DARC: Mapping Surface Topography by Ray-Casting for Effective Virtual Screening at Protein Interaction Sites

Article

pubs.acs.org/jmc

DARC: Mapping Surface Topography by Ray-Casting for Eﬀective

Virtual Screening at Protein Interaction Sites

Ragul Gowthaman,^†,○Sven A. Miller,^‡,○Steven Rogers,^§Jittasak Khowsathit,^‡Lan Lan,^‡Nan Bai,^‡

David K. Johnson,^†Chunjing Liu,^§Liang Xu,^‡,∥Asokan Anbanandam,^⊥Jeﬀrey Aube,^§,#Anuradha Roy,^∇

́

and John Karanicolas*^,†,‡

‡

§

^†Center for Computational Biology, Department of Molecular Biosciences, Center of Biomedical Research Excellence, Center for

∥

⊥

#

Cancer Experimental Therapeutics, Department of Radiation Oncology, Biomolecular NMR Laboratory, Department of

Medicinal Chemistry, and ^∇High Throughput Screening Laboratory University of Kansas, 2030 Becker Drive, Lawrence, Kansas

66045-7534, United States

S

* Supporting Information

ABSTRACT: Protein−protein interactions represent an exciting and

challenging target class for therapeutic intervention using small molecules.

Protein interaction sites are often devoid of the deep surface pockets presented

by “traditional” drug targets, and crystal structures reveal that inhibitors

typically engage these sites using very shallow binding modes. As a

consequence, modern virtual screening tools developed to identify inhibitors

of traditional drug targets do not perform as well when they are instead

deployed at protein interaction sites. To address the need for novel inhibitors of important protein interactions, here we

introduce an alternate docking strategy speciﬁcally designed for this regime. Our method, termed DARC (Docking Approach

using Ray-Casting), matches the topography of a surface pocket “observed” from within the protein to the topography

“observed” when viewing a potential ligand from the same vantage point. We applied DARC to carry out a virtual screen against

the protein interaction site of human antiapoptotic protein Mcl-1 and found that four of the top-scoring 21 compounds showed

clear inhibition in a biochemical assay. The K_ivalues for these compounds ranged from 1.2 to 21 μM, and each had ligand

eﬃciency comparable to promising small-molecule inhibitors of other protein−protein interactions. These hit compounds do not

resemble the natural (protein) binding partner of Mcl-1, nor do they resemble any known inhibitors of Mcl-1. Our results thus

demonstrate the utility of DARC for identifying novel inhibitors of protein−protein interactions.

docking methods that perform well for traditional drug targets

struggle to correctly identify compounds that disrupt protein

interactions.¹¹This problem is particularly acute because the

natural partners of these targets, proteins, cannot easily be used

for inspiration in designing small-molecule inhibitors or as

templates for ligand-based virtual screening, as is possible for

many enzymes^12,13and G protein-coupled receptors.¹⁴

In a number of cases, protein structures have been solved

both in complex with a biological protein partner and also in

complex with a small-molecule inhibitor. Comparison of the

unbound protein structure to the equivalent structure in

complex with a small molecule shows that binding is not

associated with a large conformational change, and yet, the

concave pocket on the protein surface in which the small

molecule binds is often absent in the unbound structure so that

the ﬂat surface of the unbound structure must often undergo

local rearrangement to reveal the small-molecule binding site.¹⁵

This has confounded structure-based methods for identifying

INTRODUCTION

■

Protein−protein interactions comprise the underlying mecha-

nisms for cell proliferation, diﬀerentiation, and survival;

manipulation of these interactions thus represents a promising

avenue for therapeutic intervention in a variety of settings.

Despite the urgent need for small-molecule modulators of

protein interactions, however, as recently as ﬁve years ago the

dearth of success stories meant that even the druggability of this

target class was very much in question.^1,2At the same time, the

challenges associated with these important targets help spur

reﬁnement of innovative new techniques, including fragment-

based methods,³peptidomimetics,⁴and nonpeptidic mimetics

of protein secondary structural elements.^5,6By now, these

advances have enabled discovery of inhibitors against many

diﬀerent protein−protein interactions and have led to an

impressive array of compounds in various stages of clinical trials

and preclinical optimization.^7,8

Modern methods for structure-based virtual screening have

proven eﬀective for a variety of applications, providing novel

hits to address a very wide range of “traditional” targets such as

enzymes and G protein-coupled receptors.^9,10Unfortunately,

structure-based virtual screening tools for identifying small-

molecule inhibitors of protein−protein interfaces have lagged

behind; due to their relatively exposed binding modes, the same

Special Issue: Computational Methods for Medicinal Chemistry

Received: January 27, 2015

A

DOI: 10.1021/acs.jmedchem.5b00150

J. Med. Chem. XXXX, XXX, XXX−XXX

Journal of Medicinal Chemistry

Article

Figure 1. Docking approach using ray-casting. (A) DARC ﬁrst casts a set of rays emanating from an origin within the protein (red dot), and maps

the topography of the surface pocket by monitoring intersection of these rays with the pocket (left). To evaluate the shape complementary of a given

ligand for this pocket, DARC casts the same rays (from the same origin) and monitors their intersection with the ligand (right). If ligand (in its

current position and orientation) is perfectly shape-complementary to the pocket, each ray will intersect the ligand at the same distance from the

origin as it intersected the protein surface pocket. By moving the ligand to maximize this shape complementarity, a ligand can be docked into a

protein surface pocket. (B) A schematic diagram of the complete workﬂow split into three stages: pre-DARC preparation, DARC, and post-DARC

reranking/ﬁltering of the docked models.

coordinates represent a map of the binding pocket’s topography

when viewed from the protein interior.

inhibitors: how can the protein structure be used to identify an

inhibitor, if the protein structure will change upon binding?

We recently described a method for identifying protein

surface pockets suitable for small-molecule binding, from either

cocrystal structures or from simulations starting from the

unbound protein structure.¹⁶Here, we introduce a method that

we call DARC (Docking Approach using Ray-Casting) for

matching these surface pockets to shape-complementary small

molecules. Because many inhibitors of protein interactions are

found to bind in shallow ﬂat pockets on the protein surface,^11,15

this method is speciﬁcally designed for such protein−ligand

interactions. Further, our method is intrinsically low-resolution

to allow for minor conformational variation of the protein upon

ligand binding. We anticipate this method will provide suitable

starting points for cases in which the protein structure is not

perfectly optimal for the ligand and cases in which further small

conformational changes of the protein are expected to

accompany ligand binding.

When viewed from the same vantage point (the origin), the

topography of a suitably bound ligand is expected to be highly

complementary to that of the protein’s binding pocket. To evaluate the

degree of complementary, we therefore cast the same set of rays

(deﬁned by their angles θ and ϕ) and determine the distance at which

each ray ﬁrst intersects the ligand (Figure 1a). If the ligand is indeed

complementary to the pocket, the distance to the ligand (ρ_ligand) for a

given ray will closely match the distance to the pocket (ρ_pocket) for the

corresponding ray.

We deﬁne the DARC “overlap score” as the degree to which the

pocket and ligand topographies match, by summing over all rays as

follows:

DARC overlap score

⎧

⎪

⎫

⎪

c1 × (ρ_ligand− ρ

)

if ρ

< ρ_ligand

pocket

c2 × (ρ

− ρ_ligand

if ρ_ligand< ρ

pocket

⎪

c3

c4

if ray does not intersect

ligand

⎨

⎬

⎪

= ∑ _⎪

COMPUTATIONAL APPROACH

rays

■

⎪

Computational methods are implemented in the Rosetta software

suite;¹⁷Rosetta is freely available for academic use (www.

rosettacommons.org), with the new features described here included

in the 3.6 release.

⎪

if ray does not intersect

pocket

⎪

⎩

⎪

⎭

(1)

Scoring with DARC. To map the pocket shape on the protein

surface, we use a variation of the LIGSITE¹⁸algorithm described fully

in our earlier work.¹⁶In its most basic form, this method involves

building a three-dimensional grid around a protein (or select region of

a protein) then tagging the grid points that are covered by the protein.

Next, each grid point not covered by the protein is examined to

determine whether the point is bounded on two opposing sides by the

protein along any direction of the grid, including diagonals. A

continuous group of such special points are labeled a “pocket.” We

then restrict this collection of “pocket” points to those “surface pocket”

points that immediately contact the protein: these points together

deﬁne the concave surface with which a ligand may interact.

From an “origin” point within the protein (determination of this

point is described in Supporting Information), we then cast a series of

rays toward the surface pocket; each ray connects the origin point to

one “surface pocket” point (Figure 1a). For each ray, we express the

position of the “surface pocket” point relative to the origin using

spherical coordinates (θ,ϕ,ρ_pocket); collectively, these spherical

From a physical perspective, rays that reach the pocket before the

ligand indicate underpacking between protein and ligand (condition

no. 1); alternatively, rays that reach the ligand before the pocket

indicate a steric clash (condition no. 2). Rays that do not intersect the

small molecule at all are penalized as well (condition no. 3); from a

physical perspective, this corresponds to the ligand not completely

ﬁlling the pocket. Finally, we include a series of rays that do not

intersect the pocket; intersection of these rays with the ligand indicates

that the ligand is too large for the pocket (condition no. 4).

For a given set of rays, evaluating the intersection distance of each

ray with the ligand is a task that is naturally very amenable to parallel

computing; indeed, the need to eﬃciently carry out the related “ray

tracing” operation provided motivation for development of graphics

processing units (GPUs). We have recently demonstrated that the ray-

casting approach we describe here can also be carried out on GPUs,

leading to dramatic speed enhancement relative to the analogous

calculation on a CPU alone.¹⁹

B

DOI: 10.1021/acs.jmedchem.5b00150

J. Med. Chem. XXXX, XXX, XXX−XXX

Journal of Medicinal Chemistry

Article

Determination of the values for parameters c1/c2/c3/c4 used in

these studies is described in Supporting Information.

chemical similarity to a compound in the “active” set for this

target (2D Fingerprint Tanimoto score²⁶> 0.7). To ensure

nonredundancy of the active set and nonredundancy of the

decoy set, we also removed any compound with strong

chemical similarity to another compound already present in the

same set (2D Fingerprint Tanimoto score >0.7).

For Bcl-xL, this approach yielded 27 (diverse) active

compounds, with 33 decoy compounds from the Astex set or

328 decoy compounds from the TIMBAL set. For XIAP, this

yielded 14 (diverse) active compounds, with 45 decoy

compounds from the Astex set or 425 decoy compounds

from the TIMBAL set. Although smaller than the TIMBAL

decoy sets, the Astex decoys have the advantage that these

compounds have typically been advanced through further

optimization for their respective targets; this makes the Astex

decoys less likely to exhibit (undesired) activity against Bcl-xL

or XIAP.

In addition, we also generated a third set of decoy

compounds using the DUD-E server.²⁷In this case, we

identiﬁed 50 “custom” property-matched decoys from each

active compound. Thus, the 27 compounds active against Bcl-

xL led to a set of 1350 DUD-E decoys, and the 14 compounds

active against XIAP led to a set of 700 DUD-E decoys. The

careful matching of the decoys’ physicochemical properties to

those of speciﬁc active compounds makes this a much more

challenging benchmark; at the same time, however, there is an

increased likelihood that some decoy compounds may

themselves be active against the target protein.

Screening against a Ligand-Bound Protein Structure.

The challenge in virtual screening entails identifying as many

active compounds as possible from a library (true positives)

while minimizing the number of inactive compounds

incorrectly assigned as active (false positives). A very stringent

score cutoﬀ for any method will result in few compounds being

assigned as active (such that some active compounds will be

missed), whereas relaxing this cutoﬀ will lead to many more

compounds assigned as active (such that more false positives

will be included). Because only a small fraction of the total

compound library is expected to be further evaluated or

validated experimentally, a typical goal for virtual screening is to

rank active compounds at the very top of the sorted list (good

“early” behavior), corresponding to the ability to build a small

subset of the original library that is strongly enriched in active

compounds.²⁸The “late” behavior (the fraction of the library

that must be screened to ensure no active compounds are

missed) is irrelevant for virtual screening applications because a

screening tool that oﬀers to eliminate only a small fraction of

(inactive) compounds from a large library is not particularly

useful.

