Virtual Screening for FDA-Approved Drugs That Selectively Inhibit Arginase Type 1 and 2

Arginases are often overexpressed in human diseases, and they are an important target for developing anti-aging and antineoplastic drugs. Arginase type 1 (ARG1) is a cytosolic enzyme, and arginase type 2 (ARG2) is a mitochondrial one. In this study, a dataset containing 2115-FDA-approved drug molecules is virtually screened for potential arginase binding using molecular docking against several ARG1 and ARG2 structures. The potential arginase ligands are classified into three categories: (1) Non-selective, (2) ARG1 selective, and (3) ARG2 selective. The evaluated potential arginase ligands are then compared with their clinical use. Remarkably, half of the top 30 potential drugs are used clinically to lower blood pressure and treat cancer, infection, kidney disease, and Parkinson’s disease thus partially validating our virtual screen. Most notable are the antihypertensive drugs candesartan, irbesartan, indapamide, and amiloride, the antiemetic rolapitant, the anti-angina ivabradine, and the antidiabetic metformin which have minimal side effects. The partial validation also favors the idea that the other half of the top 30 potential drugs could be used in therapeutic settings. The three categories greatly expand the selectivity of arginase inhibition.


Introduction
Arginase (E.C. 3.5.3.1) is a ureohydrolase homotrimeric binuclear manganese metalloenzyme that hydrolyses L-arginine into L-ornithine and urea [1,2] (Figure 1). In ureotelic animals, this enzyme is primarily involved in the elimination of excess ammonia through the urea cycle [3]. Interestingly, overexpression of arginase is involved in diseases such as cancer, cardiovascular, diabetes, asthma, neurodegenerative disease, and infectious disease [4][5][6][7][8]. In mammals two distinct isoforms of the enzyme exist, namely arginase type 1 (ARG1) and arginase type 2 (ARG2), which are encoded by different genes, and share approximately 60% sequence homology [9]. The isoenzymes differ in tissue distribution, intracellular location, and immunochemical characteristics [9]. ARG1 is a cytosolic enzyme expressed predominantly in the liver, where it plays a key role in urea synthesis [10]. ARG2 is a mitochondrial enzyme that is expressed in most tissues, with high levels found in the kidney and prostate, but also in blood vessels, intestines, red blood cells, and immune cells. Recent evidence suggests that ARG2 can translocate from the mitochondrion into the cytosol leading to an increase in its activity [11].
Virtual screening is a classical tool used to screen novel compounds that target a given protein of interest. Computational screening approaches have gained an advantage over high-throughput screening techniques, due to decreased time and cost to select potential compounds for further experimental validation. In the past, we have used structure-based virtual screening to identify potential ligands for Mcl-1 [53] and Pyk2 [54]. For arginase, several human experimental structures are available and are listed in Table S1.
In this study, we used molecular docking against multiple arginase conformations (3 X-ray crystallography PDB structures for each of the arginase isozymes) to virtually screen for novel ligands that can modulate the activity of arginase, both selectively, and not. In addition, we used molecular dynamics (MD) simulation to validate the proteinligand stability of the selected hits ( Figure 3). Using these techniques with FDA approved drug dataset, we identified novel small molecules that could be repurposed for cancer, Alzheimer's disease, hypertension, cardiovascular diseases, and other metabolic disorders. lifespan in mice [51], with no significant changes in plasma amino acids and ammonia levels [52]. Therefore, in our study, we aimed to identify potential FDA-approved drugs selective to ARG1 and ARG2 to ameliorate the treatment efficacy.
Virtual screening is a classical tool used to screen novel compounds that target a given protein of interest. Computational screening approaches have gained an advantage over high-throughput screening techniques, due to decreased time and cost to select potential compounds for further experimental validation. In the past, we have used structure-based virtual screening to identify potential ligands for Mcl-1 [53] and Pyk2 [54]. For arginase, several human experimental structures are available and are listed in Table S1.
