Journal of Pharmaceutical Sciences 106 (2017) 176-182

Contents lists available at ScienceDirect

Journal of Pharmaceutical Sciences

journal homepage: www.jpharmsci.org

Pharmaceutics, Drug Delivery and Pharmaceutical Technology

Application of Artificial Neural Networks in the Design and

Optimization of a Nanoparticulate Fingolimod Delivery System Based
on Biodegradable Poly(3-Hydroxybutyrate-Co-3-Hydroxyvalerate)
Shadab Shahsavari 1, Leila Rezaie Shirmard 2, 3, Mohsen Amini 4,
Farid Abedin Dokoosh 2, 5, *
1
Department of Chemical Engineering, Varamin-Pishva Branch, Islamic Azad University, Varamin, Iran
2
Department of Pharmaceutics, Faculty of Pharmacy, Tehran University of Medical Sciences, Tehran, Iran
3
Department of Pharmaceutics, Faculty of Pharmacy, Ardebil University of Medical Science, Ardebil, Iran
4
Department of Medicinal Chemistry, Drug Design and Development Research Center, Tehran University of Medical Sciences, Tehran, Iran
5
Medical Biomaterials Research Center (MBRC), Tehran University of Medical Sciences, Tehran, Iran

a r t i c l e i n f o a b s t r a c t

Article history: Formulation of a nanoparticulate Fingolimod delivery system based on biodegradable poly(3-
Received 4 March 2016 hydroxybutyrate-co-3-hydroxyvalerate) was optimized according to artificial neural networks (ANNs).
Revised 1 July 2016 Concentration of poly(3-hydroxybutyrate-co-3-hydroxyvalerate), PVA and amount of Fingolimod is
Accepted 26 July 2016
considered as the input value, and the particle size, polydispersity index, loading capacity, and entrap-
Available online 22 September 2016
ment efficacy as output data in experimental design study. In vitro release study was carried out for best
formulation according to statistical analysis. ANNs are employed to generate the best model to determine
Keywords:
artificial neural network the relationships between various values. In order to specify the model with the best accuracy and
drug delivery proficiency for the in vitro release, a multilayer percepteron with different training algorithm has been
Fingolimod examined. Three training model formulations including Levenberg-Marquardt (LM), gradient descent,
poly(3-hydroxybutyrate-co-3-
and Bayesian regularization were employed for training the ANN models. It is demonstrated that the
hydroxyvalerate)
response surface methodology predictive ability of each training algorithm is in the order of LM > gradient descent > Bayesian regu-
training algorithms larization. Also, optimum formulation was achieved by LM training function with 15 hidden layers and 20
neurons. The transfer function of the hidden layer for this formulation and the output layer were tansig
and purlin, respectively. Also, the optimization process was developed by minimizing the error among
the predicted and observed values of training algorithm (about 0.0341).
© 2016 American Pharmacists Association®. Published by Elsevier Inc. All rights reserved.

Introduction sclerosis patients at 0.5 mg daily. Furthermore, alternative for-

Fingolimod, 2-amino-2-[2-(4-octylphenyl) ethyl]-1, 3-propanediol]
as a novel immunosuppressive modulator, is characterized by a
reversible reduction of circulating peripheral blood lymphocyte
counts. In experimental allogenic organ transplantation studies,
Fingolimod exhibited mild immunosuppression; however, this
has been clinically effective only at concentrations 5-fold higher
than the dose administered normally and higher steady-state
levels will need to be achieved. The current available capsule
formulation supports daily administration with dose-proportional
pharmacokinetics to achieve active steady-state levels in multiple
* Correspondence to: Farid Abedin Dokoosh (Telephone: þ982188009440;
Fax: þ982188026734).
E-mail address: dorkoosh@tums.ac.ir (F. Abedin Dokoosh).
mulations that enable decreasing lymphocyte count with reduced
impact on other non-target tissues will be necessary.1
The polymeric nanoparticle (NP)-based drug delivery system is
ideal for controlled release drug delivery due to increased solubility
of hydrophobic drugs; enhanced half-life of drugs; improved drug
efficiency; reduced drug toxicity, and minimum side effects which
are associated with multiple drug dosing.2
Poly(3-hydroxybutyrate-co-3-hydroxyvalerate) (PHBV) is the
most commonly studied polymer of polyhydroxyalkanoates family.
It is a biodegradable, non-toxic polyester with a low production
cost; PHBV has been intensively investigated as a biomaterial for
tissue engineering and an NP-based drug delivery system.3
Artificial neural network (ANN) is a training computational
method based on simulating the neurological influence of the hu-
man brain to design the best proper model.4 This model works
according to non-linear relationships between input factors and

