Cosolvent simulations with fragment-bound proteins identify hotspots to direct lead growth

Abstract

In drug design, chemical groups are sequentially added to improve a weak-binding fragment into a tight-binding lead molecule. Often, the direction to make these additions is unclear, and there are numerous chemical modifications to choose. Lead development can be guided by crystal structures of the fragment-bound protein, but this alone is unable to capture structural changes like closing or opening of the binding site and any side-chain movements. Accounting for adaptation of the site requires a dynamic approach. Here, we use molecular dynamics calculations of small organic solvents with protein-fragment pairs to reveal the nearest “hotspots.” These close hotspots show the direction to make appropriate additions and suggest types of chemical modifications that could improve binding affinity. Mixed-solvent molecular dynamics (MixMD) is a cosolvent simulation technique that is well established for finding binding “hotspots” in active sites and allosteric sites of proteins. We simulated 20 fragment-bound and apo forms of key pharmaceutical targets to map out hotspots for potential lead space. Furthermore, we analyzed whether the presence of a fragment facilitates the probes’ binding in the lead space, a type of binding cooperativity. To the best of our knowledge, this is the first use of cosolvent MD conducted with bound inhibitors in the simulation. Our work provides a general framework to extract molecular features of binding sites to choose chemical groups for growing lead molecules. Of the 20 systems, 17 systems were well mapped by MixMD. For the three not-mapped systems, two had lead growth out into solution away from the protein, and the third had very small modifications which indicated no nearby hotspots. Therefore, our lack of mapping in three systems was appropriate given the experimental data (true-negative cases). The simulations are run for very short timescales, making this method tractable for use in the pharmaceutical industry.

Graphical Abstract

1. INTRODUCTION

In the last twenty-five years, fragment-based drug discovery (FBDD) has emerged as a viable alternative to high-throughput screening of very large molecule libraries.¹ Fragments tend to contain roughly 15 non-hydrogen (or heavy) atoms,² and this limited size allows for a smaller molecular library that is more tractable for screening. In FBDD, chemical groups are added iteratively to the fragment to improve affinity and turn it into a lead molecule. Size-dependent metrics such as ligand efficiency are used to monitor the importance of each chemical group added onto the fragment. Around 150 fragment-to-lead growth programs were analyzed to find out the median of heavy atoms counts (HAC) in fragment and lead molecules, which is 15 and 28, respectively.² On average, around 13 heavy atoms are added over the course of turning a fragment into a lead molecule. Through FBDD, four drugs have been developed,^3–6 and many are in the pipeline.

Given their small size, fragments usually only fill part of their binding sites, and there is usually space to provide more ligand-target contacts. These additional subsites are referred to as hotspots. Hotspots can be identified through the cosolvent simulation, where we simulate the protein molecule with water and small, organic molecules.⁷ An array of cosolvent simulation methods – such as MDmix,⁸ SILCS,⁹ MixMD^10,11 – are available. Here, we use MixMD, our in-house developed cosolvent method, where we simulate proteins in water with 5% v/v miscible probe molecules. Previously, we have shown that hotspots in active sites, allosteric sites,¹¹ and cryptic sites¹² can be mapped by MixMD.

Fragment to lead generation is an evolution where the molecule is continuously optimized with respect to its potency, physicochemical properties, and ligand efficiency. Typically, lead molecules can be grown in multiple directions, and the choice of chemical groups will dictate its potency. However, all possible lead compounds cannot be synthesized and evaluated. The infinite choices of molecules could be narrowed down with the help of MixMD. The interaction between probes and the protein can guide the choice of chemical groups and the direction of lead growth.

In the present work, we have simulated fragment-bound systems to identify the nearby hotspots for lead space. The choice of simulating fragment-bound systems was made because proteins might go through conformational changes¹³ upon binding a fragment, and those changes might further facilitate the probes’ binding in the lead space. This favorable binding of probe molecules in the presence of a fragment molecule is referred to as cooperativity. We have also simulated the apo states of these systems to identify whether all the hotspots can be identified without fragments to lead the way.

Here, we simulated 20 fragment-bound systems with the aim of mapping the nearby druggable hotspots. We simulated the protein in water with the following probes: pyrimidine (PYR) to map sites through aromatic or hydrophobic interactions, acetonitrile (ACN) to map sites through generic interactions, and isopropanol (IPA) to map sites through hydrogen bonds (HB) to the hydroxyl group or hydrophobic interactions with the isopropyl moiety. Below, we discuss our ability to map hotspots in 17 of 20 systems. Of course, the question arises whether the fragments were needed in the simulations to find the hotspots, so we also simulated the apo states of each protein to identify cases with cooperativity of binding the probes. This work should shed light on the utility of fragment-bound systems for mapping hotspots and directing fragment to lead growth.

2. METHODS

2.1. Structure preparation and parameter generation.

A total of 20 systems was selected to investigate the mapping of lead space, which was obtained from fragment-to-lead publications published from 2015 to 2018.^14–17 The systems with a coordinated metal ion, a cofactor, a peptide linker, a modified amino acid, or a dimer with more than 500 amino acids were excluded from this study. Table 1 contains the target names, the PDB¹⁸ identification codes of the fragment- and lead-bound protein complexes, and the activity change of going from fragment to lead for each system. Systems 1, 3, and 14 (in Table 1) exist in the dimeric form in fragment-bound complexes. A different multimeric state for System 1 (Coagulation factor XIA) was generated using PyMOL, which had better contact space between monomeric units than the biounit available in the PDB database.¹⁸ The water molecules, ions, and any other molecule except for the fragment were removed from the structure. The hydrogens were added using the Protonate 3D module in MOE,¹⁹ including titratable residues. The amide side chains in asparagines and glutamines and the imidazole ring of histidines were flipped as needed to optimize the HB interactions with the nearby residues.

Table 1.

List of all 20 systems investigated for fragment-to-lead mapping. BACE1 and PAF-AH are on the list twice for exploring different binding subsites with independent lead-development campaigns.

No. and ref	Fragment PDB ID	Lead PDB ID	Protein name	Frag to Lead ΔHAC	ΔAffinity (Frag/Lead)a
135	4CR5	4CRF	Coagulation factor XIA (FXIa)	27	140000
236	5FNQ	5FNU	Kelch-like ECH-associated protein 1 (KEAP1)	27	>770000
337	4A9H	6FFG	Bromodomain-containing protein 2 (BRD2)	24	2100
438	5A50	5A83	ATAD2 bromodomain (ATAD2)	24	>7500b
539	3R59	4J5C	Cyclophilin D (CypD)	22	>25000b
640	5YE8	5YEA	Platelet-activating factor acetylhydrolase (PAF-AH)	20	71000b
741	6AXQ	6AY3	CREB-binding protein (CBP)	19	10000
842	5JAH	5LZ9	Platelet-activating factor acetylhydrolase (PAF-AH)	17	> 8300
943	3UDH	3UDQ	Beta-secretase 1 (BACE1)	16	1400
1044	5T4U	5T4V	Bromodomain and PHD finger-containing protein 1 (BRPF1)	15	610b
1145	6D9X	6DAS	WD repeat-containing protein 5 (WDR5)	15	>320000
1246	3VBW	3VBQ	Serine/threonine-protein kinase pim-1 (S/T Pim-1)	15	12766
1347	5UEP	5UER	Bromodomain-containing protein 4 (BRD4)	13	760
1448	5F1J	5EYR	HTH-type transcriptional regulator EthR (EthR)	10	140 b
1549	4ZSM	4ZSP	Beta-secretase 1 (BACE1)	10	2200
1650	4ZLY	4Z3V	Bruton’s tyrosine kinase (BTK)	10	875
1751	5F25	5F1H	Bromodomain-containing protein 9 (BRD9)	8	650
1852	6F20	6F22	MutT Homologue 1 (MTH1)	7	12000a
1953	5K03	5K0L	Catechol O-methyltransferase (COMT)	7	920
2054	5MW3	5MW4	Dot1-like protein (Dot1L)	4	>2400000

^aAffinity values may be ratios of IC₅₀ or K_d, taken from the fragment-to-lead reviews^14–17 or the Binding MOAD database.⁵⁵

^bAffinity of a closely related compound.

