Available on line at

Association of the Chemical Engineers AChE

[Link]/CICEQ
Chemical Industry & Chemical Engineering Quarterly 15 (4) 227−236 (2009) CI&CEQ

SVETLANA IBRIĆ ARTIFICIAL INTELLIGENCE IN PHARMA-

ZORICA DJURIĆ CEUTICAL PRODUCT FORMULATION:
JELENA PAROJČIĆ
NEURAL COMPUTING*
JELENA PETROVIĆ
Institute of Pharmaceutical Techno- The properties of a formulation are determined not only by the ratios in which
logy, Faculty of Pharmacy, the ingredients are combined but also by the processing conditions. Although
Belgrade, Serbia the relationships between the ingredient levels, processing conditions, and
product performance may be known anecdotally, they can rarely be quantified.
REVIEW PAPER In the past, formulators tended to use statistical techniques to model their for-
mulations, relying on response surfaces to provide a mechanism for optimaza-
UDC 615.4:004.032.26
tion. However, the optimization by such a method can be misleading, espe-
DOI: 10.2298/CICEQ0904227I cially if the formulation is complex. More recently, advances in mathematics
and computer science have led to the development of alternative modeling and
data mining techniques which work with a wider range of data sources: neural
networks (an attempt to mimic the processing of the human brain); genetic al-
gorithms (an attempt to mimic the evolutionary process by which biological
systems self-organize and adapt), and fuzzy logic (an attempt to mimic the abi-
lity of the human brain to draw conclusions and generate responses based on
incomplete or imprecise information). In this review the current technology will
be examined, as well as its application in pharmaceutical formulation and pro-
cessing. The challenges, benefits and future possibilities of neural computing
will be discussed.
Key words: artificial neural networks; pharmaceutical formulation; genetic
algorithms; fuzzy logic; optimization.

The development of a commercial product, sible to conceptualize. To date, statistics has been
whether it is a relatively simple formulation (e.g. a used as one approach to this problem. This method
capsule, tablet or oral liquid) or a controlled release has the advantage of generating clearly expressed
formulation (e.g. an implant), is always a time-con- models with associated confidence factors. However,
suming and complicated process. Generally an initial for more than three or four inputs, statistical ap-
formulation consisting of one or more drugs mixed proaches rapidly become unwieldy, so that the for-
with various ingredients (excipients) is prepared, and, mulator is tempted to oversimplify the problem (for ex-
as development progresses, the choice of these and ample, restricting a study to three input variables) in
their levels, as well as the manufacturing process, are order to model it. Statistics also often require the as-
changed and optimized as a result of intensive, time- sumption of a functional form (for example, linearity)
consuming experimentation. These iterations, in turn, in order to generate a model and such assumptions
result in the generation of large amounts of data, the can be inappropriate for complex tasks like formu-
processing and understanding of which is challeng- lation.
ing. In reality, the formulator has to work in a design In recent years, it has been shown that neural
space that is multi-dimensional and virtually impos- networks can provide an alternative approach [1].
Neural networks are mathematical constructs that are
capable of “learning” relationships within data, with no
Corresponding author: S. Ibrić, Institute of Pharmaceutical Tech-
nology, Faculty of Pharmacy, Vojvode Stepe 450, 11221 Belgrade, prior knowledge required from the user. The neural
Serbia. network makes no assumptions about the functional
E-mail: ibric@[Link]
form of the relationships; it simply generates and
Paper received: 1 July, 2009
Paper accepted: 15 July, 2009 assesses a range of models to determine one that will
th
*A part of this study was presented as a section lecture at the 8 best fit the experimental data provided to it. As such,
Simposium „Novel Technologies and Economics Development“,
University of Niš, Faculty of Technology, Leskovac, October 21- increasingly, artificial neural networks (often referred
–24, 2009.

