Drug Discovery: Exploring the Utility of Cluster Oriented Genetic
Algorithms in Virtual Library Design
B. Sharma & I. Parmee
ACDDM Lab
(www.ad-comtech.co.uk/ACDDM_Group.htm)
University of West of England,
Bristol BS16 1QY
bhuvan.sharma, ian.parmee@uwe.ac.uk
Abstract- In silico combinatorial library design
involves the identification of molecules that have a
greater probability of exhibiting desired biological
activity when subjected to in vitro screening (assaying)
against a particular biological target. The paper
introduces the integration of cluster-oriented genetic
algorithms (COGAs) with such machine-based library
design environments.
COGAs have a proven capability to identify high-
performance regions of complex, continuous design
spaces relating to engineering design problems.
Modifications to the basic COGA approach are
described that allow a transfer of this capability from
continuous variable parameter space to the highly
discrete spaces described by reactants across reagent
libraries. Results relating firstly to the identification of
optimal molecules and secondly to the focussing of
reagent libraries in terms of high-performance
reactants are presented. Single objective optimisation
and focussing are initially considered before moving
on to multiple objective satisfaction.
1 Introduction
Drug design and discovery is a systematic, serial process
of identification and modification of chemical structure to
achieve desired results against biological targets
associated with a particular disease. Tradionally, the
process involves the development of a biochemical assay
for a biological target of interest and the subsequent
screening of large numbers of drug-like organic chemical
compounds in a high-throughput manner to identify hit
compounds. Such hit compounds usually possess weak
biological activity that requires improvement by a process
of medicinal chemistry optimisation first into a lead series
with robust properties and then into a drug candidate that
is suitable for evaluation in human clinical trials. This
process requires the optimisation of multiple parameters
including biological activity, selectivity for the biological
target of interest over related proteins, pharmacokinetics,
pharmacodynamics and pharmaceutical properties.
In modern drug discovery extensive use is being made of
in silico techniques to find hit molecules through virtual
screening and to then aid their subsequent optimisation. A
large compound collection in the form of virtual libraries
is described by the reagents required for their synthesis as
shown in figure 1. Reagents (inputs to the reaction
equation in figure 1) are the chemical reactants required to
make a set of molecules. Examples of reagents could be
M. Whittaker & A. Sedwell
Evotec OAI
151 Milton Park
Abingdon
Oxon, OX14 4SD
mark.whittaker, alistair.sedwell@evotecoai.com
carboxylic acids, amines, aldehydes etc. All molecules
that contain the key reactive functional group for a
reagent would constitute possible reactants across that
reagent. In this sense, in a variable parameter space
reagent refers to a particular variable parameter
(dimension) whereas reactants are the available values /
molecules across that dimension.
R1
OH + H2N R2
R1 H
N
R2
O O
Carboxylic Acid + Primary Amine ! Secondary Amide
Figure 1. Sample reaction scheme
The number of possible compounds in a virtual library
can be in the order of many millions and this number
combinatorially increases with number of included
reagents.
These libraries may then be subjected to in silico
screening to identify compounds (hits) that exhibit
activity against the biological target during the subsequent
assay. In silico models such as molecular docking,
pharmacophore matching, QSAR (Quantitative Structure-
Activity Relationship), and chemical similarity are
utilised as objective functions to identify possible high-
performance compounds that have a greater probability of
exhibiting desired biological activity during in vitro
assaying.
The size of the virtual libraries coupled with
computational expense relating to most of the objective
functions rule out exhaustive search (i.e. complete
enumeration of all library members from the
corresponding chemical reagents). What is therefore
required is a search process that rapidly samples and
identifies as many high performance molecules as
possible within available time limitations. The
development and integration of appropriate search
techniques during in-silico screening could significantly
enhance the drug discovery process by both improving
the hit rates during assaying and reducing drug design
cycle times.
In collaboration with Evotec OAI, Cluster-oriented
Genetic algorithms (COGAs) have been integrated as a
potential user-interactive search and exploratory tool with
Evotec OAI’s existing drug design software, EVOSeek™.
COGA [1] has a proven capability within engineering
design environments described by continous variables to
identify regions of high performance (HP) solutions. This
can be achieved with no apriori knowledge of the problem
space in terms of possible number of local optima and the
settings of niche radii, sharing factors etc.
