## Introduction

Fragment-based lead discovery (FBLD) has rapidly become a well-established technique in early drug development1. Several lead candidates developed using FBLD have already reached clinical trials, resulting in two FDA approved drugs2. In contrast to high-throughput screening (HTS), where large numbers (~105–106) of drug-sized molecules are tested experimentally, FBLD focuses on smaller libraries (typically 1000–5000 compounds) with molecules of low molecular weight (<300 Da)2, 3. By limiting the size of the molecules in the library, fragment screening achieves a much broader coverage of chemical space than HTS due to the astronomical number of possible drug-like molecules. The low molecular complexity of fragments also reduces the probability for steric mismatches with the receptor, leading to the discovery of ligands that optimally complement subpockets of the binding site4, 5. Consequently, screening of fragment libraries often delivers high hit-rates and diverse starting points for lead development2, 3. However, the ligands that emerge from fragment screening typically have low affinities and, in the second step of FBLD, these compounds need to be optimized to yield potent and selective lead candidates.

Fragment-to-lead optimization has proved to be a very challenging step in FBLD2. Prioritization of fragments for optimization is often guided by ligand efficiency (defined as the free energy of binding divided by the number of heavy atoms of the compounds6) and access to atomic resolution information regarding binding modes7. Recently, more intricate criteria, e.g. based on the thermodynamic binding profiles of the fragments, have also been suggested to be an important factor in ﻿t﻿he selection of starting points for optimization8. Two main strategies for fragment-to-lead optimization, “linking” and “growing”, have been proposed2. Although “linking” of compounds occupying different subpockets of a binding site has been successful in some cases, “growing” of the fragment by iterative additions of smaller chemical groups has become more widely used2, 3. In either case, access to high-resolution crystal structures of fragments bound to the target often makes crucial contributions to the optimization process7. Whereas computational methods for structure-based ligand design are routinely used for drug-sized molecules9, applications of such approaches to fragment optimization have been more scarce10. The fact that fragments are weak ligands, only occupy a small fraction of the binding site, and may have multiple binding modes adds extra levels of complexity that are challenging to predict with simplified models such as empirical scoring functions. Furthermore, scoring functions developed for computer-aided ligand design have been parameterized based on drug-like compounds, and it has been suggested that these may not be suitable for fragment ligands11, 12. Molecular dynamics (MD) simulations in combination with alchemical free energy methods, which explicitly consider contributions to binding from conformational flexibility and interactions with water molecules, could provide a rigorous approach to guide fragment optimization13, but this technique has only recently been applied to FBLD10. Accurate predictions of relative binding affinities for analogs to ligands identified by fragment screening could improve the efficiency of FBLD, further establishing this approach as a groundbreaking strategy for early drug development.

In this work, the utility of MD combined with alchemical free energy methods in fragment optimization was explored for the human A2A adenosine receptor (A2AAR), a G protein-coupled receptor (GPCR) relevant for drug development14 against Parkinson’s disease15 and cancer16. Multiple high-resolution crystal structures of the A2AAR have recently been determined17, 18 and numerous fragment-sized ligands have been identified to this target19,20,21, making it an ideal test case for evaluating a computational approach. Calculation of relative binding affinities using MD simulations in combination with the free energy perturbation (MD/FEP) method was first benchmarked retrospectively for 23 fragment-sized compounds22, 23. The MD/FEP technique was also used to assess multiple binding modes and predict the thermodynamic signatures governing changes in binding affinity, which are both factors of major interest in the optimization process. In a second step, MD/FEP was applied prospectively to predict relative affinities for 12 fragment-sized compounds with unknown binding affinities, followed by experimental evaluation of these in pharmacological assays. In light of the results, the feasibility of using MD simulations in combination with alchemical free energy methods as a tool in fragment-to-lead optimization will be discussed.