Docking with DARC. Determining the optimal bound pose of a

small molecule within a pocket (i.e., “docking”) entails varying the

ligand position and orientation to minimize the DARC score eq 1.

Although the ﬁrst derivative of the DARC score can be calculated with

respect to the ligand position and orientation, we instead carry out a

derivative-free minimization using the particle-swarm optimizer

implemented within the Rosetta software suite.¹⁷The ligand internal

degrees of freedom are held ﬁxed during this search; to account for

ligand ﬂexibility, we prebuild ligand “conformers” using OMEGA²⁰⁻²²

and independently dock each of these. The ﬁnal DARC score for a

given ligand is taken to be the minimum DARC score from among

each of its docked conformers.

Application of DARC for virtual screening is summarized though

the workﬂow in Figure 1b. The starting pocket shape derives from a

protein crystal structure and can be an unbound structure, a bound

structure in which the ligand has been discarded, or a structure

generated from simulation (e.g., the “pocket optimization” simulations

we have described for identifying low-energy pocket-containing

conformations¹⁶).

RESULTS

■

Building a Virtual Screening Benchmark. Comparing

DARC to other methods for virtual screening requires

development of a benchmark to evaluate the performance of

each method for correctly identifying known inhibitors from

among “decoy” compounds. Assembling a large benchmark set

is complicated by the relatively small number of available crystal

structures of small-molecule inhibitors of protein interactions

that have been solved in complex with their protein targets.

Rather than assemble many tests with a very small number of

active compounds (known inhibitors), we instead built two

separate benchmark tests around protein interaction sites for

which numerous chemically diverse inhibitors have been

reported: Bcl-xL’s interaction with Bak peptides and XIAP’s

interaction with Smac/Diablo peptides. These two proteins

(and their homologues) were excluded from parametrization of

DARC (see Table S1 in Supporting Information) to ensure that

DARC would not inadvertently be trained to perform well

against these two protein targets.

When generating a set of “decoy” compounds for this

screening experiment, it is important that these compounds be

suitably matched to the “active” compounds. If the decoy

compounds are systematically diﬀerent in some physicochem-

ical property, for example, a docking method may successfully

pick out active compounds simply due to some implicit bias

(e.g., by simply picking the largest compounds, or the most

hydrophobic, etc.). On the other hand, it is important to

consider that decoy compounds are not necessarily inactive but

rather “presumed inactive”; typically, they have not been

explicitly tested experimentally against the target proteins in the

benchmark to ensure they are not active. If decoy compounds

are too similar to active compounds in chemical structure,

certain decoys may themselves be active, which will confound

analysis of the results (because the “right answer” in the

benchmark is to label decoy compounds as inactive).

We started from the crystal structures of Bcl-xL and XIAP,

each solved in complex with an inhibitor that had been

excluded from our library of active compounds. In addition to

DARC, we used ﬁve other methods to score and rank each

compound in this benchmark experiment. Four of these

(DOCK,^29,30AutoDock,³¹rDock,³²and PLANTS³³) are

receptor-based screening tools that we used to dock each

compound in the protein interaction site of each protein. The

ﬁfth (ROCS³⁴⁻³⁶) is a ligand-based screening tool, which ranks

compounds on the basis of how well their three-dimensional

structure mimics the volume and chemical features of a

template: in this case, the structure of the inhibitor solved in

complex with the target protein. Parameters used for each of

To generate well-matched decoy compounds for these

benchmarks, we drew from either (i) a set of 85 drug-like

compounds that comprise the Astex Diverse Set,²³or (ii) a set

of 6912 compounds reported to inhibit a diﬀerent protein−

protein interaction than our protein targets, compiled from the

TIMBAL database.^24,25

To reduce the likelihood that “decoy” compounds might be

active, we removed any decoy compounds that exhibited strong

C

DOI: 10.1021/acs.jmedchem.5b00150

J. Med. Chem. XXXX, XXX, XXX−XXX

Journal of Medicinal Chemistry

Article

Figure 2. Virtual screening benchmark experiment using an inhibitor-bound protein structure. (A) This receiver operating characteristic (ROC) plot

compares the performance of various methods (DARC, DOCK,^29,30AutoDock,³¹rDock,³²PLANTS,³³and ROCS³⁴⁻³⁶) for predicting whether a

given compound is active against a particular protein interaction site. Drug-like “decoy” compounds were drawn from the Astex diverse set,²³then

further ﬁltered to remove any compounds that are similar in chemical structure to an active compound or any other decoy compound. The results

are presented on a semilog plot to highlight the “early” performance of the methods; the gray curve indicates the random retrieval of compounds

(i.e., a random predictor). Left: discriminating 27 diverse compounds active against Bcl-xL from among 33 matched decoy compounds. Right:

discriminating 14 diverse compounds active against XIAP from among 45 matched decoy compounds. (B) DARC-docked models of representative

active compounds against Bcl-xL (left) and XIAP (right). These compounds were not chemically similar to those used in solving these crystal

structures (such compounds were excluded from this benchmark).

these other packages are included in the Supporting

Information.

considerations of shape complementarity alone cannot be

suﬃcient to reliably predict binding aﬃnity or details of the

interactions, it is striking that for these two benchmark tests the

simple approach employed by DARC successfully distinguishes

active from inactive compounds with similar accuracy as more

sophisticated methods. We further note that the performance of

DARC in this benchmark is remarkably robust to the location

of the “origin” point from which rays emanate nand provided a

reasonable approach is used to deﬁne this point (Figure S3 in

Supporting Information). Thus, these observations support the

hypothesis that the crude pocket shape on the protein surface

provides a strong restriction on the chemical “shape space” of

inhibitory compounds.

Screening against a “Pocket Optimized” Protein

Structure. The fact that a protein interaction site is likely to

undergo a conformational change upon inhibitor binding¹⁵

makes the choice of receptor conformation of particular interest

for this target class, especially given that the performance of

most approaches decreases considerably when an unbound

starting structure is used in place of a ligand-bound starting

structure.³⁷A variety of methods exist for incorporating

receptor ﬂexibility, typically by docking against multiple

receptor conformations from either crystal structures³⁸⁻⁴²or

simulation.⁴³⁻⁴⁵

Given the rank order of each compound using each of the

three methods, we prepared separate receiver operating

characteristic (ROC) curves for both Bcl-xL and XIAP,

corresponding to an experiment in which we screen a library

comprised of active compounds and Astex decoys (Figure 2a).

Each point on such a plot corresponds to a given score cutoﬀ,

and each point indicates the fraction of active compounds

collected at this cutoﬀ as a function of the inactive compounds

incorrectly assigned as active at this cutoﬀ. On this plot, the

critical “early” behavior described earlier corresponds to a steep

rise at the leftmost part of the plot: a useful method for

screening must pick out as many active compounds as possible

while accumulating a very small fraction of the inactive

compounds. All three methods outperform the behavior that

would result from randomly ranking the compounds in the

library (gray curve), especially in this “early” region. We also

observe similar results when carrying out the same screening

experiment using decoy compounds drawn from the TIMBAL

set rather than from the Astex set (Figure S1 in Supporting

Information). With the exception of ROCS applied to XIAP,

meanwhile, none of the methods perform at a comparable level

when used to distinguish known active compounds from their

closely matched decoys in the DUD-E set (Figure S2 in

Supporting Information).

Rather than start from unbound crystal structures of Bcl-xL

and XIAP, which lack the surface pocket required for inhibitor

binding, we instead started from conformations generated using

the “pocket optimization” approach we have recently

described.¹⁶Brieﬂy, we developed a biasing potential that

drives conformational sampling in molecular mechanics

simulations toward conformations containing surface pockets

suitable for small-molecule binding. This biasing potential

As seen in representative examples of docked active

compounds (Figure 2b), the DARC-generated models exhibit

clear shape complementarity that allowed these compounds to

be (correctly) identiﬁed as true ligands; experimentally derived

structures of these compounds in complex with their protein

partners have not been reported to date. While low-resolution

D

DOI: 10.1021/acs.jmedchem.5b00150

J. Med. Chem. XXXX, XXX, XXX−XXX

Journal of Medicinal Chemistry

Article

Figure 3. Virtual screening benchmark experiment using protein structures from “pocket optimization” simulations. (A) This benchmark experiment

involves discrimination of the same active versus “decoy” compounds as in Figure 2, however, this time compounds were screened against a protein

conformation generated via “pocket optimization” simulations initiated from an unbound crystal structure (instead of protein conformations from an

inhibitor-bound crystal structure).¹⁶The results are again presented on a semilog plot to highlight the “early” performance of the methods; the gray

curve indicates the random retrieval of compounds (i.e., a random predictor). (B) Models of representative active compounds docked by DARC to

the “pocket optimized” conformation of Bcl-xL (left) or XIAP (right). (C) The performance of DARC in this benchmark is essentially equivalent

when screening against a known ligand-bound structure (solid black line), the lowest-energy “pocket optimized” conformation (i.e., as described in

the previous panels) (solid red line), or any of ﬁve other low-energy “pocket optimized” conformations (dashed red lines); this observation

demonstrates the insensitivity of the method to details of the protein structure.

operates solely on simple geometric considerations so that the

shapes of the resulting surface pockets are inﬂuenced solely by

properties of the protein surface and not knowledge of any

particular ligand. Starting from unbound crystal structures, we

carried out simulations to generate “pocket optimized”

conformations of Bcl-xL and XIAP: conformations containing

a surface pocket, but without the precise details that would be

encoded in a ligand-bound structure (our application of this

approach to generate these protein conformations is described

in Supporting Information).

We rescreened the same compound library described earlier,

this time using a “pocket optimized” conformation of Bcl-xL or

XIAP (Figure 3a). We did not include ROCS in this stage of

the benchmark because there is no structure of a cognate ligand

available for this protein conformation that could be used as a

template. Unsurprisingly, the performance of DOCK, Auto-

Dock, rDock, and PLANTS all suﬀer dramatically in this more

challenging regime. In contrast, the performance of DARC is

remarkably similar to the earlier experiment, demonstrating that

the low-resolution nature of the underlying calculations makes

DARC relatively robust to slight mismatches between protein

and ligand. Indeed, representative examples of active

compounds docked to these alternate conformations demon-

strate that the overall pocket shapes remain complementary to

these ligands (Figure 3b). We note that these observations also

hold when carrying out the same experiment using decoy

compounds drawn from the TIMBAL set rather than the Astex

set (Figure S4 in Supporting Information), but that again all of

these methods exhibit diminished performance against the

DUD-E set (Figure S5 in Supporting Information).

In summary, slight diﬀerences in these protein conformations

lead to diminished performance from other receptor-based

docking methods in this regime, but DARC performs equally

well in this experiment as in the previous benchmark that used

protein conformations from ligand-bound crystal structures.

This ﬁnding is conﬁrmed by extending this experiment to

E

DOI: 10.1021/acs.jmedchem.5b00150

J. Med. Chem. XXXX, XXX, XXX−XXX

Journal of Medicinal Chemistry

Article

Figure 4. DARC can be used to identify novel inhibitors of human Mcl-1. Screening was carried out against the protein interaction site of Mcl-1, and

21 of the top-scoring compounds were tested in a biochemical assay. (A) Docked models generated by DARC for four compounds emerging from

our screen. Compounds were screened either against an experimentally derived peptide-bound structure of Mcl-1 or against an Mcl-1 conformation

generated via “pocket optimization” simulations.¹⁶(B) The chemical structures of the compounds in each model are shown. We note that some of

these have reactive functional groups: such compounds could be removed from the library prior to screening in applications where such moieties are

undesirable. (C) The most likely activity of each compound was predicted using the similarity ensemble approach (SEA).⁷⁵None of these

compounds are assigned as likely inhibitors of Bcl-2-family proteins, underscoring their lack of similarity to known inhibitors of this family. (D) Each

of these four compounds inhibit Mcl-1’s interaction with a FITC-labeled cognate peptide, as determined via ﬂuorescence polarization. CHAPS

detergent (0.1%) was included in this assay, to ensure that these results were not due to compound aggregation; we also observe equivalent

inhibition in the absence of CHAPS. Each curve is ﬁt to a one-site inhibition model, with a single free parameter in the ﬁtting. Direct binding of each

of these compounds to Mcl-1 was subsequently conﬁrmed via biolayer interferometry (see Supporting Information).

consider a series of ﬁve more distinct low-energy conformations

for each protein, as generated by the “pocket optimization”

approach; we ﬁnd that the performance of DARC in this

benchmark is notably insensitive to the precise details of the

protein conformation (Figure 3c).