The successful transfer of COGA technology to
combinatorial chemical space offers significant potential
in terms of the ability to identify groupings of HP
reactants and hence support the focussing of reagent
libraries or the identification of individual HP molecules.
This library focussing and optimisation capability is
therefore the motivation for the research described in the
following sections. This work represents a proof-of-
concept relating to the potential of COGA integration.
Section 2 is a brief review of library design literature.
Section 3 introduces the COGA methodology. Section 4
describes initial experimentation to determine basic
COGA performance Section 5 concentrates upon
identufying optimal molecules from a test library whereas
Section 6 presents the introduction of a tabu element to
the basic COGA to enable focussing of this library.
Scaleability issues are investigated in section 7 before
moving on to multi-objective satisfaction in section 8.
2 Background
A review of drug design approaches and strategies can be
found in Tsinopoulos and McCarthy [2]. Computer Aided
Drug Design found application in the early 1990s in terms
of the modelling of structure-activity relationships
(SARs). The introduction of evolutionary computing (EC)
techniques to modelling strategies began to emerge a
little later. Milne [3] reports of only five published papers
employing evolutionary algorithms between 1989 and
1992. However between 1993 and 1997 more than 210
EC-based papers appeared as reviewed in [4,5,6,7]. The
first published application of EC to combinatorial library
design was by Sheridan et al. [8] utilising a measure of
chemical similarity as an objective function. Papers by
Singh et al. [9] and Weber et al. [10] however utilised the
results from in vitro biological activity to provide a fitness
measure for a GA-based search thus providing a proof-of-
concept of the potential of a GA to direct reactant
selection for chemical synthesis.
Gillet et. al. [11] developed a GA based technique
(SELECT) to optimise virtual libraries against a single
diversity objective using a distance based diversity matrix.
A weighted sum approach for multiple objectives was
refined [12] via the introduction of a multi objective
genetic algorithm (MoSELECT). Wright et al. [13]
developed a different selection scheme in MoSELECT-II
for the optimization of library size and configuration.
Brown et al. [14], using a GA based approach called
GALOPED, also addressed size and configuration
whereas Pickett et al. [15] introduced Monte Carlo
approaches to achieve similar objectives.
There is little comparison between previous work and
the aims and objectives of the research described in the
paper. Our objective, initially using chemical similarity, is
to optimise and focus a library as opposed to Gillet’s
objective of generating a diverse library as described in
[11]. Whereas we introduce search, optimisation and
multi-objective satisfaction in large reagent libraries
other works [8,9,10] have attempted single objective
optimisation on significantly smaller libraries, This, plus
the absence of suitable benchmark problems, makes
comparison of our results with other works very
difficult.
3 Cluster Oriented Genetic Algorithms
(COGA)
Cluster Oriented Genetic Algorithms, initially developed
by Parmee in the early 1990s [1], provide the means to
identify high-performance (HP) regions of complex
conceptual engineering design spaces and enable the
extraction of information from such regions [16,17].
COGAs identify HP solution regions through the on-line
adaptive filtering of solutions generated by a genetic
algorithm. Further work resulted in several variations of
COGA and also identified and illustrated the manner in
which the COGA approach can be utilised to generate
highly relevant design information relating to single and
multi-objective problem domains [17,18,19].
COGA comprises two primary components: the
diverse search engine which utilises a highly exploratory
genetic algorithm to search the design space and the
adaptive filter (AF) which extracts solutions from each
COGA generation and stores them within a Final Cluster
Set (FCS). The AF scales solution fitness in terms of
distance from the mean (figure 2) and only solutions that
lie above a pre-defned threshold value, Rf, are copied to
the FCS. By reducing the severity of Rf, more HP
solutions albeit with a lower average fitness can enter the
FCS. The user can therefore vary the filter setting in
order to identify regions ranging from succinct groupings
of very high performance solutions to larger regions of
high and lower performance solutions. Design space
exploration is enhanced in the underlying search engine
via variable mutation regimes [1] or Halton injection
sequences [20]. Sufficient HP regional set-cover (in terms
of number of solutions) can be achieved to allow
significant qualitative and quantitative design information
to be extracted.
It is not possible within the space available to provide a
more detailed description of COGA. However, anyone
wishing to replicate the research in this paper can refer to
well-documented COGA development in many papers
available at http://www.ad-comtech.co.uk/Parmee-
Publications.htm.