## Results

### Mapping binding site subpockets using free energy calculations for fragment ligands

The first set of 10 pairs of adenine derivatives (Table 1) probed opportunities for growing fragments into the ribose-recognizing site (pocket A, Fig. 1) and mainly involved substitutions in the N9-position of the adenine scaffold (Table 1). The average unsigned error for this set of relative binding free energies was 0.66 kcal/mol, resulting in strong correlation with experimental binding data (Fig. 3, R2 = 0.75). Relative free energies involving compounds 1 and 11 were not included in the analysis of correlation with experimental data as reliable Ki values could not be determined for these compounds due to their lack of binding at the highest tested concentration (Ki > 100 μM). However, it should be noted that MD/FEP correctly predicted the direction of the shift in binding free energies in both cases. One example of successful fragment growth into pocket A was observed for addition of a methyl group in the N9-position of the adenine scaffold (compounds 3 and 4). This resulted in a large improvement of the binding affinity (2.0 kcal/mol), which was also reproduced by the calculated free energy change of 2.4 kcal/mol. Interestingly, substituents larger than two heavy atoms in pocket A typically reduced binding affinities. For example, replacement of a 9-ethyl substituent by 2-hydroxyethyl, isobutyl or propyl moieties led to losses of binding affinity, which were also captured the MD/FEP calculations.

The second compound set explored the effects of substituents in the C8-position of the adenine scaffold, which extended into pocket B. A bromine in this position (compound 3) led to a 58-fold increase in affinity compared to the unsubstituted compound 2. This effect was also captured by the free energy calculations, which predicted an improvement in binding corresponding to 2.3 kcal/mol between compounds 2 and 3, in close agreement with the experimental value (2.4 kcal/mol). Interestingly, changes in affinity from addition of a bromine in the C8-position were dependent on the substituent in the N9-position. This interdependency of the two substituents was also captured by the calculated binding free energies for this subset of four compound pairs (23, 5, 9, and 1417). Another series of compounds with an ethyl group in the N9-position and varying substituents in C8-position (9 and 1823) was also considered. A 10-fold increase of binding affinity compared to compound 9 was obtained for the C8-methyl substituent (compound 18) and addition of a hydroxyl group in the same position (compound 20) also resulted in improved binding. The relative binding free energies for a majority of the considered pairs exploring pocket B were within 1 kcal/mol of the experimental value (Fig. 3). However, for the pair consisting of 8-alkoxy-9-methyladenine derivatives 23 and 21, there was a large discrepancy between the published experimental affinity and relative binding free energy calculated from MD simulations. The experimentally determined values23 indicated a >1000-fold loss of binding affinity, as compound 23 showed no significant binding (reported Ki value > 100 μM), whereas the calculated value suggested that the two compounds had similar affinity. As this was a major outlier among the considered compound pairs, compound 23 was retested in a radioligand binding assay at the A2AAR. The Ki value was determined to be 95 nM for compound 23 (Supplementary Table S1), leading to a relative free energy of −0.8 kcal/mol, which was in better agreement with the prediction and resulted in a strong correlation with experiment for the second compound set (R2 = 0.75, Fig. 3).

The ability of MD/FEP to predict changes in affinity was further highlighted by the strong correlation between experimental and predicted binding free energies for the full set of 18 compound pairs (R2 = 0.78, Fig. 3). To assess the influence of experimental uncertainties on this result, the correlation was calculated for 1000 random selections of either the maximal or minimal Ki value obtained from the 95% confidence interval, which resulted in R2 = 0.75 with a standard deviation of 0.1. It should also be noted that the correlations between the experimentally determined relative binding affinities and trivial size-descriptors, e.g. the difference in the number of heavy atoms (R2 = 0.10) or predicted 1-octanol/water partition coefficients (AlogP, R2 = 0.44), were low. In order to compare our results to an empirical scoring function, the adenine derivatives were also docked to the orthosteric site using the GLIDE docking program25 and binding free energies were calculated with the standard precision (SP) scoring function for the 18 compound pairs. All of the docked compounds reproduced the binding mode expected from crystal structures of the A2AAR in complex with adenine-based ligands. The correlation with experimentally determined relative binding free energies for GLIDE-SP (R2 = 0.42, Supplementary Figure 2) was similar to that obtained for ALogP and significantly lower than for MD/FEP.