http://dx.doi.org/10.1016/j.xphs.2016.07.026
0022-3549/© 2016 American Pharmacists Association®. Published by Elsevier Inc. All rights reserved.
 S. Shahsavari et al. / Journal of Pharmaceutical Sciences 106 (2017) 176-182
pharmaceutical outputs through data iterative training and opti-
mization of the outcomes to minimize errors.5
ANNs have many usages in the field of pharmaceutics; they have
been employed for modeling the responses in drug delivery sys-
tems and for evaluating the ANN model with multiple linear
regression.6,7 ANNs used as a parallel system consist of interactions
between a huge number of simple calculation elements, nominated
node or neurons, and fulfill intricate information processing and
learn by models.8 The main important example of this model is the
new structure of the data processing system. It is the combination
of a large number of highly bound processing origin (neurons)
functioning in order to solve specific problems. An ANN is
employed for a special application, such as design identification or
information assortment, by means of learning action. Learning in
biological systems consists of regulations to the synaptic interfaces
within the neuron networks.9
In this work, based on the suitable properties of PHBV,
Fingolimod-loaded PHBV NPs are formulated to control the release
of hydrophobic Fingolimod by encapsulating it within hydrophobic
PHBV NPs. Moreover, a feed forward, multi-layer perceptron (MLP)
type of ANN is discussed for prediction of the mechanism of release
of Fingolimod NPs.
Materials and Methods
Materials
PHBV with 3 wt.% PHV and polyvinyl alcohol (PVA) (average
molecular weight 30,000-70,000) were purchased from Sigma-
Aldrich (St. Louis, MO). Fingolimod was synthesized at the Danish
Pharmaceutical Development Company. Analytical grade LiChro-
solv acetonitrile was purchased from Merck (Darmstadt, Germany)
and other chemicals used for the analytical methods were of
analytical grade and purchased from Sigma Chemical Company (St.
Louis, MO).
Preparation and Characterization of Nanoparticles
Fingolimod-loaded PHBV NPs were prepared by emulsification
and solvent evaporation technique which can be useful for poorly
water soluble drugs.
Different formulation variables such as concentration of PHBV
and PVA, amount of Fingolimod, and speed of magnetic stirrer (300,
600, 1000 rpm) varied, and the effect on particle size, polydispersity
index (PdI), loading capacity, and loading efficacy were considered.
Only one parameter was changed at a time in each set of experi-
ments. Briefly, chloroform (1 mL) containing PHBV (14.1 mg) as oil
phase with Fingolimod (5 mg) was sonicated for 2 min, and oil
phase was added drop wise to 5 mL PVA (as aqueous phase). The
probe ultrasonication was continued for 10 min and then the
mixture was gently stirred at room temperature for 4 h to allow
complete evaporation of the organic solvent. The obtained sus-
pension was centrifuged at 14,000 rpm for 30 min, the NPs were
settled down and lyophilized for further studies, and the super-
natant was used for determination of loading capacity and loading
efficacy and also study of drug released from NPs.
Experimental Design
Further than the information achieved by preliminary studies,
which defined the most significant factors, optimum levels of the
PHBV concentration (A), PVA concentration (B), and Fingolimod
amount (C) that can significantly affect the particle size, PdI,
loading, and encapsulation efficiency were analyzed using the
response surface methodology. Size and zeta potential of particles
177