The COGA approach was primarily developed for
search and exploration across design spaces described by
continuous variable parameters which tend to
predominate in engineering design. Typical COGA output
from continuous design spaces comprises clusters of
solutions describing high performance regions. The
spread and distribution of these high quality solutions can
offer a wealth of information relating to the characteristics
of the search space and the complex relationships between
variable and objective space as demonstrated in Parmee
et. al. [17,19, 21].
3.1 Representation
One of the challenges in the research described in the
following sections has been to modify the COGA
approach in order to ensure a similar utility to that proven
in engineering design when searching the discrete
combinatorial problem spaces that are an all pervading
aspect of drug design.
Given the very positive results from the engineering
design domain the initial asssumption was that application
of COGAs would significantly support the identification
of ‘best’ reactants. The chemist could interact with the
evolutionary process by varying the adaptive filter to
identify either succinct groupings of high-performance
molecules or larger collections of lesser performance
molecules in terms of any chosen in silico objective
function (e.g. QSAR, chemical similarity etc).
Binary representation has generally been utilised within
the original COGA algorithms. However binary
representation for reactants of each reagent revealed a
number of potential problems. For instance, directly
mapping binary strings onto the integer space comprising
the numbered address of each reactant molecule of a
reagent grouping results in illegal solutions and / or a
degree of redundancy. Problems relating to the crossover
of binary strings, the generation of non-feasible solutions
and a subsequent requirement for chromosome repair
were also inherent.
Figure 2: The adaptive filter (AF)
A straightforward integer encoding was therefore adopted
where integer values represent an index of a reactant’s
location in the proprietary database. The chromosome in
our integer representation scheme has a length (number of
genes) equal to the number of dimensions (reagents) of
the virtual library. A gene can then take an integer value
between one and the maximum number of reactant
molecules across that dimension. This number can
directly be transposed to the index of that reactant
molecule in the chemical database of molecules. The
phenotype then is the product molecule of the reaction
involving the reactant molecules represented by each gene
in the chromosome.
4 Initial Experimentation and Results
A test virtual library from the reaction scheme of Figure 1
comprising amines and acids was initially chosen to
assess the performance of COGA utilizing Tanimoto
Similarity as a simple test objective function. The two
reagents each possess 400 reactants creating a search
space of 160,000 possible solutions. To allow a true
evaluation of COGA performance and proof-of-concept
all the product molecules in this virtual library were
enumerated and their chemical similarity to a specific
drug molecule (methotrexate) calculated. This allows the
top 0.5% solutions to be plotted as shown in figure 3
which illustrates a typical distribution of high-
performance solutions against which COGA output can be
compared.
The plot in Figure 3 indicates that:
• good solutions can be distributed across the entire
range of a virtual library;
• reactants producing a high percentage of good
solutions can be easily identified.
With these results in mind it was intended to determine
the potential of the COGA approach in terms of:
• optimisation i.e. the identification of a number of
very high performance molecules;
• focussed library i.e. the identification of high
performance reactants that provide a focussed
combinatorial compound library the members of
which include a significant number of high
performance molecules.
Figure 3. Top 0.5% solutions in test library identified by
exhaustive search and enumeration
It was initially assumed that appropriate settings of the
adaptive filter threshold of COGA would result in the
achievement of each of these objectives. High filter
settings would provide smaller numbers of HP solutions
whereas low settings would identify much larger numbers
of high performance solutions with a lower average
fitness which would also support the identification of high
performance reactants i.e. reactants generally exhibiting
high-performance across all possible combinations.
Before investigating these initial assumptions both
variable mutation COGA (vmCOGA) and Halton
injection COGA (hiCOGA) were investigated with the
test virtual library from Figure 1 to determine their
comparative performance. Performance criteria relates to
each COGA’s ability to identify, within their final
clustering sets (FCSs), the greatest number of the
enumerated top 0.5% solutions from the exhaustive
search. The degree of robustness of each approach was
also considered to be a significant criteria. The overall
objective of this comparative study was to identify which
approach to concentrate further effort on.