For the system setup and the dynamics simulations, AMBER18²⁰ was used. To obtain the parameters of each fragment molecule, the geometry optimization of the fragment molecule was performed using the B3LYP/6-31G(d)^21–24 method in the Gaussian 16 program.²⁵ The electrostatic potential charges were calculated by using the Merz-Kollman radii.²⁶ Later, the Restrained Electrostatic Potential (RESP) charges of the fragment molecule were obtained from the antechamber module.²⁷ The remaining simulation parameters were obtained using the second-generation Generalized AMBER force field (GAFF2).²⁸

2.2. MixMD simulation protocol.

The AMBER ff14SB²⁹ and GAFF2 force fields were used to generate parameters for protein molecules and ligands, respectively. The probe molecules were wrapped around the protein with a 7.0 Å layer and a packing fraction of 0.5. Subsequently, the water molecules were added to obtain the 5% v/v concentration of probes with respect to water molecules. The sodium or chloride ions were added as necessary to neutralize each system. The PYR, ACN, and IPA probes were used in our work to identify key aromatic, hydrophobic, and HB interactions as these probes were successfully used in previously published works.^11,12,30

The minimizations were performed in two steps, and in both minimization steps, 2500 steps of the steepest-descent and 7500 steps of the conjugate gradient were carried out. In the first minimization step, both the backbone and side chains of the protein and the fragment were restrained with 10.0 kcal/mol-Ų, which allowed only the probes and water molecules to minimize. In the second minimization step, only the protein backbone and the fragment molecule were restrained with 10.0 kcal/mol-Ų, which allowed the side chains of the protein to minimize along with the probes and water molecules. The simulation box was heated from 10 K to 300 K for 200 ps with a restrain of 5.0 kcal/mol-Ų on the protein’s backbone. The equilibration steps were performed in five steps where we lowered the restraint on the protein backbone linearly as we progressed. In the last equilibration step, the protein was simulated with no restraint. Finally, 20 ns of production simulation was carried out, using constant pressure and temperature (NPT) without any restraint on the protein backbone or ligand. The Berendsen barostat³¹ and Andersen thermostat³² were used. The SHAKE algorithm³³ was used to constrain the covalent bonds to hydrogen atoms at their equilibrium length, allowing us to simulate with a 2-fs time step.

The cosolvent simulations were carried out for all three probes separately, and 10 independent simulations of 20 ns were performed for each probe. The choice to use short simulations was deliberate to underscore the speed of MixMD mapping, and a total 30 simulations were carried out for each fragment-bound and apo form of a system. The simulation analysis was performed on the last 10 ns of each run. The root mean square deviation (RMSD) of the protein backbone was computed to check whether the protein remains stable during simulation or not. Radial distribution function (RDF) calculations were performed for each system to verify there was no probe aggregation, and the probes’ movement across the water box were visualized. To identify the probes’ binding sites, a grid was overlaid on the simulation box with 0.5 Å spacing. The grid command from the Cpptraj³⁴ program was used to compute the raw bin count of each probe’s center of mass (COM) in each voxel, which reflects the number of snapshots where a probe stays bound. The raw bin counts were normalized to Z-scores by subtracting with the average COM occupancy and dividing that by the standard deviation, where average and standard deviation were computed from all binned data across the whole grid. Z-score is measured in units of standard deviations from the mean (sigma values). More details about the grid analysis can be found here.^11,30 Our focus is to obtain the probe’s binding hotspots around the fragment, as our work is focused on mapping the lead space.

2.3. Convergence:

As per the protocol, we performed 10 independent runs of MD for 20 ns. This could raise the question of whether we sampled long enough to obtain accurate mapping of conformational space to reliably predict the hotspots for lead growth. To answer this question, we extended the MixMD simulations from 20 ns to 100 ns for systems 8, 9, and 13. We used the last 50 ns (i.e., 50-100 ns) to perform the grid analysis to obtain the hotspots around the fragment molecule. Figure S1.A–C(i) displays the hotspots from 10-20 ns simulation window, while S1.A–C(ii) displays the hotspots from 50-100 ns simulation window. Excellent agreement is seen between the maps. Hotspots have remained in the same location or shifted only one grid point away (0.5 Å). Figure S1 also shows that all the sites exhibit similar Z-scores in the 50-100 ns simulation window. This shows that running only 10 independent runs of 20 ns is sufficient to identify lead growth hotspots reliably.

3. RESULTS AND DISCUSSION

FBDD starts with finding a fragment molecule and then adding chemical groups to increase its potency while monitoring the ligand efficiency (affinity per HAC). There are a myriad of choices of chemical groups and directions to grow the fragment. Here, we show that MixMD has the ability to direct those modifications. We carried out cosolvent simulations of 20 fragment-bound proteins and identified the location of nearby probe-binding hotspots. These hotspots show excellent agreement with lead compounds, which demonstrates the relevance and druggability of those MixMD hotspots.

Table 1 lists all 20 systems and the difference in HAC between the chosen lead and fragment molecules. Theoretically, the larger the change in HAC, the more lead space will be available for probes to map. We have divided the systems into three categories with respect to the ΔHAC: a large change is ΔHAC > 20, an intermediate change is 10 < ΔHAC ≤ 20, and a small change is ΔHAC ≤ 10. Below, we will discuss large changes first (5 systems), then intermediate changes (8 systems), and lastly the small changes (7 systems).

We searched all crystal structures associated with each UniProt ID of the proteins, and we used them to identify lead-space binding. For a successful mapping of the lead space, the Z-score of the COM maxima must be of 50 sigma or higher, and the centers of the hotspots must be within 2 Å of any heavy atom of lead molecule. The Z-score of each hotspot in the lead space is labeled and presented in Table 2. In the figures, the probe-binding regions are represented through a mesh grid, and the occupancy maxima in the hotspots are shown in small spheres. In all of our figures, the hotspots for PYR, ACN, and IPA probes are shown in purple, orange, and blue, respectively. In the following sections, we discuss the interaction responsible for each hotspot in the lead space, and we note if nearby hotspots exist, which could be important for lead design. We listed the residues crucial for facilitating only PYR and IPA probe’s binding. The PYR probe forms aromatic interactions and hydrophobic interactions while the IPA probe forms hydrophobic interactions and its hydroxyl- group can form an HB. However, ACN is a small-sized probe and does not make specific interactions to enlist residues. The ACN probes primarily form hydrophobic interaction, but it also wanders around the binding site. Therefore, we did not list the residues facilitating the binding for the ACN probe. Given the molecular features of the probe interactions, one could imagine using them as pharmacophore constraints in virtual screening to identify appropriate second-generation molecules.

Table 2.

Z-scores for each probe obtained from hotspot mapping in the lead space for both fragment-bound and apo forms of the protein complexes. The list of systems is arranged in decreasing order of AHAC between lead and fragment molecules.