were determined using photon correlation spectroscopy (Malvern
Instruments, Malvern, UK).
The mathematical relevance between the responses and inde-
pendent variables were modeled by a second-order polynomial
equation.
In order to graphically show the interactions between the var-
iables and the response, three-dimensional (3D) surface plots were
used in this study.
In Vitro Drug Release Study
The cumulative release of Fingolimod from the PHBV NP in non-
biological conditions was done during 30 days by a dialysis bag
(cutoff 12 kDa) in phosphate buffered saline (PBS) at pH 7.4 and
37 C (in sink condition).
At specified time intervals (30 min; 1, 2, 4, 8, 12, 24, 48, 72, and
120 h; 1, 2, 3, and 4 weeks), 1 mL of the medium was removed for
analysis and fresh PBS of an equal volume was displaced. The
in vitro release of Fingolimod was measured in triplicate. The
samples were analyzed by HPLC method by using C8, 125  4.6 mm,
5 mm column, acetonitrile and phosphate buffer pH 3 (45:55) as
mobile phase and UV detection of 220 nm.
Prediction and Optimization Functions of ANN
Commercially available software, MATLAB® R2008a (Math-
Works, Inc., Natick, MA), was applied to write mathematical code
for training and measuring the ANN developed and used for
formulation optimization.
MATLAB is a mercantile software developed by MathWorks, Inc.
It is a mathematical software package for theoretical and simu-
lating numerical calculation.10 Backpropagation, an abbreviation
for “backward propagation of errors,” is a usual method of training
ANNs applied in conjunction with an optimization procedure such
as gradient descent (GD). The method calculates the gradient of a
loss function with respect to all the weights in the network. Also,
the multi-layer network is sometimes referred to as a back-
propagation network. However, the backpropagation technique
that is used to compute gradients and Jacobians in a multilayer
network can also be applied to many different network
architectures.11
A neural network includes an interconnected group of artificial
neurons, working in union to solve particular problems. ANNs, like
human, learn by example. The neuron has 2 methods of operations:
the training/learning mode and the using/testing mode. In original
cases, an ANN is an adjusted system that converts its structure
based on external or internal information that goes through the
network in the learning phase. Recent neural networks are non-
linear statistical data modeling tools. They are commonly used to
model complex relationships among inputs and outputs or to find
models in data. MLP learning algorithm is an administered learning
algorithm. It is one of the most important progresses in neural
networks. This learning algorithm is used in multilayer feed-
forward networks including processing elements (neurons) with
continuous recognizable activation functions (tan-sigmoid and
log-sigmoid). Given a set of training input-output pair, this algo-
rithm prepares a procedure for changing the weights in a back-
propagation network to classify an input correctly. The sense for
this weight update algorithm is mainly the GD method applied in
case of simple perceptron networks with recognizable units. This is
a procedure where the error is propagated back to covered unit. The
aim of the neural network is to train the net to attain a balance
between the net's ability to reply (memorization) and its capability
to give reasonable responses to the input that is alike but not
identical to one that is applied in training.12
178 S. Shahsavari et al. / Journal of Pharmaceutical Sciences 106 (2017) 176-182

