0.2 M trisodium citrate, 14% (w/v) PEG 4000 and 10% (v/v) glycerol), and 0.5 μL seed stock (diluted suspension of crushed crystals). For ligand soaking, crystals were transferred to a drop containing 0.1 M sodium acetate pH 5.0, 0.2 M trisodium citrate, 15% (w/v) PEG 4000, 20% (v/v) glycerol, 10 % DMSO (v/v), and 1 to 5 mM ligand. After 5 or 24 h incubation at 278 K, crystals were flash frozen in liquid nitrogen. Diffraction data were collected at 100 K at BESSY beamline 14.1 or 14.2 operated by Helmholtz-Zentrum Berlin ⁴¹ and processed with XDSapp ⁴². Structures were phased through molecular replacement with Phaser ⁴³, using hsAC apo (pdb ID 4CLL ²⁵) as a search model. Ligand files were generated with eLBOW from the PHENIX suite ⁴⁴ or with PRODRG ⁴⁵. Model building was done with Coot ⁴⁶ and refinements with PHENIX. Refinement statistics are shown in Table 1.

Table 1 –

X-ray crystallography statistics

Compound	TDI-9066 (2)	TDI-10228 (7)	TDI-10512 (22)	TDI-10962 (47)
Diffraction data collection
Resolution rangea	32.50 – 1,90 (1,97 – 1,90)	32.37 – 2.10 (2.18 – 2.10)	44.38 – 2.00 (2.07 – 2.00)	28.67 – 2.10 (2.17 – 2.10)
Space group	P63	P63	P63	P63
Unit cell [a = b, c] (Å)	99.29, 99.40	98.91, 99.75	99.14, 99.60	99.31, 98.85
Unique reflectionsa	42660 (3656)	32292 (3199)	36970 (3263)	32328 (3183)
Multiplicitya	6.4 (3.8)	8.4 (7.8)	10.4 (5.8)	8.3 (7.0)
Completeness (%)a	95.7 (68.8)	99.6 (98.5)	95.6 (64.5)	99.8 (98.6)
I/sigma(I)a	8.8 (0.4)	9.1 (0.6)	6.7 (0.3)	13.6 (0.8)
R-measa	0.168 (3.351)	0.216 (3.224)	0.255 (4.024)	0.138 (2.482)
CC1/2a	0.997 (0.13)	0.996 (0.15)	0.994 (0.09)	0.998 (0.27)
Refinement
Reflections used in refinementa	42659 (3019)	32232 (3162)	35922 (2420)	32311 (3174)
Number of TLS groups	4	3	4	3
Rwork / Rfreeb	0.191 / 0.233	0.210 / 0.233	0.203 / 0.230	0.207 / 0.244
Number of non-hydrogen atoms
protein	3793	3618	3602	3644
ligands	105	66	89	85
solvent	214	121	142	116
r.m.s. deviations
bonds (Å)	0.01	0.01	0.02	0.01
angles (°)	0.9	1.0	1.6	1.0
Average B-factor
protein	43.7	54.1	50.1	54.7
ligands	58.1	56.8	56.4	61.4
solvent	43.4	44.2	42.4	46.5

^aHighest resolution shell shown in parentheses.

^b5% of reflections were omitted from refinement for R_free calculation.

In Vitro Cyclase Activity Assay

Assays for sAC activity using purified protein were performed in 100 μL reactions containing 4 mM MgCl2, 2 mM CaCl2, 1 mM ATP, 40 mM NaHCO3, 50 mM Tris pH 7.5, and 3 mM DTT. Each reaction contained ~1,000,000 counts of α−32P labeled ATP. Generated cAMP was purified using sequential Dowex and Alumina chromatography as previously described ⁴⁷.

LogD determination

LogD7.4, which is a partition coefficient between 1-octanol and aqueous buffer pH 7.4, of the compounds was measured on the chromatographic procedure whose condition was developed based on a published method ⁴⁸.