The top 0.5% of the possible 160,000 solutions of the
test virtual library comprised 800 molecules. An adaptive
filter setting of 0.9 was initially introduced and a typical
plot of the solutions from a hiCOGA’s FCS is shown in
figure 4. Similar typical plots from vmCOGA’s FCS have (a)
also been generated. The results indicated that both
vmCOGA and hiCOGA identify significant numbers of
high performance solutions across the library. Fifty runs
each of hiCOGA and vmCOGA were executed in each of
which a population size of 100 individuals running over
200 generations was utilised. For each run the number of
FCS solutions that match those in the top 0.5% of the test
library were extracted and are shown in figure 5a and 5b.
On average hiCOGA identified ~200 solutions out of the
best 800 (top 0.5%), whereas vmCOGA identifed ~150 (b)
solutions. It is also evident from the plots that hiCOGA
also proved to be more robust with the standard deviation Figure 5: Numbers of HP solutions matching the test set for 50
of matching solutions being significantly less than that of runs of (a) hiCOGA and (b) vmCOGA
the vmCOGA. Robustness is of particular importance in
4
terms of integrating COGA techniques with the in silico
3.5
drug design processes within Evotec OAI’s EVOseek™
count ( x10^ 3)

3
Software.
2.5
2
1.5
1
0.5
0
-1.5 -1 -0.5
0 0.5 1 1.5 2
Rf setting
Figure 6a: Variation of solution count in FCS at differing Rf
settings
0.56
0.55
0.54
avgfitness

Figure 4: Plot of solutions from the FCS of a typical hiCOGA
search of the test library
Further experimentation involving variation of the number
of Halton injections per generation and variation of the
vmCOGA’s mutation regimes further indicated that, in
terms of discrete space search, hiCOGA performs better
both in terms of the identification of high performance
solutions and robustness.
5 Optimisation
For the enumerated 400 x 400 amine-acid test library the
fitness (chemical similarity against methotrexate) range is
between 0.0003 and 0.5812. The identification of small
numbers of very high performance molecules can be
achieved by appropriate setting of the AF filter.
0.53
0.52
0.51
0.5
0.49
0.48
-1.5 -1 -0.5 0 0.5 1 1.5 2
Rf setting
Figure 6b: Variation of average solution fitness in FCS at
differing Rf settings
Figure 6a shows the variation of numbers of HP solutions
in the FCS at different filter settings. Figure 6b shows the
the average solution fitness within the FCS for different
filter settings. The results represent average hiCOGA FCS
fitness of 50 runs of 200 generations each at each filter
setting with a population size of 100. A typical result from
a single 200 generation run of hiCOGA with a filter
setting of 2.0 would be an FCS containing circa 45
solutions with an average fitness of 0.56. Dropping the
 Figure 7: User interface from EVOseek™ showing top best performing molecules
filter to the previous 0.9 setting results in circa 400
solutions in the FCS with an average fitness of 0.54.
This optimisation process has now been integrated with
Evotec OAI’s EVOseek™ software with an appropriate
user-interface that allows the chemist to explore reagent
space via use of different objective functions and the
setting of contstraints (e.g. molecular weight, calculated
lipophilicity, hydrogen bond donor and acceptor counts).
The chemist can request visualisation of a number of the
top performing molecules from the optimisation process
which are then presented in a tabular form as shown in
figure 7. These have been generated from the integrated
software on a live (i.e. non-enumerated) library
comprising two reagents viz. haloalkane and amine.
Fitness was calculated using tanimoto chemical similarity
against a molecule known to be present in the library.
6 Generating Focussed Libraries
Returning to the plots of the fully enumerated test set
(figure 3) of Section 4 and the results from the initial
hiCOGA implementation (figure 4); these indicate the
existence of high-performance (HP) axes relating to
individual reactants. These significantly differ from the
typical HP regions identified in continuous design space
[16,17,21]. The identification of such axes allows the
chemist to develop a focussed library of reactants that
provide an enrichment of high performance molecules.
Experimentation showed that both hiCOGA and
vmCOGA FCS’s contain some HP axes but it is apparent
when comparing output to the enumerated test set that
many others were not identified. It appears that the
explorative nature of the underlying GAs was not
sufficient to avoid convergence upon a subset of available
HP axes.
Further experimentation showed that increasing
exploration via increased Halton injection or mutation
was not the answer as HP axis identification did not
improve and it appeared likely that an appropriate balance
between exploration and exploitation does not exist that
will overcome the problem. An alternative search and
exploration strategy that rapidly identifies a HP axis but
then moves on to discover other axes in unexplored
regions of search space was required. One way to achieve
this is to introduce some form of memory into the COGA
process. In this respect elements of tabu search seemed to
offer some utility in formulating a new approach called
tabu-COGA (taCOGA).