### Assessment of alternative binding modes and thermodynamic signatures of fragment binding

The use of MD/FEP in fragment optimization could be limited by the uncertainty associated with ligand binding modes if a crystal structure of the complex is not available. As fragments are small, such compounds can bind in a large number of orientations and it may be challenging to rank these with more simplified models, e.g. molecular docking scoring functions12, 26. The calculations for 9-methyl adenine derivatives 2 and 3 were extended to explore two alternative binding modes identified by the molecular docking study carried out by Lambertucci et al.22 (Fig. 4A and B). Both proposed binding modes predicted a hydrogen bond between the exocyclic nitrogen of the adenine moiety and the side chain oxygen of Asn253. The first binding mode involved an additional hydrogen bond between the N7 of the adenine-moiety and the side chain nitrogen donor of Asn253, leading to an orientation that was essentially identical to that observed in the crystal structure of the A2AAR in complex with adenosine (Fig. 1)18. In the alternative orientation, a hydrogen bond with the N1 atom of the adenine moiety was instead obtained, resulting in a second distinct binding mode. The relative binding affinity of the two poses (Fig. 4A and B) was calculated by alchemically transforming one binding mode into the other via an intermediate compound (24) to assess their probability (Supplementary Figure 3). The calculated free energies suggested that the pose derived from the binding mode of adenosine in A2AAR crystal structures, which was also used in the MD/FEP calculations, was favored by 6.5 ± 0.1 kcal/mol. The population of the alternative binding mode was thus predicted to be very low and would not influence the experimentally measured binding affinity significantly.

The enthalpic and entropic components of the binding free energy are increasingly attracting interest in drug discovery as these can provide more information on the driving forces of ligand binding27. Although the experimental binding free energy differences for compound 3 relative to compounds 2 and 4 were accurately reproduced by the MD/FEP calculations, it was not clear from visual inspection why the addition of a single heavy atom resulted in such a large change in binding affinity in both cases. To further quantify the change in binding free energy, it was decomposed into enthalpy and entropy components using a relationship analogous to the van’t Hoff equation. MD/FEP calculations were carried out for the two compound pairs at 13 different temperatures between 270 and 330 K. The enthalpy and entropy components could then be derived from the slope and intercept of the relation between ΔΔGbind/T and 1/T (Fig. 4C and D)28. These calculations demonstrated that the predicted affinity increase for compound 3 relative to compound 2 was driven by entropy (−TΔΔSbind = −7.1 kcal/mol), which was counteracted by an unfavorable enthalpy contribution (ΔΔHbind) of  +4.8 kcal/mol. In contrast, the gain in affinity for compound 3 relative to compound 4 was enthalpy driven (ΔΔHbind = −5.6 kcal/mol and −TΔΔSbind =  + 3.4 kcal/mol). Overall, there were only small differences in receptor structure between the three complexes based on the MD simulations, suggesting that changes in the solvent network could be responsib﻿le for the distinct thermodynamic profiles. MD snapshots of the water molecules in the binding site were clustered to identify hydration sites in the vicinity of the ligands using the algorithm developed by Young et al.29. Comparison of the solvent networks revealed that introduction of the 9-methyl group (compound 2) displaced an ordered water molecule in pocket A (Fig. 4E) whereas the 8-bromine (compound 3) replaced a different ordered water in pocket B (Fig. 4F). The large increases in binding affinity hence appeared to be connected to changes in binding site solvation in both cases, but were the result of different thermodynamic profiles.

### Prospective predictions for three nonpurine fragment series

To further challenge the MD/FEP method, calculations were extended to 12 fragment-sized nonpurine heterocycles of unknown affinity, which represented three series with varying levels of modeling difficulty (Fig. 5). The first two compound series were adenine-like and had substituents that explored pocket B. Compound 25, [1,2,4]triazolo[1,5-a][1,3,5]triazin-7-amine, was a substructure of the A2AAR antagonist ZM241385 (Fig. 1). Similar to the adenine derivatives described in previous sections, this scaffold represented a case with high confidence regarding the fragment binding mode and predictions were made for three compounds in this series (2527). The second scaffold, derivatives of [1,2,4]triazolo[1,5-a]pyridin-8-amine 28, was less similar to adenine, but was assumed to maintain hydrogen bond interactions with Asn253 (Fig. 5), and relative affinities were predicted for three 2-alkyl analogs (29–31) that probed pocket B. The third series of five variously substituted derivatives of N-(benzo[d]thiazol-2-yl)acetamide 32 was unrelated to adenine. This represented the most challenging series, as the binding mode of this scaffold was unknown. After the MD/FEP predictions had been completed, the 12 compounds were evaluated experimentally using radioligand binding assays, and Ki values were determined for the ligands that showed >50% displacement at 300 μM (Supplementary Table S2). The computational and experimental results for the three series of fragments are summarized in Fig. 5.

