ith the acquisition of InteractiveSimulations, Inc., last April, MDL added toits arsenal of discovery informatics toolsa powerful analysis package for chemists.SCULPT brings interactive 3D visualizationand manipulation to the lab bench,enabling scientists to conduct on-the-fly3D studies. What structural features make two different molecules activate thesame receptor? What conformations arepossible for a library core structure? WithSCULPT, scientists can get the answersthey need quickly, without relying on com-plicated molecular modeling packages.

At the American Chemical Society meet-ing in New Orleans, LA, in August, MDLannounced the first significant enhance-ment release of SCULPT since acquiring theproduct. SCULPT 3.0 provides severalimportant, customer-requested updates tothe software, according to Mark Surles,MDL’s director of decision support.

“To build SCULPT 3.0, we listened closelyto what chemists told us about theirwork,” Surles said. “In response, weadded an array of new features to SCULPT

and integrated the software more tightlywith other MDL tools. The result is ananalysis package that helps scientistscommunicate, as well as make, decisions.”

An interactive, 3D structural analysistool for the scientist’s desktop, SCULPT hasalways offered unprecedented capabilities.Access to 3D conformations, combinedwith the ability to align compounds andcompare their similarities and differences,can help scientists

• Discover how compound propertiesaffect binding

• Avoid synthesizing compounds withsteric or electrostatic criteria that donot match active compounds

• Choose compounds that sample theshape and location of hydrogen bondsand functional groups

SCULPT 3.0, however, manages toimprove on “unprecedented.” In additionto adding new 3D tools, enhancementsin the latest version of SCULPT make allof the software’s visualization and align-ment tasks easier and more intuitive.And the software also includes many

new features that help scientists better communicate their results to colleagues.

For instance, SCULPT 3.0 includes add-ins that make analyzing entire hitlists ofcompounds as easy as clicking a button.Scientists performing a search usingISIS simply click a button to transfertheir search results to SCULPT. SCULPT 3.0converts the compounds to 3D and

4 Molecular Connection Fall 1999

Using SCULPT to Gain Competitive Insights

A

B

C

D

continued on page 8

SCULPT 3.0Brings 3D Visualization

to the Lab Bench

SPECIAL REPORT

eurocrine Biosciences (SanDiego, CA), like so many pharmaceuticaland biotechnology companies, has had to adapt its informatics infrastructure tomeet the demands presented by combi-natorial chemistry and high-throughputscreening. New databases and softwareprograms are now available to supportchemists at the bench. And, to assist with refining library core structures andcombinatorial library design, Neurocrinehas put more emphasis on computational

chemistry. Yet, as in most organizations,Neurocrine still faces the challenge oftightly integrating its computational activities with the workflow and goals of its project teams.

According to Scott Struthers, a seniorscientist and project leader at Neurocrine,the ties between synthetic chemistry and computational chemistry today arestronger than ever. Yet the tools availableto synthetic chemists can be woefullyinadequate. “Too often, it seems that synthetic chemists fall back on physicalmolecular modeling kits to appreciate3D,” Struthers said. “But just looking at2D structures or physical models cangive scientists a lot of misconceptionsabout what works. Organizations needsoftware that can bridge the gap betweensynthesis and modeling so that all themembers of a project team can startthinking in three dimensions from thevery beginning.”

SCULPT serves as this bridge, offeringscientists a way to conduct real-time,interactive investigations in three dimen-sions. Since Neurocrine licensed SCULPT in1998, its chemists have been able to takea whole new look at the compounds andlibraries being designed in the company’squest for new ways to treat neuropsychi-atric, neuroinflammatory and neurode-generative, and neuroendrocrine diseasesand disorders.

“Before we licensed SCULPT, there just

wasn’t a convenient way for our medicinalchemists to look much beyond a 2Dstructure,” Struthers said. “With SCULPT,our scientists can test—on the fly—theirown ideas about what conformations amolecule might have.”

According to Struthers, the researcherat the bench is willing to leave the com-plex modeling and diversity analysis tothe professional computational chemistin the project team. But even if theywanted to do some of these calculationsthemselves, two things stand in theirway: hardware and software. “Molecularmodeling packages often require achemist to sit down at a UNIX box, andmost scientists aren’t interested in foolingaround with UNIX-based software whenthey are used to working on a PC,”Struthers noted.