Collectively, these experiments point to diﬀerent regimes in

which each class of method is expected to prove superior.

Detailed docking methods, such as DOCK, AutoDock, rDock,

and PLANTS, are expected to out-perform DARC if the

protein structure is solved in complex with a ligand of similar

chemotype to the desired compounds. By contrast, DARC

allows greater discrimination if the protein conformation is not

quite optimal for the ligand; this can occur when one wishes to

discover inhibitors with a radically diﬀerent chemotype than

known ligands or when the protein conformation derives from

simulation.

Identifying Novel Mcl-1 Inhibitors Using DARC. To

evaluate the performance of DARC in a realistic application, we

next applied this approach to screen for novel classes of

inhibitors for Mcl-1, a member of the Bcl-2 family of proteins.

Individual members of this protein family can serve either a

pro- or antiapoptotic role and interact with one another

through a structurally conserved binding motif. Small-molecule

inhibitors of Bcl-2/Bcl-xL have shown promising eﬃcacy in

overcoming chemo/radioresistance in various tumor models

including prostate cancer.⁴⁶⁻⁵⁰One such compound is ABT-

737, a potent Bcl-2/Bcl-xL inhibitor;⁵⁰a more recent Bcl-2-

selective derivative of this compound is currently in clinical

trials.⁵¹Recent studies have shown that cancer cells resistant to

ABT-737 have high levels of Mcl-1 and that knockdown of Mcl-

1 promotes ABT-737-induced apoptosis.^{35,48,49,52,53}Taken

together, these observations motivate the pressing need for

development of potent and selective inhibitors of Mcl-1 in

treating of a variety of cancers to be used either as a single agent

or in combination with inhibitors of Bcl-2/Bcl-xL.⁵⁴

The strong evidence supporting the potential impact of

eﬀective Mcl-1 inhibitors has spurred intense eﬀorts using a

variety of complementary approaches,^54,55including high-

throughput screening of standard libraries using a ﬂuorescence

polarization competition assay^56,57and of an sp³-rich library

using diﬀerential scanning ﬂuorimetry.⁵⁸Aiming to build upon

the success of the ABT-737 series for Bcl-2, several groups have

since applied similar NMR-based fragment screening to Mcl-1;

fragment hits were prioritized on the basis of ligand eﬃciency

(binding free energy per heavyatom), ﬁrst yielding weakly

binding compounds that were subsequently merged or

elaborated to give inhibitors with improved potency at the

expense of some ligand eﬃciency.⁵⁹⁻⁶²Inspired by the helical

conformation adopted by Mcl-1’s interaction partners, a

number of groups have grafted the interaction partners’ side

chains onto small-molecule scaﬀolds that allow them to be

presented in a similar geometry⁶³⁻⁶⁵or have used “stapled”

F

DOI: 10.1021/acs.jmedchem.5b00150

J. Med. Chem. XXXX, XXX, XXX−XXX

Journal of Medicinal Chemistry

Article

a

Table 1. DARC Screening Hits Conﬁrmed Experimentally As Inhibitors of Mcl-1

a

The complete set of DARC hits (including those that did not inhibit Mcl-1) are included as Table S2 in Supporting Information. K_ivalues were

determined via a ﬂuorescence polarization competition assay (see Supporting Information), and then bio-layer interferometry was used to directly

conﬁrm binding to Mcl-1 (see Supporting Information).

derivatives of the helical peptides.⁶⁶⁻⁶⁸Such approaches

predicated on mimicry of a known helical binding partner

can rapidly lead to potent inhibitors, but the large chemical

scaﬀolds required to recapitulate the helical geometry diminish

from their ligand eﬃciency.

bound structure of Mcl-1 and also separately using the “pocket

optimized” conformation of Mcl-1.

The docked models produced by DARC are intrinsically low

resolution because they rely only on matching the protein−

ligand topography. To achieve further high-resolution discrim-

ination, we therefore included an additional ﬁnal step (Figure

1b): for each of the top ranked 10% of the complexes produced

by DARC we carried out fullatom gradient-based minimization

in Rosetta, using the standard Rosetta fullatom energy

function¹⁷and ligand parameter deﬁnitions.⁷³All internal

dihedral angles of the protein were included as degrees of

freedom, along with the ligand position and orientation relative

to the protein. We then ﬁltered the resulting models to remove

those with no intermolecular hydrogen bonds or an abundance

of buried unsatisﬁed polar groups. From the remaining models,

we then purchased the top-scoring 21 compounds on the basis

of DARC score (Table S2 in Supporting Information). Of these

21 compounds, 11 were identiﬁed by screening against the

(peptide-bound) crystal structure of Mcl-1, and the other 10

were from screening against the “pocket optimized” structure.

The structural basis for the favorable scores in our DARC

screen is evident from the models of these compounds in

complex with Mcl-1 (Figure 4a). In each case the ligand, in its

modeled conformation, exhibits exquisite shape complemen-

tarity for the protein surface; this is unsurprising given that the

DARC score is expressly built to identify the ligand position,

orientation, and conformation that will maximize similarity to

the topography of the protein surface. Nonetheless, there is also

clear diversity among the models; the ligand shapes are

diﬀerent from one another, and they ﬁll the protein surface

Like other members of the Bcl-2 family, Mcl-1’s interaction

partners bind to an exposed hydrophobic groove on the protein

surface.⁶⁹We recently applied our “pocket optimization”

approach to this protein surface and generated ensembles of

low-energy Mcl-1 conformations that present surface pockets

suitable for small-molecule binding.⁷⁰Among these ensembles,

we ﬁnd conformations similar to those observed in crystal

structures of Mcl-1 bound to diverse small-molecule ligands,

and we also observe “distinct” conformations that have not yet

been observed in any experimentally derived structures.⁷⁰We

elected to use DARC to screen chemical libraries for

compounds that would complement one of two Mcl-1

conformations: either an experimentally derived conformation

or a conformation derived from “pocket optimization”

simulations. At the time we initiated these studies, no unbound

structure was available,⁷¹so for the former we used a peptide-

bound crystal structure (PDB ID 2pqk).⁶⁹For the latter, we

used the lowest-energy pocket-containing conformation gen-

erated from 1000 independent trajectories.

Drawing from the ZINC database,⁷²we compiled a small

virtual library corresponding to 62442 highly diverse com-

pounds with drug-like properties (MW ≤ 500 Da, xlogP ≤ 5,

etc., see Supporting Information) available for immediate

purchase from a commercial vendor. We then applied DARC to

separately dock and rank each compound using the peptide-

G

DOI: 10.1021/acs.jmedchem.5b00150

J. Med. Chem. XXXX, XXX, XXX−XXX

Journal of Medicinal Chemistry

Article

pocket in distinctly diﬀerent ways. In the case of M0, for

example, a hydrophobic groove on the Mcl-1 surface is neatly

complemented by the spatial arrangement of the two aromatic

rings in the selected ligand conformation.

particular, these four initial hits are less active than the most

potent compounds arising from screens of libraries designed to

mimic the side chain interactions of helical cognate

peptides;⁶³⁻⁶⁵however, the DARC hits exhibit greater ligand

eﬃciency due to the extensive size of the compounds designed

for mimicry of the helix.

The selected compounds also represent a diverse array of

diﬀerent chemical scaﬀolds (Figure 4b), with no evident

similarity to one another. It must be noted that of the

compounds shown, only M7 (a natural product) is devoid of

potentially reactive functional groups. This is, of course, a

reﬂection of the screening library: in a typical drug discovery

application, such compounds should be removed from the

library prior to screening because their potential for advance-

ment is likely to be extremely limited.⁷⁴For the purpose of

evaluating the performance of DARC in this virtual screening

experiment, however, we did not exclude such compounds.

By visual inspection, the structures of these compounds bear

no obvious resemblance to any known inhibitors of Mcl-1 or to

any known inhibitors of any other Bcl-2 family members. To

systematically identify biologically active compounds most

resembling these compounds, we applied the similarity

ensemble approach (SEA); this method uses ligand similarity

to identify likely receptors for a query ligand.⁷⁵In most cases,

SEA identiﬁes chemical scaﬀolds with clear resemblance to our

queries, but these related compounds have each been described

in the context of very diﬀerent activities (Figure 4c); even

delving further into each list, we did not ﬁnd similarity to other

inhibitors of Bcl-2 family members. This observation highlights

the fact that these compounds would have been very diﬃcult to

pick out using ligand-based screening due to their lack of

similarity to any known inhibitors of this protein family.

We next examined the ability of each compound to inhibit

the interaction of Mcl-1 with one of its binding partners. We

used a ﬂuorescein-labeled peptide derived from the Noxa

protein which exhibits an increase in polarization upon Mcl-1

binding (Figure S6a in Supporting Information). Upon

addition of a known Mcl-1 inhibitor, AT-101 (i.e., R-

(−)-gossypol), we observe a dose-dependent decrease in

polarization that conﬁrms this compound competes with

Noxa for Mcl-1 binding and is consistent with previously

reported data⁷⁶(Figure S6b in Supporting Information). Using

this assay we found that six of the 21 compounds from our

computational screen appeared to inhibit the interaction

between Mcl-1 and its cognate peptide with K_ivalues ranging

from 1.2 to 21 μM (Figure 4d). These results were not aﬀected

by the presence of absence of detergent (CHAPS), suggesting

that compound aggregation was not contributing to the

observed inhibition. We then used an orthogonal label-free

assay based on biolayer interferometry (BLI) to test for direct

binding of these compounds to Mcl-1 and through this assay

conﬁrmed the activity for four of these six compounds (see

Figure S7 in Supporting Information): M0, M1, M5, and M7

(Table 1). In all four cases we found that inhibition did not

increase over time in the competition assay and that the BLI

signal showed a dissociation phase (decreased amplitude) upon

transfer into buﬀer lacking the compound of interest; together,

these imply that all four compounds bind in a reversible

manner.

Superposition of the modeled complexes for each of the

DARC hits with the structure of Mcl-1 bound to a cognate

peptide reveals that these compounds are decisively not

recapitulating the interactions of the helical peptide (Figure

5a). While the cognate peptide uses a cluster of aliphatic

residues at its N-terminus and aromatic residues at its C-

terminus, compound M0 (as a representative example)

occupies only the regions of the Mcl-1 surface engaged by

the N-terminus of the peptide. Given that it is not constrained

to ﬁt onto the backbone of the helix, this compound also ﬁts

more deeply into the binding pocket on the Mcl-1 surface. In

contrast, the most potent of the helix mimetics was designed to

recapitulate the complete set of interactions provided by the

helical template.⁶³The side chains that comprise these

interactions are separated by a span of 15 residues in sequence,

and while this led to an inhibitor with K_iof 0.2 μM, the large

chemical scaﬀold needed to present these interactions reduced

the ligand eﬃciency of this compound to 0.21 kcal/mol·

heavyatom.

While the underlying methodology used in our DARC screen

did not explicitly select for compounds that bind in place of the

N-terminus of the cognate peptide, each of the resulting hits

complemented this region. This is in stark contrast to the

binding modes observed in crystal structures of Bcl-xL solved in

complex with potent inhibitors such as ABT-737 and WEHI-

539; each of these bind to the region of the protein surface used

by the C-terminus of the cognate peptide, and thus have

relatively little overlap with the models of the DARC hits

(Figure 5b). In the time since we carried out our computational

screen, crystal structures of Mcl-1 solved in complex with four

distinct classes of inhibitor have been reported^58−60,77and in

each case the ligand occupies the region of the protein surface

corresponding to the N-terminus of the cognate peptide. Thus,

the lack of DARC hits complementing the protein surface used

by inhibitors of Bcl-xL may indicate a lack of druggability for

the analogous surface of Mcl-1.