Table 1
Box-Behnken Experimental Design

Formula Independent Variables Dependent Variables

No.
X1 X2 X3 Particle Size (nm) Polydispersity Index Loading (%) Entrapment Efficiency (%)

1 1.50 0.25 5.00 268 0.23 22.53 73.80

2 1.50 1.13 2.75 286 0.20 10.19 56.00
3 0.50 2.00 2.75 495 0.48 7.63 5.06
4 0.50 1.13 5.00 300 0.23 19.96 14.32
5 1.50 1.13 2.75 218 0.13 10.19 65.00
6 0.50 1.13 0.50 368 0.27 4.17 12.00
7 1.50 2.00 5.00 292 0.24 13.00 51.48
8 1.50 2.00 0.50 351 0.30 0.01 51.52
9 2.50 1.13 5.00 245 0.17 14.41 58.00
10 0.50 0.25 2.75 350 0.40 16.50 35.86
11 1.50 1.13 2.75 239 0.19 10.19 67.00
12 2.50 1.13 0.50 261 0.21 0.42 17.80
13 1.50 1.13 2.75 281 0.19 10.19 67.86
14 1.50 0.25 0.50 324 0.30 2.10 57.55
15 1.50 1.13 2.75 241 0.19 10.19 67.86
16 2.50 0.25 2.75 330 0.40 6.52 56.73
17 2.50 2.00 2.75 270 0.23 2.89 24.12

In this research, a type of ANN called MLP was used for the
prediction of the mechanism of preparation of Fingolimod NPs.
MLP is one of the neural network architecture which was most
studied and consists of an input layer, one or more latent layers, and
an output layer. The neurons take input values in the input layer
and transfer them on to the neurons in the hidden layer without
any measurement.
The connectivity among the neurons is nominated weighting
variables. Weighting variables determine the strength of the input
data. Delta rule (learning algorithm in MLP) or backward propa-
gation is applied for optimization the weights during the learning
process. In the training data batch, ANN reads the input and output
values and moderates the value of the weighted links to minimize
the difference among the predicted and observed values. To achieve
the specified area of accuracy, the error is minimized through many
training cycles.
The primary parameter in training networks is train operation. A
trained ANN model can be applied to predict the response by
providing a batch of input amounts. To train an ANN model, the
data provided from experiments are categorized into 3 sets:
training, test, and validation sets. The training set is put on to train
the ANN model by adjusting the link weights of network model,
which consists of data covering all experimental space.
To forecast or optimize various kinds of controlled release for-
mulations, many researchers have used different types of ANN
models and learning algorithms.

Figure 1. The 3D response surface plots of Fingolimod nanoparticles obtained from optimum formulation for (a) particle size, (b) PdI, (c) loading capacity, and (d) entrapment
efficiency.
 S. Shahsavari et al. / Journal of Pharmaceutical Sciences 106 (2017) 176-182 179

Figure 2. % Drug release of Fingolimod from the PHBV nanoparticles for optimum formulation.

In this study, various algorithms for training ANN was used as X .

below: GD, Bayesian regularization (BR), and Levenberg-Marquardt
(LM) training model.
Gradient Descent
In this research, GD local search method which is the easiest and
popular training optimization algorithm was used. The procedure
operates by numerating and gradient calculating the output errors
and also optimizing the weights and biases.13
Bayesian Regularization
BR is a mathematical process which converts a nonlinear
regression into a statistical issue in the form of series regression.
This training model also automatically sets optimum values for
accurate function parameters and brings forward a smooth
network using the mean sum squared error.14
Levenberg Marquardt
The LM algorithm is a curve-fitting method and the most highly
used optimization algorithm that applies the minimum of a several
function which is considered as the sum of squares of the errors
among the function and the measured data parts.15
LM16 is indeed a compound of 2 minimization procedures:
the GD method for updating the values to the largest depletion
of the least squares and the Gauss-Newton model for consid-
ering the least squares function and finding the minimum of the
quadratic.
In this research, the focus was on the learning case known as
directed learning, in which a set of input or output data are avail-
able. Thereafter, the ANN has to be trained to make the favorable
output in accordance with the examples. In order to follow a su-
pervised training, the method for measuring ANN output error
between the real and the expected output is needed.
A proper formula is the mean squared error (MSE) which is
presented in Equation 1.
MSE ¼ ðyi  zi Þ2 n (1)
where yi is the predicted value, zi the observed value, and n the
number of dataset.
In this research, the data from in vitro release study were used to
model ANN. For this purpose, time was applied as the input and the
percent of drug released from NPs was used as the expected output
of the network. The algorithms applied in the MATLAB program
(mbackprop) consist of few numbers of iterations to speed up the
process. The training set was applied to train the network, and the
test set was employed to specify the level of generalization pro-
duced by the training set and to supervise overtraining the
network, with a corresponding MSE.
Results and Discussion
Characterization of Nanoparticles
The particle size, PdI, loading capacity, and loading efficiency
of NPs were measured by dynamic light scattering using a
Malvern Zetasizer ZS (Worcestershire, UK) (Table 1). Every
sample was measured 3 times and the data were represented as
mean ± SD.17
Experimental Design
The Box-Behnken design was used to discriminate the optimum
level of the independent variables, including PHBV concentration
(%), PVA (%), and Fingolimod amount (mg), affecting the particle
size, PdI, loading, and encapsulation efficiency of Fingolimod-
loaded NPs.
Particle Size and Polydispersity Index
ANOVA results indicated that only PHBV concentration (X1) has
an effect on the size of NPs (p ¼ 0.0008); however, PVA