Results

Due to the modest binding affinity of LRE1 (1473 nM) and its lack of promising growth vectors, the first stage in the hit-to-lead campaign was a scaffold hop to look for alternative chemotypes. We generated a series of ideas and predicted their binding affinity using FEP+, maintaining the aminopyrimidine headgroup which lies in the bicarbonate binding site and the thiophene ring which lies in a hydrophobic pocket. Using FEP+, we identified 2 as a compound with increased binding affinity (141 nM), improved lipophilic efficiency ⁴⁹, and more amenable growth vectors (see Table 2).

Table 2 –

FEP+ predicted and experimental sAC pIC₅₀s for the scaffold hop from 1 to 2.

Compound	1	2
pIC50 (FEP+)	5.83*	7.67
pIC50 (experimental)	5.83	6.85
logD (pH 7.4)	3.03	2.52
Lipophilic Efficiency	2.80	4.33

^*1 was used as a single FEP+ reference compound and edges were run for 20 ns.

We solved the crystal structure of 2 in complex with sAC (see Figure 2) which, consistent with prediction, shows a very high degree of overlap with the binding mode of LRE1.

Figure 2 –

Overlaid crystal structures of 1 (grey) and 2 (green). Proteins are shown as ribbons with wire atoms and ligands are shown as balls and sticks.

Our next step was to explore the chemical space around 2. We placed particular emphasis on replacement of the thiophene ring as it can undergo facile CYP mediated oxidation to electrophilic species that can readily react with nucleophiles such as glutathione. It is important to note that the majority of SAR for this step was generated without running FEP+ prospectively. Rather, we used a combination of molecular docking and simple medicinal chemistry transformations. However, a retrospective analysis is shown in Table 3, with calculated values determined that are consistent with those generated experimentally.

Table 3 –

FEP+ predicted and experimental sAC pIC₅₀s for a set of 2 analogues. All compounds were used as FEP+ reference compounds and all edges were run for 20 ns. The mean unsigned error in prediction is 0.97 kcal/mol.

Compound Structure	Compound ID	pIC50 (FEP+)	pIC50 (experimental)
	2	7.20	6.85
	3	6.39	6.26
	4	7.82	7.30
	5	6.87	6.74
	6	7.12	7.12
	7	6.51	7.00
	8	6.55	6.77
	9	3.92	6.15
	10	7.46	7.22
	11	7.07	6.92
	12	5.85	6.96
	13	5.27	6.40
	14	6.06	6.89
	15	6.41	6.80
	16	6.86	6.59
	17	6.66	6.8
	18	6.38	6.15
	19	7.68	6.62
	20	8.41	6.39
	21	9.03	7.00

Overall, the mean unsigned error for the compounds in Table 3 is 0.97 kcal/mol. However, it is worth noting that the carboxylic acid compound 9 is poorly predicted (mean unsigned error of 2.23 kcal/mol) and this skews the results. This error could be due to the known difficulties of alchemical calculations in which the net charge on the molecule is altered ⁵⁰ or a change in the protonation state of an active site residue such as Asp47 or Asp99. Other compounds are very well predicted. We were gratified that simple changes, such as thiophene to meta-fluoro phenyl (7), were well tolerated. We solved the crystal structure of 7 in complex with sAC (see Figure 3) which shows the bioisoteric replacement of a thiophene ring with a phenyl ring.

Figure 3 –

Overlaid crystal structures of 2 (green) and 7 (orange). Proteins are shown as ribbons with wire atoms and ligands are shown as balls and sticks.

After identifying the phenyl ring as a simple thiophene replacement, we focused on improving the binding affinity and selected the ortho position of the phenyl as a good growth vector based on analysis of the crystal structure. Table 4 shows the FEP+ results for a set of compounds designed and tested around 3.