In contrast to Mcl-1 inhibitors derived by mimicry of the

cognate helix, the most promising hits emerging from fragment

screens initially proved less potent than those identiﬁed by

DARC but had superior ligand eﬃciency.⁵⁹⁻⁶²One representa-

tive class of fragments inhibited Mcl-1 with K_iof only ∼100

μM, albeit with a promising ligand eﬃciency of about 0.3 kcal/

mol·heavyatom.⁵⁹NMR-derived models of the binding mode

for this class of compound suggested that they induce a

conformational change in Mcl-1, which in turn produces a very

deep pocket required for binding; upon merging with another

fragment class, crystallography conﬁrmed that the larger

compound also induced this conformational change.⁵⁹In

retrospect, each of the Mcl-1 crystal structures in complex

with an inhibitor (solved after completion of our screen)

revealed a conformational change that allowed the ligand to

access a deep pocket that was not evident from the unbound or

peptide-bound structures^58−60,77(Figure 5c). The observation

that extensive ligand burial is related to high ligand eﬃciency

(across many small-molecule inhibitors of protein−protein

interactions)¹¹is thus consistent with the impressive ligand

eﬃciency of these (deeply buried) Mcl-1 fragment hits.

Given the chemical structures of these four most potent

compounds, these observed K_ivalues correspond to ligand

eﬃciency (binding free energy per heavyatom) of 0.21 to 0.27

kcal/mol·heavyatom, which is typical of other promising small-

molecule inhibitors of protein−protein interactions reported in

the literature.⁸With respect to other Mcl-1 inhibitors in

H

DOI: 10.1021/acs.jmedchem.5b00150

J. Med. Chem. XXXX, XXX, XXX−XXX

Journal of Medicinal Chemistry

Article

However, such conformations were not among those included

in our DARC screen, and accordingly the DARC hits could not

have taken advantage of the deep pockets presented upon

conformational rearrangement of the protein.

Evaluating the DARC Model of M0 Binding. An

important beneﬁt that can arise from the use of structure-

based modeling (such as DARC) is the prospect of immediately

using the model of the complex to guide design of derivatives

once an initial hit compound is validated experimentally. To

test whether a DARC model can indeed serve this purpose, we

next sought to test whether the tighter-binding derivatives

could be designed from the DARC model. Despite the fact that

M0 (2′,4-dihydroxy-3,4′,6′-trimethoxychalcone) was the least

potent of the four hits in our original screen and concerns of

promiscuity because similar chalcones have been found to

exhibit other diverse activities in biological assays, the

straightforward synthetic accessibility of the chalcone series

prompted us to select this scaﬀold for further investigation.

While this compound would not necessarily be an optimal

starting point for advancement into a drug or biological probe,

then, selecting of this compound as a starting point allowed us

to make and test more derivatives than would have been

possible with the other three validated DARC hits.

Starting from compound M0, we therefore we sought to ask

whether such a model could be used to explain the structure−

activity relationship (SAR) observed across a series of related

compounds. Because our goal was to gauge the veracity of the

modeled complexes that led to selection of these compounds,

and not simply to identify the most potent derivatives, we

further restricted our study to compounds very close to the

parent compound, M0; we reasoned that these would best

allow us to test our model of the structure, reducing the

likelihood of conformational changes leading to drastic and

unexpected SAR.

Before designing a series of analogues, however, we noted

that other chalcones are known to form covalent adducts to

proteins (the chalcone α,β-unsaturated carbonyl system is a

Michael acceptor and thus may react with unpaired cysteine

side chains).^78,79We therefore carried out analysis via surface

plasmon resonance (Figure S8 in Supporting Information) and

HSQC chemical shift mapping (Figure S9 in Supporting

Information) to further characterize this interaction (see

Supporting Information). Taken together, the stoichiometric

Hill coeﬃcient⁸⁰and insensitivity to the presence of

detergent⁸¹in the ﬂuorescence polarization assay, the fact

that the SPR response did not exhibit superstoichiometric

behavior,⁸²and the observation of a small number of distinct

chemical shift diﬀerences in the HSQC spectrum⁸³all provide

evidence that the observed activity of M0 is not due to

compound aggregation that deactivates Mcl-1 but results from

speciﬁc binding. The NMR experiment also supports two

features of our model for M0 binding: ﬁrst that the M0 binding

mode is not completely overlapping with that of the Bcl-xL

inhibitors (Figure 5b) and second that M0 binding does not

induce extensive protein conformational changes as seen in

other Mcl-1 inhibitors (Figure 5c).

Figure 5. Evaluating DARC’s model of the M0/Mcl-1 complex. (A)

Though M0 was identiﬁed in a screen against a peptide-bound

conformation of Mcl-1, M0 (peach) is distinctly not mimicking the

interactions of the cognate peptide (light gray). In the DARC-

generated model, M0 competes with the N-terminus of the peptide for

Mcl-1 binding but ﬁts more deeply into the Mcl-1 surface pocket than

the peptide side chains. (B) This M0 binding mode is distinctly

diﬀerent from that of known Bcl-xL inhibitors such as ABT-737

(green), which instead overlap with the C-terminus of the peptide-

binding site. The surface of the Bcl-xL protein is shown (gray), solved

in complex with ABT-737. (C) Other recently described inhibitors of

Mcl-1, including the representative example shown here (dark pink),⁵⁹

bind at a similar location to M0; however, each distinct series of

inhibitors induce their own unique Mcl-1 conformational change to

allow deeper ligand burial. (D) The locations of two M0 positions that

play key roles in our model of the complex: the 6′-methoxy group

(yellow) and the unsubstituted 5-position (green). (E) A model of the

D1, the most potent M0 derivative, in complex with Mcl-1. This

compound preserves the crucial 6′-methoxy group, and embellishes

the 5-position with a chlorine substitution to improve packing. Closely

analogous compounds to D1 that lack the methoxy group at the 6′-

position (D3) or that lack the chlorine at the 5-position (D5) exhibit

reduced activity.

We then proceeded to produce a variety of M0 derivatives

using a simple aldol condensation reaction (Figure S10 in

Supporting Information). We were concerned that designing

these analogues from our model of the M0/Mcl-1 complex

might bias the resulting series toward certain regions of

chemical space and thus not provide a fair assessment of our

model. To evaluate the structure−activity relationship more

I

DOI: 10.1021/acs.jmedchem.5b00150

J. Med. Chem. XXXX, XXX, XXX−XXX

Journal of Medicinal Chemistry

Article

a

Table 2. Chemical Analogues of M0

J

DOI: 10.1021/acs.jmedchem.5b00150

J. Med. Chem. XXXX, XXX, XXX−XXX

Journal of Medicinal Chemistry

Article

Table 2. continued

K

DOI: 10.1021/acs.jmedchem.5b00150

J. Med. Chem. XXXX, XXX, XXX−XXX

Journal of Medicinal Chemistry

Article

Table 2. continued

L

DOI: 10.1021/acs.jmedchem.5b00150

J. Med. Chem. XXXX, XXX, XXX−XXX

Journal of Medicinal Chemistry

Article

Table 2. continued

a

We designed and synthesized a series of 27 M0 derivatives and tested each of these for inhibition of human Mcl-1 using a ﬂuorescence polarization

competition assay (see Supporting Information). This series of analogues was designed without consideration of DARC’s model of the M0/Mcl-1

complex to allow unbiased evaluation of the model.

objectively, we therefore designed a set of compounds by

selecting the most readily available starting materials based only

upon the chemical structure of M0 and without consideration

of our model of the complex. Using the same ﬂuorescence

polarization competition assay described earlier, we charac-

terized a total of 27 analogues of M0 (Table 2, Figure S11 in

Supporting Information). As described below, the SAR

deduced from this series presents a compelling narrative

when interpreted using our model of M0 binding (Figure 5d).

The most potent of these M0 analogues are D1 and D2: they

each inhibit Mcl-1 about 7-fold more potently than M0.

Compounds D1 and D2 diﬀer only by a single methoxy versus

ﬂuorine substitution at the 3-position; M0 harbors a methoxy

here. This methoxy group is exposed in our model of the M0/

Mcl-1 complex, consistent with the observation from D1 versus

D2 that its replacement with ﬂuorine does not aﬀect activity.

Compounds D1 and D2 together share three diﬀerences

relative to M0: both have a chlorine at the 5-position, both have

a methoxy group at the 5′-position, and both lack the 2′-

hydroxyl of M0. The latter two positions are exposed in our

model of M0 binding and are thus assumed not to strongly

aﬀect activity. In contrast, the 5-position (unsubstituted in M0)

faces toward the protein and points toward a small cavity

(Figure 5d), thus the improved packing resulting from this

substitution may explain the slightly enhanced potency of D1/

D2 relative to M0.

The only other two compounds in this series with potency

exceeding that of M0 are D3 and D4; this equipotent pair again

diﬀers only by the same ﬂuorine versus methoxy substitution

that distinguished D1 and D2. Relative to D1/D2, compounds

D3/D4 lack the 5′- and 6′-methoxy groups. While the 5′-

position is unsubstituted and exposed in our model of the M0/

Mcl-1 complex, the 6′-methoxy group is deeply buried in a

hydrophobic cavity (Figure 5d); loss of this interaction may

explain the reduced potency of D3/D4 relative to D1/D2.

Across this series of 27 M0 analogues, only compounds D1−

D4 proved more potent than M0. While the SAR described

above is consistent with our model of M0 binding, these results

are also notably inconsistent with the behavior expected if these

compounds were simply carrying out thia-Michael additions.

Kinetic studies using a model thiol (cysteamine) demonstrate

that the presence of the 6′-methoxy group decreases reactivity,

and the 2′-hydroxy group increases reactivity.⁸⁴Through the

comparisons above, we ﬁnd that the most potent compounds

for inhibition of Mcl-1 correspond to those expected to be least

reactive, further implying that the inhibition we observe is not

due to reactivity of these compounds with Mcl-1’s unpaired

cysteine side chain.

Through the results presented in Table 2 we demonstrate

that chalcones are not “privileged scaﬀolds”^85,86for inhibition

of Mcl-1, an assertion that is best supported by speciﬁc

examples. Compounds D7/D12/D13/D14 lack the key 6′-

methoxy group described earlier (Figure 5d) and accordingly

are less potent than M0. Compound D5 preserves each of the

M0 functional groups that interact with Mcl-1 in our model

(i.e., the 6′-methoxy group and the 4-hydroxyl group) and

accordingly exhibits a very similar inhibition constant as M0.

Compound D8 maintains the substituents of D2 on one ring

but lacks the substituents on the other ring (including the 6′-

methoxy group) and is thus less potent. Compounds D9/D10/

D11 maintain the 6′-methoxy group but harbor a variety of

alternate substituents on the other ring, making them less

potent as well. Compounds harboring extra fused rings (D20/

D21/D22/D23) are inevitably less potent, as are compounds in

which the chalcone linkage has been replaced with a ﬂavone

(D24/D25/D26/D27). Overall, the assertion that chalcones

are not privileged scaﬀolds for inhibition of Mcl-1 or generic

helix mimetics is supported by the observation that most

analogues presented here, designed without consideration of

our model of binding, inhibit Mcl-1 less potently than the

parent compound M0. In summary then, it is the precise

complementary of select compounds for the surface of Mcl-1

that dictate their activity and not some property of this

chemical scaﬀold.

Ultimately, direct structural evidence will be required to

conﬁrm that these compounds are indeed binding via the

designed pose. Unfortunately, our eﬀorts to crystallize Mcl-1 in

complex with a member of this chemical series have not yet

proven successful. Validating the models that lead to selection

of active compounds through structural biology will be

important not only for retrospective evaluation of the DARC

models but also as a starting point for inspiring design of

subsequent analogues. Accordingly, we used DARC to compare

our model of the M0/Mcl-1 complex to a model in which M0

was replaced with D1 (Figure 5e). We ﬁnd that DARC scores

D1 even more favorably than M0 due to the improved packing

resulting from substitution of chlorine at the 5-position. Thus,

we expect that compounds with even better shape comple-

mentarity for Mcl-1 than those selected in our initial screen,

compounds such as D1, can be identiﬁed by screening larger

chemical libraries with DARC. By further analogy to D1, we

anticipate that such compounds will also exhibit superior

M

DOI: 10.1021/acs.jmedchem.5b00150

J. Med. Chem. XXXX, XXX, XXX−XXX

Journal of Medicinal Chemistry

Article

potency relative to the initial hits described in this ﬁrst screen.