No.	Protein name	Fragment PDB ID	Lead-space mapping in fragment-bound systems			Lead-space mapping in apo systems
			PYR	ACN	IPA	PYR	ACN	IPA
1	Coagulation factor XIA (FXIa)	4CR5	101, 58	54	236, 53	76	70	119, 54
2	Kelch-like ECH-associated protein 1 (KEAP1)	5FNQ	221, 178	82	175, 178	-	66, 51	74
3	Bromodomain-containing protein 2 (BRD2)	4A9H	196	130	309	71	63	129
4	ATAD2 bromodomain (ATAD2 BRD)	5A50	-	64	-	-	-	-
5	Cyclophilin D (CypD)	3R59	129	384	441	222	328	371
6	Platelet-activating factor acetylhydrolase (PAF-AH)	5YE8	129	-	75, 50, 133	68	123	60, 278
7	CREB-binding protein (CREBBP)	6AXQ	135, 90	104, 73	90, 84, 65, 63	50, 59	-	-
8	Platelet-activating factor acetylhydrolase (PAF-AH)	5JAH	245, 69	54	54, 112	191	-	139
9	Beta-secretase 1 (BACE1)	3UDH	128	-	153, 59	160, 63	-	-
10	Bromodomain and PHD finger-containing protein 1 (BRPF1)	5T4U	-	-	-	58	-	-
11	WD repeat-containing protein 5 (WDR5)	6D9X	327, 194	171, 124	94, 125	114, 214	150	188, 85
12	Serine/threonine-protein kinase pim-1	3VBW	-	51, 54	59	-	126	-
13	Bromodomain-containing protein 4	5UEP	381, 63	223, 85	413	112		72, 63
14	HTH-type transcriptional regulator EthR	5F1J	166	97	222	107	102	68
15	Beta-secretase 1 (BACE1)	4ZSM	315	328	400	283	100	144
16	Bruton’s tyrosine kinase (BTK)	4ZLY	156	-	-	-	-	-
17	Bromodomain-containing protein 9 (BRD9)	5F25	-	-	-	-	-	-
18	MutT Homologue 1 (MTH1)	6F20	92	126, 60	71	-	-	-
19	Catechol O-methyltransferase (COMT)	5K03	-	-	-	-	-	-
20	Dot1-like protein (Dot1L)	5MW3	75	-	-	-	-	-

Also, we generated the apo form of each protein by deleting the fragment. We compare the mapping of lead space in fragment-bound and apo form of protein molecules. If probes only map the lead space in fragment-bound form, not in the apo form, then presence of fragment likely assists the binding of probes in the lead space, a type of cooperativity. The fragment molecules may induce conformational changes or optimize the HB network to allow better mapping of lead space. We found three systems that showed cooperative binding of probes where the fragment was needed to map the lead-space hotspot.

3.1. Mapping of lead space in systems with large change (ΔHAC > 20).

3.1.1. Cyclophilin D (CypD).

CypD catalyzes the cis-trans isomerization of proline’s peptide bond.³⁹ CypD regulates the mitochondrial permeability transition pore (mPTP) opening, which retains calcium ions and reactive oxidative species that induces cell death.^56–58 Therefore, the opening of mPTP can be targeted to prevent cell death under oxidative stress or facilitate cell death to control irregular cancer cell division.⁵⁸ Ahmed-Belkacem et al. designed a non-peptidic inhibitor through the FBDD approach, and this inhibitor shows antiviral activity against HCV, HIV, and HCoV-229E coronavirus.³⁹ The lead molecule consists of amino-benzyl, pyrrolidine, and thio-benzyl groups (Figure 1Ai). The amino-benzyl group from the lead molecule overlaps with the fragment; therefore, the probes can only map the pyrrolidine and thio-benzyl groups. The binding space of the lead molecule is spread into two neighboring cavities. In the first cavity, the amino-benzyl group binds, while pyrrolidine and thio-benzyl groups bind in the second cavity, and both of these cavities are connected through a linker made up of a urea moiety. From the simulations of the fragment-bound CypD complex, we observe that all three probes bind into the second cavity. The PYR, ACN, and IPA probes map the lead space with Z-scores of 129, 384, and 441 sigma, respectively (Figure 1Aii, Table 2). The second cavity is surrounded by Phe60, Gln63, Met61, Phe113, and Leu122 residues. The binding of the PYR probe is stabilized by Phe60 and Phe113 through π-π stacking interactions and by Leu122 through hydrophobic interactions. The binding of IPA probe is assisted by Leu122 and Phe113 through hydrophobic interactions. Moreover, the hydroxyl group of IPA probe forms a HB with the side chain of His126, Asn102, and Gln63. The PYR and IPA probes map the second cavity and extract the molecular features such as aromatic and hydrophobic interactions. This suggests that one could add either an aromatic or hydrophobic group to occupy the second cavity. The chosen lead molecule (shown in cyan) contains a thio-benzyl group (an aromatic group) to occupy the second cavity, which matches well with extracted molecular features by probes (Figure 1Aii).

Figure 1.

Mapping the lead space for (A) CypD, (B) KEAP1, (C) BRD2, and (D) FXIa. A-D (i) The two-dimensional representations of the lead molecules. A-D (ii) The fragment and lead molecules are shown in yellow and cyan, respectively. The hotspots mapping lead space where grid mesh represents the area occupied by the probes, and the sphere in the hotspot denotes the site of maximum occupancy. The probes PYR, ACN, and IPA hotspots are shown in purple, orange, and blue, respectively. The PDB identifiers for fragment- and lead-bound proteins are listed in Table 1.

3.1.2. Kelch-like ECH-associated protein 1 (KEAP1).

KEAP1 regulates the nuclear factor erythroid 2-related factor 2 (NRF2), which protects cells from oxidative damage by regulating the level of oxidative species.³⁶ Upon increase in oxidative species, NRF2 dissociates from KEAP1, moves to the cytoplasm, and produces antioxidant response elements, which increases the production of cytoprotective proteins.⁵⁹ The KEAP1-NRF2 pair has been shown to play a role in cancer and neurodegenerative disorders;⁵⁹ therefore, this pair is considered an important therapeutic target. Davies et al. performed fragment screening and found that three fragments bind contiguously in different subsites of the KEAP1 active site.³⁶ The binding of these fragments is assisted by Arg483, Tyr525, and Ser602 residues. The transformation of fragments to lead is done by merging the fragments while keeping their key interactions intact. This lead molecule can be divided into three moieties: benzothiazepine, benzotriazole, and phenyl group (Figure 1Bi). We simulated the fragment-bound KEAP1 system (PDB ID – 5FNQ). The fragment molecule overlaps with the phenyl group of the lead molecule. Therefore, probes can only map benzothiazepine and benzotriazole moieties. Both moieties contain an aromatic group, and PYR is successfully able to map benzothiazepine and benzotriazole moieties with Z-scores of 178 and 221 sigma, respectively (Figure 1Bii, Table 2). To remove any ambiguities in further discussion, we will use the probe name along with its Z-score with a hyphen such as PYR-178. The PYR-178 probe’s binding on benzothiazepine moiety is driven by π-π stacking interactions with Tyr334 and Phe577, respectively. Additionally, the PYR-221 probe is present on benzotriazole moiety, and it forms π-π stacking interaction with Tyr525. The IPA probe also maps both benzothiazepine and benzotriazole moieties with the Z-scores of 175 and 178 sigma, respectively. The IPA-175 probe hovers over the benzothiazepine moiety and forms hydrophobic interaction with Tyr334, Tyr572, Phe577, and alkyl part of Arg380 and Arg415 side chains. The hydroxyl- group of IPA-175 probe forms a HB with Asn414 and Ser602 side chains. The IPA-178 probe also maps the benzotriazole moiety and forms hydrophobic interactions with Tyr525, Ala556, and Tyr572. The KEAP1 binding site contains multiple aromatic residues such as Phe577, Tyr334, and Tyr525. The PYR and IPA probes were able to extract the aromatic and hydrophobic molecular features of the protein’s binding site, which matches well with the moieties of the lead molecules.

3.1.3. Bromodomain-containing protein 2 (BRD2).

Bromodomains bind to acetylated lysines attached to a histone tail to regulate the transcription. Law et al. chose a quinoline fragment and grew it into a lead molecule specific to BRD2.⁶⁰ The fragment-bound structure (PDB ID – 4A9H) is a dimer structure, while the chosen lead-bound complex (PDB ID – 6FFG) is a monomer. We aligned the lead-bound complex to each unit of fragment-bound dimer complex to identify the probes’ mapping. The lead molecule contains the tetrahydro-pyridine, dihydroquinoxaline, and hydroxy-methyl benzyl groups (Figure 1Ci). The dihydroquinoxaline group overlaps with the fragment, and that leaves the tetrahydro-pyridine and hydroxy-methyl benzyl groups to be mapped by probes. The PYR, ACN, and IPA probes map the hydroxy-methyl benzyl group with Z-score of 196, 130, and 309 sigma, respectively (Figure 1Cii). The PYR-196 probe forms π-π stacking interaction with Trp97 and the fragment with hydrophobic interaction with Pro98 and Ile162. Moreover, the PYR-196 probe forms pi-sulfur interaction with the side chain of Met165. The IPA-309 probe forms hydrophobic interaction with Trp97, Pro98, and Ile162. The PYR probe does map a nearby site to the tetrahydro-pyridine group shown by the mesh grid (in purple). But the hotspot maximum is within 3Å, not in the 2Å cutoff, which we chose as a criterion to classify the system as mapped. The PYR and IPA probes map the hydroxy-methyl benzyl group, which shows that both probes extracted correct molecular features for the BRD2 pocket.

