See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/318301185

Artificial Neural Network based Particle Size Prediction of Polymeric

Nanoparticles

Article  in  European journal of pharmaceutics and biopharmaceutics: official journal of Arbeitsgemeinschaft fur Pharmazeutische Verfahrenstechnik e.V · July 2017
DOI: 10.1016/j.ejpb.2017.06.030

CITATIONS READS
12 898

3 authors, including:

John Youshia Mohamed Ehab Ali

University of Bonn University of Bonn
9 PUBLICATIONS   82 CITATIONS    14 PUBLICATIONS   147 CITATIONS   

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Predicting particle size of polymeric nanoparticles View project

Spray freeze Drying View project

All content following this page was uploaded by John Youshia on 21 December 2017.

The user has requested enhancement of the downloaded file.

 Artificial Neural Network based Particle Size Prediction of
Polymeric Nanoparticles

John Youshia1,2, Mohamed Ehab Ali1,3 and Alf Lamprecht1,4

1
Department of Pharmaceutics, Institute of Pharmacy, University of Bonn, Bonn, Germany.
2
Department of Pharmaceutics and Industrial Pharmacy, Faculty of Pharmacy, Ain Shams
University, Cairo, Egypt.
3
Department of Industrial Pharmacy, Faculty of Pharmacy, Assiut University, Assiut, Egypt.
4
FDE (EA4267), University of Burgundy/Franche-Comté, Besançon, France.

1
Abstract
Particle size of nanoparticles and the respective polydispersity are key factors influencing their
biopharmaceutical behavior in a large variety of therapeutic applications. Predicting these attributes
would skip many preliminary studies usually required to optimize formulations. The aim was to
build a mathematical model capable of predicting the particle size of polymeric nanoparticles
produced by a pharmaceutical polymer of choice. Polymer properties controlling the particle size
were identified as molecular weight, hydrophobicity and surface activity, and were quantified by
measuring polymer viscosity, contact angle and interfacial tension, respectively. A model was built
using artificial neural network including these properties as input with particle size and
polydispersity index as output. The established model successfully predicted particle size of
nanoparticles covering a range of 70-400 nm prepared from other polymers. The percentage bias
for particle prediction was 2%, 4% and 6%, for the training, validation and testing data,
respectively. Polymer surface activity was found to have the highest impact on the particle size
followed by viscosity and finally hydrophobicity. Results of this study successfully highlighted
polymer properties affecting particle size and confirmed the usefulness of artificial neural networks
in predicting the particle size and polydispersity of polymeric nanoparticles.

Keywords: Polymeric nanoparticles; particle size; prediction; artificial neural network; contact
angle; interfacial tension; viscosity; in-silico.

List of Abbreviations

ANN Artificial neural network PLA Poly(lactic acid)

EC Ethyl cellulose PLGA Poly(lactic-co-glycolic acid)
IFT Interfacial tension PVA Poly(vinyl alcohol)
MW Molecular weight PVAc Poly(vinyl acetate)
NPs Nanoparticles RMSE Root mean square error
PCL Poly(ε-caprolactone) RSM Response surface methodology
PDI Polydispersity index θc Contact Angle

1. Introduction
Nanoparticles (NPs) have proven their efficient use as drug delivery carriers in a wide range
of therapeutic applications [1,2]. This is based on unique properties allowing for the enhancement
of drug penetration across biological barriers [3,4] and drug targeting towards malignant [5,6] and
inflamed tissues [7]. Approaches such as passive targeting with modified surface properties by
using different surfactants or stealth NPs by polyethylene glycol decoration involve significant
changes to surface properties. Similarly, active targeting can be accomplished by decorating the
surface of nanoparticles with targeting moieties and ligands [8]. Most of these phenomena have
been found to be size-dependent, for example; enhanced oral drug absorption [9], selective
targeting towards tumors [10] or inflamed tissues [11]. Therefore, controlling the particle size of

2
NPs and its distribution is of crucial importance. Size distribution is usually defined by the
polydispersity index (PDI), which specifies the uniformity and stability of NPs and should be
within 0.01 to 0.5 [12]. Until now, when a certain particle size with a narrow size distribution is
aimed for, this has been done by empirical approaches and trial and error. Accordingly, developing
a mathematical model that can predict the particle size and PDI of polymeric NPs obtained from
various types of polymers would be very beneficial, as it will save time and money by preserving
polymers, chemicals and materials normally consumed during the optimization phase.
The main statistical and modelling tools used for optimizing and predicting characteristics
of NPs are response surface methodology (RSM) [13,14] and artificial neural network (ANN) [15–
18]. Both approaches were already compared to each other with the results demonstrating the
superiority of ANN to RSM in data fitting and prediction capabilities [19–22]. This was attributed
to the limitation of RSM to quadratic functions only unlike ANN, which can handle a broader range
of functions and find relationships between independent and dependent variables with no prior
specific mathematical equation or function [19]. ANN learns by example, where a data set is used
for building the model termed training data and then the efficiency of established model is checked
against new data termed testing data.
Previous studies using ANN focused mainly on investigating the process parameters
affecting particle size and examined a limited number of polymers without relating polymer
properties to the obtained particle size [16–18]. In these cases, the developed models were used to
characterize and optimize the factors affecting the NPs preparation process. However, thorough
evaluation of the prediction power of these models was not the primary focus in these studies, as
the test data were relatively small and confined to the training range of the model.
Here, ANN was utilized to develop a mathematical model capable of predicting the particle
size and PDI of polymeric nanoparticles manufactured from a larger choice of pharmaceutical
polymers with various properties. In order to achieve this goal, polymer properties affecting
particle size and PDI were precisely identified, quantified and then used as an input for the model.
Afterwards, the model was tested comprehensively against data located inside and outside the
borders used to train it. Furthermore, the evaluation of the established model involved completely
new polymers, which were not included in the training data.