In this study, we used molecular docking against multiple arginase conformations (3 X-ray crystallography PDB structures for each of the arginase isozymes) to virtually screen for novel ligands that can modulate the activity of arginase, both selectively, and not. In addition, we used molecular dynamics (MD) simulation to validate the protein-ligand stability of the selected hits ( Figure 3). Using these techniques with FDA approved drug dataset, we identified novel small molecules that could be repurposed for cancer, Alzheimer's disease, hypertension, cardiovascular diseases, and other metabolic disorders.

Materials and Methods
Database selection and ligand preparation: FDA-approved drug datasets (DrugBank FDA only, version 2018-02-20) were retrieved in Mol2 format from the non-commercial ZINC15 database (https://zinc.docking.org/ (accessed on 7 February 2022) [55]. We then used the Open Babel package (version 2.3.2) Pittsburgh, PA, USA [56] to convert file format from Mol2 to pdbqt format, followed by Vina Split [57] to obtain 2115 FDA-approved drugs for virtual screening.

Materials and Methods
Database selection and ligand preparation: FDA-approved drug datasets (DrugBank FDA only, version 2018-02-20) were retrieved in Mol2 format from the non-commercial ZINC15 database (https://zinc.docking.org/ (accessed on 7 February 2022) [55]. We then used the Open Babel package (version 2.3.2) Pittsburgh, PA, USA [56] to convert file format from Mol2 to pdbqt format, followed by Vina Split [57] to obtain 2115 FDA-approved drugs for virtual screening.
Target preparation and grid generation: Three High resolution X-ray crystallography structures [PDB IDs 4HWW(1.30 Å) [58], 4HXQ (1.45 Å) [58], 3SJT (1.60 Å) [59]] of ARG1, and three X-ray crystallography structures [PDB ID 4HZE (1.60 Å) [58], 4I06 (1.80 Å) [58], 4IXU (1.90 Å) [60]] of ARG2 were selected among the multiple conformations of these proteins listed in Supplementary Table S1. Protein receptors for virtual screening were prepared further using AutoDockTools (version 1.5.7), California, CA, USA [61], where polar hydrogen atoms were added and Gasteiger charges were assigned. The search space for ligand docking was figured based on the coordinates of the native ligand from the crystal structure. AutoGrid plugin of PyMOL (version 1.8.4.0) New York, NY, USA [62] was used to build a cubic grid box measuring 25 × 25 × 34 Å with 0.375 Å spacing at the binding site of the above-mentioned arginase structures. The X, Y, and Z, coordinates of the center of the docking box were  [57] were as follows: The exhaustiveness of the global search was 8, the maximum number of binding modes to generate was 9, and the maximum energy difference between the best and worst binding mode displayed as 3 kcal/moL, thus limiting the number of poses. For each ligand, only the best pose was retained. In all cases, the binding site on arginase was defined and limited by a box measuring 25 × 25 × 34 Å 3 around different X, Y, and Z, center coordinates.
Protein-Ligand interaction analysis: Protein-ligand interaction analysis was performed using PLIP (Protein-Ligand Interaction Profiler) [63] to analyze non-covalent interactions between the arginase and the best docking pose of the selected hit showing fewer side effects.