Table 2
In Vitro Data Release of Fingolimod From the PHBV Nanoparticle in Optimum Formulation

Time (day) 0.02 0.04 0.08 0.16 0.32 0.5 1

Drug release (%) 3.58 ± 0.74 5.36 ± 2.57 7.44 ± 1.54 10.30 ± 4.26 12.38 ± 3.51 23.60 ± 1.97 24.31 ± 1.48
Time (day) 2 3 4 5 6 7 10
Drug release (%) 24.98 ± 1.25 25.21 ± 0.74 25.59 ± 1.19 26.87 ± 0.82 29.54 ± 4.51 42.44 ± 3.90 46.35 ± 0.89
Time (day) 15 20 23 25 27 30
Drug release (%) 50.84 ± 2.63 58.12 ± 1.38 60.98 ± 1.22 62.52 ± 3.45 64.87 ± 1.06 65.6 ± 0.53
180 S. Shahsavari et al. / Journal of Pharmaceutical Sciences 106 (2017) 176-182

Table 3
Training Parameters of Various Structures With GD Algorithm

Case Number Transfer Function Transfer Function Number Correlation Coefficient MSE of
of Layer of Hidden Layer of Output Layer of Neuron Training Set
Training Test Validation

A 10 tansig purelin 20 0.8879 0.8963 0.8312 0.0874

B 10 tansig purelin 30 0.8295 0.8653 0.7485 0.1205
C 10 tansig purelin 40 0.7692 0.7962 0.7340 0.1341
D 10 logsig purelin 20 0.9193 0.9773 0.8545 0.0521
E 10 logsig purelin 30 0.8793 0.8891 0.8389 0.0920

concentration (X2) and Fingolimod amount (X3) have no effect on
particle size (their p values are higher than 0.05). The plot for 3D
response surface of particle size alteration of optimum formulation
is shown in Figure 1a.
It was demonstrated that the concentration of PHBV is the most
significant factor that affected PdI (p ¼ 0.0052). As shown in 3D
plots (Fig. 1b), reducing the concentration of both PHBV and PVA
leads to a decrease in PdI.
Loading and Encapsulation Efficiency
3D response surface plot loading capacity and entrapment
efficiency of optimum formulation are illustrated in Figures 1c
and 1d.
Regression analysis of variance revealed that concentration of
PHBV, PVA, and amount of Fingolimod affected on loading (p <
0.0001), and concentration of PHBV and amount of Fingolimod
affected the loading efficacy.
In Vitro Drug Release Study
In vitro release of Fingolimod from the PHBV NP was studied in
PBS at pH 7.4 and 37 C according to United States Pharmacopeia
and the results were ready during 30 days as shown in Figure 2.
Moreover, data of in vitro release for NPs in optimum condition are
shown in Table 2. Experiments were administrated 3 times and
data are represented as mean ± SD.
The in vitro drug release profile showed a triphasic pattern, an
initial burst release in the first 24 h followed by sustained release
for up to 4 weeks.
The first stage (burst effect) is caused by drug adsorbed on the
surface of nanospheres, and the second stage (6 days) appears to
follow a slow release, because of drug embedding inside the mo-
lecular chain of the polymer. In the last phase, over 22 days, due to
diffusion of drug via the channels made by the bulk degradation,
drug release follows a linear relationship with time.
ANN Modeling
A neural network is composed of an input and output layer with
various numbers of layers and neurons, that is, a backpropagation
network was chosen for the purposes of this study. Time of release
was used as the input to the network and corresponding to that,
expected output including percent of drug released was applied,
using 20 experimental data generated from Box-Behnken design.
The training group was put for training the network and
learning models with different training functions.18 In order to
determine the amount of extension submitted by sets and evalu-
ated qualification of trained network, test batch was applied. A