The relative binding affinities for substituted triazolo-triazin-amine derivatives 26 and 27 compared to unsubstituted 25 were calculated using the same protocol as for the series of adenine derivatives. The MD/FEP calculations predicted that both compound 26 (0.6 kcal/mol) and 27 (4.3 kcal/mol) had higher affinity than compound 25, which was also confirmed experimentally. Compounds 26 and 27 had Ki values of 78.5 and 20 μM, respectively, which were large improvements over compound 25 that only displayed 34% radioligand displacement at 300 μM. For the second series of analogs, the calculated relative binding free energy (ΔΔGcalc = 0.4 kcal/mol) correctly predicted that the 2-methyl substituted triazolopyridine 29 (31% at 300 μM) had higher affinity than compound 28, which showed close to negligible radioligand displacement at 300 μM (17%). However, the 2-ethyl and 2-isopropyl substituted compounds 30 (49% at 300 μM) and 31 (223 μM) were incorrectly ranked relative to compound 29 (31% at 300 μM). Both compounds 30 and 31 were predicted to be weaker than compound 29 by 0.9 and 2.2 kcal/mol, respectively. The discrepancies for triazolopyridines 30 and 31 were intriguing considering the close agreement with experiment for the adenine-like ligands. To understand the origin of these erroneous predictions, the MD simulation trajectories for the ethyl-substituted compound 30 and methoxy-substituted compound 21 were inspected visually. The main difference between compounds 30 and 21 was found to be the torsional angles of the substituents protruding into pocket B. For compound 21, the methoxy group primarily sampled angles that were within the plane of the adenine moiety, whereas the ethyl group did not align with the plane of the aromatic ring. Torsion angle scans for compound 30 using density functional theory (DFT) revealed large errors in the force field parameters both in the location of the minimum and the energy barrier height of the potential energy curve (Fig. 5E), whereas there was reasonable agreement between OPLSAA_2005 and DFT for compound 21 (Supplementary Figure 4). MD/FEP calculations were then repeated for compounds 30 and 31 using a force field term for the torsion that reproduced the DFT calculations. The calculated relative binding free energies (to compound 29) changed from −0.9 to  + 0.1 kcal/mol for compound 30 and from −2.2 to + 0.7 kcal/mol for compound 31. These shifts in calculated values resulted in accurate ranking of the two ligands relative to 29 (Supplementary Table S2).

For the third series of compounds, 2-acetamido-benzothiazole (compound 32) was the core scaffold and had a Ki value of 79 μM. 4-Hydroxy-N-(4-methoxy-7-morpholinobenzo[d]thiazol-2-yl)-4-methylpiperidine-1-carboxamide (tozadenant), a compound that has been in clinical trials for the treatment of Parkinson’s disease30, could essentially be considered as a superstructure of this fragment. However, tozadenant could not be accommodated in the crystal structure used for the adenine derivatives in a manner that allowed for hydrogen bonding with residue Asn253. Hence, prior to experimental evaluation of this fragment series, an alternative binding site conformation based on a different A2AAR crystal structure24 was used. In this conformation, alternative side chain rotamers for His264 and Glu169 lead to a more open binding pocket, which could accommodate tozadenant and compound 32. This binding pose for compound 32 was found to be stable in MD simulations, and a representative snapshot was used as starting point for the FEP calculations (Fig. 5C). The effects of adding a methyl group at three different positions (4, 5 and 6) of the benzothiazole ring of compound 32 were evaluated computationally (Fig. 5C, compounds 3335). Improvements of affinity corresponding to 0.3 and 0.4 kcal/mol were predicted by MD/FEP for compounds 33 and 34 respectively, whereas a large loss of binding was obtained for compound 35. The predictions for compounds 33 and 34 agreed reasonably well with the subsequently determined 7-fold increases of affinity. Compound 35 was the weakest ligand of the three analogs with only a two-fold increase of affinity, but the MD/FEP calculations had predicted a loss of binding free energy in this case (2.7 kcal/mol). To further optimize compound 33, MD/FEP calculations were carried out for the 4-methoxy-substituted compound 36. The experimentally determined 44-fold increase of affinity (corresponding to 2.3 kcal/mol, Fig. 5F) compared to compound 32 was partially captured by the MD/FEP calculations, which predicted a 0.8 kcal/mol improvement in binding free energy. Hence, whereas the direction of the change in binding affinity was correct, the magnitude of the improvement in affinity was underestimated. To investigate if prediction accuracy could be improved by increasing sampling, we retrospectively extended the simulations by doubling the production time for the transformation between compounds 36 and 32, which resulted in improved agreement with experimental data (ΔΔGbind = 1.3 kcal/mol).