To generate protein parameter files and ligand parameter files VMD (Visual molecular dynamics; version 1.9.3), Illinois, IL, USA [68] and CHARMM-GUI Ligand Reader and Modeler module [69] were used, respectively. As manganese ions were undefined in the CHARMM36 forcefield, we used Mg 2+ ion parameters instead of Mn 2+ ions of ARG1 and ARG2, as routinely recommended by Nagagarajan et al. [70]. Solvation and auto-ionization of the protein-ligand complex were performed using VMD. A standard TIP3 water model was added as a solvent with periodic boundary conditions, and a padding distance of 10 Å for each dimension [71]. Na + and Cl − ions were used to neutralize the system. Up to date, the CgenFF version of CHARMM-GUI [72,73] was used to generate topology and parameter files for the selected ligands. Periodic boundary conditions were applied. The particle-mesh Ewald method [74] was used for electrostatic interactions of the system. Minimization was done for 1000 steps at a constant temperature of 298.15 K. Position restraints were applied on Mn 2+ ion before carrying out equilibration and production run using NPT (constant particle number, pressure, and temperature) ensemble for 50 ns. Three sets of simulations [denoted by Set A (ARG1 (s1a-s5a), ARG2 (s1a-s5a)); Set B (ARG1 (s1b-s5b), ARG2 (s1b-s5b)); Set C (ARG1 (s1c-s5c), ARG2 (s1c-s5c))] were performed for 50 ns each. VMD-Hbonds plugin was used to count strong hydrogen bond (Donor-Acceptor distance 4.0 Å, Angle cutoff 30 • ) formation between protein and ligand complex throughout the trajectory. RMSD trajectory plugin tool in VMD was used to calculate the RMSD of protein and ligand.

Results
Validation of molecular docking assay: Remarkably, our molecular docking assay was validated using two approaches: (1)  in three distinct arginase structures. The predicted K d of ARG1 was taken as the average of the top poses in PDBs 4HZE, 4IXU, and 4I06, while the predicted K d of ARG2 was calculated using PDBs 4HWW, 3SJT, and 4HXQ. A plot of the predicted and experimental dissociation constant (K d ) is shown in Figure S1. The experimental and predicted K d correlate well, with correlation coefficients (R) of 0.72 for ARG1, and 0.77 for ARG2. These two approaches validate our screening methodology (Table S2). Thus, the predictive power of our in-silico assay was proven in a small dataset and paved the way for further analysis on large datasets.
Less remarkably, the predicted and experimental selectivity of these compounds did not correlate at all. with R = 0.0023. The ARG1 selectivity was taken as the ratio of dissociation constant (K d ) of ARG1 divided by ARG2, and the ARG2 selectivity vice versa. The selectivity plot ( Figure S2) of experimental and predicted selectivity shows a weak correlation, and as a result, we resorted to using MD simulations for potential selectivity.
Virtual docking with FDA-approved drug data set: Virtual docking of 2115 FDA-approved drugs on the six arginase structures, three of ARG1 (PDB IDs: 4HWW, 4HXQ, 3SJT) and three of ARG2 (PDB IDs: 4HZE, 4IXU, 4I06) took~36 h (Table S3). Notably, the top pose of each ligand is oriented similarly in all three arginases, as indicated by visual inspection, and low RMSDs, (except in less than 2% of the cases) (Table S11). Table 1 lists the top 15 FDA-approved drugs, both non-selective, ARG1 selective, and ARG2 selective. The top 15 non-selective drugs are ranked according to the total of their average binding energy. Interestingly, most of the drugs are used to treat inflammation, headache, cancer, metabolic disorders, and type 2 diabetes which could become clinically relevant for drug repurposing. Table 1 also lists the top 15 FDA-approved drugs that are potentially selective to ARG1 and ARG2. The difference between the average free binding energies (∆∆G) of three ARG1 structures and three ARG2 structures is used to rank the top drugs. The top drugs that show potential selectivity for ARG1 are associated with treating fungal and bacterial infection, arthritis, colitis, cancer, depression, and hypertension (Table S4). And the top ARG2 selective drugs are used to treat pain, hypertension, anxiety, infection, heart failure, and malaria among others (Table S4 for indications). Of these, some are associated with severe side effects (i.e., valrubicin, rocuronium, etc.), and are unsafe for repurposing towards arginase inhibition. Some of the ligands were selected based on their relatively low toxicity, and adverse effects. For example, the antihypertensive drugs candesartan, and irbesartan which are angiotensin II receptor blockers (ARB), the thiazide-like diuretic indapamide, the antiemetic rolapitant which is an NK1-receptor antagonist, the anti-angina ivabradine which bind binds to HCN4 receptors, the antihypertensive amiloride that is an epithelial sodium channel (ENaC), the antidiabetic metformin, and the antifungal triazole isavuconazole. Furthermore, metformin was selected because its pharmacokinetic action is not completely understood. These approved drugs have minimal side effects and were further analyzed for their potential repurposing as arginase inhibitors. In particular, we mapped the interaction networks of the selected ligands and performed MD simulations of some with ARG1 (PDB ID 3SJT) and ARG2 (PDB ID 4HZE) structures to check the stability of the protein-ligand interaction.
