Footprint-based identification of viral entry inhibitors targeting HIVgp41

Abstract

A targeted virtual screen to the N-helix hydrophobic pocket on HIVgp41 was performed using DOCK followed by re-ranking with a new footprint-based scoring function which employed native gp41 C-helix residues as a reference. Of ca. 500,000 small molecules screened, 115 were purchased, and 7 hits were identified with favorable binding (K_i), cell-cell fusion (IC₅₀), and cytotoxicity (CC₅₀) profiles. Three of the seven active compounds would not have been discovered without the use of the footprints, demonstrating the utility of the method for structure-based design when a known reference compound or substrate is available.

The HIV glycoprotein (HIVgp41) is a validated viral entry drug target with one FDA approved inhibitor, a 36 a.a. peptide known as T20 (FUZEON).¹ Despite the success of T20 as a first-in-class fusion inhibitor, there are known resistance problems, it is expensive to produce (ca. USD $25,000 per person per year), and it must be delivered via injection.^1–5 An effective small molecule gp41 inhibitor would be an important addition to the arsenal of compounds used to treat HIV/AIDS. A highly conserved hydrophobic pocket, adjacent to the putative T20 binding interface,⁶ was previously identified as a likely drug target.⁷ Active small molecules, presumably targeting the pocket, have been reported (see Cai and Jiang⁸ for a recent review), although specific binding poses have not yet been confirmed by crystallography. Compounds have been discovered using traditional high-throughput, NMR, and computer-based screening.^{8, 9} Most small molecule leads have been in the μM range although some have been reported with activities that are sub-μM.^{10, 11} The extent of the HIV/AIDS pandemic necessitates continued screening and development of new inhibitors targeting gp41.

In this work, we have taken a computational screening approach which employs a novel molecular footprint comparison rescoring method¹² recently implemented by our laboratory into the program DOCK6.5.¹³ Footprint comparisons employ per-residue decompositions of the standard DOCK energy score to compare docked molecules with a known reference as illustrated in Figure 1. Balius et al¹² have shown that footprint-guided selection enhances success for reproducing known binding geometries (poses). We hypothesize that screening for compounds which make footprints similar to a reference known to bind gp41 will facilitate discovery of anti-gp41 inhibitors.

Figure 1.

(top) Key C-helix residues of the gp41 reference and associated VDW footprint in green. (bottom) Comparison between a docked ligand (magenta) and the reference (green).

The gp41 reference used for rescoring in this study is based on the native gp41 C-helix reduced to four key sidechains, Trp117, Trp120, Asp121 and Ile124 (gp41 numbering), that interact with the conserved hydrophobic pocket. Figure 1 shows the van der Waals (VDW) footprint derived from the crystallographic reference 1AIK¹⁴ (green), in comparison to the VDW footprint from a representative docked ligand (magenta), taken from the virtual screen. The periodicity in the footprints reflects the α-helical nature of the target and the strongest peaks indicate residues which interact most favorably with either the reference, the docked molecule, or both. Importantly, this example demonstrates how footprint overlap can be used as a means to identify compounds that make similar interactions with gp41 as a known reference, in terms of both a.a. position and energetic strength.

Throughout this study footprint overlap is quantified using a Pearson correlation coefficient and termed footprint similarity (FPS) score, although other metrics such as Euclidian distance may be used.¹² In addition to the VDW example in Figure 1 (FPS_VDW = 0.97), electrostatic interactions (FPS_ES) computed from Coulombic energies scaled by a distance dependent dielectric, as well as the combined sum (FPS_VDW+ES), were also employed. In particular, significant electrostatic interactions in this system are observed between Asp121 (C-helix) and Lys63a (N-helix chain a) which are known to be important for binding.¹⁵ By using footprint comparisons to help guide the search, molecules making similar interactions can be selected over others.