## Discussion

The focus of this work was to evaluate using MD simulations in combination with free energy calculations as a tool for fragment optimization. Three key results emerged from calculations of relative binding affinities for fragments ligands of the A2AAR, a GPCR that has been intensively studied as a drug target14. First, there was a strong correlation between calculated and experimental relative binding free energies for ligands based on an adenine scaffold. Remarkably, the direction of the shift in binding free energy was correctly predicted for all of the adenine derivatives that had an absolute experimental free energy change >0.5 kcal/mol. Second, the potential of MD/FEP calculations to assess alternative binding modes and predict thermodynamic signatures of fragment binding was demonstrated, which could be used to tailor ligand properties during optimization. Finally, prospective predictions for three compound series and evaluation of these in pharmacological assays highlighted opportunities and challenges for the use of MD/FEP calculations in FBLD.

The potential of MD/FEP to guide fragment optimization was clearly demonstrated by the excellent results obtained for the series of adenine derivatives22, 23. Substantial changes in affinity could be achieved by introduction of substituents in the C8- and N9-positions of this scaffold. These effects were not obvious by visual inspection of the complexes and changes in affinity did not correlate with trivial descriptors such as heavy atom count. The fact that addition of diverse substituents to the same subpocket improved binding affinities likely reflects a complex interaction network involving structural water molecules, polar and non-polar side chains. These effects were accurately captured by the MD/FEP calculations, but not by docking in combination with an empirical scoring function. As previously demonstrated by Warren et al., empirical scoring functions are more suitable for screening of large chemical databases to prioritize compounds for experimental testing rather than ranking closely related ligands by affinity31. The improved accuracy for MD/FEP may be due to explicitly taking into account water molecules, induced fit effects and associated entropic contributions to the binding9. The performance of docking scoring functions can also be further improved for specific targets by considering the effects of specific water molecules22, 32, 33, but such protocols may not be readily transferred to other ligand series or targets. Consideration of the enthalpic and entropic components of the binding free energy has recently been suggested to be an important metric to guide fragment optimization8. In addition to improved predictions of relative affinities, MD/FEP calculations make it possible to characterize the driving forces behind a change in free energy, which has previously been applied successfully to study ion hydration28 and enzyme catalysis34. In this work, we used the same approach to investigate the large differences in binding observed for two compound pairs from the series of adenine-based ligands. The affinity gains obtained for a substituent in the C8-position of adenine were found to be associated with a large increase of entropy and displacement of a binding site water molecule in pocket B, which appeared to be a classic example of the hydrophobic effect. Interestingly, the same hotspot has previously been identified based on MD-derived maps of the solvent network in the A2AAR binding site35, 36. In contrast, the addition of methyl substituent in the N9-position, which involved displacement of a water molecule from pocket B, led to a decrease of entropy and the improvement of the binding affinity was instead driven by enthalpy. Large improvements of affinity due to the addition of a single heavy atom, which has been referred to as the “magic methyl” effect, may hence have completely different thermodynamic origins. Although the predictions of the entropy and enthalpy contributions to the relative binding free energies will need to be further tested by comparison to experimental data, our results suggest that MD/FEP is not limited to guiding affinity optimization, but can also be used to tailor the thermodynamic profile of ligands.