227
S. IBRIĆ et al.: ARTIFICIAL INTELLIGENCE IN PHARMACEUTICAL PRODUCT… CI&CEQ 15 (4) 227−236 (2009)

to as ANNs) are used to model a complex behavior in cessing unit is known as a perceptron, a feed-forward
problems like pharmaceuticals formulation and pro- system in which the transfer of data is in the forward
cessing. direction, from inputs to outputs, only. A neural net-
The models generated by neural networks allow work consists of many neurons organized into a struc-
“what if” possibilities to be investigated easily. How- ture called the network architecture. Although there
ever, their capabilities are enhanced substantially by are many possible network architectures, one of the
combining them with other technologies. For exam- most popular and successful is the multilayer percep-
ple, using genetic algorithms for optimization, to- tron (MLP) network. This consists of identical neurons
gether with neural networks models, has proved to be all interconnected and organized in layers, with those
exceptionally powerful when the formulator must de- in one layer connected to those in the next layer so
velop a formulation to meet stringent, often conflict- that the outputs in one layer become the inputs in the
ing, objectives. The objectives for the optimization subsequent layer. Data flow into the network via the
can easily and intuitively be defined using another ar- input layer, pass through one or more hidden layers,
tificial intelligence technology, fuzzy logic. Fuzzy logic and finally exit via the output layer (Fig.1). In theory,
has proved to be especially valuable when conflicting any number of hidden layers may be added, but in
properties (for example, hard tablets that disintegrate practice multiple layers are necessary only for those
quickly) are desired. More recently, efforts have been applications with extensive nonlinear behavior, and
made to integrate the technologies even more tightly, they result in extended computation time. It is gene-
creating new methodologies like neurofuzzy logic, rally accepted that the performance of a well-design-
which combines the ability of neural networks to ed MLP model is comparable with that achieved by
“learn” from data, with fuzzy logic’s capacity to ex- classic statistical techniques.
press complex concepts in a simple fashion. These Unlike conventional computer programs, which
techniques are capable of “mining” the information are explicitly programmed, supervised neural net-
directly from data, presenting it in the form of easy to works are “trained” with previous examples. The net-
understand, actionable rules that can guide the for- work is presented with example data, and the weights
mulator’s future work. of inputs feeding into each neuron are adjusted ite-
ratively until the output for a specific network is close
NEURAL NETWORKS to the desired output. The method used to adjust the
weights is generally called back propagation, because
Like humans, neural networks learn directly from the size of the error is fed back into the calculation for
input data. The learning algorithms take two main the weight changes. There are a number of possible
forms. Unsupervised learning, where the network is back propagation algorithms, most with adjustable
presented with input data and learns to recognize pat- parameters designed to increase the rate and degree
terns in the data, is useful for organizing amounts of of convergence between the calculated and the de-
data into a smaller number of clusters. For supervised sired (actual) outputs. Although training can be a re-
learning, which is analogous to “teaching” the net- latively slow process, especially if there are large
work, the network is presented with a series of match- amounts of data, once trained, neural networks are
ing input and output examples, and it learns the re- inherently fast in execution.
lationships connecting the inputs to the outputs. Su-
pervised learning has proved most useful for the for- FUZZY LOGIC
mulation, where the goal is to determine cause-and-
effect links between inputs (ingredients and process- Conventional logic demands that a proposition is
sing conditions) and outputs (measured properties). either true or false. This maps onto a conventional set
The basic component of the neural network is theory, so that a hypothesis lies either in the “true”
the neuron, a simple mathematical processing unit set, or lies wholly outside it. That is, the membership
that takes one or more inputs and produces an out- function in the “true” set is either 1 (the hypothesis is
put. For each neuron, every input has an associated true) or 0 (the hypothesis lies outside the “true” set,
weight that defines its relative importance, and the and is false). In real life, though, these black-and-white
neuron simply computes the weighted sum of all the concepts may be of little utility. An oft-cited example
outputs and calculates an output. This is then mo- is the definition of a comfortable room temperature. If
dified by means of a transformation function (some- a temperature of 20 °C is defined as “comfortable”,
times called a transfer or activation function) before conventional logic would dictate that 19 or 21 °C,
being forwarded to another neuron. This simple pro- which lie outside this set, are “uncomfortable”). A very

228
S. IBRIĆ et al.: ARTIFICIAL INTELLIGENCE IN PHARMACEUTICAL PRODUCT… CI&CEQ 15 (4) 227−236 (2009)

Figure 1. Diagram of a multilayer perceptron with one hidden layer.

complex set of rules would be required to define Fuzzy logic can be especially useful in descry-
“comfortable” using conventional logic. Fuzzy logic al- bing target properties for optimizations. For example,
lows more intuitive statements and descriptors to be the formulator might be seeking a tablet disintegration
used, so is more akin to the linguistic format adopted time of 300 s, i.e. any value less than 300 s has a
by formulators. Fuzzy logic is based on the concept of desirability of 1 (i.e. 100%). But a tablet which disin-
fuzzy sets introduced in the 1960s by Lotfi Zadeh [2]. tegrates in 310 s is not entirely undesirable (as crisp
For fuzzy sets, membership functions are not res- logic would insist), and instead might be assigned a
tricted to be 0 or 1, but can take any continuous value desirability value of 0.9.
between these limits. In the context of comfortable Fuzzy logic is also widely used in the area of
room temperature, for example, a temperature of 17 process control, because it allows rules to be expres-
°C might have a membership of 0.4 in the “hot” set sed in a simple linguistic form IF (A) THEN (B) with an
and 0.6 in the “cold” set, as illustrated in Figure 2. associated confidence function that is related to the

Figure 2. Fuzzy logic representation of room temperature.