3.1.4. Coagulation factor XIA (FXIa):

FXIa generates thrombin to increase the amplification phase of the coagulation process, but excess thrombin generation can result in thrombosis.³⁵ Therefore, FXIa inhibitors reduce the thrombin generation in the amplification phase without affecting its generation in the initiation phase. Fjellstrom et al. chose the quinoline fragment bound in the S1 pocket and turned it into a lead molecule occupying the S1, S1’, and S2’ sites.³⁵ We generated an alternate dimer structure as the deposited dimer structure in PDB database had few contacts between its monomeric units. The lead molecule contains hydroxy-quinoline, chloro-quinoline, and benzyl groups (Figure 1Di). The chloro-quinoline group overlaps with the fragment, allowing probes to map only the hydroxy-quinoline and benzyl groups. The lead-bound protein complex is a monomer and we overlaid it with each monomeric units of fragment-bound complex. Here, we are presenting the probes mapping from the first monomeric unit. The PYR, ACN, and IPA probes map the benzyl group with 101, 54, and 236 sigma values, respectively (Figure 1Dii). The PYR-101 hotspot makes π-π stacking interaction with His57 and hydrophobic interaction with Leu39 and Cys40-Cys58 disulfide bridge. The IPA-236 hotspot makes hydrophobic interaction with Leu39 and His57, and it may form an HB with backbone oxygen of His57. The PYR and IPA probes map the hydroxy-quinoline groups with 58 and 53 sigma values, respectively. The PYR-58 hotspot has π-π stacking interaction with Tyr143 and hydrophobic interactions with Leu39, Ile151, and Lys192. The IPA-53 hotspot forms hydrophobic interactions with Leu39, Ile151, Tyr143 and the hydroxyl group of IPA-53 forms an HB with the backbone oxygen of His38. The lead molecule branches out in two directions, and PYR and IPA probes were able to map aromatic groups in both directions.

3.2. Mapping of lead space in systems with intermediate change (10 < ΔHAC ≤ 20)

3.2.1. Platelet-activating factor acetylhydrolase (PAF-AH).

This protein hydrolyzes the ester bond of platelet-activating factor (PAF) at the sn-2 position and other oxidized phospholipids.⁶¹ The increase in PAF-AH level has been associated with many diseases such as cardiovascular disease, atherosclerosis, diabetic macular edema, and many others.^40,62 Woolford et al. carried out the virtual screening and found that fragments bind contiguously in three subsites.⁴² They transformed the fragments into the lead molecule through fragment merging and further increased its potency through structure-based optimization. The PDB identifier for fragment-bound and lead-bound complexes is 5JAH and 5LZ9, respectively. The lead molecule can be divided into phenyl group, γ-lactam, and sulfone ring (Figure 2Ai). We simulated the PAF-AH system with fragment bound in the active site, and the fragment overlaps with the γ-lactam ring of the lead molecule. Therefore, probes will only be able to map the phenyl group and sulfone ring of the lead molecule. The PYR probe maps the phenyl group with the Z-score of 245 sigma, and the PYR-245 hotspot’s binding is supported by multiple aromatic residues such as Phe110, Phe156, Tyr160, and Phe357 through π-π stacking interactions (Figure 2Aii). The PYR probe also maps the sulfone ring with the Z-score of 69 sigma (Figure 2Aii), and forms π-π stacking interaction with Trp298. The ACN probe maps the phenyl ring with a Z-score of 54 sigma. The IPA probe maps the phenyl ring with the Z-score of 112 sigma (Figure 2Aii), and the binding of IPA-112 hotspot is assisted by Leu107, Ala155, Leu159, and Ala355 through hydrophobic interactions. The IPA probe also maps the sulfone ring with the Z-score of 54 sigma (Figure 2Aii), and its binding is stabilized by Phe322, Trp298, and Tyr324. The lead space contains a phenyl group and sulfone ring, which is mapped with PYR and IPA probes with remarkably high Z-scores. This indicates MixMD can extract the molecular features correctly.

Figure 2.

Mapping lead space for (A) & (B) PAF-AH, (C) CBP, (D) BRD4, (E) S/T Pim-1, and (F) WDR5. The PDB identifiers for fragment-bound proteins (A) and (B) are 5JAH and 5YE8, respectively. A-F(i) Two-dimensional representations of the lead molecules. A-F(ii) lead-space mapping where grid mesh represents the area occupied by the probes, and the location of the hotspot sphere represents the site of maximum occupancy (labeled with Z-score). A-B(iii) Extra hotspots beyond the lead molecule for both PAF-AH systems. A(iii) Lead molecule from PDB ID - 5YEA (shown in magenta) overlaid with 5JAH-5LZ9 fragment-lead pair. B(iii) Lead molecule from PDB ID – 5LZ9 (shown in green) overlaid with 5YE8-5YEA fragment-lead pair. Coloring is the same as Figure 1.

Liu et al. carried out a fragment screen and choose a fragment with sulfonamide group.⁴⁰ Later, they grew the fragment into a lead molecule with a single-digit nM-level binding affinity.⁴⁰ The PDB identifiers associated with the fragment-bound and lead-bound crystal structures are 5YE8 and 5YEA, respectively. The lead molecule consists of carboxy-benzene, benzonitrile, and chloro-benzene groups (Figure 2Bi). The benzonitrile group overlaps with the fragment; hence, lead space contains only carboxy-benzene and chloro-benzene groups. The chloro-benzene group is placed in a cylindrical-shaped cavity filled with aromatic and hydrophobic residues. Therefore, both PYR and IPA maps the chloro-benzene group with the Z-score of 129 and 50 sigma, respectively (Figure 2Bii). The PYR-129 probe wanders around the cavity and makes π-π stacking interactions with Phe104, Phe125, and Phe357 and hydrophobic interactions with Leu107, Leu111, Leu121, Leu369, and Leu371. Moreover, the IPA-50 probe makes hydrophobic interaction with Leu107, Leu111, Leu121, Phe304, and Leu369. In addition, the IPA probe maps hotspots around chloro-benzene group and fragment with the Z-score of 75 and 133 sigma, respectively. The carboxy-benzene do not make much contact with the protein as it protrudes out of active site. Therefore, no probe is present in the vicinity of carboxy-benzene. The cavity occupying the chloro-benzene group is mapped by both PYR and IPA probes with high Z-scores, which shows that both probes can extract molecular features correctly.

From the probe mapping of fragment-bound Lp-PLA₂ (PDB ID – 5JAH), we found three extra hotspots around the chloro-benzene moiety with Z-scores of 99 (mapped by PYR) and 99 and 58 (mapped by IPA) sigma (Figure 2Aiii). These extra hotspots hint that the lead molecule can further be grown in this direction, which was done by Liu et al. They synthesized a lead molecule (PDB ID – 5YEA) that overlaps with the PYR-99 and IPA-99 hotspots and make π-π stacking interaction with Phe357 and nearby hydrophobic residues. The cylindrical-shaped binding site extends beyond these PYR-99 and IPA-99 hotspots. The ACN and IPA probe maps additional hotspots in the cylindrical-shaped binding site with Z-scores of 90 and 58 sigma, respectively (Figure 2Aiii). The IPA-58 hotspot forms hydrophobic interactions with Leu124, Phe125, Thr361 (hydrophobic part of side chain), Ile365, and Leu369. Similarly, from the simulation of fragment-bound Lp-PLA₂ (PDB ID – 5YE8), we found that all three probes - PYR, ACN, and IPA map the opposite side of the lead molecule with 468, 158, and 146 sigma values, respectively (Figure 2Biii). Such strong mapping by all three probes suggests that this subsite is crucial and should be utilized for lead growth in this direction. The lead molecule (PDB ID – 5LZ9) from Woolford’s study overlaps with these subsites, suggesting the site’s druggability and the MixMD’s ability to map these druggable sites. The binding of all probes at this site was driven by the π-π interactions with Trp298 and Phe322, and hydrophobic interactions with Leu153.