Despite the improved potency relative to M0, we nonetheless

note that compound D2 does not exert the desired biological

eﬀect using an in vitro cellular assay (see Supporting

Information, Figure S12). This underscores the need to test

candidate inhibitors in cellular assays as well because

biochemical activity may not necessarily translate to cellular

activity. Indeed, careful analysis has shown that a number of

other inhibitors, each with comparable binding aﬃnity for

Mcl-1 as that of compound D2, also appear not to act on Mcl-1

in cells.^57,87,88

electrostatic complementarity into DARC lead to improved

performance in virtual screens against protein surfaces that

display surface polar groups.⁹⁹

Approaches for Identifying Inhibitors of Protein−

protein Interactions. In many cases, artiﬁcial ligands that

bind to a speciﬁc protein surface bear some resemblance to an

endogenous ligand that also binds to that surface; analysis of

precedented drug targets shows that knowledge of a substrate,

product, or eﬀector with “drug-like” physicochemical properties

is a good predictor of druggability for a given protein surface.¹⁰⁰

Analogues of endogenous ligands can also provide a starting

point for designing inhibitors.¹⁰¹In the case of protein

interaction sites, however, the natural ligand is not a small

molecule and thus does not provide an obvious template from

which to start. This has driven advances in new methodologies

for identifying inhibitors of these “nontraditional” targets, most

notably including mimicry of key interacting groups of the

protein partner^{63−65,102−104}and fragment-based ap-

proaches.^59,105−107

In addition to chemical scaﬀolds that mimic secondary

structural elements,⁶³⁻⁶⁵exciting new computational ap-

proaches have facilitated identiﬁcation of compounds that

instead mimic the geometric orientation of key “hotspot” or

“anchor” side chains in a protein−protein interface.^102−104

While either mimicry-based approach may provide a robust

starting point for recapitulating the interactions of a protein

partner using a small-molecule scaﬀold, both are likely to

encounter the same intrinsic limitation: ligand eﬃciency is

unlikely to exceed that of the interacting groups from the

protein partner, which in turn is usually less than that those

achieved by small-molecule inhibitors that do not explicitly

mimic interactions of the protein partner.¹⁵Accordingly, an

important advantage of de novo structure-based screening

methods that do not rely on mimicry, such as DARC, is the

potential to identify inhibitors that achieve superior ligand

eﬃciency through deeply buried interactions not available to a

protein-based ligand.

In contrast, fragment-based approaches often prioritize small

compounds with very high ligand eﬃciency from the outset:

this allows exploration of deep crevices that would not

necessarily be evident from the structure of the protein−

protein complex. Here, the challenge often lies in elaborating

initial fragment hits (by growing, merging, or linking them)

into larger compounds that maintain these highly productive

interactions.^108,109In our DARC screen against Mcl-1, we

focused on compounds larger than traditional fragments; this

led to initial hits with easily detectable activity, albeit at some

expense of ligand eﬃciency. Much as the “multiple solvent

crystal structures” method¹¹⁰uses very small fragments to

probe the protein surface for productive interactions with

isolated functional groups, screening a commercially available

library of “prototype” compounds, as we have done here, allows

rapid evaluation of speciﬁc protein−ligand interactions

predicted in silico and can provide potential chemical scaﬀolds

for further optimization.

Screening against Multiple Protein Conformations.

The compounds identiﬁed as “hits” by DARC, and by

structure-based screening methods in general, naturally depend

on the conformation of the protein target. With respect to

inhibitors of protein−protein interactions in particular, a

number of examples have been shown to bind concave surface

pockets that are absent in the corresponding unbound protein

structures; these “cryptic” pockets are revealed by local

DISCUSSION AND CONCLUSION

■

Screening on the Basis of Surface Topography.

Because the protein surfaces of “traditional” drug targets have

evolved to bind some cognate small-molecule ligand, they

typically include a deep pocket or groove that can be targeted

by inhibitors. In contrast, protein interaction sites typically do

not include such surface features and accordingly small-

molecule inhibitors acting at these sites rely on shallower

bound poses.¹¹Because these shallow binding modes prove

challenging to predict by conventional docking approaches,

modern virtual screening tools exhibit diminished performance

when used to predict inhibitors of protein interactions (relative

to their performance when applied to “traditional” drug

targets).¹¹Here we present DARC, an alternate approach to

docking and virtual screening. By matching the topography of

the protein surface to the buried face of the ligand, we ﬁnd that

DARC outperforms popular canonical tools (DOCK, Auto-

Dock, rDock, and PLANTS) when screening for inhibitors of

protein−protein interactions.

Given the importance of shape complementation for binding,

several other fast approaches have been described to rapidly

evaluate poses and enable large-scale virtual screening. These

include methods based on Fourier correlation theory,^89,90

spherical harmonics,^91,92geometric hashing,⁹³and negative

images.⁹⁴⁻⁹⁶Most recently, others have used variations of ray-

casting to compare internal pockets in proteins to one

another.^97,98In addition to their intended usage, a subtle, but

very important, diﬀerence between the latter work and DARC

is the origin from which rays emanate. This other study casts

rays from within the pocket (at the center of mass), resulting in

a shape description that is most useful when the pocket is

mostly (or completely) enclosed by the protein.^97,98In

contrast, DARC casts rays that originate from “behind” the

pocket (i.e., inside the protein); this instead emphasizes shape

complementarity at the deepest regions of the binding groove

and is thus better suited for describing the shallow bound poses

typical of small-molecule inhibitors acting at protein interaction

sites.

While shape complementarity is clearly a necessary feature of

ligands that will bind to a protein surface, this alone obviously

cannot be suﬃcient. In the present study, we used DARC to

optimize and evaluate surface shape complementarity without

consideration of electrostatic complementarity or solvation

eﬀects; for this reason, we included a “re-ranking” step in our

virtual screen against Mcl-1 to ﬁlter out models with no

intermolecular hydrogen bonds or an abundance of unsatisﬁed

buried polar groups. In retrospect, the hydrophobicity of the

Mcl-1 surface pocket provided a convenient testing ground for

the ability of DARC to identify shape-complementary ligands

without the complication of polar groups on the protein

surface. In the future, however, we expect that incorporation of

N

DOI: 10.1021/acs.jmedchem.5b00150

J. Med. Chem. XXXX, XXX, XXX−XXX

Journal of Medicinal Chemistry

Article

conformational changes associated with inhibitor binding.¹⁵In

the case of Bcl-2 family members, the protein surface can also

adopt a number of diﬀerent conformations to accommodate

ligands with radically diﬀerent shapes and chemical proper-

ties.⁷⁰

this compound but will not inhibit other family members that

cannot adopt such conformations. By identifying “common”

pockets harbored on the surfaces of multiple family members, it

may be possible to identify inhibitors designed to act against

multiple family members (pan-inhibitors). “Pocket optimiza-

tion” simulations also reveal highly unique pockets for each

family member that are sampled by one family member but are

not accessible to any other family member.⁷⁰Thus, DARC’s

ability to identify ligands that precisely address a given family

member’s unique “signature” pockets may also provide a means

to identify highly selective compounds at a very early stage of

development.

Variation in the protein structure is most commonly included

in virtual screening by ﬁrst generating an ensemble of relevant

conformations from crystal structures³⁸⁻⁴²or simulation⁴³⁻⁴⁵

and then separately screening against each of these

conformations. Nonetheless, identifying the optimal conforma-

tions to include in these ensembles is still a challenging

task.^111−113By preferentially exploring conformations that

contain a surface pocket suitable for binding some

(unspeciﬁed) ligand,¹⁶the “pocket optimization” approach

provides a natural complement to DARC screening. In

principle, each of these low-energy pocket-containing con-

formations represents a starting point for screening, and

capturing the potential diversity of pockets on the protein

surface will be essential for fully realizing the available diversity

of potential inhibitors.

METHODS

■

Detailed descriptions of computational and experimental methods are

provided in the Supporting Information, along with sample command-

lines and instructions for using DARC.

ASSOCIATED CONTENT

* Supporting Information

■

S

Detailed descriptions of computational and experimental

methods; additional supporting experimental results; sample

command-lines and instructions for using DARC (PDF).

Molecular formula strings (CSV). The Supporting Information

is available free of charge on the ACS Publications website at

DOI: 10.1021/acs.jmedchem.5b00150.

In our Mcl-1 screen, DARC identiﬁed inhibitors both when

screening against the peptide-bound conformation and when

screening against the pocket-optimized conformation. Due to

the conformations selected for these screens, however, our

models did not take advantage of deep pockets that have been

observed in recent inhibitor-bound crystal structures of

Mcl-1^58−60,77(Figure 5c); the particular pocket-optimized

conformation used in our screen was much more similar to

the peptide-bound conformation, making these deep pockets

unavailable to DARC. Careful analysis showed that similar

conformations to those observed in the inhibitor-bound crystal

structures were indeed represented in the ensemble produced

by pocket optimization⁷⁰but were not included here because

we only screened against a single pocket-optimized con-

formation. In the future, then, inclusion of additional pocket-

containing protein conformations may allow identiﬁcation of

inhibitors that bind more deeply and exhibit improved ligand

eﬃciency.

Advantages to Structure-Based Virtual Screening.

Enhanced and reliable tools for structure-based virtual

screening will enable development of new tool compounds,

particularly in academic settings where the costs associated with

large-scale biochemical (or phenotypic) screening can be

prohibitive. However, the ability to precisely complement a

speciﬁc surface pocket also makes structure-based virtual

screening particularly attractive in a variety of other contexts

that may prove challenging for traditional biochemical

screening. These include building “conformation-selective”

inhibitors (such as compounds that are sensitive to the

phosphorylation state of a kinase activation loop¹¹⁴), targeting

sites that enable unique binding kinetics (such as a newly

discovered pocket on ERK1/2¹¹⁵), and building allosteric

inhibitors that address both the wild-type and drug-resistant

isoforms of a target (such as BCR-ABL¹¹⁶or HIV-1

protease¹¹⁷).

AUTHOR INFORMATION

Corresponding Author

*Phone: 785-864-8298. E-mail: johnk@ku.edu.

■

Author Contributions

^○R.G. and S.A.M. contributed equally.

Notes

The authors declare no competing ﬁnancial interest.

ACKNOWLEDGMENTS

■

We thank Amy Keating for providing the plasmid encoding

Mcl-1, and we thank Anne Cooper and Philip Gao for

assistance with protein production. We are grateful to OpenEye

Scientiﬁc Software (Santa Fe, NM) for providing an academic

license for the use of OMEGA and QuacPac. We are grateful to

the developers of the DOCK, AutoDock, rDock, and PLANTS

software for providing academic licenses for the use of these

programs. This work was supported by grants from the

National Institute of General Medical Sciences

(1R01GM099959, 5P50GM069663, 8P30GM103495, and

8P20GM103420) and the National Center for Research

Resources (5P30RR030926 and 5P20RR017708). This work

was also supported by the National Science Foundation

through XSEDE allocation MCB130049, the University of

Kansas Undergraduate Research Awards (S.M.), and the Alfred

P. Sloan Fellowship (J.K.).

REFERENCES

■

In light of the plasticity of many protein interaction sites,¹⁵

these tools may also prove particularly useful for designing

inhibitors that exhibit speciﬁc selectivity proﬁles. We have

shown that the activity of a given compound against various

members of the Bcl-2 family can be predicted from whether or

not each protein samples a complementary surface pocket.⁷⁰In

other words, a given compound will inhibit those Bcl-2 family

members that include conformations to suitably accommodate

(1) Domling, A. Small molecular weight protein-protein interaction

antagonists: an insurmountable challenge? Curr. Opin. Chem. Biol.

2008, 12 (3), 281−91.

(2) Whitty, A.; Kumaravel, G. Between a rock and a hard place? Nat.

Chem. Biol. 2006, 2 (3), 112−8.

(3) Murray, C. W.; Rees, D. C. The rise of fragment-based drug

discovery. Nat. Chem. 2009, 1 (3), 187−92.

(4) Akram, O. N.; DeGraff, D. J.; Sheehan, J. H.; Tilley, W. D.;

Matusik, R. J.; Ahn, J. M.; Raj, G. V. Tailoring peptidomimetics for

O

DOI: 10.1021/acs.jmedchem.5b00150

J. Med. Chem. XXXX, XXX, XXX−XXX

Journal of Medicinal Chemistry

Article

targeting protein-protein interactions. Mol. Cancer Res. 2014, 12 (7),

967−78.

molecules disrupting protein-protein interfaces: the TIMBAL data-

base. Chem. Biol. Drug Des. 2009, 74 (5), 457−67.