In addition, sophisticated molecularmodeling packages do more than synthetic chemists need. “Syntheticchemists are most interested in seeingwhat their compound or design idealooks like in 3D and comparing that withhypothesized bioactive conformations or pharmacophores,” Struthers said.“Their questions are straightforward: ‘AmI on the right track with this library corestructure?’ ‘Will this functional group go into the same orientation as the sidechain on the active compound?’ ‘How

Molecular Connection Fall 1999 5

Neurocrine’s scientists can use SCULPT

to get ideas—quickly. A competitor’spublication (A) describes the smallmolecule antagonist at right as mimickingthe peptide hormone at left. To investigatethese claims, Neurocrine scientists useSCULPT to flexibly fit the peptide hormoneto the antagonist. They begin by buildingthe small molecule in ISIS/Draw and past-ing it into SCULPT. They then use SCULPT totether the features hypothesized to corre-spond in the two molecules, freezing thepeptide in a previously known, favorableconformation (B). Using minimization, theyalign the two molecules, but find that onepair of features cannot be reconciled (C,indicated by the arrow). The scientiststhen thaw the peptide conformation,allowing both the peptide and small mole-cule to find a commonly accessible con-formation overlaying the hypothesizedcommon features (D), which can be usedto suggest additional functional groupsthat may improve the small molecule’saffinity. Total time for a first glance at acompetitor’s 3D thinking? Less than fiveminutes.

“SCULPT helps scientists see that something

that looks good on paper is impossible in 3D space.

They can use the software to look for possibilities

and to quickly rule out impossibilities.”

Scott Struthers, Senior Scientist and Project Leader, Neurocrine Biosciences

continued on page 8

SCULPTUncoversPossibilities (and Impossibilities)

ynthetic chemists have traditionallyviewed pharmacophore develop-

ment and optimization as part of thecomputational chemistry domain. But apresentation at last August’s meeting ofthe American Chemical Society in NewOrleans, LA, revealed that high-through-put chemistry is turning the explorationof pharmacophores into an essential tool in library design.

“[P]harmacophores increasingly assumethe role of filters for real and virtuallibraries of structures,” state MDLresearchers Dr. Douglas R. Henry, Dr. AliG. Özkabak, and Dr. Mitchell A. Miller, incollaboration with Dr. Osman F. Güner ofMolecular Simulations, Inc. “It is likely[pharmacophores] will increasingly beused as filters for alternative drug actions,side effects, and pharmacokinetic andmetabolic property predictions.”

Henry and his colleagues propose a simplified method for the development and optimization of 3D property-basedpharmacophores, one that can be per-formed without specialized molecularmodeling software. Their technique relies on two MDL tools—SCULPT andCheshire—to identify and characterizepotential pharmacophore candidates and uses a custom ISIS/Host program tooptimize the final pharmacophore.

This article summarizes the approach.Further details can be found in a forth-coming book edited by Güner, to be published by International University Line (http://www.iul-press.com). Formore information on the chapter, titled“Development and Optimization ofProperty-based Pharmacophores,” or the book, contact the publisher.

An Approach for Bench Chemists

According to the authors, the nature ofpharmacophore development has changedas new tools have been introduced tosupport scientists. “With the introduction

of flexible 3D searching and large struc-tural databases, pharmacophore devel-opment becomes, in modern terminology,a problem in structural data mining,”they write. Automated algorithms abound,as do an array of scientific informationmanagement tools.

To illustrate how the average benchresearcher might engage in pharma-cophore development, the authors combined several commonly availabletools with which most scientists wouldalready be familiar. Their techniqueencompasses four steps:

1. Flexible superposition of active structures onto a rigid or docked analog using SCULPT

2. Identification of relevant functional andphysicochemical property groups usingthe chemical perception and manipula-tion language Cheshire for ISIS

3. Hierarchical clustering of the identifiedgroups to visually guide the selectionof an initial pharmacophore

4. Optimization of the pharmacophoreusing a custom ISIS/Host applicationprogram

Superpositioning with SCULPT

Henry and his co-authors note thatbecause relatively few docked ligands havebeen characterized in public databases ofcrystal structures, such as the ProteinData Bank, superposition and comparisonof known active analogs are the best firststeps in pharmacophore development.Alignment algorithms can provide stronginsights into the steric and electrostaticproperties important to binding.