229
S. IBRIĆ et al.: ARTIFICIAL INTELLIGENCE IN PHARMACEUTICAL PRODUCT… CI&CEQ 15 (4) 227−236 (2009)

set membership. To understand how this is used for only two, LOW and HIGH would suffice. In some
the process control, consider a simple example of a cases, it is appropriate to add more; for example, a
fan heater governed by 4 rules, summarized in Figure problem showing a quadratic dependency would re-
3. These rules map onto the four fuzzy sets COLD, quire at least LOW, MEDIUM and HIGH in order that
COOL, WARM and HOT. So, for example, if the room it be properly represented. Where data are scarce,
temperature is 18 °C, then by Rule 2 the fan speed is relatively few membership functions should be used.
medium, with truth level 0.7, and by Rule 3 the fan As the number and complexity of the inputs in-
speed is low, with truth level 0.3. The process of de- creases, the rules become more complicated, and
fuzzification then allows the correct and smooth ad- this can make them difficult to understand.
justment of the fan speed.
EVOLUTIONARY COMPUTING
NEUROFUZZY LOGIC
Evolutionary computing is a general term that
Because fuzzy logic allows objectives to be ex- describes computational processes in which solutions
pressed in simple terms, it complements neural net- evolve, using rules of inheritance, recombination (or
work modelling. In the case of neurofuzzy logic, as cross-over), mutation and selection. One particular
the name suggests, the fuzzy logic is tightly coupled subset of this, evolutionary algorithms, has found the
with a neural network. application in the formulation research.
Neurofuzzy logic combines the ability of neural
networks to learn from data with fuzzy logic’s ability to GENETIC ALGORITHMS
express complex concepts intuitively. This creates a
degree of transparency for the otherwise “black box” Genetic algorithms were pioneered in the 1970s
neural network models, leading to the term “grey box by John Holland [4]. They provide a search technique
modelling” being applied for these methods. Neuro- which is particularly suited to optimization; a trial po-
fuzzy logic has proved to be exceptionally suited to pulation is assumed, and this evolves in an iterative
data mining, since it not only can develop good mo- process. During this process, an initial population of
dels from data, but it also has the capability of expres- solutions is generated, and the fitness of each mem-
sing these as linguistic IF…THEN rules. ber of the population is assessed. The fittest solutions
The neurofuzzy architecture is in essence a then become the “parents” of the next generation.
neural network with two additional layers for fuzzifi- Allowing some recombination and mutation introdu-
cation of inputs and defuzzification of outputs [3]. The ces a further degree of novelty into the population so
modeling capabilities of neurofuzzy systems depend that the genetic algorithm is more likely to find a glo-
on the number, shape and distribution of the fuzzy bal optimum solution. It is this ability to find the global
membership input functions. In the simplest case, optimum in a complex design space which renders

Figure 3. Fuzzy logic for process control of room temperature.

230
S. IBRIĆ et al.: ARTIFICIAL INTELLIGENCE IN PHARMACEUTICAL PRODUCT… CI&CEQ 15 (4) 227−236 (2009)

genetic algorithms so useful, especially when com- for the case where the disintegration time should lie
pared with more directed searches like conjugate gra- between 240 and 360 s, becoming progressively less
dient and steepest descent methods. desirable as it moves farther away from this region.
One requirement for genetic algorithms is that a Genetic programming, generally regarded as a
criterion of “fitness” can be defined. This can vary subset of genetic algorithms, is the most recent of the
from problem to problem. For multi-dimensional opti- techniques reviewed here, having been widely popu-
mization, it has proved useful to define an objective larized only in the 1990s primarily by Koza [5]. As yet,
function which is a weighted sum of the desirability of it has had a limited use in pharmaceutical formulation,
each of the properties. The use of weights in the sum but it shows great promise since it has the learning
allows some properties to assume more importance capabilities similar to that of neural networks but the
than others, and the fittest solutions are those that transparency associated with a straightforward ma-
best meet the overall objectives. thematical expression. In genetic programming, each
In defining the desired values of the properties, solution is a “tree”, in which each tree node has an
fuzzy logic (discussed previously) provides a useful operator function and each terminal node is an ope-
framework. Two typical desirability functions are shown rand. These trees provide an alternative way of re-
in Figure 4. One is for the case where disintegration presenting equations; Figure 5 shows one such ex-
time of a tablet is most desirable below 240 s, and ample. An initial population of solutions is assumed,
completely undesirable above 360 s. The second is and as with other evolutionary methods, the fitness of

Figure 4. Desirability functions for tablet disintegration time (a) below 240 s and (b) between 240 and 360 s.