Application of the MD/FEP technique to fragment-sized molecules has several advantages from a methodological standpoint. Molecular mechanics force fields are likely more accurate for fragment- than for drug-like compounds as parameters are typically developed based on fragment-sized molecules37. Furthermore, it should be more feasible to reach convergence of the free energy calculations for fragments as such molecules typically have fewer degrees of freedom than drugs. Comparison of our results to a recent study that applied the MD/FEP technique to two series of lead-like A2AAR ligands supports this idea38. The lower correlation with experiment obtained for two series of adenine-derived ligands may reflect that these were of lead-like size and interacted with the flexible extracellular loops whereas the fragments considered in this work had limited conformational flexibility and extended into a relatively rigid TM region. Interestingly, two recent benchmarks of binding free energy calculations for a large number of soluble targets showed a similar trend10, 39. It should be noted that access to information regarding the binding mode for a representative ligand was likely a key contributor to the accuracy of the MD/FEP calculations in all cases. Hence, if a high-resolution structure of a representative complex is available, MD/FEP calculations can be a valuable technique for ligand optimization and the approach appears to be particularly suited for fragments.

The prospective predictions carried out for three different fragment series revealed potential pitfalls for the use of MD/FEP to predict ligand binding affinities. In this case, the compounds had affinities in the high micromolar to millimolar range, which closely mimicked the scenario encountered in fragment-to-lead optimization3. In agreement with the results obtained for the adenine derivatives, the predictions were excellent for the fragments derived from a ligand co-crystallized with the A2AAR. The second fragment series illustrated the importance of high quality torsional force field parameters for small molecule ligands. The MD/FEP calculations based on OPLSAA_2005 parameters were not in agreement with experimental data, whereas a DFT-derived torsional potential resulted in correct ranking of the compounds. This result suggests that force field parameters should be used with caution even for fragment-sized molecules. As we focused only on congeneric series of compounds with mainly non-polar substituents, the performance of the partial atomic charges from OPLSAA_2005 was not assessed in this work. For example, consideration of different heterocyclic compounds could involve large changes in charge distribution, which may not be accurately represented by empirical partial charges. Fortunately, as fragments will have a small number of atoms and rotatable bonds, torsions and partial charges could in principle be derived using ab initio methods prior to the MD/FEP calculations to further enhance modeling accuracy. In line with these ideas, particular focus was put on improvements of torsional potentials and partial atomic charges in the recently released OPLS3 force field for organic molecules40. The last series of fragments illustrated challenges associated with lack of crystal structure information regarding fragment binding modes. Modeling of the binding mode involved consideration of several binding site conformations and required expert knowledge regarding ligand recognition by the target. Encouragingly, the direction of the change in binding free energy was correctly predicted in three out of four cases, but the magnitude of the shift in affinity was not always captured. Considering the many uncertainties involved in modeling of fragment binding modes, fragment optimization in the absence of a crystal structure should be considered to be very challenging. In these instances, the use of metadynamics41 and MD/FEP in combination with mutagenesis studies42, 43 to identify ligand binding modes hold promise. Binding modes that reproduce the initial structure-activity relationships could be used to guide compound selection in the following rounds of optimization.

The major advances made in molecular and structural biology for GPCRs44, 45 make it possible to apply FBLD to numerous targets of therapeutic interest. In the case of the A2AAR, fragment screening against stabilized receptor constructs by biophysical methods19, 20 and computationally using molecular docking21 have led to the discovery of diverse starting points for development of lead compounds. The determination of multiple high-resolution crystal structures of GPCRs in complex with fragments24, 46 provides exciting opportunities to apply computational methods in FBLD for GPCRs. Our results demonstrate that the MD/FEP approach can contribute to efficient optimization of fragment hits, which is key for successful use of FBLD in drug development. The combination of molecular docking screening for fragment identification and efficient ligand optimization via MD/FEP has the potential to become a powerful addition to the toolbox of methods used in fragment-based drug discovery.