There are two general approaches to superpositioning. Atom- and bond-matching methods identify similaritiesamong compounds, often using maximalcommon substructure (MCSS) compar-isons. MCSS is most successful amongwell-characterized, drug-sized moleculeswith a common structural framework. The other approach uses steric and electrostatic field-driven optimizations to determine the potential alignment ofstructures. The most well-known fieldmethod is Smith and Kearsley’s SEALtechnique, which is designed to simulatethe binding of two small structures to anactive site even when the details of thereceptor site are unknown.

The authors selected SCULPT to handlesuperpositioning because of the program’scombination of flexibility, power, andusability. Because the program implementsboth MCSS and SEAL methodologiesautomatically and in combination with simplified force-field molecularrelaxation, scientists can take a real-timelook at how a flexible, mobile structurerelates to a template structure (seeFigure 1). Other interactions, such as vander Waals and Coulombic interactionsand constrained energy minimization, arealso performed rapidly by SCULPT, givingscientists real-time updates of atomcoordinates. As a result, scientists canquickly obtain a set of superpositionedactive structures for further evaluation.

6 Molecular Connection Fall 1999

Figure 1: Example overlap of several ACEinhibitors using SCULPT’s “Paste Overlap” functionality.

SPECIAL REPORT

MDL Software SimplifiesPharmacophore Development

Characterizing Receptor SiteInteractivity

The next step in pharmacophoredevelopment is identifying the atoms andfunctional groups that could potentiallyinteract with complementary atoms andgroups in the receptor site. Henry andhis co-authors note that, computationally,all pharmacophores used in 3D searchingare topological—most software programsstore and search structures not as prop-erties, but as connection tables. Chemists,however, prefer to describe functionalgroups by their properties—a hydrogenbond donor, for instance, instead of itstopological equivalent N-H, O-H, or S-H.

With its JavaScript-like command and control syntax, associative arrays,and transparent access to all the atom,bond, fragment, ring, and collection

information associated with a molecule,Cheshire for ISIS can give scientistsenormous insight into the propertiespossessed by a set of actives. In this particular paper, the authors developeda Cheshire script to count rotatablebonds and identify aromatic ring centroidsand H-bond donors in the superimposedstructures from SCULPT. Cheshireproduced a table listing the location andtype of each possible pharmacophorepoint in the ensemble. The authors notethat, with a little more scripting, Cheshireperception could have been combinedwith property calculation. In this case,the result would have been a combinedset of possible pharmacophore pointsand properties that would have providedmore in-depth information for developingthe initial query.

Initial Pharmacophore Generation

Armed with the table of pharma-cophore points from Cheshire for ISIS,Henry and his colleagues were ready tolet their initial pharmacophore take shape.The authors write that most scientistsprefer to be guided by intuition andexperience at this stage; yet, in this agewhere high throughput is almost manda-tory, automated techniques can be quiteappealing. The technique proposed inthis paper offers a way to combineautomation and intuition.

The authors first hierarchically clus-tered the pharmacophore points usingsingle-linkage or Ward’s linkage algo-rithms. The results are represented as adendrogram, which scientists can visuallytrim to include as many and whicheverpoints appear to be meaningful for thepharmacophore. The dendrogram shownin Figure 2, for instance, shows fivewell-separated clusters from the pharma-cophore points generated by the Cheshirefor ISIS scripts. Visual examination allowsscientists to winnow the list to includethose functional properties most worthyof further optimization.

Pharmacophore Optimization

Traditional optimization techniquesrequire molecular modeling and X-raycrystallography. But Henry and his co-authors propose mining a 3D structuraldatabase as a less expensive, faster wayof producing a more selective pharma-cophore that will generate higher search

Molecular Connection Fall 1999 7

100.00

84.74

69.48

54.22

Similarity

Pharmacophore Points

AR AR HBA HBD HBD HBD HBD HBA HBA HBA HBA HBA HBA HBA HBA HBD HBD HBD HBD

Figure 2: Hierarchical clustering of putative pharmacophore points for a set of superimposed ACE inhibitors.Points are for aromatic ring centroids (AR), H-bond donors (HBD), and H-bond acceptors (HBA).