Figure 5. Genetic programming tree representation of equation.

231
S. IBRIĆ et al.: ARTIFICIAL INTELLIGENCE IN PHARMACEUTICAL PRODUCT… CI&CEQ 15 (4) 227−236 (2009)

each member is assessed. The population then evol- INTEGRATED SOFTWARE

ves allowing crossover (whereby parts of trees are
swapped) and mutation; these operations are illus- The technologies described above are well suit-
trated in Figure 6. The evolution is biased so that the ed to data mining and modelling, but in their raw form
fittest solutions are emphasized in successive gene- require a degree of expertise. To be truly useful to
rations, leading to increased improvement in the fit of product formulators, the technologies described above
the model to the training data. need to be integrated into packages that use sensible
In the same way as for other genetic algorithms, default values for the parameters, and that incorpo-
a criterion of fitness needs to be defined. The simp- rate all of the essential tools. For example, to develop
lest criterion would simply minimize the mean-squar- a package aimed at producing optimized formula-
ed error between the calculated and actual values, tions, the modelling capability of neural networks
but this could result in an overly complex, and poten- combines well with the optimization provided by ge-
tially over-fitted, model. Therefore, it is often appro- netic algorithms. As discussed above, fuzzy logic
priate to use a model assessment criterion (such as complements this by providing a useful framework for
structural risk minimization) to penalize those solu- defining the objectives for the optimization in a clear
tions whose added complexity does not return signi- and intuitive way.
ficant new knowledge. In addition to the specific neural and evolutio-
Genetic programming currently suffers from the nary technologies, it is useful also to integrate some
disadvantage that it is time consuming, and its appli- basic statistics (both for examining the data and for
cation is less well understood in the formulation do- assessing the quality of the models, using ANOVA
main than are neural networks. Nonetheless they are Analysis of Variance, statistics), to provide a visuali-
attractive possibilities for future work, because they zation capability. Such integrated packages are now
can produce “transparent” models. available commercially and are proving useful in the
pharmaceutical industry. An early exemplar of this is
CAD/Chem [6]; more recently, since CAD/Chem is no

Figure 6. Crossover (a) and mutation (b) for genetic programming trees.

232
S. IBRIĆ et al.: ARTIFICIAL INTELLIGENCE IN PHARMACEUTICAL PRODUCT… CI&CEQ 15 (4) 227−236 (2009)

longer available, INForm from Intelligensys has been tive models from each neural network varied with res-
developed. A data mining package based on neuro- pect to the optimum network architecture and training
fuzzy logic, FormRules, is also commercially availab- algorithm. No significant differences were found in the
le, with integrated visualization as well as statistical predictive ability of these models.
techniques to assess the quality of models. Recently, the same data have been analysed by
Shao et al. [16] who compared neurofuzzy techniques
APPLICATIONS with neural networks. They found that neurofuzzy me-
thods developed models that were almost as good as
Over the past fifteen years the technology des- neural networks (as determined using analysis of va-
cribed above has been used extensively to model and riance statistics). However, the neurofuzzy results
optimize formulations from simple to very complex. It had the advantage that they also generated rules that
has also been used to a lesser degree in pharma- were presented in a simple understandable format.
ceutical processing. Rocksloh et al. [17] used neural networks suc-
Oral formulations – immediate release cessfully to optimize the crushing strength and disin-
tegration time of a high-dose plant extract tablet. In
The earliest study in this field was reported by
another study using both neural networks and genetic
Turkoglu et al. [7] who modeled a direct compression
algorithms, Do et al. [18] showed the advantages of
tablet formulation containing hydrochlorothiazide in
combining these technologies in the formulation of
order to maximize tablet strength and select the best
antacid tablets. Chen et al. [19] used neural networks
lubricant. In another study, Kesavan and Peck [8] mo-
to predict the drug content and hardness of intact
deled a tablet formulation of caffeine in order to relate
tablets of ophylline mixed with microcrystalline cellu-
both formulation (diluent type and concentration, bin-
lose from their near-infrared spectra. The model pro-
der concentration) and processing variables (type of
ved better than a statistical model generated with the
granulator, method of addition of binder) with granule
same data. The superiority of neural network models
and tablet properties (friability, hardness, and disinter-
over statistical models has also been found by Sathe
gration time). Both these investigations showed that
and Venitz [20] this time for predicting the dissolution
neural networks performed better than conventional
of 28 diltiazem immediate release tablet formulations.
statistical methods. Subsequently, the data of Kesa-
In a recent paper, Lindberg and Colbourn [21] work-
van and Peck were reanalyzed using a combination
ing in Sweden and the UK used neural networks, ge-
of neural networks and genetic algorithms [9]. This
netic algorithms and neurofuzzy to analyze historical
showed that the optimum formulation depended both
data from three different immediate release formula-
on the relative importance placed on the output pro-
tions. In all cases, the models generated performed
perties and on the constraints applied both to the le-
satisfactorily in producing tablets with specific desired
vels of the ingredients and to the processing variab-
properties.
les. Many optimum formulations could be produced,
Apart from immediate release tablet formula-
depending on the “trade-offs” that could be accepted
tions, neural networks have also been applied to mo-
for different aspects of product performance. The
delling the immediate release capsule formulations
same data have been studied using neurofuzzy com-
[22], rapidly disintegrating or dissolving tablets [23]
puting [10]. Useful rules were automatically gene-
and a novel oral microemulsion formulation of rifamy-
rated, highlighting the most important factors for each
cin and isoniazid for the treatment of children during
property and their interdependencies. As expected,
the continuation phase of tuberculosis [24]. Solid dis-
the rules for friability are the inverse of those for tablet
persion formulations of ketoprofen have recently been
hardness, while the rules for disintegration time in-
modelled using both neural networks and neurofuzzy
volved the diluent itself, the binder concentration, the
with good predictability [25]. The study has also been
method of addition of binder (wet or dry) and the me-
extended by the addition of a microemulsion formu-
thod of granulation.
lation [26].
In a series of papers, Bourquin et al. [11–14]
In a detailed evaluation of both neural networks
highlighted the advantages and disadvantages of
and guided evolutionary simulated annealing for the
neural networks in formulating the immediate release
modeling and optimization of a tablet coating formu-
tablets. The data generated by these authors have
lation, Plumb et al. [27,28] concluded that, for conflict-
been used by Plumb et al. [15] to compare three dif-
ing properties such as crack propagation and film
ferent neural network programs and four classes of
opacity that displayed highly curved responses with
training algorithms in terms of capability of generating
respect to the formulation inputs (e.g. pigment particle
predictive models. They found that the most predict-

