LEAD DISCOVERY IN THE WEB

Iaakov Exman

and Michal Pinto

Software Engineering Dept.,

Pharmaceutical Engineering Dept.

Jerusalem College of Engineering, POB 3566, Jerusalem, 91035, Israel

Keywords: Lead, discovery, Lead-&-Search protocol, Knowledge representation, Linearized components.

Abstract: The Web is a huge and very promising source of medical drug leads. But, conventional search with generic

search engines, does not really obtain novel discoveries. Inspired by the drug discovery approach, we add

the idea of a "lead" to the search process. The resulting Lead-&-Search protocol avoids the trap of repeated

fruitless search and is domain independent. The approach is applied in practice to drug leads in the Web. To

serve as input to generic search engines, multi-dimensional chemical structures are linearized into strings,

which are sliced into keyword components. Search results are reordered by novelty criteria relevant to drug

discovery. Case studies demonstrate the approach: linearized components produce specific search outcomes,

with low risk of semantic ambiguity, facilitating reordering and filtering of results.

1 INTRODUCTION

Development of new medical drugs, from initial

concepts to commercial marketing, is an extremely

long process. Any path to shorten it is of great value.

Often, new drug development starts from leads –

molecules with desirable activity – later improved

by modifying groups of atoms (called fragments), to

increase activity and decrease undesirable side-

effects. We use the term component for fragment.

The Web is a huge potential source of leads to

new drugs. But, conventional search with a generic

engine, rarely results in unexpected discoveries.

Inspired by medical drug development, we added

the "lead" idea to the search process. One obtains a

generic Lead-&-Search protocol, applicable to any

Web discovery process, not just drugs.

The Lead-&-Search protocol alternates between

lead proposal and search phases, as needed, thereby

avoiding sterile repetitive search.

Our Web lead discovery approach is applied to

novel drug development. Much of the information

on potential drugs is multi-dimensional. This is an

obstacle to generic search engines limited to linear

keyword strings. We directly use linearized

structures sliced into components as search inputs.

2 LEAD-&-SEARCH – THE

PROTOCOL

Lead-&-Search is a completely generic protocol.

Though inspired by new drug development from

leads, it is applicable to any kind of Web discovery.

The Lead-&-Search protocol goal is to overcome

the problem of repeated fruitless search – analogous

to a local minimum in an optimization problem – by

random jumping to another lead. It aims to converge

at successful discovery results, by alternate steps of

proposing a lead and performing search.

The iterative Lead-&-Search process is

schematically depicted in Figure 2.1. Two spaces are

involved: a semantic space containing potential

leads – concepts used as inputs in a search engine

working in a search space – and the search space

containing result sets found in the Web.

The Lead-&-Search protocol starts from an

initial lead (Lead-1 in Fig. 2.1), as input to search.

The protocol should be independent of the initial

lead quality.

If search is fruitful, one may refine this Lead

with the acquired knowledge. If search does not

produce results as desired, one jumps to another lead

(Lead-2) and retries search.

The Lead-&-Search loop is repeated until it is

either successful or ends by a termination criterion.

The Lead-&-Search protocol is detailed in Fig. 2.2

in pseudo code. In order to apply the

471

Exman I. and Pinto M..

LEAD DISCOVERY IN THE WEB.

DOI: 10.5220/0003098304710474

In Proceedings of the International Conference on Knowledge Discovery and Information Retrieval (KDIR-2010), pages 471-474

ISBN: 978-989-8425-28-7

 2010 SCITEPRESS (Science and Technology Publications, Lda.)

Figure 2.1: Schematic Lead-&-Search Protocol– It

alternates between proposing leads in a semantic space

and performing search in the search space, until it

succeeds or terminates.

Lead-&-Search protocol in practice one needs: a) to

set values to criteria (search-termination and lead-

timeout);

to determine algorithms to choose and

refine a lead, perform search and random jumps.

Figure 2.2: Lead-&-Search Protocol in pseudo-code –

After an initiation, the loop alternates between search and

random jumps to other leads, if necessary.

3 LEAD & SEARCH SOFTWARE

ARCHITECTURE

The overall Lead & Search software architecture and

behavior is schematically seen in Fig. 3.1. A control

unit makes decisions, depending on the software

state and on current results. It decides whether to:

• Jump to another lead and perform search;

• Refine the drug-lead and repeat search;

• Terminate the whole process.

The Lead-&-Search software system interacts with

unmodified generic search engines. Search results

are reordered and filtered by an analyzer module.

The drug-lead repository provides input to the

search engine. Lead-&-Search success means:

1. The drug-lead is gradually refined;

2. New leads are discovered.

Figure 3.1: Lead-&-Search Software Behavior Units – The

control module decides whether to jump to another lead

and search, to refine the lead, or to end the whole process.