3.2.2. CREB-binding protein (CBP):

CBP along with adenoviral E1A binding protein (P300) acetylates the histone binding tail, which makes genes adopt an open structure and that facilitates transcription.⁶³ Inadequate level of CBP/P300 pair has been linked with several types of cancer^41,63,64 The CBP bromodomain binds at the acetylated lysine site on the histone tail, which activates gene expression. Bronner et al. identified various fragments bound to the acetylated lysine site of CBP and hybridized them with previously identified tetrahydroquinoline scaffold⁶⁴ to increase the potency of inhibitor.⁴¹ The tetrahydroquinoline scaffold binds to the LPF shelf, composed of Leu1109, Pro1110, and Phe1111.

The lead molecule can be divided into three parts: indole, tetrahydroquinoline, and pyrazole (Figure 2Ci). The fragment overlaps with the indole group; therefore, we will observe mapping only on the tetrahydroquinoline and pyrazole groups. The PYR probe maps the tetrahydroquinoline moiety with the Z-score of 135 and 90 sigma (Figure 2Cii). The PYR-135 probe forms interactions with Pro1110, Arg1173, and Val1174, while the PYR-90 probe binds by making hydrophobic interactions with Leu1109 and Pro1110. The ACN probe also maps the tetrahydroquinoline moiety at the same locations as the PYR probe with the Z-score of 104 and 73 sigma (Figure 2Cii). The IPA probe maps the tetrahydroquinoline moiety at three places with the Z-scores of 90, 84, and 65 sigma and the pyrazole moiety with the Z-score of 63 sigma (Figure 2Cii). The Val1174 stabilizes the binding of the IPA-90 probe through hydrophobic interaction. The hydroxyl group of IPA-84 probe form an H B with the side chain of Gln1113, and the methyl groups of IPA-84 form hydrophobic interaction with Leu1109. The IPA-65 hotspot’s binding is stabilized by the hydrophobic interactions formed by Pro1114, Leu1119, and Leu1120. Moreover, the IPA probe maps the pyrazole moiety with the Z-score of 63 sigma. The IPA-63 probe’s binding is assisted by the hydrophobic interactions formed with Pro1106 and Leu1109. The binding site of CBP contains many hydrophobic residues. The IPA probe can capture the hydrophobic features well around the lead molecules along with the PYR probe.

3.2.3. Bromodomain containing protein 4 (BRD4).

BRD4 binds to acetylated lysine located in histones to activate transcription. BRD4 recruits the positive transcription elongation factor complex (p-TEFb), which is a target for chronic lymphocytic leukemia.⁶⁵ Therefore, BRD4 is considered an anti-cancer target⁴⁷. We obtained the fragment- and lead-bound complexes from the PDB database. The lead molecule contains isoindole, phenoxy-, and phenyl sulfonamide groups (Figure 2D (i)). The fragment molecule overlaps with the isoindole and phenyl group of phenyl sulfonamide moiety, which leaves phenoxy and sulfonamide groups to be mapped by probes. The PYR, ACN, and IPA probes map the phenoxy group with the Z-score of 381, 223, and 413 sigma, respectively (Figure 2Dii). The PYR-381 and IPA-413 bind to the same location and interact with the same set of residues. The PYR-381 forms π-π stacking interactions with the Trp374 and fragment and forms hydrophobic interactions with the Pro375, Leu385, Glu438 (alkyl part), Val439, and Met442. The IPA-413 forms hydrophobic interactions with the Trp374, Pro375, Leu385, Glu438 (alkyl part), Val439, and Met442. The PYR and ACN probe map the sulfonamide with the Z-score of 63 and 85 sigma, respectively (Figure 2Dii). The PYR-63 probe forms hydrophobic interactions with the Asp381 (alkyl part), Val382, and Leu385. The PYR and IPA probes extract aromatic or hydrophobic molecular features currently as they bind to the phenoxy-group cavity of the lead molecule. The ACN probe maps the sulfonamide group, which shows that ACN probe can identify the HB donor molecular feature.

3.2.4. Serine/threonine protein kinase pim-1 (S/T Pim-1).

The S/T Pim-1 plays an important role in cell cycle progression,⁶⁶ cell proliferation, and regulation of MYC transcription activity.⁶⁷ Moreover, S/T Pim-1 has been used to target the tumorigenesis selectively. Good et al. performed docking with a HB constraint and selected a fragment molecule forming an HB with Lys67.⁴⁶ They grew the fragment into a lead molecule which maintains the HB constraint and interacts with Phe49 from the P loop.⁴⁶ The Lys67 residue is positively charged, and it makes a salt-bridge interaction with Glu89, which weakens the HB interaction with the fragment. In a few runs, this HB is broken, and the fragment’s orientation changes, but the fragment remains in the active site. The fragment breaches into the lead space, which decreases the probes’ mapping in the lead space to an extent. Moreover, Phe49 from the P loop enters the lead space, reducing the possibility of probes’ mapping in the lead space. The lead molecule contains thiazolidinediones, phenyl, pyrazine, and cyclohexyl groups (Figure 2Ei). In the crystal structure, the fragment molecule overlaps with the phenyl group and partially overlaps with the thiazolidinediones ring. The ACN and IPA probe maps the cyclohexyl moiety with the Z-score of 51 and 59 sigma, respectively (Figure 2Eii). The IPA-59 probe makes hydrophobic interaction with Phe49, Val52, and Lys67 (alkyl part). The ACN probe maps the thiazolidinediones moiety with the Z-score of 54 sigma.

3.2.5. WD repeat containing protein 5 (WDR5):

WDR5 modifies multiple histones by methylation at lysine residues to activate transcription.⁴⁵ WDR5 forms protein-protein interactions with Mixed-lineage leukemia 1⁶⁸ and RbBP5 and Ash2L.⁶⁹ WDR5 has become an important target for mixed-lineage leukemia, breast, bladder, pancreatic, and colorectal cancers.⁴⁵ Wang et al. designed a non-peptidic inhibitor for the WDR5, where it contains fused pyrrole and imidazole rings, indane group, and benzamide group (Figure 2Fi). The fragment molecule completely overlaps the fused pyrrole and imidazole rings and partly overlaps the indane group. All three probes map the benzamide group at two locations: one on amide group and another one on phenyl group. The PYR, ACN, and IPA probes map the amide group with the Z-score of 327, 124, and 94 sigma (Figure 2Fii). The PYR-327 forms π-π stacking interaction with Phe133, Tyr260, and fragment and forms hydrophobic interaction with Ile305. The PYR-327 probe forms pi-sulfur interaction with the side chain of Cys261. The IPA-94 probe forms hydrophobic interaction with Tyr260 and Phe133. The PYR, ACN, and IPA probes map the phenyl group of benzamide group with the Z-scores of 194, 171, and 125 sigma, respectively (Figure 2Fii). The PYR-194 probe forms π-π stacking interaction with Phe133, Phe149, and Tyr191 and form hydrophobic interactions with the Pro173, Ser175(alkyl group), Cys261. The IPA-125 form hydrophobic interaction with Phe133, Phe149, Tyr191, Pro173, Ser175(alkyl group), and Cys261. The lead space is filled with aromatic residues, and the PYR probe is successfully able to identify and map the lead space.