233
S. IBRIĆ et al.: ARTIFICIAL INTELLIGENCE IN PHARMACEUTICAL PRODUCT… CI&CEQ 15 (4) 227−236 (2009)

size, pigment concentration and film thickness), clas- propylmethyl cellulose with lactose and cornstarch.
sical experimental designs to map the experimental The release profiles of theophylline were characte-
space are inappropriate for neural network modeling. rized as the sum of the fast and slow release frac-
However, if a pseudo-random design was used, it tions. To build the ANN model, the amounts of con-
was possible to model and optimize the film coating, trose, cornstarch and compression pressure were se-
predicting formulations that were either crack resis- lected as causal factors; the initial weight of theophyl-
tant or that were fully opaque. Similar formulations line, the rate constant in the fast release fraction and
have also been studied using neurofuzzy computing the release rate constant in the slow release fraction
[29], where rules were generated relating both the were chosen as response variables. The results pre-
opacity and crack resistance to the input variables. In dicted by the trained ANN model agreed well with the
line with expectations, the technique discovered for observed values. Assuming that the release rate of
itself that for maximum opacity the films needed to be theophylline in the GI tract is equal to the absorption
thick with a high pigment concentration and a small rate, the rate constant in the fast release fraction and
pigment size. the release rate constant in the slow release fraction
were used as absorption rate constants. The plasma
Controlled release oral formulations
concentrations profiles were simulated based on the
Chen et al. [30] used an artificial neural network simulated plasma concentration profiles. The optima-
(ANN) and pharmacokinetic simulations in the design zation of the controlled release theophylline tablets
of controlled-release formulations. Seven formulation was performed by a generalized distanced function
variables and three other tablet variables (moisture, method using the optimal release parameters.
particle size and hardness) for 22 tablet formulations A generalized regression neural network (GRNN)
of a model drug were used as the ANN model inputs. was used in the design of extended-release aspirin
In vitro cumulative percentage of drug released at 10 tablets by Ibric et al. [33,34]. Ten aspirin matrix tablet
different sampling time points were used as outputs. model formulations were prepared with Eudrgit RS
The ANN model was developed and trained from the PO [33]. The amount of Eudragit RS PO and com-
input and the output data sets using CAD/Chem soft- pression pressure were selected as causal factors. In
ware. The trained ANN model was used to predict vitro dissolution-time profiles at four different samp-
optimal formulation compositions based on two desi- ling time points, as well as coefficients n (release
red in vitro dissolution-time profiles and two desired in order) and log k (release constant) from the Peppas
vivo release profiles. The authors assumed that the equation were estimated as release parameters. A
dissolution is the rate-limiting step in the in vivo ab- set of release parameters and causal factors were
sorption of the drug that the fraction of the drug ab- used for training. The optimized GRNN model was
sorbed in vivo is linearly related to the in vitro disso- used to predict the formulation and process factors for
lution of the drug. Three out of four predicted formula- the optimized formulations, which would give the de-
tions showed very good agreement between the ANN sired in vitro drug release profiles. The two optimized
predicted and the observed in vitro release profiles formulations were then prepared and tested in vitro.
based on difference factor, f1, and similarity factor, f2. The comparison between the GRNN predicted and
Zupančić Božić et al. [31] developed an ANN observed in vitro profiles, and estimated coefficients
model to optimize diclofenac sodium sustained re- indicated that there is no difference between the pre-
lease matrix tablets. Formulation variables including dicted and experimentally observed drug release pro-
concentrations of cetyl alcohol, polyvinylpyrolidone K files for the two tested formulations based on the dif-
30 and magnesium stearate, and sampling time were ference factor, f1 and similarity factor, f2. The same
chosen as inputs. Twelve hidden nodes were inclu- authors applied GRNN for predicting the drug stabi-
ded in the hidden layer. The percentage of the drug lity, and in vitro–in vivo correlation [35,36].
released at each sampliing time point was used as In coated tablets, the controlling mechanism for
the output. A trained ANN model was employed to drug release is generally the film applied to the tablet,
predict release profile and optimize the formulation although in some circumstances the release may be
composition based on the percentage of the drug re- controlled in addition by the tablet core formulation. In
leased. a detailed study of 125 formulations for small tablets
Tokayama et al. [32] developed a simultaneous prepared from a model drug embedded in a hydro-
optimization technique in which the ANN model was philic matrix and then coated with an enteric polymer,
used to optimize controlled release theophylline tab- Leane et al. [37] were able to apply various input fea-
lets prepared with controse, the mixture of hydroxy- ture selection algorithms, including genetic algo-