Finally, codeine causes side effects such as addiction, and budesonide is an inhaled glucocorticoid. Nevertheless, preliminary analyses were also performed on them.
MD simulation for protein-ligand stability: Three sets of molecular dynamics simulations up to 50 ns were carried out, to assay ligand interactions and arginase selectivity (candesartan, irbesartan, isavuconazole, codeine, and metformin). Hydrogen bond formation and RMSD analyses of protein-ligand complexes were assessed to evaluate stability (Tables S6-S9). Average RMSD and standard deviation of protein and ligand for each set of the simulation were provided in Table S10. As a control, the binding of the inhibitor nor-NOHA to ARG1 and ARG2 was also simulated [70,75,76]. nor-NOHA remained in the binding site for 50 ns and approx. 16 ns (of 50 ns calculated), respectively (data not shown). Remarkably, candesartan and irbesartan occupied the binding site of ARG1 for the entire length of the simulation (50 ns). In comparison, candesartan and irbesartan remained in the binding site of ARG2 less than 37 ns and 13 ns, respectively, suggesting they are selective for ARG1. This limits, our earlier results which show irbesartan selectivity towards ARG2, and both irbesartan and candesartan could potentially favor ARG1. Interestingly, isavuconazole and metformin remained in the binding site of both arginase for 50 ns, potentially suggesting non-selective arginase binding, thus reinforcing our molecular docking results that isavuconazole and metformin are non-selective. Codeine stayed up tõ 8.5 ns with ARG1 before drifting away from the binding pocket to distances obstructive for binding interactions, and 50 ns with ARG2 thus potentially preferring binding to ARG2 in agreement with our molecular docking prediction. Figures 5 and 6 show hydrogen bond formation and RMSD of protein-ligand (Set A) during our MD simulation (Set B and Set C hydrogen bond formation and RMSD of the protein-ligand complex are provided in Tables S6-S10). Thus, our selectivity predictions should be taken with a grain of salt, and require experimental validation. Notably, the amino-rich, metformin forms multiple hydrogen bonds in the binding site with arginase and contributes to complex stability ( Figure 5).
(candesartan, irbesartan, isavuconazole, codeine, and metformin). Hydrogen bond formation and RMSD analyses of protein-ligand complexes were assessed to evaluate stability (Tables S6-S9). Average RMSD and standard deviation of protein and ligand for each set of the simulation were provided in Table S10. As a control, the binding of the inhibitor nor-NOHA to ARG1 and ARG2 was also simulated [70,75,76]. nor-NOHA remained in the binding site for 50 ns and approx. 16 ns (of 50 ns calculated), respectively (data not shown). Remarkably, candesartan and irbesartan occupied the binding site of ARG1 for the entire length of the simulation (50 ns). In comparison, candesartan and irbesartan remained in the binding site of ARG2 less than 37 ns and 13 ns, respectively, suggesting they are selective for ARG1. This limits, our earlier results which show irbesartan selectivity towards ARG2, and both irbesartan and candesartan could potentially favor ARG1. Interestingly, isavuconazole and metformin remained in the binding site of both arginase for 50 ns, potentially suggesting non-selective arginase binding, thus reinforcing our molecular docking results that isavuconazole and metformin are non-selective. Codeine stayed up to ~8.5 ns with ARG1 before drifting away from the binding pocket to distances obstructive for binding interactions, and 50 ns with ARG2 thus potentially preferring binding to ARG2 in agreement with our molecular docking prediction. Figures 5 and 6 show hydrogen bond formation and RMSD of protein-ligand (Set A) during our MD simulation (Set B and Set C hydrogen bond formation and RMSD of the protein-ligand complex are provided in Tables S6-S10). Thus, our selectivity predictions should be taken with a grain of salt, and require experimental validation. Notably, the amino-rich, metformin forms multiple hydrogen bonds in the binding site with arginase and contributes to complex stability ( Figure 5).