Parameter Initial Value (mean) Optimized Value (mean)Optimizing function 1

D1 ( Å ) 3.5-4.5 (4.04) 1.8-4.8 (3.30) D2 ( Å 5.0-8.0 (6.94) 5.9-7.9 (6.90)D3 ( Å ) 2.0-4.5 (2.96) 2.6-5.9 (4.25)D4 ( Å ) 5.0-7.0 (6.09) 3.5-7.6 (5.55)Function value 0.57 0.63Iterations 53Yield (active, MDDR) 404 422

Optimizing function 2D1 ( Å ) 3.5-4.5 (4.04) 2.0-4.2 (3.10)D2 ( Å ) 5.0-8.0 (6.94) 5.1-9.0 (7.05)D3 ( Å ) 2.0-4.5 (2.96) 3.1-6.4 (4.75)D4 ( Å ) 5.0-7.0 (6.09) 2.7-6.1 (4.40)Function value 0.68 0.79Iterations 36Yield (active, MDDR) 404 425

[N,O,S] [N,O,S]

HH

DONORDONOR

[O,N,S]

D32-4.5 Å

D13.5-4.5 Å

D25-8 Å

ACCEPTORACCEPTOR

[P,C]

D45-7 Å

[O,N,S]

[P,C]

Figure 3: The table shows optimization results for the ACE inhibitor pharmacophore indicated. While optimization in general increased both the yield and the specificity of thepharmacophores, Function 2 converged more rapidly and performed slightly better than Function 1.

continued on next page

yields. Miller wrote an ISIS/Host applica-tion to perform repeated searches overa 3D database, each time modifying a set of constraint parameters in the pharmacophore to produce higher yield-ing and/or more selective search results.

Figure 3 (see page 7) shows the train-ing and prediction results that the authorsobtained for optimizing an ACE inhibitorpharmacophore against a set of 200known inhibitors in MDL’s MDDR data-base. The initial pharmacophore retrieved404 of the 471 possible ACE inhibitors, a

good showing, particularly when com-pared to the results obtained using a pub-lished ACE inhibitor, which retrieved just356 structures. The resulting optimizedpharmacophores showed tightening ofsome ranges and some shifts of the meanvalues. Comparing the optimization func-tions shows the second function, devel-oped by Güner and presented at the 1998Charleston Conference, converged morerapidly and yielded a pharmacophore withslightly better prediction performance.Finally, in each of the optimizations withthis data set, both the yield and the speci-ficity increased as a result of optimization.

Conclusions

The authors’ analyses, excerpted here,were intended to be illustrative of a technique. The results themselves weresecondary. Yet this simple approachdemonstrates that pharmacophore development can be done speedily withthe assistance of industry-standard, readily available tools. The authors hopethat this type of database mining will provide scientists with more flexibilityduring the early stages of drug development.❖

8 Molecular Connection Fall 1999

aligns them automatically so that theirstructures can be readily compared. Thistight integration between SCULPT andother MDL software makes it easier forscientists to explore hypotheses duringthe course of their research.

SCULPT 3.0 users will also benefit frombeing able to

• Automatically convert 2D struc-

tures to 3D models. SCULPT 3.0 auto-matically generates low-energy, 3Dconformations for structures pastedinto SCULPT from ISIS/Draw, ISIS/Base,or Chemscape Chime. The conversionscan also be applied to entire lists ofmolecules saved as SDfiles.

• Automatically align compounds

to each other or to bound ligands.

Compound superpositioning and comparison reveal how compoundsmay interact with receptor sites.SCULPT 3.0 uses two different alignmentalgorithms: a maximal common

substructure (MCSS) methodology,which identifies structural similaritiesbetween compounds, and the SEALapproach, which uses steric and electrostatic field-driven optimizationsto determine the potential alignment of structures.

• View high-quality 3D models and

surfaces. Significantly improvedgraphics in SCULPT 3.0 provide production-quality images and newvisualization features. Scientists canquickly determine the orientation of and distance between functionalgroups, the size and volume of com-pounds or important Rgroups, andthe position and potential influence ofhydrogen bond donors and acceptorson the conformation. SCULPT also generates solvent-accessible surfacesaround one or more compounds, col-ored by the charge of the contributingatoms. Other available SCULPT 3.0visualizations include protein ribbonrepresentations, side-chain character-istics, and ball-and-stick models.