A comprehensive virtual screen of ca. 500,000 compounds was performed using DOCK6.5. The docking database was taken from the ZINC library (http://zinc.docking.org) of commercially-available compounds for virtual screening.¹⁶ As shown in Figure 2, each compound was flexibly docked to the gp41 receptor grid (DOCK FLX protocol)¹⁷ and the single lowest-energy pose was retained. The docking grid was generated from the gp41 crystal structure 1AIK¹⁴ using only the trimeric N-helix core. After docking, the FPS scores were calculated between each energy minimized compound and the C-helix derived reference (Figure 2). The top 100,000 molecules (energy score) were then clustered, using MACCS fingerprints,¹⁸ as implemented in the program MOE¹⁹ using a tanimoto coefficient of 0.75. The best scoring member of each cluster (i.e. clusterhead) was retained for subsequent steps to increase the overall chemical diversity of the compounds ultimately selected for purchase.

Figure 2.

Virtual screening and experimental evaluation workflow.

The resultant clusterheads (ca. 56,500) were then rank- ordered using four different scoring methods: (i) standard DOCK score (DCE_VDW+ES), (ii) van der Waals footprint similarity score (FPS_VDW), (iii) electrostatic footprint similarity score (FPS_ES), and (iv) the combined footprint sum (FPS_VDW+ES). The top 500 compounds obtained by each method were retained (Figure 3) and after visual inspection a subset from each of the ranked lists was selected for purchase and experimental testing. An effort was made during the selection process to choose a diverse set of molecules with favorable properties among several categories.

Figure 3.

Histograms of (a) DCEVDW+ES score, (b) molecular weight, (c) number of rotatable bonds, (d) FPSVDW+ES score, (e) FPSVDW score, and (f) FPSES score from the different ensembles of 500 molecules obtained from each of the four ranking methodologies: DCE_VDW+ES = black, FPS_{VDW+ES =} orange, FPS_VDW = green, FPS_{ES = blue.}

Importantly, each group of 500 top ranked molecules obtained by different scoring methods (DCE_VDW+ES = black, FPS_VDW+ES = orange, FPS_VDW = green, FPS_ES = blue) has a different set of characteristics as shown by the histograms in Figure 3 which plot (a) DCE_VDW+ES scores, (b) molecular weight, (c) number of rotatable bonds, (d) FPS_VDW+ES scores, (e) FPS_VDW scores, and (f) FPS_ES scores. As expected, molecules selected using a particular scoring function will have the most favorable scores, relative to other sets, when viewed in the accompanying histogram. For example, molecules picked using DCE_VDW+ES generally have more favorable DCE_VDW+ES scores (Figure 3a, black line). Likewise, molecules picked using FPS_VDW+ES have more favorable FPS_VDW+ES scores (Figure 3d, orange line). Not surprisingly, molecules chosen using FPS_VDW+ES (orange line) also show enhanced populations in the FPS_VDW and FPS_ES histograms. As a baseline, the computed DCE_VDW+ES energy score for the reference (Figure 1) is −46.36 kcal/mol which interestingly is at the intersection (Figure 3a) of the DCE (upper) and FPS (lower) score ranges.

Compounds chosen using DCE_VDW+ES generally have more favorable DCE_VDW+ES scores than those chosen using FPS methods which is understandable when one considers that molecules chosen using DCE_VDW+ES also have higher MW (Figure 3b black vs other colors), a known scoring bias problem in DOCK. In contrast, use of FPS yields a wider range of MW, which on average, are also lower: DCE_VDW+ES = 483.2 g/mol vs FPS_VDW = 456.8 g/mol, FPS_ES = 415.1 g/mol, FPS_VDW+ES = 432.2 g/mol. In general, use of FPS also results in molecules with lower numbers of rotatable bonds which indicate more rigid, potentially more drug-like, lead molecules (Figure 3c black vs other colors).

Following selection protocols, a total of 115 compounds were purchased. Compounds were tested in a fluorescence-based binding assay (K_i),²⁰ and those that displayed activity were evaluated in follow-up cell-cell fusion (IC₅₀), and cytotoxicity assays (CC₅₀) as previously described in Zhou et al.¹¹ Encouragingly, 24 molecules showed binding with K_i values < 15 μM. Importantly, 18 out of the 24 were selected using FPS scoring. Ultimately, 7 compounds were identified as having favorable properties (low K_i, low IC₅₀, high CC₅₀) in all three assays and are termed here “actives“ (Figure 4, Table 1). It should be emphasized that 3 of the 7 actives (SB-D10, C01, and H02) would not have been chosen for experimental testing without the use of footprints because their DCE_VDW+ES scores, while close to that of the reference, were not favorable enough to be within the top 500 molecules. For example, SB-C01 ranked 27,427th by DCE_VDW+ES score and thus in a traditional virtual screen would have been discarded. Through its similarity in electrostatic interactions to the reference the rank dropped to 247th in the list of FPS_ES clusterheads and was ultimately selected for purchase.