3.3. Mapping of lead space in systems with small change (ΔHAC ≤ 10).

3.3.1. Beta-secretase 1 (BACE1).

BACE1 cleaves the amyloid precursor protein (APP) to create β-amyloid peptides (Aβ).^12,70 These Aβ peptides oligomerize and form plaques, which are found in the brain of Alzheimer’s patients. To halt the progression of Alzheimer’s disease, researchers have designed multitudes of inhibitors. Table 1 includes BACE1 twice, where lead molecules in both BACE1 structure explore different subsites. Winneroski et al. started with aminothiazine as an initial fragment and designed a set of small molecule inhibitors to exploit the S3 subsite of BACE1.⁴⁹ This example corresponds to a small change in HAC, seen in the structures for system 15. The aminothiazine fragment forms HBs with catalytic Asp32 and Asp228 residues. The lead molecule contains an aminothiazine group and chloro-benzyl group, connected through an amide group. The aminothiazine group overlaps with the fragment; therefore, chloro-benzyl group is available to be mapped by probes and it occupies the S3 subsite of BACE1 active site. All three probes PYR, ACN, and IPA, map the chloro-benzyl group with the Z-score of 315, 328, and 400 sigma, respectively (Figure 3Aii). The binding of the PYR probe is assisted through its interaction with the glycine-rich loop, ranges from residue 10 to 13 and two hydrophobic residues Leu30 and Ile110. The hydroxyl group of IPA probe forms an HB with the backbone oxygens of the residue 10-13 loop and Gly230 and makes hydrophobic interactions with Leu30, Leu110, and Ala335.

Figure 3.

Lead-space mapping for (A) & (B) BACE1, (C) EthR. The PDB identifiers for fragment-bound proteins (A) and (B) are 4ZSM and 3UDH, respectively. A-C (i) Two-dimensional representation of the lead molecules. A-C (ii) Lead-space mapping of all three systems where grid mesh represents the area occupied by the probes, and the location of hotspot maxima are shown with a small sphere. A-B (iii) Extra hotspots beyond the lead molecule for BACE1 systems. A (ii) The PDB identifiers for the fragment and lead molecules are 4ZSM and 4ZSP, respectively. A(iii) Lead molecules with PDB ID – 3UDQ (shown in light blue) and 4FCO (shown in pink) overlaid with 4ZSM-4ZSP fragment-lead pair. B (ii) The PDB identifiers for the fragment and lead molecules are 3UDH and 3UDQ, respectively. B(iii) Lead molecule from PDB ID – 4ZSP (shown in green) overlaid with 3UDH-3UDQ fragment-lead pair. Coloring is the same as in Figure 1.

Efremov et al. chose a spiro-pyrrolidine molecule as a fragment through fragment screening. The subsequent lead molecule contains the spiro-pyrrolidine and dioxothiochromane group and both groups are connected through the carboxylate group (Figure 3Bi). The spiropyrrolidine group overlaps with the fragment; hence, probes can only map the dioxothiochromane group. The dioxothiochromane group occupies the S2’ subsite, which is on the opposite side of the binding site than the S3 subsite. This example corresponds to an intermediate change in HAC, seen in the structures for system 9. Both PYR and IPA probes bind in S2’ subsite and maps the dioxothiochromane moiety with the Z-score of 128 and 153 sigma, respectively (Figure 3Bii). The PYR-128 probe makes π-π stacking interactions with Tyr71 and Tyr198 side chains, hydrophobic interactions with Val69 and Ile126, and an HB with the side chains of Arg128 and Tyr198. The IPA-153 probe forms hydrophobic interaction with Val69 and Ile126 and forms HBs with the backbone oxygen of Gly34 and Pro70. The IPA probe also maps close to the S1’ subsite with the Z-score of 59 sigma. The PYR probe maps the benzene ring of dioxothiochromane group, and the IPA probe maps the sulfone ring. The molecular features in the lead align well with the features exhibited by the probes.

The lead molecule in 4ZSP and 3UDQ occupies the S3 and S2’ subsites, respectively. In our MixMD simulations, the probes mapped all S1’, S2’, and S3 subsites in both fragment-bound structures of BACE1 (PDB IDs – 4ZSM and 3UDH). In the case of fragment(4ZSM)-lead(4ZSP) pair, the lead space occupies the S3 subsite, so the mapping at S1’ and S2’ subsites will be considered as extra hotspots. The S1’ subsite is mapped by IPA with the Z-score of 99 sigma, and the S2’ subsite is mapped by PYR with 272 and IPA with 133 sigma values (Figure 3Aiii). The S2’ subsite is occupied by the lead molecule bound to another BACE1 complex (PDB ID – 3UDQ), and we have already discussed the interactions made by probes bound to the S2’ subsite. Therefore, we will only mention the interactions made by probes bound to the S1’ subsite. The IPA-99 probe present in S1’ subsite forms a HB with the side chain of Asp228 and hydrophobic interactions from Tyr198, Ile226, Thr329, and Val332.

For the fragment(3UDH)-lead(3UDQ) pair, the lead space maps the S2’ subsite; therefore, hotspots lying in S1’ and S3 subsite will be considered as extra hotspots. In 3UDH, the S1’ subsite is mapped by PYR and IPA probes with the Z-score of 73 and 59 sigma, respectively. While the S3 subsite is mapped by PYR, ACN, IPA probes with the Z-scores of 339, 242, and 460 sigma, respectively (Figure 3Biii). The S1’ subsite was mapped as an extra hotspot in case of fragment(4ZSM)-lead(4ZSP) pair, and the PYR-73 and IPA-59 probes interact to same set of residues what we found in case of 4ZSM. The S3 subsite was mapped by the lead molecule present in PDB ID – 4ZSP and we have already discussed the probes interactions with protein residues.

3.3.2. HTH-type transcriptional regulator (EthR)

Tuberculosis claims more than a million lives worldwide as per the WHO report in 2020.⁷¹ Several antibiotics such as isoniazid, ethambutol, rifampicin, and many others have been used to treat tuberculosis in the last 50 years;⁴⁸ however, side effects and drug resistance associated with these drugs led researchers to continue working on finding new antibiotics. Ethionamide, a novel TB drug, is a prodrug that needs to be activated by EthA (ethionamide activator), and the presence of EthA is regulated as a transcriptional repressor EthR(ethionamide regulator).⁴⁸ Willand et al. showed that inhibiting EthR reduces the dose level of ethionamide significantly as its high doses has been linked to side effects.⁷² Nikiforov et al. carried out fragment screening and used the fragment merging approach to synthesize the inhibitors for EthR.⁴⁸

The lead molecule consists of pyrrolidine, piperidine, and phenyl groups. The pyrrolidine and piperidine groups overlap with the fragment; hence, lead space contains only the phenyl group (Figure 3Ci). The PYR, ACN, and IPA map the phenyl group with 166, 97, and 222 sigma values, respectively and these probes occupy the same spot in the lead space (Figure 3Cii). The PYR-166 probe forms π-π stacking interaction with Trp103 and Tyr148 and hydrophobic interactions with Leu90 and Ala91. The hydroxyl group of IPA-222 probe makes an HB with the side chain hydroxyl group of Tyr148 and backbone oxygen of Met102. The IPA-222 probe forms hydrophobic interactions with the side chain of Leu87, Leu90, and Ala91. The lead space is surrounded by aromatic groups such as Trp103 and Tyr148 and hydrophobic groups such as Leu90 and Ala91. The PYR and IPA probes were able to map lead space with a high Z-score; therefore, probes were able to extract the molecular features of EthR’s active site.

3.3.3. MutT Homologue 1 (MTH1) – An unusual case:

When we superimpose the fragment-bound and lead-bound complexes of MTH1, the end-to-end length of the bound fragment and lead molecules are very similar (Figure 4ii), which means there is little lead space for probes to bind. However, our simulations for this system were complicated by the fact that the fragment moved significantly from the binding site in the great majority of the simulations (Figure S2), which left both the fragment space and lead space open for probes to bind. In all other systems in this work, each fragment’s binding pose did not change or breach into the lead space. For both the fragment-bound and apo simulations of MTH1, we received very similar mapping, showing that MixMD can accurately map the binding site’s fragment and lead hotspots (Figure 4iii). However, we cannot determine whether the presence of the fragment in its binding site would facilitate the probes’ binding in the lead space.

Figure 4.