Figure 3. Scatter plots of observed versus ANN predicted factorial variables: (a) GD, (b) BR, and (c) LM.
 S. Shahsavari et al. / Journal of Pharmaceutical Sciences 106 (2017) 176-182 181

Table 4
Training Parameters of Various Structures With BR Algorithm

Case Number Transfer Function Transfer Function Number Correlation Coefficient MSE of
of Layer of Hidden Layer of Output Layer of Neuron Training Set
Training Test Validation

A 10 tansig purelin 20 0.7849 0.7487 0.7839 0.2019

B 15 tansig purelin 20 0.8209 0.8581 0.8056 0.1912
C 20 tansig purelin 20 0.9541 0.8829 0.8455 0.0544
D 20 logsig purelin 20 0.9134 0.8815 0.8392 0.0698
E 15 logsig purelin 20 0.7685 0.7392 0.7620 0.1983

final check on validation of the trained network was completed by
using verification data sets. Therefore all the data were classified
into 3 groups including training (70%), testing (15%), and verifying
classes (15%).
To intercept satiation of the transfer function in MATLAB pro-
cess, approval of the data was scaled up to 0-1. To attain minimum
MSE, the number of hidden nodes was chosen because there was a
growth in error with increasing number of nodes. Also, different
number of hidden layers was examined for each training
algorithms.
ANN Model Training Using Gradient Descent Algorithm
In GD algorithm which was perceptually modeled as an easiest
training method in ANN, there was no significant difference in
proficiency by decreasing or increasing the number of hidden
layers; therefore the number of hidden layer was fixed to 10. Log-
sigmoid and tangent sigmoid (tansigmoid) was chosen as a hidden
layer and hyperbolic (purelin) as an output layer for transfer
function. Several training with GD training algorithms are dis-
played in Table 3.
It was seen that with the increasing number of neurons by more
than 20 in both logsigmoid and tansigmoid hidden layer, the
amount of error was increased and R2 was reduced. Overall, log-
sigmoid has shown better performance results.
The observed versus predicted values using GD methodology is
shown in Figure 3a, where R2 ¼ 0.9023.
ANN Model Training Using Bayesian Regularization Algorithm
In accordant with BR training algorithm, tansig and logsig were
chosen as a transfer function of hidden layer and purlin as an
output layer, respectively, with 20 neurons. With increasing num-
ber of hidden layers (up to 20) for tansig transfer function of the
hidden layer, the minimum amount of error has been obtained.
When logsig was selected with 20 layers, the development of cor-
relation coefficient has not been observed.
As in Figure 3b, the observed versus predicted values using BR
methodology eventuates R2 ¼ 0.8326. The modeling outcomes for
various structures with BR algorithm are reported in Table 4.
ANN Model Training Using Levenberg-Marquardt Algorithm
In this research, LM algorithm which is commonly applied for
mean square problems and approaching functions was also used as
a training function with 15 layers, and therefore tansig and logsig
have been selected as a transfer function of hidden layer and purlin
as a transfer function of output layer. With the increment in the
number of neurons (up to 20), the amount of error increased in
tansig hidden layer transfer function. Moreover, tansig with 15
layers and 20 neurons showed better performance compared to
logsig.
As shown in Figure 3c, the observed versus predicted values
using LM methodology eventuates R2 ¼ 0.9291.
The training parameters of various networks with the LM
training model are demonstrated in Table 5.
Performance Criteria
ANN operates based on receiving the input and output in-
formation in the training group and changing a value of the
weighted link to diminish the variation among the perceived
and predicted values. Because of some training cycles called
epoch, the error is minimized to achieve acceptable network
precision.
In this research, many numerical iterations like epoch with
minimum MSE (optimum epoch) were selected. The analogy be-
tween 3 GD models was also built by recording the end of the
central processing unit time that passed.
The process of minimizing errors in network is adjustable; in
contrast, the error in response surface methodology depends on
experimental values and cannot be progressed, even in corre-
spondence with best optimization.19
Table 6 demonstrates the statistic measures and performance
indexes for every 3 training algorithm. As displayed in Table 6, LM
algorithms have shown better average for MSE compared with GD
and BR (0.0341 vs. 0.0521 and 0.0544). Moreover, LM has shown the
smaller average mean prediction error toward the others (0.2571
vs. 0.2782 and 0.3545).
As explained previously, the training ended when minimum
root-mean-square error on the test dataset was obtained. The
number of training epochs and the time elapsed for total
epochs indicated when the training ended, which varied
drastically between GD modes.
LM outperformed GD and BR in terms of prediction and gener-
alization capability. On the other hand, LM demonstrated to be less
biased and more precise, in comparison with GD and BR.