Figure 4.

Predicted binding modes for 7 active compounds (magenta) in comparison with the reference sidechains (green). Activities in μM determined using fluorescence binding (K_i), cell-cell fusion (IC₅₀), and cytotoxicity (CC₅₀) assays as previously described in Zhou et al.¹¹

Table 1.

Primary selection method and virtual screening scores.

IDa	Methodb	DCEVDW+ES	FPSVDW+ES	FPSVDW	FPSES
SB-D10	FPSES	−43.85	1.90	0.91	0.99
SB-D04	DCEVDW+ES	−50.86	1.90	0.91	0.98
SB-C01	FPSES	−41.86	1.90	0.90	0.99
SB-A05	DCEVDW+ES	−51.51	1.82	0.84	0.98
SB-C09	DCEVDW+ES	−51.48	1.97	0.98	0.99
SB-H11	DCEVDW+ES	−52.10	1.84	0.86	0.99
SB-H02	FPSVDW+ES	−44.33	1.96	0.97	0.99
RG-1 (29)	DCEVDW+ES	−51.93	1.67	0.89	0.78
RG-2 (21)	FPSVDW+ES	−43.54	1.96	0.98	0.99
RG-3 (33)	FPSVDW	−42.25	1.08	0.99	0.09
RG-4 (32)	FPSES	−43.81	1.88	0.88	0.99
SB-avg		−48.00	1.90	0.91	0.99
RG-avg		−45.38	1.65	0.94	0.71
CH-ref		−46.36	2.00	1.00	1.00

^aSB prefix designates active molecules, RG prefix designates ensembles based on the four different ranking methods. CH-ref designates reference molecule based on C-helix sidechains.

^bDCE scores in kcal/mol, FPS scores are unitless.

It should be emphasized that during compound selection, priority was also given if a ligand showed favorable scores in more than one category. As a result, each of the four Ranked Groups (RG prefix, Table 1), for the most part, also showed reasonable high average scores across all scoring methods. Only the FPS_VDW group (RG-3) yielded poor FPS_ES scores (0.09) and consequently lower FPS_VDW+ES scores (1.08). Notably, on average, the 7 actives (SB-avg, Table 1) have more favorable scores than the total group of 115 tested (RG-avg in Table 1) in three out of four categories (DCE_VDW+ES, FPS_VDW+ES, and FPS_ES) and more closely approach the reference values (CH-ref, Table 1).

Of the actives, SB-C09 was ranked in the top 500 by DCE_VDW+ES, FPS_ES, and FPS_VDW+ES methods. If the group size is expanded to the top 2000, then SB-C09, SB-D04, SB-A05, and SB-H11 are all ranked in at least one additional category. Given that these four compounds are all active provides evidence that an intersection of good scores from the different rankings methods (i.e. consensus scoring) is a reasonable protocol for prioritizing compounds. Within the top 2000 ranked using DCE_VDW+ES, 68 ligands are in the top FPS_VDW, 181 are in the top FPS_ES, and 104 are in the top FPS_VDW+ES.

As highlighted in Figure 4, the 7 actives have a number of intra and intermolecular structural features in common. All have a formal charge of −1 and five have a carboxylic acid which is predicted to interact with Lys63a of gp41 in a manner analogous to Asp121 of the reference. This salt-bridge is considered to be an important feature of many inhibitors.¹⁵ The actives are also dominated by the presence of planar aromatic rings which occupy positions in the pocket normally occupied by the reference sidechains (Figure 4 green vs magenta). In particular, all compounds have significant ring overlap with Trp117. Prior work by our laboratory²¹ quantified the importance of Trp117 and Trp120 which together make up the bulk of the VDW interactions in the pocket.