MixMD mapping in case of MTH1. A(i) Two-dimensional representation of the lead molecule. A(ii) lead space mapping in fragment-bound protein complexes. A(iii) lead space mapping in the apo form of MTH1. A(ii)-(iii) grid mesh represents the area occupied by the probes, and the location of spheres denote the site of maximum occupancy (Z-score). Coloring is the same as in Figure 1.

3.4. Cooperative binding of probes in the presence of a fragment

In the present work, we carried out the MixMD simulations on 20 systems in fragment-bound form with the aim to identify whether the presence of a fragment would assist the binding of probes in lead space. This cooperative binding of probes could occur because of conformational change, HB network, or hydrophobic interaction due to the presence of the fragment. To identify the systems with definite cooperative binding between fragment and probes, we have simulated the apo (fragment-free) form of all 20 systems. Upon comparing the lead-space mapping in fragment-bound and apo forms of all systems, we found three systems where probes did not map the lead space in the apo form but did map the lead space in the fragment-bound form. Moreover, we see a clear trend that probes bind with higher sigma values in fragment-bound form rather than in apo form (Table 2). We also found that the fragment space was always mapped by MixMD in the 20 apo systems (Table S1 and text in the supplemental information). In the following cooperative systems, we will discuss the interactions made by probes present in the lead space of both fragment-bound complexes along with the interactions made by probes present in the fragment binding site in the apo form.

3.4.1. Bruton’s tyrosine kinase (BTK).

BTK signals the B-cell antigen receptor and activates B-cells.⁵⁰ Therefore, BTK is considered an important target for autoimmune disorders such as rheumatoid arthritis, lupus, and multiple sclerosis.⁷³ Smith et al. designed a selective BTK inhibitor by using the fragment growth and structural optimization approach.⁵⁰ The lead molecule contains two moieties: cinnoline and indazole (Figure 5Ai). The cinnoline unit overlaps with the fragment; therefore, probes will be able to map only the indazole unit of the lead molecule. In the apo form of BTK, all three probes map the fragment pocket (Figure 5 Aiii). The PYR probe maps the fragment pocket at two locations with the Z-score of 358 and 63 sigma. The ACN and IPA probes map only at one site with the Z-score of 65 and 112 sigma, respectively (Figure 5Aiii). However, not a single probe maps the lead space in the apo form of BTK (Figure 5Aiii). This occurs because the P-loop (residue 411 to 414) breaches into the lead space and that will hinder the binding of probes in lead space. As a result, we do not see any mapping on the indazole unit in apo form. Furthermore, we do not observe any hotspot in the lead space of the fragment-bound from of BTK. However, we see a PYR hotspot with the Z-score of 156 sigma in the fragment-bound form, away from the P-loop but close to the indazole unit (Figure 5Aii). The binding of this PYR-156 probe is stabilized by π-π stacking interaction with the fragment and the hydrophobic interaction with Leu528.

Figure 5.

Cooperative binding between fragment and probes in BTK, ATAD2, and Dot1L. A-C (i) Two-dimensional representation of the lead molecules. A-C (ii) lead space mapping in fragment-bound protein complexes A-C (iii) lead space mapping in the apo form of protein complexes. In A-C (ii) & A-C (iii) grid mesh represents the area occupied by the probes, and the location of small sphere denotes the site of maximum occupancy (Z-score). Coloring is the same as in Figure 1.

We observed several hotspots around the fragment, which could be crucial for drug design (Figure 6). We searched all the protein entries deposited in the PDB database for BTK and found three lead molecules (PDB ID – 5T18, 5VFI, and 6E4F), which perfectly overlap with the extra hotspots (Figure 6). The lead molecule bound in the 5T18 structure (shown in green) overlap with extra hotspots at two locations: quinazoline moiety overlaps with PYR-65 and IPA-51, and isopropyl alcohol moiety overlaps with PYR-193 and ACN-65 hotspots (Figure 6). The PYR-65 and IPA-51 probes make hydrophobic interaction with Cys481, Leu483, Asn484, and alkyl part of Arg525 side chain. While the PYR-193 probe makes an HB with the amino-group of the fragment, makes hydrophobic interaction with Leu408, and π-π stacking interaction with Tyr476. The lead molecule bound in the 5VFI structure (shown in yellow) overlaps with new hotspots PYR-87 and two hotspots PYR-193 and ACN-65 (Figure 6), which also overlapped with the lead molecule bound in 5T18. The PYR-87 hotspot makes π-π stacking interaction with Phe413 and hydrophobic interaction with Ile432, Ile472, and Leu542. The lead molecule bound in the 6E4F structure (shown in magenta) overlaps with a new hotspot ACN-105 (Figure 6). The ACN-105 probe forms hydrophobic interaction with Leu432, Leu460, Leu472, and Ile542. All three lead molecules branch out in different directions than the chosen lead molecule (PDB ID – 4Z3V). Moreover, all three lead molecules have better IC₅₀ – less than 1 nM than the chosen lead molecule (4 nM). These facts suggest that a potent lead molecule could be grown in multiple directions. This also highlights the robustness of MixMD, which was able to map these druggable extra hotspots in addition to mapping the lead space of structure 4Z3V.

Figure 6.

Extra hotspots beyond the lead molecule of BTK overlaid with other multiple lead molecules from PDB IDs – 5VFI (shown in yellow), 5TI8 (shown in green), and 6E4F (shown in magenta).

3.4.2. ATAD2 bromodomain.

ATAD2 plays a role in activating various transcription factors such as MYC, estrogen/androgen receptors, and many others.³⁸ Several cancer cells overexpressed ATAD2, and its expression level has been linked to cancer cells’ progression.³⁸ ATAD2, like other bromodomains, binds to the acetyl-lysine group; however, ATAD2 contains Arg1077 in place of Tryptophan from the WPF shelf in BRD4, which makes it difficult to optimize the binding of inhibitors. Demont et al. identified quinoline as a suitable fragment as it would bind acetyl-lysine site, and they synthesized a lead molecule with μM-level inhibitor³⁸ and it was further optimized to nM-level inhibitors.⁷⁴

The lead molecule is composed of four groups: pyridine, quinolinone, piperidine, and thianedioxide (Figure 5Bi). The quinoline moiety overlaps with the fragment; therefore, the lead space consists of the pyridine, piperidine, and thianedioxide moieties and are available to be mapped by probes. In the apo systems, all three probes map quinolinone moiety. In the apo form, the PYR, ACN, and IPA probes map the quinolinone core with the Z-scores of 180, 421, and 194 sigma, respectively (Table 2). However, not a single probe maps the lead space in the case of the apo system (Figure 5Biii). When the fragment is present in the binding site, then the ACN probe is able to map the thianedioxide moiety of the lead molecule (Figure 5Bii). The ACN is a small-sized probe, wanders around the thianedioxide moiety, and maps the thianedioxide moiety with the strongest hotspot of 64 sigma values (Figure 5Bii). The ACN probe forms HB with the side chain of Arg1077 and makes hydrophobic interaction with Val1008, Leu1073, and Ile1074.

3.4.3. Dot1-like protein (Dot1L).

Dot1L adds a methyl group to a lysine residue of histone, which changes the transcription. Aberration in the methylation process causes acute leukemia.⁵⁴ Mobitz et al. identified a fragment that binds to an induced pocket next to the SAM binding site.⁵⁴ The neighboring pocket was screened with the first fragment bound to the induced pocket site. Later, both fragments were linked and further optimized to increase the potency of the lead molecule. We simulated Dot1L with both fragments bound in their sites with the aim if we could map the linker. However, only the PYR probe maps the induced pocket with the Z-score of 75 sigma, and the linker region remains unmapped (Figure 5Cii). The PYR-75 probe forms π-π stacking interaction with the fragment, Phe131, and Phe243 and hydrophobic interactions with Pro130 and Ser269 (alkyl part).

3.5. Systems with no mapping in lead space

MixMD was unable to map the lead space in a few systems, and there are multiple reasons that can account for this possibility. As we have noted above, a larger ΔHAC will mean there is more space for the probes to bind. This is one of the prime reasons why most of the unmapped systems fall into the category with an HAC difference of less than 10.