(5) Cummings, C. G.; Hamilton, A. D. Disrupting protein-protein

interactions with non-peptidic, small molecule alpha-helix mimetics.

Curr. Opin. Chem. Biol. 2010, 14 (3), 341−6.

(25) Higueruelo, A. P.; Jubb, H.; Blundell, T. L. TIMBAL v2: update

of a database holding small molecules modulating protein-protein

interactions. Database 2013, 2013, bat039.

(6) Whitby, L. R.; Boger, D. L. Comprehensive peptidomimetic

libraries targeting protein-protein interactions. Acc. Chem. Res. 2012,

45 (10), 1698−709.

(7) Cierpicki, T.; Grembecka, J. Targeting protein-protein

interactions in hematologic malignancies: still a challenge or a great

opportunity for future therapies? Immunol Rev. 2015, 263 (1), 279−

301.

(26) Willett, P. Similarity-based approaches to virtual screening.

Biochem. Soc. Trans. 2003, 31, 603−606.

(27) Mysinger, M. M.; Carchia, M.; Irwin, J. J.; Shoichet, B. K.

Directory of useful decoys, enhanced (DUD-E): better ligands and

decoys for better benchmarking. J. Med. Chem. 2012, 55 (14), 6582−

94.

(28) Nicholls, A. What do we know and when do we know it? J.

(8) Arkin, M. R.; Tang, Y.; Wells, J. A. Small-Molecule Inhibitors of

Protein-Protein Interactions: Progressing toward the Reality. Chem.

Biol. 2014, 21 (9), 1102−14.

Comput.-Aided Mol. Des. 2008, 22 (3−4), 239−255.

(29) Lang, P. T.; Brozell, S. R.; Mukherjee, S.; Pettersen, E. F.; Meng,

E. C.; Thomas, V.; Rizzo, R. C.; Case, D. A.; James, T. L.; Kuntz, I. D.

DOCK 6: combining techniques to model RNA-small molecule

complexes. RNA 2009, 15 (6), 1219−30.

(9) Kar, S.; Roy, K. How far can virtual screening take us in drug

discovery? Expert Opin. Drug Discovery 2013, 8 (3), 245−61.

(10) Lionta, E.; Spyrou, G.; Vassilatis, D. K.; Cournia, Z. Structure-

based virtual screening for drug discovery: principles, applications and

recent advances. Curr. Top Med. Chem. 2014, 14 (16), 1923−1938.

(11) Gowthaman, R.; Deeds, E. J.; Karanicolas, J. Structural

properties of non-traditional drug targets present new challenges for

virtual screening. J. Chem. Inf. Model. 2013, 53 (8), 2073−81.

(12) Cihlar, T.; Ray, A. S. Nucleoside and nucleotide HIV reverse

transcriptase inhibitors: 25 years after zidovudine. Antiviral Res. 2010,

85 (1), 39−58.

(30) Brozell, S. R.; Mukherjee, S.; Balius, T. E.; Roe, D. R.; Case, D.

A.; Rizzo, R. C. Evaluation of DOCK 6 as a pose generation and

database enrichment tool. J. Comput.-Aided Mol. Des. 2012, 26 (6),

749−73.

(31) Morris, G. M.; Huey, R.; Lindstrom, W.; Sanner, M. F.; Belew,

R. K.; Goodsell, D. S.; Olson, A. J. AutoDock4 and AutoDockTools4:

Automated docking with selective receptor flexibility. J. Comput. Chem.

2009, 30 (16), 2785−91.

(32) Ruiz-Carmona, S.; Alvarez-Garcia, D.; Foloppe, N.; Garmendia-

Doval, A. B.; Juhos, S.; Schmidtke, P.; Barril, X.; Hubbard, R. E.;

Morley, S. D. rDock: a fast, versatile and open source program for

docking ligands to proteins and nucleic acids. PLoS Comput. Biol.

2014, 10 (4), e1003571.

(33) Korb, O.; Stutzle, T.; Exner, T. E. Empirical scoring functions

for advanced protein-ligand docking with PLANTS. J. Chem. Inf.

Model. 2009, 49 (1), 84−96.

(34) Rush, T. S., 3rd; Grant, J. A.; Mosyak, L.; Nicholls, A. A shape-

based 3-D scaffold hopping method and its application to a bacterial

protein-protein interaction. J. Med. Chem. 2005, 48 (5), 1489−95.

(35) Hawkins, P. C.; Skillman, A. G.; Nicholls, A. Comparison of

shape-matching and docking as virtual screening tools. J. Med. Chem.

2007, 50 (1), 74−82.

(36) ROCS, version 3.2.0.3; OpenEye Scientiﬁc Software, Santa Fe,

NM; http://www.eyesopen.com.

(37) McGovern, S. L.; Shoichet, B. K. Information decay in molecular

docking screens against holo, apo, and modeled conformations of

enzymes. J. Med. Chem. 2003, 46 (14), 2895−907.

(38) Claussen, H.; Buning, C.; Rarey, M.; Lengauer, T. FlexE:

efficient molecular docking considering protein structure variations. J.

Mol. Biol. 2001, 308 (2), 377−95.

(39) Cavasotto, C. N.; Abagyan, R. A. Protein flexibility in ligand

docking and virtual screening to protein kinases. J. Mol. Biol. 2004, 337

(1), 209−25.

(40) Osguthorpe, D. J.; Sherman, W.; Hagler, A. T. Exploring protein

flexibility: incorporating structural ensembles from crystal structures

and simulation into virtual screening protocols. J. Phys. Chem. B 2012,

116 (23), 6952−9.

(41) Huang, S. Y.; Zou, X. Ensemble docking of multiple protein

structures: considering protein structural variations in molecular

docking. Proteins: Struct., Funct., Genet. 2007, 66 (2), 399−421.

(42) Cosconati, S.; Marinelli, L.; Di Leva, F. S.; La Pietra, V.; De

Simone, A.; Mancini, F.; Andrisano, V.; Novellino, E.; Goodsell, D. S.;

Olson, A. J. Protein flexibility in virtual screening: the BACE-1 case

study. J. Chem. Inf. Model. 2012, 52 (10), 2697−704.

(43) Cavasotto, C. N.; Kovacs, J. A.; Abagyan, R. A. Representing

receptor flexibility in ligand docking through relevant normal modes. J.

Am. Chem. Soc. 2005, 127 (26), 9632−40.

(13) Hodgson, D. R.; Schroder, M. Chemical approaches towards

unravelling kinase-mediated signalling pathways. Chem. Soc. Rev. 2011,

40 (3), 1211−23.

(14) Jacobson, K. A. Structure-based approaches to ligands for G-

protein-coupled adenosine and P2Y receptors, from small molecules to

nanoconjugates. J. Med. Chem. 2013, 56 (10), 3749−67.

(15) Wells, J. A.; McClendon, C. L. Reaching for high-hanging fruit

in drug discovery at protein-protein interfaces. Nature 2007, 450

(7172), 1001−9.

(16) Johnson, D. K.; Karanicolas, J. Druggable protein interaction

sites are more predisposed to surface pocket formation than the rest of

the protein surface. PLoS Comput. Biol. 2013, 9 (3), e1002951.

(17) Leaver-Fay, A.; Tyka, M.; Lewis, S. M.; Lange, O. F.;

Thompson, J.; Jacak, R.; Kaufman, K.; Renfrew, P. D.; Smith, C. A.;

Sheffler, W.; Davis, I. W.; Cooper, S.; Treuille, A.; Mandell, D. J.;

Richter, F.; Ban, Y. E.; Fleishman, S. J.; Corn, J. E.; Kim, D. E.; Lyskov,

S.; Berrondo, M.; Mentzer, S.; Popovic, Z.; Havranek, J. J.; Karanicolas,

J.; Das, R.; Meiler, J.; Kortemme, T.; Gray, J. J.; Kuhlman, B.; Baker,

D.; Bradley, P. ROSETTA3: an object-oriented software suite for the

simulation and design of macromolecules. Methods Enzymol. 2011,

487, 545−574.

(18) Hendlich, M.; Rippmann, F.; Barnickel, G. LIGSITE: automatic

and efficient detection of potential small molecule-binding sites in

proteins. J. Mol. Graphics Modell. 1997, 15 (6), 359−63 89 .

(19) Khar, K. R.; Goldschmidt, L.; Karanicolas, J. Fast docking on

graphics processing units via Ray-Casting. PLoS One 2013, 8 (8),

e70661.

(20) Hawkins, P. C.; Skillman, A. G.; Warren, G. L.; Ellingson, B. A.;

Stahl, M. T. Conformer generation with OMEGA: algorithm and

validation using high quality structures from the Protein Databank and

Cambridge Structural Database. J. Chem. Inf. Model. 2010, 50 (4),

572−84.

(21) Hawkins, P. C.; Nicholls, A. Conformer generation with

OMEGA: learning from the data set and the analysis of failures. J.

Chem. Inf. Model. 2012, 52 (11), 2919−36.

(22) OMEGA, version 2.4.3; OpenEye Scientiﬁc Software: Santa Fe,

NM, 2011; http://www.eyesopen.com.

(23) Hartshorn, M. J.; Verdonk, M. L.; Chessari, G.; Brewerton, S.

C.; Mooij, W. T.; Mortenson, P. N.; Murray, C. W. Diverse, high-

quality test set for the validation of protein-ligand docking perform-

ance. J. Med. Chem. 2007, 50 (4), 726−41.

(24) Higueruelo, A. P.; Schreyer, A.; Bickerton, G. R.; Pitt, W. R.;

Groom, C. R.; Blundell, T. L. Atomic interactions and profile of small

(44) Dixit, A.; Verkhivker, G. M. Integrating ligand-based and

protein-centric virtual screening of kinase inhibitors using ensembles

of multiple protein kinase genes and conformations. J. Chem. Inf.

Model. 2012, 52 (10), 2501−15.

P

DOI: 10.1021/acs.jmedchem.5b00150

J. Med. Chem. XXXX, XXX, XXX−XXX

Journal of Medicinal Chemistry

Article

(45) Cheng, L. S.; Amaro, R. E.; Xu, D.; Li, W. W.; Arzberger, P. W.;

McCammon, J. A. Ensemble-based virtual screening reveals potential

novel antiviral compounds for avian influenza neuraminidase. J. Med.

Chem. 2008, 51 (13), 3878−94.

(46) Shoemaker, A. R.; Oleksijew, A.; Bauch, J.; Belli, B. A.; Borre, T.;

Bruncko, M.; Deckwirth, T.; Frost, D. J.; Jarvis, K.; Joseph, M. K.;

Marsh, K.; McClellan, W.; Nellans, H.; Ng, S.; Nimmer, P.; O’Connor,

J. M.; Oltersdorf, T.; Qing, W.; Shen, W.; Stavropoulos, J.; Tahir, S. K.;

Wang, B.; Warner, R.; Zhang, H.; Fesik, S. W.; Rosenberg, S. H.;

Elmore, S. W. A small-molecule inhibitor of Bcl-XL potentiates the

activity of cytotoxic drugs in vitro and in vivo. Cancer Res. 2006, 66

(17), 8731−8739.

Jin, S.; Judge, R. A.; Kovar, P. J.; Madar, D.; Nimmer, P.; Park, C.;

Petros, A. M.; Rosenberg, S. H.; Smith, M. L.; Song, X.; Sun, C.; Tao,

Z. F.; Wang, X.; Xiao, Y.; Zhang, H.; Tse, C.; Leverson, J. D.; Elmore,

S. W.; Souers, A. J. Structure-guided design of a series of MCL-1

inhibitors with high affinity and selectivity. J. Med. Chem. 2015, 58 (5),

2180−94.

(58) Fang, C.; D’Souza, B.; Thompson, C. F.; Clifton, M. C.;

Fairman, J. W.; Fulroth, B.; Leed, A.; McCarren, P.; Wang, L.; Wang,

Y.; Feau, C.; Kaushik, V. K.; Palmer, M.; Wei, G.; Golub, T. R.;

Hubbard, B. K.; Serrano-Wu, M. H. Single Diastereomer of a

Macrolactam Core Binds Specifically to Myeloid Cell Leukemia 1

(MCL1). ACS Med. Chem. Lett. 2014, 5 (12), 1308−12.