Structurally, several actives have features in common with other reported gp41 inhibitors. For example, compound SB-D10 contains an N-substituted pyrrole group (see Figure 4) similar to the NB-2 and NB-64 inhibitors reported by Jiang et al.²² And, the general size and layout of SB-D10 is roughly similar to a group of analogs reported by Katritzky et al¹⁰ which were designed around the pyrrole functionality to more completely fill the hydrophobic pocket. SB-D10 also shares significant chemical similarity with SB-D04 although interestingly the 3D poses are ca. 180° flipped (Figure 4). To probe if either compound could adopt the pose adopted by the other, a series of additional docking runs were performed. Figure 5 shows two possible low-energy poses with higher structural overlap than that illustrated in Figure 4. Interestingly, although the best alternative pose shown for SB-D04 (Figure 5, pose #1, orange) has a less favorable DCE_VDW+ES score (−41.36 vs −50.86 kcal/mol) the larger FPS_VDW (0.94 vs 0.91) and FPS_ES (0.99 vs 0.98) scores indicate improved overlap with the C-helix reference footprints (Table 1). Additional studies, including MD refinement (see discussion below), are in progress.

Figure 5.

Alternative pose comparisons for SB-D10 (magenta) and SB-D04 (orange) in the gp41 pocket (gray surface).

Another compound, SB-C01, contains an indole group which is positioned midway between Trp120 and Ile124 of the reference molecule (Figure 4). Indoles form the basis of series of gp41 inhibitors recently reported by Zhou et al,¹¹ although in a different arrangement than that reported here. Of particular interest for SB-C01 is the striking overlap of the carboxylic acid with that of the reference. Preliminary SAR studies were pursued by creating virtual analogues of SB-C01 (Table 2, M1–M8) and energy-minimizing them in the gp41 site using a harmonic restraint tethered to the predicted binding pose (Figure 4). Although the FPS_ES overlap for SB-C01 with the reference was already high (0.99), a comparison of the VDW footprints (0.90) suggested interactions with residues Leu54c, Leu57c and Trp60c which lie along the edge of the gp41 pocket could be improved. Based on visual inspection of the predicted SB-C01 pose, positions R1 and R2 (see Table 2) are natural points for SAR exploration.

Table 2.

SB-C01analogues, corresponding DCE scores, FPS scores, and movement (rmsd) after energy minimization.

	ID	R1	R2	R3	DCEVDW+ESa	FPSVDW+ES	FPSVDW	FPSES	Rmsd
	SB-C01	H	H	CH3	−41.86	1.89	0.90	0.99	0.05
	M1	CH3	H	CH3	−42.90	1.91	0.92	0.99	0.05
	M2	CH2CH3	H	CH3	−43.49	1.91	0.92	0.99	0.19
	M3	H	CH3	CH3	−41.68	1.90	0.91	0.99	0.27
	M4	H	CH2CH3	CH3	−32.98	1.66	0.66	0.99	1.03
	M5	H	H	OH	−41.79	1.89	0.90	0.99	0.06
	M6	H	H	COO−	−39.18	1.83	0.84	0.99	0.36
	M7	H	H	CH2OH	−42.02	1.88	0.89	0.99	0.16
	M8	H	H	CH2COO−	−39.02	1.85	0.86	0.99	0.51

^aDCE scores in kcal/mol, FPS scores are unitless, rmsd in Å.