• Share results with colleagues

throughout their organization. In addition to integrating better withISIS, SCULPT also connects to enter-prise reporting and communicationsystems. OLE embedding enablesSCULPT images to be stored inMicrosoft Word, email documents,and intranet Web sites as live 3Dobjects. Scientists on the receivingend of a report or email update canrotate molecules, query basic struc-tural properties, and toggle the visual-izations without having to learn oreven launch SCULPT. These featureskeep project teams connected,enabling scientists to pool their com-bined knowledge to determine thenext step to take in a research plan.

To learn more about how SCULPT 3.0can help your scientists make better decisions at the bench, visit MDL’s Website at http://www.mdli.com/sculpt,contact your MDL sales representative, or the MDL office nearest you.❖

might these diverse screening hits bindto the same receptor?’”

“SCULPT helps scientists see that something that looks good on paper isimpossible in 3D space,” Struthers continued. “They can use the software to look for possibilities and to quicklyrule out impossibilities.”

Scientists at Neurocrine have beenworking with SCULPT for over a year,using it for superpositioning and comparing screening hits and activecompounds (see “Using SCULPT to Gain

Competitive Insights,” page 4). Chemistsdesigning new molecules and librariesalso consult SCULPT to sort through theirideas to determine which make the mostsense to pursue. Even Struthers, a com-putational chemist by training, frequentlyturns to SCULPT instead of a UNIX-basedmolecular modeling tool. “SCULPT givesme answers quickly when I just need tocheck something out,” he said.

With the release of SCULPT 3.0,Struthers is particularly interested inmaking SCULPT models and researchmore widely available on Neurocrine’sintranet. The tighter integration betweenSCULPT and Chemscape Chime Pro is

enabling Neurocrine to create interactiveWeb pages containing SCULPT data.

“The Web gives us a way to shareinformation that has previously beeninaccessible—unless you are lookingover someone’s shoulder at a computerscreen,” Struthers explained. “MakingSCULPT data available throughoutNeurocrine will help project teams communicate in three dimensions. Itshould be particularly useful to theresearchers designing compounds,who should be thinking about the conformational flexibility of compoundsearlier in their investigations.”❖

Pharmacophore Developmentcontinued from previous page

SCULPT at Neurocrinecontinued from page 5

SPECIAL REPORT

SCULPT 3.0continued from page 4

Molecular Connection Fall 1999 9

he tight integration between SCULPT

and other MDL software, such asISIS/Draw, lets scientists add more powerto their searches. For instance, SCULPT canbe used to turn a set of active 2D com-pounds into a 3D conformationally flexiblesubstructure (CFS) search query, whichcan retrieve a more structurally relevantset of hits from a database search.

Figure 1 shows a group of compoundsthat displayed activity in a particularscreen. Unfortunately, the scientists do nothave a 3D model of the target site. Andwhile there seems to be some correlationamong the 2D structures, the scientistsare not sure why there is a 30 percentvariation in the activity data.

To investigate which features of thesestructures influence activity, the scientistsselect “Chemistry➔Transfer to SCULPT.”The compounds are seamlessly transferredto SCULPT, converted into 3D models,and automatically aligned to each other(see Figure 2). Several structural featuresare present in similar orientations amongthe molecules, including a six-memberedaromatic ring, a nearby H-bond acceptorand sulfur atom, and a hydrophobic methylsubstituent. The absence of and positionof the methyl substituent in Actives-3and Actives-5, respectively, may explainthese compounds’ decreased activity.

Using these observations as a guide, thescientists measure the distance and angleparameters between the functional partsof the molecules, either directly in SCULPT

or by transferring the structures back toISIS/Draw (see Figure 3). Comparing theresults and selecting “Chemistry➔Create3D” produces a 3D CFS query that aver-ages the values for each compound (seeFigure 4). This query specifies an aromaticring, a simple H-bond acceptor, a five-membered sulfur heterocycle, a methylgroup, and an exclusion sphere with distance and angle parameters reflectingthe measurements from the five original molecules. Executing the query againstMDDR-3D 99.2 (105,372 compounds)using default CFS parameters retrieves 10hits—a much more refined list than the5,863 hits retrieved from a 2D substructuresearch using the fragments upon which the 3D query was based.❖

Figure 2

Figure 1

Figure 3

Figure 4

SCULPTing Hits into 3D Queries