234
S. IBRIĆ et al.: ARTIFICIAL INTELLIGENCE IN PHARMACEUTICAL PRODUCT… CI&CEQ 15 (4) 227−236 (2009)

rithms, to evaluate the relative importance of the input those networks trained using gradient descent back-
variables. They then used a neural network to model propagation algorithms outperformed the others. The
subsets of the data, with the less significant inputs textural properties of a novel pellet formulation ca-
eliminated. As expected, the elimination of the less pable of extensive gelation and swelling in biological
significant inputs results in more generalized predict- fluids, coined “gelisphere” by its inventor, have been
tive models. In another study, Wu et al. [38] used a modeled by Pillay and Danckwerts [46] using both
neural network to model the formulation of salbutanol statistics and neural networks. The authors concluded
sulfate osmotic pump tablets, using the amount of that neural networks were a more reliable data pre-
hydroxypropyl methyl cellulose and polyethylene gly- dictor in the design of their system.
col present in the cellulose acetate coating, in addi-
tion to the coating weight, as control factors. Using BENEFITS AND ISSUES
the model, the authors predicted the release parame-
ters for 1000 formulations, from which they selected Although there is a great deal of interest in neu-
an optimum with the desired release pattern. ral computing, the quantified information on the bene-
In pelleted or multi-particulate formulations, the fits has been harder to find. From the applications
drug release mechanism can be controlled either by described above, benefits that could be seen included:
using a rate controlling matrix or by the use of films. – effective use of incomplete data sets,
The pellets are either produced by using extrusion – rapid analysis of data,
and spheronization or by layering onto sugar cores. In – ability to accommodate more data and retrain
some cases, the pellets may be tabletted. In others, the network (refine the model),
they are packed into hard gelatin capsules. Peh et al. – effective exploration of the total design space,
[39,40] applied both multi-layer perceptrons and re- irrespective of complexity,
current neural networks [41] to model successfully the – ability to accommodate constraints and prefe-
release of theophylline from a matrix controlled re- rences and
lease pellet formulation prepared using extrusion and – ability to generate understandable rules.
spheronization. In another study on pellets, this time In a survey [47] on the use of 93 neural com-
prepared using the layering technique followed by puting applications in 75 UK companies covering all
polymer film coating, Vaithiyalingam et al. [42] com- business sectors, the major benefits identified were
pared the modelling and optimization abilities of sim- the improved quality, improved response times, and
plex and neural network procedures. They concluded increased productivity. Eighty-four percent of users
that from very limited data sets, simplex optimization were satisfied or very satisfied with their systems with
was more appropriate although neural networks were only three percent expressing dissatisfaction. Busi-
“a valuable and predictive tool”. In a follow-up study, ness benefits, specifically for the domain of product
Vaithiyalingam and Khan [43] compared a response formulation (albeit for nonpharmaceuticals), have
surface methodology and neural networks for model- been given as [48]:
ling and optimizing the effect of the process and for- – enhancement of product quality and perfor–
mulation variables on the release profile of verapamil mance at low cost,
hydrochloride. In each case, the observed drug re- – shorter time to market,
lease profile of the optimized formulation was close to – development of new products,
that predicted from the model. In another study, flui- – improved customer response,
dized bed manufactured, enteric-coated, omeprazole – improved confidence and
pellets compressed into tablets were analyzed using – improved competitive edge.
neural networks [44]. From the model, the authors As this new technology moves from the realm of
were able to predict a positive correlation between the academe into practical application, there are also is-
tablet strength and the concentration of the microcrys- sues regarding the implementation of neural com-
talline cellulose used as a compression aid. However, puting. Users in the previously cited study were asked
the degradation of the omeprazole in such media was to identify where they had experienced problems.
also dependent on the microcrystalline cellulose con- Thirty-nine percent had found problems related to
centration. software and lack of development skills; this will be
A bimodal drug delivery system consisting of reduced as commercial packages come into wider
pellets coated with pectin and chitosan has recently use and there is less need for bespoke in-house sys-
been modeled using neural networkswith five different tems with their high programming and maintenance
training algorithms [45]. The authors concluded that burden.