COGA’s FCS is a repository of high fitness solutions an
analysis of which during run time could give an indication
of the presence of HP axes. After a specified initial
number of generations solutions in the FCS can be
analysed at each subsequent generation to assess the count
of solutions involving specific reactants. A threshold
count could be used to identify when a reactant is eligible
to be declared as a HP axis. Thereafter, this reactant
would become tabu and the search could be directed to
other less-visited areas of library space. Some form of
replacement strategy is therefore required that, to some
extent, re-initialises tabu solutions. Two such strategies
have been considered:
a) Reactant Replacement (RR): Here the tabu reactant is
replaced by another reactant randomly selected from
the entire search space.
b) Fitness Reassignment (FR): In this approach any
solution containing a tabu reactant is assigned a low
fitness.
The objective of taCOGA is to identify maximum number
of true HP axes within a minimum number of generations/
fitness evaluations. With this in mind determining a
robust and meaningful threshold count presented a
problem. Too low a count and individual HP solutions
could falsely render a reactant ‘high-performance’
whereas too high a count could result in very lengthy
COGA run times to identify a sufficient number of HP
axes. A preliminary experiment utilising hiCOGA with
the same test library and filter and injection settings as
utilised in the experiments of section 4 was carried out to
establish an initial threshold count for taCOGA.

Figure 8. Change in HP axes identified with increasing tabu
threshold count.
Figure 9a. Number of HP Axes for R1 and R2 identified via RR
(Reactant Replacement) strategy. Results given for 50 runs.
Figure 9b. Number of HP Axes for R1 and R2 identified via FR
(Fitness Reassignment) strategy. Results given for 50 runs.
Average results from 50 runs utilising increasing counts
indicated that, for threshold counts below 10, the variation
in the number of identified high performance axes was
higher than the degree of variation between threshold
counts greater than 10 (see figure 8). For this reason a
threshold count of 10 was initially adopted for further
investigation of the approach. However, further validation
in this area is required.
A series of experiments then assessed the exploratory
potential of the two replacement strategies (RR and FR)
with a tabu thereshold count of 10. 50 runs for each
strategy were initiated to determine the number of HP
axes along each reagent that they could identify. Results
are given in figure 9a and 9b.The reactant replacement
strategy consistently identifies a greater number of HP
axes. It is likely that the immediate reassignment of tabu
reactants is more effective than the more gradual
replacement of tabu reactants that occurs with fitness
reassignment.
7 Increasing Dimension
Having established a basic working system via the two
dimensional enumerated test library, hiCOGA was then
integrated with a live library of three reagents (primary
amines, aromatic acids, aldehydes) with a total size of 400
x 400 x 400 (64 x 106) reactant combinations. In this case
a focussed library approach (i.e. with tabu) again using
chemical similarity against methotraxate as a criteria was
introduced with a hiCOGA filter setting of 0.9. Resulting
plots of FCS solutions projected onto two dimensional
hyperplanes are shown in figure 10. The plots show
identification of a significant number of high performing
reactant axes which can be further investigated by the
chemist.
A four reagent virtual library comprising aliphatic
ketone, isocyanide, primary amine and unsaturated
carboxylic acid creating a library of 5297 x 32 x 17510 x
23060 (6.8 x 1013 ) molecules was then introduced.
Chemical similarity against a target molecule known to
have close analogs in the library, provided an objective
for an optimising hiCOGA (i.e. no tabu). Tentative fitness
range therefore was between 0.0 and 0.9. The best
solution fitness achieved is 0.761 in just 200 generations
with a population size of 100 which is considered a
promising result considering the size and nature of the
search space
8 Multi-objective satisfaction within
combinatorial libraries
Various in silico models can be used as objective
functions to ascertain molecule performance against
differing criteria. Molecules best satisfying a number of
differing criteria have a greter probability of passing the
subsequent in vitro tests. In addition to chemical
similarity other simple in silico objectives are cLog P,
Polar surface area (PSA) and Molar Refractivity. In
addition, more sophisticated Quantitative Structure
Activity Relationships (QSARs), which correlates the
activity of a compound against a particular property (e.g.
aqueous solubility, Caco-2 permeability, blood-brain-
barrier permeability, hERG channel blockade), with its
sructural features can also be used as objective functions.