Table 4 –

FEP+ predicted and experimental sAC pIC₅₀ values for a set of 3 analogues. 3 and 25 were used as the FEP+ reference compounds and edges were run for 15 ns. The mean unsigned error for non-reference compounds is 0.45 kcal/mol.

Compound ID	pIC50 (FEP+)	pIC50 (experimental)	R1	R2
22	8.34	7.66
23	7.31	7.00
24	7.01	7.24
25	6.69	6.78
26	6.49	5.92
3	6.37	6.26

Prediction and experiment are in agreement that ortho substitution is tolerated or in some cases beneficial to binding affinity. In particular, methyl ester (22) and methoxy (24) have increased binding affinity and are amenable to elaboration. We solved the crystal structure of 22 in complex with sAC (see Figure 4) which shows the ortho ester group along the designed growth vector, reinforcing the predictive values of these simulations in the design of superior analogs.

Figure 4 –

Overlaid crystal structures of 7 (orange) and 22 (magenta). Proteins are shown as ribbons with wire atoms and ligands are shown as balls and sticks.

Based on the data in Table 4, we identified ortho alkoxy groups as well as ortho alkyl groups as promising substituents and designed a further set of compounds using FEP+. Data for the most promising compound 27 are shown in Table 5.

Table 5 –

FEP+ predicted and experimental sAC pIC₅₀ values for extension of the ortho alkoxy group.

Compound	3	25	27
pIC50 (FEP+)	*6.27	*6.79	8.11
pIC50 (experimental)	6.26	6.74	8.37
logD (pH 7.4)	2.76	2.90	2.58
Lipophilic Efficiency	3.50	3.83	5.80

^*3 and 25 were used as the FEP+ reference compounds and edges were run for 15 ns. The morpholine in 27 was modelled with the amine nitrogen in both its protonated and non-protonated states. The non-protonated state was predicted to be most favorable and only this data is reported.

The sAC inhibitor 27 exemplifies a potent and druglike sAC inhibitor with a binding affinity of 4.24 nM and a much higher lipophilic efficiency (5.80) than the starting point 1 (2.80). Studying the crystal structures, we noted an opportunity to replace the phenyl group with bicyclic surrogates. For this reason, a series of bicyclic analogs were designed and assessed using FEP+ (see Table 6). Some of the bicyclic compounds were designed in the context of a difluoromethyl group in place of a methyl. This difluoromethyl group was shown to improve binding affinity (see Table 3) and is a common medicinal chemistry bioisostere that can block metabolism of exposed methyl groups.

Table 6 –

FEP+ predicted and experimental sAC pIC₅₀ values for a set of bicyclic analogues of 3.

Compound ID	pIC50 (FEP+)	pIC50 (experimental)	R1	R2
28	8.09	6.85
29	8.06	7.82
30	7.54	6.60
31	7.33	7.30
12	*6.96	6.96

^*12 was used as the FEP+ reference compound and all edges were run for 5 ns. The mean unsigned error for non-reference compounds is 0.66 kcal/mol.

We also attempted to modify the aminopyrimidine headgroup by replacing the chlorine with groups such as methyl, cyano, and methoxy. However, replacement of the chloro was predicted to be deleterious to binding affinity by FEP+ for all 14 groups explored. We also explored changes to the 5 position of the aminopyrimidine ring by introducing small groups such as fluoro, chloro, and methyl. However, all 11 changes explored were predicted to be deleterious to binding affinity by FEP+. We attempted to validate these predictions using a potent reference molecule as a starting point. Table 7 shows the predicted and experimental results for modifications to the aminopyrimidine ring of 27 at the 4 and 5 positions.

Table 7 –

FEP+ predicted and experimental sAC pIC₅₀ values for a set of analogues of 27.