235
S. IBRIĆ et al.: ARTIFICIAL INTELLIGENCE IN PHARMACEUTICAL PRODUCT… CI&CEQ 15 (4) 227−236 (2009)

However, even when commercial packages are [19] U. Chen, S. Thosor, R. Forbess, M. Kemper, R. L. Rubi-
novitz, A. J. Shukla, Drug Dev. Ind. Pharm. 27 (2001)
used, there are a number of features that should be
623-631
present before neural computing can be used to ad-
[20] P. M. Sathe, J. Venitz, Drug Dev. Ind. Pharm. 29 (2003)
vantage. The problem must be numeric in nature, and 349-355
reasonable quantities of data should be available to [21] N. O. Lindberg, E. A., Colbourn, Pharm. Tech. Eur. 16
train an adequate model. The greatest benefits are (2004) 35-39
achieved for multidimensional problems where it is [22] R. Zolotariou, J. Anwar, J. Pharm. Pharmacol. 50 (1998) 190
difficult to express any analytic model and difficult to [23] H. Sunada, Y. Bi, Powder Technol. 122 (2002) 188-198
[24] S. Agatonovic-Kustrin, B. D. Glass, M.H. Wisch, R.G.
abstract the rules by any other mechanism than neu-
Alany. Pharm. Res. 20 (2003) 1760-1765
ral computing. It helps if the problem is of practical
[25] A. Mendyk, R. Jachowicz, Expert Sys. Appl. 28 (2005)
importance, is part of the organization’s essential 285-294
activity, and meets a real business need. Pharmaceu- [26] A. Mendyk, R. Jachowicz, Expert Sys. Appl. 32 (2007)
tical formulation meets these criteria well, and neural 1124-1131
computing can be expected to provide significant be- [27] A. P. Plumb, R. C. Rowe, P. York, C. Doherty, Eur. J.
nefits in industry in the future. Pharm. Sci. 16 (2002) 281-288
[28] A. P. Plumb, R. C. Rowe, P. York, C. Doherty, Eur. J.
Pharm. Sci. 18 (2003) 259-266
REFERENCES [29] R. C. Rowe, C. G. Woolgar, Pharm. Sci. Technol. Today
2 (1999) 495-497
[1] R. C. Rowe, E. A. Colbourn, Pharm. Tech. Eur. 9 (1996)
[30] Y. Chen, T.W. McCall, A.R. Baichwal, M.C. Meyer, J.
46-55
Control. Release 59 (1999) 33–41
[2] L. A. Zadeh, Fuzzy sets. Information and Control 8 (1965)
[31] D. Zupancic Bozic, F. Vrecar, F. Kozjek, Eur. J. Pharm.
338-353
Sci. 5 (1997) 163-169
[3] K. M. Bossley, Neurofuzzy Modelling Approaches in Sys-
[32] K. Takayama, A. Morva, M. Fujikawa, Y. Hattori, Y. Oba-
tem Identification, PhD Thesis, University of South-
ta, T. Nagai, J. Control. Release 68 (2000) 175-186
ampton, UK, 1997
[33] S. Ibric, M. Jovanovic, Z. Djuric, J. Parojcic, Lj. Solomun,
[4] J. H. Holland, Adaptation in Natural and Artificial Sys-
J. Control. Release 82 (2002) 213-222
tems, The University of Michigan Press, Ann Arbor, MI,
[34] S. Ibric, M. Jovanovic, Z. Djuric, J. Parojcic, S.D. Petro-
USA (1975)
vic, Lj. Solomun, B. Stupar, Pharm. Sci. Tech. 4 (2003)