To date chemical similarity, cLog P, and QSAR models
for aqueous solubility have been utilised to investigate the
utility of the developed taCOGA approach for multi-
objective satisfaction.
The multi-objective COGA techniques (MOCOGA)
developed for searching continuous design spaces [17]
should also offer utility in discrete chemical space. The
continuous space approach involves running a COGA for
each objective with a relatively relaxed filter setting of
around 0.5. It has been shown that the resulting identified
regions give a very good indication of the degree of
conflict between the objectives under investigation
[17,19,21]. Mutually inclusive regions of high
performance relating to two or more objectives indicate a
low degree of conflict whereas mutually exclusive regions
indicate high degree of conflict i.e. significantly lower
performance solutions will have to be considered to
satisfy all objectives.
Figure 10: taCOGA FCS solutions projected on 2
dimensional hyperplanes for a 3 reagent library.
Similarly, in discrete reagent space HP reactant axes that
are common to a number of objectives can represent sets
of good compromise solutions. In order to identify such
common axes an optimisation approach is not appropriate.
Initial experimentation utilising the 400x400 enumerated
virtual library of section 4 with high Rf filter settings
illustrates identification of significantly less number of
HP axes. Also, just one common HP solution existed in
the FCSs of two selected objectives (aqueous solubility
and chemical similarity) from 50 hiCOGA runs with an
Rf setting of 0.9. However introducing the taCOGA
focussing approach results in the identification of
significant numbers high performing reactant axes
common to each objective as shown in figure 11. An
average of 25 common R1 reactants were identified.
Repeating the experiment using cLogP and chemical
similarity resulted in the identification of an average of 30
common HP R1 reactants and running taCOGA on all
three objectives identifies an average of 10 common HP
R1 reactants. In each case lesser numbers of common R2
reactants were identified.
Figure 11: Number of high performing R1 and R2 axes clusters
common to various objectives identified using tabu COGA. Results
given for 50 runs.
It is apparent from these preliminary results that taCOGA
offers a great deal of utility when integrated with multi-
objective reagent library search. This relates directly to
the focussing of libraries in terms of solutions that have a
higher probability of high performance regarding several
objectives during subsequent in vitro assaying processes.
Further analysis of the identified reagents by the chemist
and machine-based sampling of HP reactant axes can
result in further focussing of the taCOGA sets.
9 Further work
An extensive and successful study relating to the
satisfaction of molecular constraints to further promote
‘drug likeness’ has been carried out. This has resulted in
the introduction of appropriate penalties that further focus
the reagent libraries. Unfortunately there is insufficient
space to include results from the study in this paper.
User-interfaces that present library search results in a
succinct manner to the chemist and support user-
interaction with the evolutionary search and exploration
processes have been developed and are currently being
assessed within Evotec OAI.
Further research relating to user-preference and multi-
objective satisfaction is currently underway and, now that
proof-of-concept has been achieved, further development
of the basic techniques described will likely increase
performance both in terms of number and fitness of
identified solutions and in reduced drug discovery cycle-
times.
Information emerging from such optimisation and
focussing could be utilised to refine in-silico objective
functions in the similar manner to the problem
redefinition aspects of previous interactive evolutionary
design work [16,22]. Further information regarding
degree of objective conflict may also be available as has
been the case with engineering design applications [17,21]
10 Conclusions
A significant utility has been identified via the transfer of
basic COGA techniques from the engineering design
domain into drug discovery processes. This proof-of-
concept has been achieved via experimental development
of hiCOGA to enable successful integration with discrete
reagent library focussing and optimisation. A tabu-
oriented COGA (taCOGA) approach has been
successfully developed to support the identification of
larger numbers of high-performance reactants. Results are
presented from both enumerated reagent libraries and
from the integration of COGA with Evotec OAI’s library
definition and evaluation sofware. Scaleability aspects
have been successfully addressed and the manner in
which multi-objective considerations can be taken into
account has been illustrated.
All results so far indicate a significant potential for the
COGA concept providing a firm foundation for complex
reagent library search and exploration.
References
1. Parmee I.C. The Maintenance of Search Diversity for
Effective Design Space Decomposition using Cluster
Oriented Genetic Algorithms (COGAs) and MultiAgent
Strategies (GAANT), in Proc. Adaptive Computing in
Engineering Design and Control, 1996, University of
Plymouth, UK, pp 128-138.