Compound ID	pIC50 (FEP+)	pIC50 (experimental)	R1	R2
27	*8.25	8.37
32	*7.10	7.00
33	*7.05	6.92
34	5.62	~6

^*27, 32 and 33 were used as the FEP+ reference compounds and edges were run for 5 ns. All morpholine rings were modelled with the amine nitrogen in the non-protonated state. The mean unsigned error is 0.29 kcal/mol.

The predicted and experimental results are in excellent agreement and highlight the difficulty of modifying the aminopyrimidine ring. In the final stage of the campaign, we explored the pyrazole 5 position, noting the growth vector from the crystal structures (see Figures 2, 3, and 4). Table 8 shows the FEP+ results for a set of compounds modifying the pyrazole 5 position. Again, the SAR was generated using docking and traditional medchem transformations and this FEP+ analysis was run retrospectively.

Table 8 –

FEP+ predicted and experimental sAC pIC₅₀ values for a set of analogues of 25. All compounds were used as FEP+ reference compounds and all edges were run for 5 ns. The mean unsigned error for all compounds is 0.92 kcal/mol. Compounds marked with an asterisk were modelled as both protonated and deprotonated using the pKa state correction approach. In all cases the neutral molecule bound most favorably. The calculated pKa values and state populations are reported in supplementary table S1. The racemic compound marked with a dollar sign was modelled as both enantiomers using a state correction approach.

Compound ID	pIC50 (FEP+)	pIC50 (experimental)	R1
35	9.89	7.59
36	9.40	8.82
37*	9.21	9.34
38	9.13	9.08
39	9.10	8.33
40*	9.07	8.00
41*	8.72	8.22
42	8.67	8.42
43	8.34	7.49
44	8.22	7.59
45$	7.83	7.70
9	7.62	7.70
46*	7.61	8.46
47*	7.26	8.10
48	6.88	6.51
25	6.24	6.77
13	5.92	6.40
49*	5.89	6.52
8	4.50	6.15

Compounds bearing substituents at the 5 position exhibited the highest binding affinity in the series, reaching sub-nanomolar levels. To check the binmding mode, we solved the crystal structure of 47 in complex with sAC (see Figure 5) which shows the ester group along the designed growth vector.

Figure 5 –

Overlaid crystal structures of 7 (orange) and 47 (cyan). Proteins are shown as ribbons with wire atoms and ligands are shown as balls and sticks.

Whilst FEP proved highly effective in this project, it is a time-consuming and expensive technique. For this reason, we performed a retrospective analysis to identify whether computationally less demanding methods could have been used to similar effect. Thus, for each round of the FEP we also ran ensemble Glide SP docking followed by MM-GBSA rescoring for comparison purposes. The performance of different methods, compared to experiments, was quantified by the Mean Unsigned Error (MUE), the correlation coefficient (r²) and the Kendall’s rank correlation coefficient (τ) (Table 9). The scatter plots of experimental binding energies versus predicted number for the entire data set are shown in Figure 6. The FEP estimated binding affinity outperformed ensemble docking scores and MM-GBSA with the highest R² of 0.57, highest τ index of 0.61, and lowest MUE (not applied for MM-GBSA) of 0.78 kcal/mol. The ensemble docking model with 19 structures reproduce the experimental ranking reasonably well with τ = 0.47. However, the ensemble docking scores are not accurate in predicting binding affinity (MUE = 3.72 kcal/mol). MM-GBSA refinement did not further significantly improve the accuracy in ranking the molecules (τ = 0.46). Taken together, these results suggest that using FEP was significantly more effective than Glide ensemble docking or MM-GBSA would have been.

Table 9.

Summary of the FEP+ calculations in comparison with Glide SP and MM-GBSA. Each table is reported in a separate column and the overall results are given in the final column.