Simple alkylations at R1 (M1 and M2) were moderately successful in increasing the VDW footprint correlation and generating a more favorable DCE_VDW+ES score relative to the parent SB-C01. FPS_VDW scores increased from 0.90 to 0.92 while DCE_VDW+ES scores favorably increased from −1 to −2.5 kcal/mol. Importantly, the FPS_ES scores remained high and the low rmsds obtained after minimization indicates the M1 and M2 virtual analogs are as well-accommodated in the pocket as the parent SB-C01. In contrast, the same modifications at R2 (M3, M4) did not lead to noticeable improvement. Methylation (M3) had no effect while ethylation (M4) significantly reduced the FPS_VDW (0.90 → 0.66) and DCE_VDW+ES (−41.86 → −32.98 kcal/mol) values compared to SB-C01. Further, the larger movement for M4 (1.03 Å rmsd) upon minimization, compared to the other analogs tested suggests ethylation at R2 may have introduced an unfavorable interaction. In general, the analogs with the smallest rmsd movement yielded more favorable DCE_VDW+ES scores (Table 2). Initial attempts to increase favorable electrostatic interactions with Arg68c at the end of the gp41 pocket were also not fruitful. R3 modifications incorporating hydroxyl (M5, M7) or carboxylate (M7, M8) functionalities led to less favorable interactions with the exception of M7 (R3 = CH₂OH), although the difference was negligible. Further studies will be needed with an expanded palate of functional groups. More extensive SAR based on all 7 actives shown in Figure 4 are in progress.

Finally, to further probe the stability of the predicted pose of SB-C01, molecular dynamics (MD) simulations were performed of the gp41-ligand complex using the program AMBER²³ with FF94SB²⁴ (protein), GAFF²⁵ (ligand), and TIP3P²⁶ (solvent) force fields. Sixteen 2-ns production simulations, using different random seeds, were performed with weak restraints on the protein backbone after an initial energy minimization and equilibration protocol with decreasing restraints. The simulations yielded a low 2.3 Å average rmsd for the ligand providing further evidence the predicted SB-C01 pose is compatible with the gp41 pocket. From another point of view, the MD-averaged FPS_VDW and FPS_ES similarity scores (using the original docked pose as reference) remain high (FPS > 0.96) indicating the primary energetic signature also remains stable even when subjected to thermal fluctuations.

Figure 6 shows a binding site closeup from a representative MD simulation of the SB-C01/gp41 complex. Here, overlaid snapshots (N=10) reveal a coordinated movement of the carboxylic acid and indole functionalities on SB-C01 with Lys63a and Leu57c respectively in the gp41 pocket. Relative to the original pose, the carboxylic acid alternates between a salt-bridge with Lys63a and an H-bond with Gln66a. Using a salt-bridge definition of ≤ 4.0 Å (carboxylic O --- lysine N) the interaction was present 50.48% of the time. Interestingly, 9.42% of the time there is both a salt bridge with Lys63a and an H-bond with Gln66a (2.5 Å definition).

Figure 6.

Starting position (stick) and location every 200 ps thereafter of SB-C01 (orange) and key residues (cyan) from one 2 ns MD simulation.

In general, out of sixteen simulations, the most favorable FPS_ES scores were from runs in which the salt-bridge was formed more frequently. This is to be expected given that the Gln66a H-bond is not observed in the original SB-C01 docked pose. For comparison, a 9-ns fully-solvated MD simulation (following protocols outlined in McGillick et al.⁶), of the C-helix peptide inhibitor C34 used here to create the reference, yielded the analogous salt-bridge with Lys63a 66.76% of the time. Interestingly, the C34 peptide simulation did not yield H-bonds with Gln66a. Thus, a possible strategy for inhibitor development would be to prevent or reduce the apparent slippage of SB-C01 towards Gln66a observed in Figure 6 which could further increase similarity to reference. Studies to test this hypothesis are underway through the addition of functionality designed to promote VDW contact with Leu54c, Leu57c, Trp60c, and Gln66a.

In summary, in this report we performed a virtual screen of ca. 500,000 molecules against the hydrophobic pocket of HIVgp41. Using a new molecular footprinting method in conjunction with the standard DOCK energy score we selected 115 molecules for experimental testing. Seven inhibitors were identified with μM activity (K_i, IC₅₀) and reasonable cytotoxicity (CC₅₀) characteristics. Three of the seven inhibitors would not have been discovered without the use of footprints suggesting that the methodology may have utility for structure-based design. The fact that footprint-based screening leads to actives, in some cases, which share structurally related features to known gp41 inhibitors discovered via different methods^{11, 22} is particularly encouraging. Additional SAR studies are underway and extension of the method to incorporate reference-based footprints into lead refinement and de novo design and protocols are planned.