2. Tsinopoulos C., McCarthy I.P. An evolutionary
classification of the strategies for drug discovery.
Manufacturing Complexity Network Conference,
Cambridge, 2002, pp 373-385.
3. Milne G. W. Mathematics as a basis for chemistry. J.
Chem. Inf Comput Sci., 1997, 37(4), pp 639-44.
4. Parrill, A. Evolutionary and genetic methods in drug
design, Drug Discovery Today 1996, 1, 514-521,
5. Brown, R.D. and Clark, D.E. Genetic Diversity:
Application of evolutionary algorithms to combinatorial
library design. Expert Opinion on Therapeutic Patents
1998, 8, pp 1447-1460.
6. Agrafiotis D, Lobanov V.S, Salemme F; Combinatorial
Informatics in the post Genomic Era. Nat Rev Drug
Discov. 2002 May, 1(5), pp 337-46.
7. Lobanov V. S, Agrafiotis D K. Scalable Methods for
the Construction and Analysis of Virtual Combinatorial
Libraries, Combin. Chem. and High-Throughput
Screen., 2002, 5, pp 167-178
8. Sheridan R.P., Kearsley S.K. Using a genetic algorithm
to suggest combinatorial libraries. J. Chem Informatics
Comp. Sci., 1995, 35, pp 310-320
9. Singh J, Ator M.A., Jaeger E.P. et al. Application of
genetic algorithms to combinatorial synthesis: A
computational approach to lead identification, J. Am
Chem. Soc., 1996, 118, pp 1669-1676
10. Weber L, Wallabaum S, Broger C, Gubernator K:
Optimization of Biological Activity of combinatorial
Libraries by a genetic algorithm. Angew. Chem. Int. Ed.
Engl. 1995 34, pp 2280-2282
11. V. J. Gillet, P. Willett, J. Bradshaw, and D. V. S.
Green: Selecting combinatorial libraries to optimize
diversity and physical properties. J. Chem. Inf. Comput.
Sci. 1999, 39, pp 169–177
12. Gillet V, Khatib, Willet. P. Combinatorial Library
design using a multi-objective genetic algorithm. J.
Chem. Inf. Comp. Sci., 2002, 42, pp 375-385
13. Wright T, Gillet V, Green D, et al. Optimising the size
and configuration of Combinatorial Libraries. J. Chem.
Inf. Comp. Sci. 2003, 43, pp 381-390
14. Brown, R. D., Martin, Y. C. Designing combinatorial
library mixtures using a genetic algorithm. Journal of
Medicinal Chemistry, 1997, 40, pp 2304–2313.
15. Pickett S. D, McLay I. M, Clark D.E. Enhancing the
hit to lead properties of lead optimization libraries. J.
Chem Inf. Comput. Sci. 2000, 40, pp 263-272
16. Parmee I.C., Cvetkovic D, Watson A, Bonham C.
Multi-objective satisfaction within an interactivee
evolutionary design environment. Evolutionary
Computation, 2000, 8, MIT Press, pp 197-222.
17. Parmee. I.C., Bonham C. R. Towards the support of
innovative conceptual design through interactive
designer/evolutionary computing strategies. Artificial
Intelligence for Engineering Design Analysis and
Manufacturing, 2000, 14, pp 3-16.
18. Parmee I.C., Abraham J.A. User Centric Evolutionary
Design. Procs Design 2004, Dubrovnik, 2004 May..
19. Abraham J.A., Parmee I.C. User Centric Evolutionary
Design Systems – the visualisation of emerging multi-
objective design information,. Xth International
Conference on Computing in Civil and Building
Engineering, Weimar. 2004, June 02-04.
20. Bonham C. R., Parmee I. C. Improving the
peformance of cluster oriented genetic algorithms
(COGAs). IEEE Congress on Evolutionary
Computation, Washington D. C, 1996, pp 554-561
21. Parmee I.C., Abraham J.A. Supporting Implicit
Learning via the Visualisation of COGA Multiobjective
Data. IEEE Congress on Evolutionary Computation,
Portland, USA, 2004, pp 395-402.
22. Parmee I. C. Improving problem definition through
interactive evolutionary computation. Artificial
Intelligence in Engineering Design, Analysis and
Manufacture, 2002, 16(3).