Table		2	3	4	5	6	7	8	Total
Experimental Datapoints		2	20	6	3	5	4	19	59
FEP+	MUE (kcal/mol)	0.57	0.97	0.45	0.13	0.66	0.29	0.92	0.78
	R2		0.23	0.75	1.00	0.09	0.94	0.65	0.57
	Kendall τ	-	0.31	0.73	1.00	0.00	0.91	0.54	0.61
Glide SP	MUE (kcal/mol)	2.93	3.72	4.01	3.71	4.07	4.46	3.48	3.72
	R2		0.06	0.03	1.00	0.39	0.01	0.28	0.36
	Kendall τ	-	0.15	0.14	1.00	-0.60	-0.55	0.35	0.44
MM-GBSA	R2		0.22	0.10	0.93	0.11	0.03	0.39	0.47
	Kendall τ	-	0.28	0.07	1.00	0.00	0.00	0.41	0.46

Figure 6.

Correlation of experimental and calculated binding activities for all the compounds synthesized. The diagonal line of equality and ± 2 kcal/mol errors are shown as black lines. Data trendlines are shown in red.

It is also worth noting that the Glide ensemble docking and MM-GBSA approaches used all of the crystal structures that were solved in the project, not just those that were available at the time each that each compound was designed. Conversely, the FEP calculations were performed using only the crystal structure of LRE1 which was available at the outset of the project. Figure 6 presents these results graphically.

Discussion

Free energy perturbation (FEP) makes use of molecular dynamics simulation and can be used to compute free energy differences between structurally related compounds ⁵¹. Recently, FEP has progressed significantly in speed and accuracy, which has allowed it to contribute to drug discovery efforts through accurate prediction of binding free energy ^{4, 5}. Whilst binding affinity is only one aspect of an effective drug, it is an important prerequisite for efficacy and is commonly a key property that must be improved in the hit-to-lead stage of drug discovery. Here we employed FEP+ to aid in a hit-to-lead campaign to identify inhibitors of sAC. During our hit-to-lead campaign, FEP+ calculations were used to support compound design in a prospective manner. The first step was a scaffold hop from LRE1 to look for alternative chemotypes. We identified 2 with an improved binding affinity (pIC₅₀ of 6.85 vs 5.83), better lipophilic efficiency, and more amenable growth vectors. We then replaced the thiophene ring to address a safety concern, yielding 7 with similar binding affinity (pIC₅₀ of 7.00). We obtained crystallographic data with 7, illustrating the success of the design concept. We then focused on improving binding affinity by exploring changes to the phenyl ring, the aminopyrimidine headgroup, and the pyrazole ring. Calculations predicted that modifications to the aminopyrimidine headgroup would be deleterious to binding affinity, a prediction that was confirmed experimentally. However, inhibitors with high affinity were identified by growing an ortho alkoxy substituent from the phenyl ring and an alkyl chain from the pyrazole ring. The resulting compounds 27 and 37 showed nanomolar and sub-nanomolar inhibition in the sAC biochemical assay (pIC₅₀ values of 8.37 and 9.34). In addition, 27 has improved pharmacokinetic characteristics suitable for use in downstream experimental work.

Across all the maps, the mean unsigned error of the experimental and predicted binding affinities was 0.78 kcal/mol. If one excludes reference compounds, the mean unsigned error of the experimental and predicted binding affinities was 0.85 kcal/mol. We suggest that this dataset will prove useful for validation of methods to predict binding free energy and have thus made the data and structures available on Github (https://github.com/djhuggins/sAC-data-BFE-validation).

This study illustrates the effectiveness of FEP in the hit-to-lead stage of projects that are structurally enabled. In this case, FEP+ in consultation with the medicinal chemistry team allowed us to rapidly develop a 1473 nM hit into a 4.24 nM lead using a stepwise design approach, enabling proof of concept studies for a promising drug target.

Data and Software Availability

Ligprep, Forcefield Builder, Glide, MMGBSA, Preparation Wizard, and FEP+ calculations were performed with Schrödinger software (https://www.schrodinger.com/) version 2020–3. The IC50 data and input structures are available on Github (https://github.com/djhuggins/sAC-data-BFE-validation).