## Methods

### MD/FEP calculations

The MD simulations were performed using a high-resolution crystal structure of the A2AAR (PDB accession code: 4EIY, 1.8 Å)17. In a first step, a hydrated 1-palmitoyl-2-oleoyl phosphatidylcholine (POPC) membrane bilayer was first equilibrated around the A2AAR structure with periodic boundary conditions using the 4.5.5 version of GROMACS47. These simulations were setup using the GPCR-ModSim protocol48 and the OPLS all atom (OPLSAA) force field37, TIP3P waters49, and Berger lipid parameters50. All protein atoms were tightly restrained to their initial coordinates and the hydrated membrane was equilibrated for a total of 40 ns at 300 K. All MD/FEP calculations were carried out starting from the membrane equilibrated A2AAR system using the program Q51 with the same force field. Ligand parameters were obtained using the OPLSAA_2005 version implemented in the program hetgrp_ffgen (Schrödinger, LLC, New York, NY, 2017). The simulations were carried out at 310 K in a sphere of 18 Å radius centered on the ligand. All protein, water, and ligand atoms within 18 Å of the center of the sphere were explicitly included in the simulations. Atoms close to the sphere edge were restrained to their initial coordinates and atoms beyond the sphere edge were excluded from nonbonded interactions. Asp, Glu, Lys, and Arg residues within 15 Å of the sphere center were protonated according to their most probable states at pH 7 and ionizable residues closer to the sphere edge were set to their neutral state. The protonation states of the histidines in the binding site were set by manual inspection. His278, His250, and His264 were protonated at Nδ, Nε, and both nitrogen positions, respectively. The series of 2-acetamido-benzothiazole derivatives (compounds 32–36) were simulated with conformations of His264 and Glu169 obtained from an alternate crystal structure of the A2AAR24. In these simulations His264 was protonated at the Nδ position. The SHAKE52 algorithm was applied to constrain all solvent bonds and angles and the water molecules at the sphere surface were subjected to radial and polarization restraints according to the SCAAS model51, 53. A nonbonded cutoff of 10 Å was used for all atoms except the ligand, for which no cutoff was applied. Long-range electrostatic interactions were treated with the local reaction field method54. The time step was set to 1 fs and nonbonded pair lists were updated every 25 steps. In the simulations of the ligands in aqueous solution, the compound was positioned in the center of the sphere and a weak harmonic restraint was applied to a central atom (e.g. C5 of the adenine scaffold) to prevent it from approaching the sphere edge. Clustering of the water network in the binding site was carried out based on a simulation of 8 ns with the receptor and ligand restrained to their starting coordinates, from which 8000 snapshots were extracted and processed using the algorithm of Young et al.29.

The relative binding free energy for a pair of compounds was calculated in multiple steps using MD/FEP: (i) The transformation of partial charges and (ii) combined transformation of Lennard-Jones (LJ) and parameters involving covalent bonds in several MD/FEP calculations. If multiple heavy atoms were annihilated, a separate MD/FEP calculation was carried out to remove these in a step-wise manner. A soft-core potential was introduced for the atom in a first step, followed by removal of the resulting van der Waals potential42, 55. The force field parameters describing angles, bonds, and improper torsions were retained for annihilated atoms whereas the torsional potential was removed in some cases to improve convergence. The total free energy was calculated as the sum of the results obtained in each step. Each MD/FEP calculation was divided into n intermediate states that were equilibrated separately. The potential (U m ) defining each state was a linear combination of energy functions describing the start- (A) and endpoint (B) of the transformation

$${U}_{m}=(1-{\lambda }_{m}){U}_{A}+{\lambda }_{m}{U}_{B}$$
(1)

where λm varies from 0 to 1. The FEP calculations involving partial charges were performed using 11 states whereas the number of λ values used to transform LJ and bonded parameters varied from 40 to 122 steps. Each receptor-ligand complex was equilibrated for 475 ps at each λ value. In this simulation, harmonic restraints on the protein and ligand atoms were released in several steps and the temperature was gradually increased to 310 K. The equilibration step was followed by 250 ps of unrestrained simulation, from which potential energies were extracted. The same transformation was carried out in a spherical water droplet. In this case, the system was equilibrated for 350 ps, followed by 100 ps of unrestrained simulation. The free energy difference between states A and B was calculated by summing up the free energy differences of the n intermediate states using

$${\rm{\Delta }}{G}_{A\to B}^{FEP}=-kT\,\sum _{m=1}^{n-1}\,\mathrm{ln}\,{\langle {e}^{-({U}_{m+1}-{U}_{m})/kT}\rangle }_{m}$$
(2)

where 〈…〉 m represents an ensemble average on the potential U m , which is calculated from the MD simulations13. Three replicates, which were initiated from different starting velocities for the atoms in the system, were performed for each state and these were exponentially averaged in calculations of the free energy. The uncertainty of a transformation was quantified as the difference in free energy obtained by applying the FEP formula in the forward and reverse direction and was optimized by increasing the number of λ values or simulation length until convergence was obtained. The uncertainty of a calculated relative binding free energy was estimated as the standard error of the mean of three independent trajectories.