[5] J. R. Koza, Genetic Programming – On the Programming
62-70
of Computers by Means of Natural Selection. MIT Press,
Cambridge, MA, USA, 1992 [35] S. Ibrić, M. Jovanović, Z. Đurić, J. Parojčić, Lj. Solomun,
B. Lučić, J. Pharm. Pharmacol. 59 (2007) 745-750
[6] W. H. Verduin, Better Products Faster – A Practical Guide
to Knowledge-Based Systems for Manufacturers, Irwin [36] J. Parojčić, S. Ibrić, Z. Djurić, M. Jovanović, O. I. Corri-
gan, Eur. J. Pharm. Sci. 30 (2007) 264-272
Professional Publishers, USA, 1994
[37] M. M. Leane, I. Cumming, O. Corrigan, Pharm. Sci.
[7] M. Turkoglu, R. Ozarslan, A. Sakr, Eur. J. Pharm. Bio-
Techol. 4 (2003) 218-229
pharm. 41 (1995) 315-322
[38] T. Wu, W. Pao, J. Chen, R. Shang, Drug Dev. Ind.
[8] J. G. Kesavan, G. E. Peck, Pharm. Dev. Technol. 1
Pharm. 26 (2000) 211-215
(1996) 391-404
[39] K.K. Peh, C.P. Lim, S.S. Qwek, K.H. Khoti, Pharm. Res.
[9] E. A. Colbourn, R. C. Rowe, Pharm. Tech. Eur. 8 (1996)
17 (2000) 1386-1398
46-55
[40] S.S. Qwek, C.P. Lim, K.K. Peh, Int. J. Comp. Intel. Appl.
[10] R. C. Rowe, E. A. Colbourn, Pharm Tech Eur. 12 (2000)
1 (2001) 187-292
24-27
[41] W.Y. Goh, C.P. Lim, K.K. Peh, K. Subari, Neural Comput.
[11] J. Bourquin, H. Schmidli, P. van Hoogevest, H.
Appl. 10 (2002) 311-317
Leuenberger, Pharm. Dev. Technol. 2 (1997) 111-121
[42] S. Vaithiyalingam, A.A. Zaghloul, I.K. Reddes, M.A.
[12] J. Bourquin, H. Schmidli, P. van Hoogevest, H. Leuen-
Khan, Pharm. Tech. 24 (2000) 60-72
berger, Eur. J. Pharm. Sci. 6 (1998) 287-300
[43] S. Vaithiyalingam, M.A. Khan, Int. J. Pharm. 234 (2002)
[13] J. Bourquin, H. Schmidli, P. van Hoogevest, H. Leuen-
179-193
berger, Eur. J. Pharm. Sci. 7 (1998) 5-16
[44] M. Turkoglu, H. Varol, M. Celikok, Eur. J. Pharm. Bio-
[14] J. Bourquin, H. Schmidli, P. van Hoogevest, H. Leuen-
pharm. 57 (2004) 277-286
berger, Eur. J. Pharm. Sci. 7 (1998) 17-28
[45] A. Ghaffari, H. Abdollahi, M.R. Khoshayand, I. Soltani
[15] A.P. Plumb, R.C. Rowe, P. York, M. Brown, Eur. J
Bozchalooi, A. Dadgar, M. Rafiee-Tehrani, Int. J. Pharm.
Pharm. Sci. 25 (2005) 395-405
327 (2006) 126-138
[16] Q. Shao, R. C. Rowe, P. York, Eur. J. Pharm. Sci. 28
(2006) 394-404 [46] V. Pillay, M.P. Danckwerts, J. Pharm. Sci. 91 (2002)
2559-2570
[17] K. Rocksloh, F.R: Rapp, S. Abu Abed, W. Mueller, M.
Reher, G. Gauglitz, P. C. Schmidt, Drug Dev. Ind. Pharm. [47] C. Rees, Neural Computing—Learning Solutions—User
Survey. Department of Trade and Industry, London, UK,
25 (1999) 1015-1025
1996.
[18] Q. M. Do, G. V. Dang, N. Q. Le, Tap Chi Duoc Hoc 6
[48] R. C. Rowe, R. J. Roberts, Intelligent software for product
(2000) 16-19
formulation, Taylor and Francis, London, 1998.

236