Table 5
Training Parameters of Various Structures With LM Algorithm

Case Number Transfer Function Transfer Function Number Correlation Coefficient MSE of
of Layer of Hidden Layer of Output Layer of Neuron Training Set
Training Test Validation

A 15 tansig purelin 20 0.9526 0.9102 0.9178 0.0341

B 15 tansig purelin 30 0.8954 0.8912 0.9055 0.0508
C 15 tansig purelin 40 0.7758 0.8506 0.8764 0.0965
D 15 logsig purelin 20 0.8025 0.8865 0.8930 0.0732
E 15 logsig purelin 30 0.7928 0.8619 0.8853 0.0880
182 S. Shahsavari et al. / Journal of Pharmaceutical Sciences 106 (2017) 176-182
Table 6
Statistical and Comparison Between Performance Index of Three Training
Algorithms
Performance Index Training Algorithms
GD BR
Average MSE for training set 0.0521 0.0544
Average mean prediction error 0.2782 0.3545
Average number of epochs at the 231 205
end of training
Average central processing unit 8.5 7.2
time elapsed at the end of training (s)
Conclusion
In this study, Box-Behnken design-expert experimental design
was used to optimize and reach the best formulation of Fingoli-
mod NPs based on biodegradable PHBV. The consequence of
formulation variables includes concentration of PHBV, PVA, and
amount of Fingolimod. Also, NPs' characteristics was included
particle size, PdI, loading capacity, and entrapment efficiency. The
in vitro release study of optimized Fingolimod NPs was then car-
ried out.
The ANN was employed to predict the best model for in vitro
release study of NPs. The feed-forward backpropagation was
applied using MATLAB program to appraise the effect of input
variable (time) on response (amount of drug released from NPs).
Three training model formulations including LM, GD, and BR were
employed for training the ANN models. Therefore, the generaliza-
tion and predictability of the models were investigated.
It is demonstrated that the predictive ability of each training
algorithm is in the order of LM > GD > BR. In summary, optimum
formulation was achieved by LM training function with 15 hidden
layers and 20 neurons. The transfer function of the hidden layer for
this formulation and the output layer was tansig and purlin,
respectively. Also, the MSE of training was 0.0341.
Finally, the results of this research demonstrate that ANN was a
useful program for modeling and predicting a nanoparticulate
Fingolimod delivery system.
References
1. Chen J, Davis S. The release of diazepam from poly (hydroxybutyrate-hydrox-
yvalerate) microspheres. J Microencapsul. 2002;19(2):191-201.
2. Pacheco DP, Amaral MH, Reis RL, Marques AP, Correlo VM. Development of an
LM injectable PHBV microparticles-GG hydrogel hybrid system for regenerative
medicine. Int J Pharm. 2015;478(1):398-408.
0.0341 3. Masood F, Chen P, Yasin T, Fatima N, Hasan F, Hameed A. Encapsulation of