3.5.1. Bromodomain-containing protein 9 (BRD9).

The end-to-end lengths of the fragment and lead molecules are almost the same (Figure S3 Aii in the supplemental information), and ΔHAC difference is only 8. The lead contains isopropyl- and O-methoxy groups that branch out from the fragment space (Figure S3 Aii), but these moieties are very small and do not point to another nearby hotspot. As such, we found no probes binding in the lead space. In fact, this is likely to be a true-negative test case for the method. The PYR probe maps a site (shown in purple mesh grid) close to the lead within 3-4 Å distance (Figure S3 Aii), but we are using a strict criterion, which is a hotspot must be present within 2 Å of the lead molecule. Hence, BRD9 also falls in the category of unmapped systems.

3.5.2. Catechol O-methyltransferase (COMT).

The lead molecule contains thiazole, pyrazol, cyclopropyl, and methoxy-phenyl groups (Figure S3 Bi). The thiazole and pyrazol groups overlap with the fragment, leaving only cyclopropyl and methoxy-phenyl groups to be mapped by probes. However, the methoxy-phenyl group protrudes out into solution and makes minimal contact with the protein. Therefore, there is no cavity around the methoxy-phenyl group to bind, which makes this system unmapped.

3.5.3. Bromodomain and PHD finger-containing protein 1 (BRPF1).

The lead molecule contains the quinoline and cyano-substituted benzene groups (Figure S3 Ci). The fragment group overlaps with the quinoline that leaves only the cyano-substituted benzene group to be mapped by probes. The cyano-substituted benzene group protrudes out into solution and makes little contact with the protein. The Phe714 residue makes edge-to-face (T shaped) π-π stacking interaction with the cyano-substituted benzene group. The PYR probe binds in the lead space but with a Z-score of only 26 sigma, which is less than our chosen cutoff of 50 sigma. No other probe mapped the lead space; therefore, we can say MixMD could not map the lead space of BRPF1 (Figure S3 Cii). However, there is a cavity within 2-3 Å from the lead space, which could be explored further for lead growth.

3.6. Comparison of MixMD with Site Finder and SiteMap:

We chose to build this protocol using very short MD simulations to make it more tractable for use in academia and the pharmaceutical industry. With a standard cluster of GPUs, MixMD can be run overnight for most proteins. However, overnight is still much longer than some other site-mapping methods that take seconds. To show the advantage of our method, we compare MixMD with two other popular programs: Site Finder^75,76 of MOE 2019.01¹⁹ and SiteMap^77,78 from Schrödinger. The hotspot predictions for all 20 systems from Site Finder and SiteMap are given in Supplementary information (Figure S4). Many of the systems are well described by all 3 methods, but MixMD has an advantage when Site Finder and SiteMap miss nearby hotspots and when their hotspots are found over a wide swath of the protein surface, lacking focus on the appropriate lead space.

The lead molecule (in cyan) in Figure 7A is branched in two directions: aromatic and sulfonamide groups. The MixMD approach mapped the aromatic group with significantly high Z-scores (Figure 7A); however, the aromatic group was not mapped by both SiteMap and Site Finder (Figure 7B and C). The sulfonamide group is mapped by MixMD and Site Finder but not by SiteMap. Hence, SiteMap and Site Finder may miss the prediction of correct lead growth sites as they use only single, rigid protein structure to compute the cavities.

Figure 7.

Prediction of hotspots for Bromodomain-containing protein 4 (BRD4) using A) MixMD, B) Site Finder, and C) SiteMap approaches. Both directions of expanding the fragment are captured by MixMD but missed by the other approaches. A-C) the fragment and lead molecules are shown in yellow and cyan, respectively. A) grid mesh represents the area occupied by the probes, and the location of small sphere denotes the site of maximum occupancy (Z-score). Coloring scheme of MixMD is explained in Figure 1 while the coloring scheme of SiteMap and Site Finder are explained in Figure 7.

In Figure 8A, the lead molecule (in cyan) is mapped by all three probes with high Z-score values. The binding of PYR and IPA probes suggests that moiety in lead growth sites could be either aromatic or hydrophobic. Since the probes occupy a small space in the large binding site of BACE1, this will help identify the correct direction of lead growth and chemical groups to grow the fragment into a lead molecule. On the other hand, both SiteMap and Site Finder map a large number of hotspots over its rather large active site (Figure 8B and C). Therefore, these approaches do not focus on identifying the most productive direction for lead growth in this example.

Figure 8.

Prediction of hotspots for BACE1 system (PDBID – 3UDH) using A) MixMD, B) Site Finder, and C) SiteMap. The hotspots determined by MixMD are focused on the immediate lead space, whereas Site Finder and SiteMap cover the whole large binding site of BACE1, lacking focus on the appropriate space for this fragment. A-C) the fragment and lead molecules are shown in yellow and cyan, respectively. A) The grid mesh represents the area occupied by the probes, and the location of small sphere denotes the site of maximum occupancy (Z-score). Coloring scheme is the same as in Figure 1. B) The red and grey spheres represent hydrophilic and hydrophobic regions, respectively. C) The red, blue, and yellow regions represent the acceptor, hydrogen bond donor, and hydrophobic maps.

Both SiteMap and Site Finder compute the cavities quickly because they only use a single crystal structure of a protein. However, MixMD is a dynamic approach, where we simulate the protein along with water and probes (small organic molecules). This allows proteins to move and sometimes opens cryptic sites to identify novel hotspots,¹² which is not possible to obtain from SiteMap and Site Finder. The snapshots from the MD can be used to expand structure-based drug design, which is an added benefit to the approach. Lastly, scientists at Schrodinger have shown an advantage to cosolvent MD over SiteMap. Ghanakota et al. implemented mixed-solvent molecular dynamics⁷⁹ (MSMD) using a very similar approach to MixMD. They compared SiteMap’s and MSMD’s abilities to find hotspots on the protein-protein interface.⁷⁹ MSMD was able to find hotspots in multiple cases where SiteMap could not.

4. CONCLUSION

Here, we examined the use of fragment-bound proteins in MixMD simulations to identify nearby hotspots that would indicate how to grow a fragment into a lead. We were able to map hotspots in the lead space in 17 out of 20 systems. Systems that were not mapped included cases where the lead modification was oriented out into solution and a case where the lead modifications were too small to correspond to an additional hotspot (possible true-negative cases). We also conducted simulations of the apo state for each system. One of our goals was to identify systems with binding cooperativity where a hotspot was mapped in the presence of the fragment, but not in the apo state. We found three systems that had cooperativity. Moreover, the occupancies (Z-scores) of probes in the lead space were significantly higher in fragment-bound simulations than in the apo simulations, also reflecting a type of cooperativity. These findings provide support to our approach of simulating the fragment-bound systems over the apo form to map the lead space.

In addition to identifying the hotspots for the chosen lead molecules, we explored nearby hotspots around the lead molecules, also termed as extra hotspots. In the case of BTK, we found three lead molecules branching out in different directions that overlapped well with nearby hotspots. We simulated two fragment-bound systems of BACE1 where lead molecules sampled different subsites, and MixMD was able to identify all druggable subsites sampled by lead molecules in both BACE1 structures. In the apo form, we found that all three probes were able to map the fragment binding site in all 20 systems, and at least one probe was ranked in the top-four of occupancies (Table S1 in supplementary info), which points to the robustness of MixMD to map functional and druggable binding sites.

Lastly, our probes can indicate the possible use of aromatic, aliphatic, and hydrogen-bonding moieties in the lead space, but our method does not take into account growing specific lead compounds nor their synthetic accessibility. This is an area for future development.

ABBREVIATIONS

MixMD

mixed-solvent molecular dynamics

HAC

Heavy atom count

HTS

High-throughput screening

FBDD

Fragment-based drug design

PYR

Pyrimidine

ACN

Acetonitrile

IPA

Isopropyl alcohol

MD

Molecular dynamics

HB

hydrogen bond