Discussion
ARG1 and ARG2 play a key role in many diseases, such as cancer, Alzheimer's, hypertension, inflammation, etc. Despite the advances in computer-aided drug design, the clinical trial time frame is a major obstacle to developing a new commercial drug. Thus, using an FDA-approved drug repurposing approach could be advantageous to reduce the

Discussion
ARG1 and ARG2 play a key role in many diseases, such as cancer, Alzheimer's, hypertension, inflammation, etc. Despite the advances in computer-aided drug design, the clinical trial time frame is a major obstacle to developing a new commercial drug. Thus, using an FDA-approved drug repurposing approach could be advantageous to reduce the clinical trial timeframe and improve treatment. In this study, we used FDA-approved drugs to identify molecules that are (1) non-selective, (2) ARG1 selective, and (3) ARG2 selective. We selected 10 FDA-approved drugs with minimal side effects and studied protein-ligand interaction and stability using a protein-ligand interaction profiler (PLIP) followed by an MD simulation of five selected ligands. We found the antihypertensive drugs candesartan and irbesartan showed potential high affinity towards arginases. Moreover, irbesartan also showed more stability with ARG1 (50 ns) than ARG2 (less than 13 ns) despite having stronger binding energy for ARG2 (∆G ARG1(3SJT) = −7.2, ∆G ARG2(4HZE) = −8.7). We found metformin and isavuconazole are non-selective to both the arginases based on binding energies, protein-ligand interaction, and molecular dynamics. Codeine was more stable with ARG2 (50 ns), which is corroborated by the binding energy (∆G ARG1(3SJT) = −6.7, ∆G ARG2(4HZE) = −7.9). Altogether, candesartan, ibersartan, codeine, metformin, and isavuconazole qualified as suitable candidates for the development of potential arginase inhibitors. Interestingly, irbesartan, and candesartan extend a healthy lifespan in heart failure patients [77], and improve cognitive function in patients [78]. As such, our study exposes potential non-selective, and selective ARG1 and ARG2 ligands to the scientific community for further biological screening, and development of new arginase inhibitors.
As a potential limitation to this study, the AutoDock Vina binding energy is not a quantitative measure of experimental binding energy, however, they are a good measure to qualitatively estimate the relative binding affinity. Like molecular mechanics energies calculated with MM/PBSA and MM/GBSA methods, they are popular approaches to estimating the free binding energy of small ligands to biological macromolecules. Such methods have been applied to a large number of systems with varying success [79]. In this study, we used the AutoDock Vina binding energies to estimate relative binding affinities. Another limitation is our study uses computational methods, which have not been validated experimentally. Thus, additional in-vitro and in-vivo studies are required to characterize ligand binding. Moreover, experimental data and clinical trials are needed to identify the full potential of these selected candidates to serve as arginase inhibitors and improve diseases related to old age, such as hypertension, diabetes, and dementia.   Table S6: Average number of hydrogen bonds during MD simulation (for Set A, Set B, Set C); Table S7: RMSDs of protein ligand complex during MD simulation (Set A); Table S8: RMSDs of protein ligand complex during MD simulation (Set B); Table S9: RMSDs of protein ligand complex during MD simulation (Set C); Table S10: Average RMSD and standard deviation of protein and ligand for each set of simulation; Table S11