0.2571 Ellipticine in poly-(3-hydroxybutyrate-co-3-hydroxyvalerate) based nano-
379 particles and its in vitro application. Mater Sci Eng C Mater Biol Appl.
2013;33(3):1054-1060.
10.6 4. Stanwick JC, Baumann MD, Shoichet MS. Enhanced neurotrophin-3 bioactivity
and release from a nanoparticle-loaded composite hydrogel. J Control Release.
2012;160(3):666-675.
5. Duan B, Wang M. Encapsulation and release of biomolecules from CaeP/PHBV
nanocomposite microspheres and three-dimensional scaffolds fabricated by
selective laser sintering. Polym Degrad Stab. 2010;95(9):1655-1664.
6. Peh KK, Lim CP, Quek SS, Khoh KH. Use of artificial neural network to predict
drug dissolution profiles and evaluation of network performance using simi-
larity factor. Pharm Res. 2000;17:1384-1388.
7. Barmplexis P, Kanaze FI, Kachrimanis K, Georgarakis E. Artificial neural net-
works in the optimization of a nimodipine controlled release tablet formula-
tion. Eur J Pharm Biopharm. 2010;74:316-323.
8. Pham DT. An introduction to artificial neural networks. In: Bulsari AB, ed. Neural
Networks for Chemical Engineering. Amsterdam: Elsevier; 1995 Chapter 1.
9. Shahsavari S, Vasheghani-Farahani E, Ardjmand M, Dorkoosh F. Modeling
of drug released from acyclovir nanoparticles based on artificial neural
networks. Lett Drug Des Discov. 2014;11(2):174-183.
10. The MathWorks Inc. Pro-matlab for Sun Workstations, User's Guide. Natick, MA:
The MathWorks Inc.; 1990.
11. Chaibva F, Burton M, Walker RB. Optimization of salbutamol sulfate dissolution
from sustained release matrix formulations using an artificial neural network.
Pharmaceutics. 2010;2:182-198.
12. Jain S. Brain cancer classification using GLCM based feature extraction in
artificial neural network. Int J Comput Sci Eng Technol. 2013;4(7):966-970.
13. Rumelhart D, Hinton GE, Williams RJ. Learning representations by back-
propagation errors. Nature. 1986;233:533-536.
14. MacKay DJC. A practical Bayesian framework for back propagation network.
Neural Comput. 1992;4(3):448-472.
15. Hecht-Nielsen R. Kolmogorov's mapping neural network existence theorem.
San Diego, CA: Proc First IEEE Int Joint Conf Neural Networks; 1987:11-14.
16. Levenberg KA. Method for the solution of certain non-linear problems in least
squares. Q Appl Math. 1994;2:164-168.
17. De Campos AM, Sa nchez A, Alonso MJ. Chitosan nanoparticles: a new
ophthalmic vehicle. Int J Pharm. 2001;224:159-168.
18. Zhang L. Parallel Training Algorithms for Analogue Hardware Neural Nets. Ph.D.
thesis. Australia: School of software engineering and Data communication,
Queensland University of Technology; 2007.
19. Shahsavari S, Bagheri G, Mahjub R, et al. Application of artificial neural net-
works for optimization of preparation of insulin nanoparticles composed of
quaternized aromatic derivatives of chitosan. Drug Res. 2014;64:151-158.