4 SEARCH – WITH LINEARIZED

STRUCTURES

Generic search engines are powerful, but only accept

linear input strings. An important issue is which

kinds of information can be effectively used in this

way; see (Konyk, 2008) and (Searls, 2005).

Medical drugs are based on an active

substance, represented by a molecule. There is a

variety of knowledge types about a substance:

• Text – substance name and properties;

• Images – diverse physical spectra;

• 2-Dimensional (2-D) formula;

• 3-Dimensional (3-D) model.

Text does not deserve special considerations.

Spectral information images say, NMR (Nuclear

Magnetic Resonance) see e.g. (Homans, 2004), are

too complex to serve as direct search inputs.

Molecular structures – 2-D or 3-D – are graphs

with edges connecting discrete entities (atoms or

groups of atoms). They are amenable to linearized

forms useful as search inputs.

SMILES is a "Simplified Molecular Input Line

Entry" proposed by Weininger et al. (

Weininger

1988), (Weininger, 1989). It is an unambiguous

string describing 2-D or 3-D structures. There is an

open standard (OpenSMILES, 2007).

SMILES is obtained by depth-first tree traversal

of the molecule graph. The graph is trimmed

ead-&-

earch – Protocol

//Initiation phase

Get initial-Lead; Set Current-Lead = initial-

Lead;

//Loop

While (Search-Termination = False)

Do {While (Lead-Timeout = False)

Do {Perform Search;

If Search-Termination-Criterion Achieved

Then Set Search-Termination = True;

Return;

If Lead-Timeout-Criterion Reached

Then Set Lead-Timeout = True;

Else Refine Current-Lead;}

If Lead-Timeout = True

Then {Random Jump to Another-Lead;

Set Current-Lead = Another-Lead;

Set Lead-Timeout = False

;}

}

Semantic Space

ace

Lead-1

Region-1

Region-2

Lead-2

KDIR 2010 - International Conference on Knowledge Discovery and Information Retrieval

472

(removing hydrogen atoms implicit by the graph

skeleton); cycles are broken obtaining a spanning

tree (numeric suffixes mark linked nodes of broken

cycles). Parentheses mark tree branching points.

5 VALIDATION

Preliminary validation is provided by case studies

with different types of physiological activity. Search

with sliced linearized components obtains drugs and

potential leads among the results.

5.1 Case Study 1: Vancomycin

Vancomycin is an antibiotic of molecular formula

. Fig. 5.1 displays its SMILES string

(the bold [red] component is the search input). The

corresponding 2-D structure is seen in Fig. 5.2.

Table 1 shows search results for three search

engines (bing, google and yahoo). For the small

result sets obtained (between 15 and 30 results) there

are relatively many potential substances of interest.

Figure 5.1: Vancomycin SMILES Linear String –The

fragment in bold [red] was used to search potential leads.

Table 1: Vancomycin Search Results.

# Drug/Lead

rank

notes

1 Chloro-

orienticin

bing/2 Glycopeptide Antibiotic

2 telavancin google/

Semisynthetic

Vancomycin

3 methicillin google/

Not used anymore

4 chloro-

eremomyci

yahoo/

Vancomycin family

antibiotic (synonym to 1)

Result Set sizes: bing = 15; google = 28; yahoo = 30; Search rank

format is: "engine name / numeric rank of result", say "bing/2".

5.2 Case Study 2: Midazolam

Midazolam, a hypnotic-sedative drug, has molecular

formula C

ClFN

Fig. 5.3 displays a SMILES

string, corresponding to the 2-D structure in Fig. 5.4.

Figure 5.2: Vancomycin 2-D Structure – It shows atoms

(say O=Oxygen), groups of atoms (CH

=methyl), linked

by chemical bonds (graph edges), and hexagonal rings.

Figure 5.3: Midazolam SMILES String –The fragment in

bold [red] was used to search potential leads.

Figure 5.4: Midazolam 2-D Structure – It shows

hexagonal and other rings and atoms (N=Nitrogen,

F=Fluor, Cl=Chlorine), linked by chemical bonds.

The component C2N1C3=C(C=C(C=C3)Cl)

stressed in bold in Fig. 5.3 served as search input.

Results are seen in Table 2 for the same engines.

The result sets (between 11 and 138 results) are still

relatively small. The number of potential substances

of interest is of the same order of magnitude.

Table 2: Midazolam Search Results.

# Drug/Lead

rank

notes

deracyn yahoo/1 generic name: adinazolam

4-hydroxy-

alprazolam

bing/1,

google/2

analog of alprazolam

alprazolam google/4

sedative to treat insomnia

Result Set sizes: bing = 11; google = 138; yahoo = 41;

5.3 Case Study 3: Nelarabine

Nelarabine is a chemoterapy drug used to treat

leukemia. Its molecular formula is C

Fig.