(47) Wang, G.; Nikolovska-Coleska, Z.; Yang, C. Y.; Wang, R.; Tang,

G.; Guo, J.; Shangary, S.; Qiu, S.; Gao, W.; Yang, D.; Meagher, J.;

Stuckey, J.; Krajewski, K.; Jiang, S.; Roller, P. P.; Abaan, H. O.; Tomita,

Y.; Wang, S. Structure-based design of potent small-molecule

inhibitors of anti-apoptotic Bcl-2 proteins. J. Med. Chem. 2006, 49

(21), 6139−42.

(48) van Delft, M. F.; Wei, A. H.; Mason, K. D.; Vandenberg, C. J.;

Chen, L.; Czabotar, P. E.; Willis, S. N.; Scott, C. L.; Day, C. L.; Cory,

S.; Adams, J. M.; Roberts, A. W.; Huang, D. C. The BH3 mimetic

ABT-737 targets selective Bcl-2 proteins and efficiently induces

apoptosis via Bak/Bax if Mcl-1 is neutralized. Cancer Cell 2006, 10 (5),

389−99.

(49) Konopleva, M.; Contractor, R.; Tsao, T.; Samudio, I.; Ruvolo, P.

P.; Kitada, S.; Deng, X.; Zhai, D.; Shi, Y. X.; Sneed, T.; Verhaegen, M.;

Soengas, M.; Ruvolo, V. R.; McQueen, T.; Schober, W. D.; Watt, J. C.;

Jiffar, T.; Ling, X.; Marini, F. C.; Harris, D.; Dietrich, M.; Estrov, Z.;

McCubrey, J.; May, W. S.; Reed, J. C.; Andreeff, M. Mechanisms of

apoptosis sensitivity and resistance to the BH3 mimetic ABT-737 in

acute myeloid leukemia. Cancer Cell 2006, 10 (5), 375−88.

(50) Oltersdorf, T.; Elmore, S. W.; Shoemaker, A. R.; Armstrong, R.

C.; Augeri, D. J.; Belli, B. A.; Bruncko, M.; Deckwerth, T. L.; Dinges,

J.; Hajduk, P. J.; Joseph, M. K.; Kitada, S.; Korsmeyer, S. J.; Kunzer, A.

R.; Letai, A.; Li, C.; Mitten, M. J.; Nettesheim, D. G.; Ng, S.; Nimmer,

P. M.; O'Connor, J. M.; Oleksijew, A.; Petros, A. M.; Reed, J. C.; Shen,

W.; Tahir, S. K.; Thompson, C. B.; Tomaselli, K. J.; Wang, B.; Wendt,

M. D.; Zhang, H.; Fesik, S. W.; Rosenberg, S. H. An inhibitor of Bcl-2

family proteins induces regression of solid tumours. Nature 2005, 435

(7042), 677−681.

(51) Souers, A. J.; Leverson, J. D.; Boghaert, E. R.; Ackler, S. L.;

Catron, N. D.; Chen, J.; Dayton, B. D.; Ding, H.; Enschede, S. H.;

Fairbrother, W. J.; Huang, D. C.; Hymowitz, S. G.; Jin, S.; Khaw, S. L.;

Kovar, P. J.; Lam, L. T.; Lee, J.; Maecker, H. L.; Marsh, K. C.; Mason,

K. D.; Mitten, M. J.; Nimmer, P. M.; Oleksijew, A.; Park, C. H.; Park,

C. M.; Phillips, D. C.; Roberts, A. W.; Sampath, D.; Seymour, J. F.;

Smith, M. L.; Sullivan, G. M.; Tahir, S. K.; Tse, C.; Wendt, M. D.;

Xiao, Y.; Xue, J. C.; Zhang, H.; Humerickhouse, R. A.; Rosenberg, S.

H.; Elmore, S. W. ABT-199, a potent and selective BCL-2 inhibitor,

achieves antitumor activity while sparing platelets. Nat. Med. 2013, 19

(2), 202−208.

(59) Friberg, A.; Vigil, D.; Zhao, B.; Daniels, R. N.; Burke, J. P.;

Garcia-Barrantes, P. M.; Camper, D.; Chauder, B. A.; Lee, T.;

Olejniczak, E. T.; Fesik, S. W. Discovery of potent myeloid cell

leukemia 1 (Mcl-1) inhibitors using fragment-based methods and

structure-based design. J. Med. Chem. 2013, 56 (1), 15−30.

(60) Petros, A. M.; Swann, S. L.; Song, D.; Swinger, K.; Park, C.;

Zhang, H.; Wendt, M. D.; Kunzer, A. R.; Souers, A. J.; Sun, C.

Fragment-based discovery of potent inhibitors of the anti-apoptotic

MCL-1 protein. Bioorg. Med. Chem. Lett. 2014, 24 (6), 1484−8.

(61) Rega, M. F.; Wu, B.; Wei, J.; Zhang, Z.; Cellitti, J. F.; Pellecchia,

M. SAR by interligand nuclear overhauser effects (ILOEs) based

discovery of acylsulfonamide compounds active against Bcl-x(L) and

Mcl-1. J. Med. Chem. 2011, 54 (17), 6000−13.

(62) Prakesch, M.; Denisov, A. Y.; Naim, M.; Gehring, K.; Arya, P.

The discovery of small molecule chemical probes of Bcl-X(L) and Mcl-

1. Bioorg. Med. Chem. 2008, 16 (15), 7443−9.

(63) Ding, X.; Li, Y.; Lv, L.; Zhou, M.; Han, L.; Zhang, Z.; Ba, Q.; Li,

J.; Wang, H.; Liu, H.; Wang, R. De novo design, synthesis and

evaluation of benzylpiperazine derivatives as highly selective binders of

Mcl-1. ChemMedChem 2013, 8 (12), 1986−2014.

(64) Cao, X.; Yap, J. L.; Newell-Rogers, M. K.; Peddaboina, C.; Jiang,

W.; Papaconstantinou, H. T.; Jupitor, D.; Rai, A.; Jung, K. Y.; Tubin, R.

P.; Yu, W.; Vanommeslaeghe, K.; Wilder, P. T.; MacKerell, A. D., Jr.;

Fletcher, S.; Smythe, R. W. The novel BH3 alpha-helix mimetic JY-1−

106 induces apoptosis in a subset of cancer cells (lung cancer, colon

cancer and mesothelioma) by disrupting Bcl-xL and Mcl-1 protein-

protein interactions with Bak. Mol. Cancer 2013, 12 (1), 42.

(65) Kazi, A.; Sun, J.; Doi, K.; Sung, S. S.; Takahashi, Y.; Yin, H.;

Rodriguez, J. M.; Becerril, J.; Berndt, N.; Hamilton, A. D.; Wang, H.

G.; Sebti, S. M. The BH3 alpha-helical mimic BH3-M6 disrupts Bcl-

X(L), Bcl-2, and MCL-1 protein-protein interactions with Bax, Bak,

Bad, or Bim and induces apoptosis in a Bax- and Bim-dependent

manner. J. Biol. Chem. 2011, 286 (11), 9382−92.

(66) Stewart, M. L.; Fire, E.; Keating, A. E.; Walensky, L. D. The

MCL-1 BH3 helix is an exclusive MCL-1 inhibitor and apoptosis

sensitizer. Nat. Chem. Biol. 2010, 6 (8), 595−601.

(67) Muppidi, A.; Doi, K.; Edwardraja, S.; Drake, E. J.; Gulick, A. M.;

Wang, H. G.; Lin, Q. Rational design of proteolytically stable, cell-

permeable peptide-based selective Mcl-1 inhibitors. J. Am. Chem. Soc.

2012, 134 (36), 14734−7.

(68) Muppidi, A.; Doi, K.; Ramil, C. P.; Wang, H. G.; Lin, Q.

Synthesis of cell-permeable stapled BH3 peptide-based Mcl-1

inhibitors containing simple aryl and vinylaryl cross-linkers. Tetrahe-

dron 2014, 70 (42), 7740−5.

(69) Fire, E.; Gulla, S. V.; Grant, R. A.; Keating, A. E. Mcl-1-Bim

complexes accommodate surprising point mutations via minor

structural changes. Protein Sci. 2010, 19 (3), 507−519.

(70) Johnson, D. K.; Karanicolas, J. Selectivity by small-molecule

inhibitors of protein interactions can be driven by protein surface

fluctuations. PLoS Comput. Biol. 2015, 11 (2), e1004081.

(52) Chen, S.; Dai, Y.; Harada, H.; Dent, P.; Grant, S. Mcl-1 down-

regulation potentiates ABT-737 lethality by cooperatively inducing

Bak activation and Bax translocation. Cancer Res. 2007, 67 (2), 782−

91.

(53) Lemmen, C.; Lengauer, T.; Klebe, G. FLEXS: a method for fast

flexible ligand superposition. J. Med. Chem. 1998, 41 (23), 4502−20.

(54) Belmar, J.; Fesik, S. W. Small molecule Mcl-1 inhibitors for the

treatment of cancer. Pharmacol. Ther. 2015, 145, 76−84.

(55) Bajwa, N.; Liao, C.; Nikolovska-Coleska, Z. Inhibitors of the

anti-apoptotic Bcl-2 proteins: a patent review. Expert Opin. Ther. Pat.

2012, 22 (1), 37−55.

(56) Abulwerdi, F.; Liao, C.; Liu, M.; Azmi, A. S.; Aboukameel, A.;

Mady, A. S.; Gulappa, T.; Cierpicki, T.; Owens, S.; Zhang, T.; Sun, D.;

Stuckey, J. A.; Mohammad, R. M.; Nikolovska-Coleska, Z. A novel

small-molecule inhibitor of mcl-1 blocks pancreatic cancer growth in

vitro and in vivo. Mol. Cancer Ther. 2014, 13 (3), 565−75.

(57) Bruncko, M.; Wang, L.; Sheppard, G. S.; Phillips, D. C.; Tahir, S.

K.; Xue, J.; Erickson, S.; Fidanze, S.; Fry, E.; Hasvold, L.; Jenkins, G. J.;

(71) Liu, G.; Poppe, L.; Aoki, K.; Yamane, H.; Lewis, J.; Szyperski, T.

High-quality NMR structure of human anti-apoptotic protein domain

Mcl-1(171−327) for cancer drug design. PLoS One 2014, 9 (5),

e96521.

(72) Irwin, J. J.; Sterling, T.; Mysinger, M. M.; Bolstad, E. S.;

Coleman, R. G. ZINC: A Free Tool to Discover Chemistry for

Biology. J. Chem. Inf. Model. 2012, 52 (7), 1757−1768.

Q

DOI: 10.1021/acs.jmedchem.5b00150

J. Med. Chem. XXXX, XXX, XXX−XXX

Journal of Medicinal Chemistry

Article

(73) Davis, I. W.; Baker, D. RosettaLigand docking with full ligand

and receptor flexibility. J. Mol. Biol. 2009, 385 (2), 381−92.

(74) Baell, J. B.; Holloway, G. A. New substructure filters for removal

of pan assay interference compounds (PAINS) from screening libraries

and for their exclusion in bioassays. J. Med. Chem. 2010, 53 (7), 2719−

40.

(75) Keiser, M. J.; Roth, B. L.; Armbruster, B. N.; Ernsberger, P.;

Irwin, J. J.; Shoichet, B. K. Relating protein pharmacology by ligand

chemistry. Nat. Biotechnol. 2007, 25 (2), 197−206.

(76) Wolter, K. G.; Verhaegen, M.; Fernandez, Y.; Nikolovska-

Coleska, Z.; Riblett, M.; Martin de la Vega, C.; Wang, S.; Soengas, M.

S. Therapeutic window for melanoma treatment provided by selective

effects of the proteasome on Bcl-2 proteins. Cell Death Differ. 2007, 14

(9), 1605−1616.

(77) Tanaka, Y.; Aikawa, K.; Nishida, G.; Homma, M.; Sogabe, S.;

Igaki, S.; Hayano, Y.; Sameshima, T.; Miyahisa, I.; Kawamoto, T.;

Tawada, M.; Imai, Y.; Inazuka, M.; Cho, N.; Imaeda, Y.; Ishikawa, T.

Discovery of potent Mcl-1/Bcl-xL dual inhibitors by using a

hybridization strategy based on structural analysis of target proteins.