The enthalpy (ΔΔHbind) and entropy (ΔΔSbind) contribution to the relative binding free energy were calculated from a relationship analogous to the van’t Hoff equation28. Each FEP transformation was carried out at 5 K intervals between 270 and 330 K to calculate the temperature dependence of the relative binding free energy (ΔΔGbind). The entropy and enthalpy components were then calculated from the slo﻿pe and ﻿interc﻿ept of the relation between ΔΔGbind/T and 1/T:

$$\frac{{\rm{\Delta }}{\rm{\Delta }}{G}_{bind}}{T}=\frac{1}{T}{\rm{\Delta }}{\rm{\Delta }}{H}_{bind}-{\rm{\Delta }}{\rm{\Delta }}{S}_{bind}$$
(3)

### Molecular docking and ALogP calculations

The docking calculations were carried out with GLIDE25 (version 6.0, Schrödinger, LLC, New York, NY, 2017) using the standard precision (SP) protocol and default settings. The same A2AAR crystal structure (PDB code 4EIY17) and protonation states for the ionizable residues as in the MD simulations were used. Prior to docking, all non-protein atoms (e.g. water molecules) were removed and the binding site was defined based on the co-crystallized ligand. ALogP values were calculated using Maestro (Release 2017–1: Maestro, Schrödinger, LLC, New York, NY, 2017).

### Torsion scans

Potential energy profiles for the ethyl and methoxy substituents of compounds 21 and 30 were calculated using Gaussian0956. The torsional scans were performed using B3LYP/DFT with the 6-311++G(d,p) basis set57,58,59,60 and a full geometry optimization was carried out at each point. The corresponding force field potential energies were calculated based on the geometries obtained from the DFT optimization.

### Radioligand binding and functional assays

Radioligand binding assays for compounds 2536 were performed as previously described61 using membrane preparations from Chinese hamster ovary (CHO) or human embryonic kidney (HEK)293 cells stably expressing the human A1, A2A or A3AR. The following radioligands: antagonist [3H]DPCPX (0.5 nM); antagonist [3H]ZM241385 (1.0 nM); and agonist [125I]AB-MECA (0.2 nM) were used for A1, A2A, and A3ARs, respectively. Binding parameters were calculated using Prism 6 software (GraphPAD, San Diego, CA, USA). IC50 values obtained from competition curves were converted to Ki values using the Cheng-Prusoff equation. Data were expressed as mean ± standard error. The radioligand binding assays for compounds 18, 20, 21, and 23 were performed as previously described23. Each Ki value was determined from at least three independent experiments.

Four adenine derivatives (compounds 5, 19, 22, and 23) were evaluated in functional assays. CHO cells, stably transfected with the human A2AAR and transiently with firefly luciferase biosensor, were grown adherently and maintained in Dulbecco’s Modified Eagles Medium with nutrient mixture F12 (DMEM/F12 with phenol red), supplemented with 10% FBS (Fetal Bovine Serum), 100 U/mL penicillin, 100 µg/mL streptomycin, 2.5µg/mL amphotericin, 1 mM Sodium pyruvate, and 0.1 mg/mL Geneticin (G418) at 37 °C, and aerated with 5% CO2: 95% O2. Cells were harvested in CO2 independent media and counted in a Neobauer chamber. The desiderate cell number was incubated in equilibration medium containing a 3% v/v GloSensor cAMP reagent stock solution, 10% FBS and 87% CO2 independent medium. After 2 hours of incubation cells were dispensed in wells of 384 well plate and when a steady-state basal signal was obtained, different concentrations of antagonists were added. After 10 min, 1 µM of 5′-N-ethylcarboxamidoadenosine (NECA), the reference agonist, was injected. Responses were expressed as percentage of the Maximal Relative Luminescence Units (RLU). Inhibition-response curves were fitted by a non-linear regression using a Prism 4.0 program (GraphPad Software, San Diego, CA, USA).