CC1=NC=C2N1C3=C(C=C(C=C3)Cl)C(=NC2)C4=CC=

CC=C4F

CC1C(C(CC(O1)OC2C(C(C(OC2OC3=C4C=C5C=C3OC6

=C(C=C(C=C6)C(C(C(=O)NC(C(=O)NC5C(=O)NC7C8=

CC(=C(C=C8)O)C9=C(C=C(C=C9C(NC(=O)C(C(C1=CC(

=C(O4)C=C1)Cl)O)NC7=O)C(=O)O)O)O)CC(=O)N)NC(=

O)C(CC(C)C)NC)O)Cl)CO)O)O)(C)N)O.Cl

LEAD DISCOVERY IN THE WEB

473

5.5 displays a SMILES string for this substance. Fig.

5.6 shows its 2-D structure.

Figure 5.5: Midazolam SMILES String – This is a linear

string representing Nelarabine. The fragment in bold [red]

was used to search potential leads.

Figure 5.6: Nelarabine 2-D Structure – This formula

displays hexagonal and pentagonal rings and atoms and

groups (say NH

=Amino), linked by chemical bonds.

The component COC1=NC(N)=NC2=C1N=C

highlighted in bold [red] in Fig. 5.5 served as search

input. Result sets (249 up to 429 results) seen in

Table 3 are manageable. The number of substances

of interest is of the same order of magnitude.

Table 3: Nelarabine Search Results.

# Drug/Lead

rank

notes

6-O-Methyl

Guanosine

yahoo/1 guanine derivative

used in drug design

alfuzosin google/2 treats benign prostatic

hyperplasia

6-methoxy-9-

methyl-9H-purine

yahoo/7 substance for drug

design

Result Set sizes: bing = 429; google = 276; yahoo = 249;

6 DISCUSSION

Some preliminary conclusions from case studies are:

a) one can control the jump size between

consecutive leads in the Lead-&-Search

protocol, by controlling the leads' overlap;

b) randomly

sliced SMILES strings give small

result sets, being improbable combinations of

letters; the risk of semantic ambiguity is low;

c) one expects an approximately inverse

proportional relation between components'

string size and result set size;

d) direct search of drug names, say Vancomicyn,

is too weak to be of value for discovery.

The Lead notion has been used within a closed

computational framework (Wise, 1983) and (Exman,

1988), but has not been yet applied to the Web.

Another linear molecular naming system is

InChI (

McNaught 2006). SMILES is more readable

than InChI, but conveys less information. Our choice

of SMILES can be changed, if proved necessary.

In order to demonstrate the actual efficiency of

the approach for drug discovery, an extensive

investigation of a variety of drug families is needed.

6.1 Main Contribution

Our main contribution is the randomized "Lead"

proposal phase added to "Search", forming the

"Lead-&-Search" protocol, a powerful discovery

mechanism in the Web

REFERENCES

Exman, I. and D. H. Smith – "Get a Lead & Search: A

Strategy for Computer-Aided Drug Design'', in Symp.

Expert Systems Applications in Chemistry, ACS, 196

National Meeting, Los Angeles, p. COMP-69, (1988).

Homans, S. W., “NMR Spectroscopy Tools for Structure-

Aided Drug Design”, Angewandte Chemie Int. Ed.

Vol. 43, pp. 290–300, (2004).

Konyk, M., A. De Leon and M. Dumontier, "Chemical

Knowledge for the Semantic Web", in A. Bairoch, S.

Cohen-Boulakia, and C. Froidevaux (eds.): DILS,

LNBI 5109, pp. 169-176, Springer, Berlin (2008).

McNaught, Alan, "The IUPAC International Chemical

Identifier: InChI", Chemistry Int., Vol. 28 (6) (2006).

OpenSMILES Standard – http://www.opensmiles.org/

Draft (November 2007).

Searls, D. B., "Data integration: challenges for drug

discovery", Nature Reviews Drug Discovery 4, 45-58

(January 2005).

Weininger, D., "SMILES, a chemical language and

information system. 1. Introduction to methodology

and encoding rules", J. Chem. Inf. Comput. Sci. Vol.

28. pp. 31-36 (1988).

Weininger, D., Weininger, A., Weininger, J.L. "SMILES.

2. Algorithm for generation of unique SMILES

notation", J. Chem. Inf. Comput. Sci, 29, pp. 97-101

(1989).

Wise, M., R. D. Cramer, D. Smith and I. Exman -

“Progress in 3-D Drug Design: the use of Real Time

Colour Graphics and Computer Postulation of

Bioactive Molecules in DYLOMMS" – in J. Dearden,

(ed.) Quantitative Approaches to Drug Design, Proc.

European Symp. on "Chemical Structure-

Biological Activity: Quantitative Approaches". Bath

(U.K.), pp. 145-146., Elsevier, Amsterdam, 1983.

COC1=NC(N)=NC2=C1N=CN2C1OC(CO)C(O)C1O

KDIR 2010 - International Conference on Knowledge Discovery and Information Retrieval

474