J. Med. Chem. 2013, 56 (23), 9635−45.

(78) Srinivasan, B.; Johnson, T. E.; Lad, R.; Xing, C. Structure-

activity relationship studies of chalcone leading to 3-hydroxy-4,3′,4′,5′-

tetramethoxychalcone and its analogues as potent nuclear factor

kappaB inhibitors and their anticancer activities. J. Med. Chem. 2009,

52 (22), 7228−35.

(79) Battenberg, O. A.; Yang, Y.; Verhelst, S. H.; Sieber, S. A. Target

profiling of 4-hydroxyderricin in S. aureus reveals seryl-tRNA

synthetase binding and inhibition by covalent modification. Mol.

BioSyst. 2013, 9 (3), 343−51.

(80) Shoichet, B. K. Interpreting steep dose-response curves in early

inhibitor discovery. J. Med. Chem. 2006, 49 (25), 7274−7.

(81) Ryan, A. J.; Gray, N. M.; Lowe, P. N.; Chung, C. W. Effect of

detergent on ″promiscuous″ inhibitors. J. Med. Chem. 2003, 46 (16),

3448−51.

(82) Giannetti, A. M.; Koch, B. D.; Browner, M. F. Surface plasmon

resonance based assay for the detection and characterization of

promiscuous inhibitors. J. Med. Chem. 2008, 51 (3), 574−80.

(83) Coan, K. E.; Maltby, D. A.; Burlingame, A. L.; Shoichet, B. K.

Promiscuous aggregate-based inhibitors promote enzyme unfolding. J.

Med. Chem. 2009, 52 (7), 2067−75.

(84) Amslinger, S.; Al-Rifai, N.; Winter, K.; Wormann, K.; Scholz, R.;

Baumeister, P.; Wild, M. Reactivity assessment of chalcones by a

kinetic thiol assay. Org. Biomol. Chem. 2013, 11 (4), 549−54.

(85) DeSimone, R. W.; Currie, K. S.; Mitchell, S. A.; Darrow, J. W.;

Pippin, D. A. Privileged structures: applications in drug discovery.

Comb Chem. High Throughput Screening 2004, 7 (5), 473−493.

(86) Welsch, M. E.; Snyder, S. A.; Stockwell, B. R. Privileged scaffolds

for library design and drug discovery. Curr. Opin. Chem. Biol. 2010, 14

(3), 347−61.

(87) Varadarajan, S.; Vogler, M.; Butterworth, M.; Dinsdale, D.;

Walensky, L. D.; Cohen, G. M. Evaluation and critical assessment of

putative MCL-1 inhibitors. Cell Death Differ. 2013, 20 (11), 1475−84.

(88) Eichhorn, J. M.; Alford, S. E.; Hughes, C. C.; Fenical, W.;

Chambers, T. C. Purported Mcl-1 inhibitor marinopyrrole A fails to

show selective cytotoxicity for Mcl-1-dependent cell lines. Cell Death

Dis. 2013, 4, e880.

(89) Katchalski-Katzir, E.; Shariv, I.; Eisenstein, M.; Friesem, A. A.;

Aflalo, C.; Vakser, I. A. Molecular surface recognition: determination

of geometric fit between proteins and their ligands by correlation

techniques. Proc. Natl. Acad. Sci. U. S. A. 1992, 89 (6), 2195−9.

(90) Gabb, H. A.; Jackson, R. M.; Sternberg, M. J. Modelling protein

docking using shape complementarity, electrostatics and biochemical

information. J. Mol. Biol. 1997, 272 (1), 106−20.

filter for large virtual throughput screening. J. Mol. Graphics Modell.

2002, 20 (4), 313−28.

(93) Schneidman-Duhovny, D.; Inbar, Y.; Nussinov, R.; Wolfson, H.

J. PatchDock and SymmDock: servers for rigid and symmetric

docking. Nucleic Acids Res. 2005, 33, W363−W367.

(94) Lee, H. S.; Lee, C. S.; Kim, J. S.; Kim, D. H.; Choe, H.

Improving virtual screening performance against conformational

variations of receptors by shape matching with ligand binding pocket.

J. Chem. Inf. Model. 2009, 49 (11), 2419−28.

(95) Ebalunode, J. O.; Ouyang, Z.; Liang, J.; Zheng, W. Novel

approach to structure-based pharmacophore search using computa-

tional geometry and shape matching techniques. J. Chem. Inf. Model.

2008, 48 (4), 889−901.

(96) Hoffmann, B.; Zaslavskiy, M.; Vert, J. P.; Stoven, V. A new

protein binding pocket similarity measure based on comparison of

clouds of atoms in 3D: application to ligand prediction. BMC Bioinf.

2010, 11, 99.

(97) Li, B.; Turuvekere, S.; Agrawal, M.; La, D.; Ramani, K.; Kihara,

D. Characterization of local geometry of protein surfaces with the

visibility criterion. Proteins: Struct., Funct., Genet. 2008, 71 (2), 670−83.

(98) Chikhi, R.; Sael, L.; Kihara, D. Real-time ligand binding pocket

database search using local surface descriptors. Proteins: Struct., Funct.,

Genet. 2010, 78 (9), 2007−28.

(99) Gowthaman R., Lyskov S., Karanicolas J. DARC 2.0: Improved

docking and virtual screening at protein interaction sites. Manuscript

in preparation.

(100) Fauman, E. B.; Rai, B. K.; Huang, E. S. Structure-based

druggability assessment–identifying suitable targets for small molecule

therapeutics. Curr. Opin. Chem. Biol. 2011, 15 (4), 463−8.

(101) Naylor, E.; Arredouani, A.; Vasudevan, S. R.; Lewis, A. M.;

Parkesh, R.; Mizote, A.; Rosen, D.; Thomas, J. M.; Izumi, M.; Ganesan,

A.; Galione, A.; Churchill, G. C. Identification of a chemical probe for

NAADP by virtual screening. Nat. Chem. Biol. 2009, 5 (4), 220−6.

(102) Koes, D.; Khoury, K.; Huang, Y.; Wang, W.; Bista, M.;

Popowicz, G. M.; Wolf, S.; Holak, T. A.; Domling, A.; Camacho, C. J.

Enabling large-scale design, synthesis and validation of small molecule

protein-protein antagonists. PLoS One 2012, 7 (3), e32839.

(103) Koes, D. R.; Camacho, C. J. Small-molecule inhibitor starting

points learned from protein-protein interaction inhibitor structure.

Bioinformatics 2012, 28 (6), 784−91.

(104) Metz, A.; Schanda, J.; Grez, M.; Wichmann, C.; Gohlke, H.

From determinants of RUNX1/ETO tetramerization to small-

molecule protein-protein interaction inhibitors targeting acute myeloid

leukemia. J. Chem. Inf. Model. 2013, 53 (9), 2197−202.

(105) Tsao, D. H.; Sutherland, A. G.; Jennings, L. D.; Li, Y.; Rush, T.

S., III; Alvarez, J. C.; Ding, W.; Dushin, E. G.; Dushin, R. G.; Haney, S.

A.; Kenny, C. H.; Malakian, A. K.; Nilakantan, R.; Mosyak, L.

Discovery of novel inhibitors of the ZipA/FtsZ complex by NMR

fragment screening coupled with structure-based design. Bioorg. Med.

Chem. 2006, 14 (23), 7953−7961.

(106) Petros, A. M.; Huth, J. R.; Oost, T.; Park, C. M.; Ding, H.;

Wang, X.; Zhang, H.; Nimmer, P.; Mendoza, R.; Sun, C.; Mack, J.;

Walter, K.; Dorwin, S.; Gramling, E.; Ladror, U.; Rosenberg, S. H.;

Elmore, S. W.; Fesik, S. W.; Hajduk, P. J. Discovery of a potent and

selective Bcl-2 inhibitor using SAR by NMR. Bioorg. Med. Chem. Lett.

2010, 20 (22), 6587−91.

(107) Frank, A. O.; Feldkamp, M. D.; Kennedy, J. P.; Waterson, A.

G.; Pelz, N. F.; Patrone, J. D.; Vangamudi, B.; Camper, D. V.;

Rossanese, O. W.; Chazin, W. J.; Fesik, S. W. Discovery of a potent

inhibitor of replication protein a protein-protein interactions using a

fragment-linking approach. J. Med. Chem. 2013, 56 (22), 9242−50.

(108) Villemagne, B.; Flipo, M.; Blondiaux, N.; Crauste, C.;

Malaquin, S.; Leroux, F.; Piveteau, C.; Villeret, V.; Brodin, P.;

Villoutreix, B. O.; Sperandio, O.; Soror, S. H.; Wohlkonig, A.;

Wintjens, R.; Deprez, B.; Baulard, A. R.; Willand, N. Ligand efficiency

driven design of new inhibitors of Mycobacterium tuberculosis

transcriptional repressor EthR using fragment growing, merging, and

linking approaches. J. Med. Chem. 2014, 57 (11), 4876−88.

(91) Morris, R. J.; Najmanovich, R. J.; Kahraman, A.; Thornton, J. M.

Real spherical harmonic expansion coefficients as 3D shape descriptors

for protein binding pocket and ligand comparisons. Bioinformatics

2005, 21 (10), 2347−55.

(92) Cai, W.; Shao, X.; Maigret, B. Protein-ligand recognition using

spherical harmonic molecular surfaces: towards a fast and efficient

R

DOI: 10.1021/acs.jmedchem.5b00150

J. Med. Chem. XXXX, XXX, XXX−XXX

Journal of Medicinal Chemistry

Article

(109) Murray, C. W.; Blundell, T. L. Structural biology in fragment-

based drug design. Curr. Opin. Struct. Biol. 2010, 20 (4), 497−507.

(110) Mattos, C.; Bellamacina, C. R.; Peisach, E.; Pereira, A.; Vitkup,

D.; Petsko, G. A.; Ringe, D. Multiple solvent crystal structures:

probing binding sites, plasticity and hydration. J. Mol. Biol. 2006, 357

(5), 1471−82.

(111) Sperandio, O.; Mouawad, L.; Pinto, E.; Villoutreix, B. O.;

Perahia, D.; Miteva, M. A. How to choose relevant multiple receptor

conformations for virtual screening: a test case of Cdk2 and normal

mode analysis. Eur. Biophys. J. 2010, 39 (9), 1365−72.

(112) Korb, O.; Olsson, T. S.; Bowden, S. J.; Hall, R. J.; Verdonk, M.

L.; Liebeschuetz, J. W.; Cole, J. C. Potential and limitations of

ensemble docking. J. Chem. Inf. Model. 2012, 52 (5), 1262−74.

(113) Sorensen, J.; Demir, O.; Swift, R. V.; Feher, V. A.; Amaro, R. E.

Molecular docking to flexible targets. Methods Mol. Biol. 2015, 1215,

445−469.

(114) Hari, S. B.; Perera, B. G.; Ranjitkar, P.; Seeliger, M. A.; Maly,

D. J. Conformation-selective inhibitors reveal differences in the

activation and phosphate-binding loops of the tyrosine kinases Abl and

Src. ACS Chem. Biol. 2013, 8 (12), 2734−43.

(115) Chaikuad, A.; Tacconi, E. M. C.; Zimmer, J.; Liang, Y.; Gray,

N. S.; Tarsounas, M.; Knapp, S. A unique inhibitor binding site in

ERK1/2 is associated with slow binding kinetics. Nat. Chem. Biol.

2014, 10 (10), 853−860.

(116) Gray, N. S.; Fabbro, D. Discovery of allosteric BCR-ABL

inhibitors from phenotypic screen to clinical candidate. Methods

Enzymol. 2014, 548, 173−188.

(117) Ung, P. M.; Dunbar, J. B., Jr.; Gestwicki, J. E.; Carlson, H. A.

An allosteric modulator of HIV-1 protease shows equipotent inhibition

of wild-type and drug-resistant proteases. J. Med. Chem. 2014, 57 (15),

6468−78.

S

DOI: 10.1021/acs.jmedchem.5b00150

J. Med. Chem. XXXX, XXX, XXX−XXX

Article Doi

DARC: Mapping Surface Topography by Ray-Casting for Effective Virtual Screening at Protein Interaction Sites

DOI: 10.1021/acs.jmedchem.5b00150

Source and publish data:

Authors:

Article abstract of DOI:10.1021/acs.jmedchem.5b00150

Full text of DOI:10.1021/acs.jmedchem.5b00150

Products guided by the article

R&D Labs maybe for 7401-39-0

Relevant to this article

Hot Product