2. Materials and Methods

2.1. Materials
Ethyl cellulose (EC) with an ethoxy content of 48-49.5% but different molecular weights
(MWs) and consequently viscosity grades (Ethocel® standard 4, 7, 10 and 45 premium) was a kind
donation from Colorcon (Dartford, England). Acid terminated poly(lactic acid) (PLA) (Purasorb
PDL 02A) was a gift from Purac Biomaterials (Gorinchem, The Netherlands). Poly(vinyl acetate)
(PVAc) (Vinnapas B17 special) was kindly granted by Wacker Chemie AG (Burghausen,
Germany). Ammonio Methacrylate Copolymer, Type B (Eudragit® RS PO) was a kind sample
from Evonik (Darmstadt, Germany). Poly(vinyl alcohol) (PVA) (Poval® 40–88) was a gift from
Kuraray (Frankfurt, Germany). Acid terminated poly(DL-lactide-co-glycolide) (PLGA) 50:50 of
different MWs and viscosities (Resomer® RG 502 H MW 7,000-17,000, Resomer® RG 503 H MW
24,000-38,000 and Resomer® RG 504 H MW 38,000-54,000) and ester terminated PLGA 50:50
(Resomer® RG 505 MW 54,000-69,000) were purchased from Evonik (Darmstadt, Germany).
Poly(ε-caprolactone) (PCL) (Mn 10000 and Mn 45000) were purchased from Sigma-Aldrich
(Steinheim, Germany). All other chemicals were of analytical grade or equivalent purity (For
further details on the characteristics of EC and PLGA polymer types see supplementary materials
Tables 1 and 2).
3
 2.2. Preparation of polymeric nanoparticles
Polymeric nanoparticles were prepared using the emulsification solvent evaporation
method [23] replacing dichloromethane with ethyl acetate. Briefly, 100 mg of the respective
polymer was dissolved in ethyl acetate to form the organic phase, while the aqueous phase was
composed of PVA in different concentrations (0.05-1.5%). The aqueous phase was added to the
organic phase and the mixture was ultrasonicated using a probe ultrasonicator (Sonoplus HD 2200,
Bandelin, Berlin, Germany) at an amplitude of 50% for 3 minutes. Preliminary experiments
revealed that varying the emulsification energy input did not significantly affect neither the particle
size nor PDI (Data not shown). After emulsification, the organic solvent was evaporated using a
rotary evaporator (Rotavapor RE 120, Büchi, Flawil Switzerland) under reduced pressure to form
the polymeric NPs. The volume of the aqueous phase was kept constant at 10 ml while the volume
of ethyl acetate, in which the polymer was dissolved, was changed (2-9 ml) altering the solvent to
water ratio (S:W) to elucidate its effect on the particle size.

2.3. Particle size and PDI of polymeric nanoparticles

The particle size and PDI of the nanoparticles were determined by dynamic light scattering
technique (Nanopartica SZ-100, Horiba, Kyoto, Japan) at a fixed angle of 90° at 25 °C using 1.5
ml polymethyl methacrylate cuvettes. The samples were diluted with distilled water before
measuring to avoid multiple scattering.

2.4. Determination of polymer solutions viscosity

Viscosity of polymer solutions in ethyl acetate was measured using a rotation viscometer
(Haake Rheostress 1, Thermo Fisher Scientific, Karlsruhe, Germany). 100 mg of the respective
polymer was dissolved in 5.5 ml of ethyl acetate (1.81% w/v) and its viscosity was measured at
shear rate of 100 s-1 at 20 °C.

2.5. Determination of polymer contact angle

The static contact angle (θc) of different polymers was measured by the drop shape analysis
technique (DSA100, Krüss, Germany). Firstly, polymer films were fabricated using the solvent
casting method [24]. The polymer was dissolved in ethyl acetate to simulate the NPs preparation
procedure and then the solution was cast on a glass slide and placed in a hood overnight at room
temperature to evaporate the organic solvent. To ensure complete drying, the glass slides were
further dried under vacuum at 25 °C overnight (VDL23, Binder, Germany). Secondly, the θ c
between the obtained dried polymer films and water was determined using sessile drop method at
room temperature. A constant volume of water (8 μl) was deposited on the surface of the polymer
film and the contact angle value was calculated from the recorded droplet image using sessile drop
fitting method.

2.6. Interfacial tension measurements

Interfacial tension (IFT) between water and either pure or polymer-containing ethyl acetate
was also determined using drop shape analysis technique (DSA100, Krüss, Germany) applying
pendant drop method. A water drop was formed through a needle of diameter 1.8 mm in the organic
phase, which was placed in an optical glass cuvette (40 × 40 × 30 mm). IFT was then calculated
from the recorded droplet image of the suspended water drop in ethyl acetate.

4
 2.7. Artificial neural network model
A typical ANN consists of three connected layers; input layer composed of the independent
factors, output layer represented by the responses and a hidden layer(s) in-between made of a
certain number of nodes connecting the input layer to the output one. The input layer had 5 factors;
viscosity of polymer solutions in ethyl acetate, θc between the polymer film and water, IFT between
ethyl acetate polymer solutions and water, concentration of PVA and finally S:W. The output layer
had 2 responses; particle size and PDI of the polymeric NPs. There was one hidden layer with 10
nodes (Figure 1). ANN was built using Visual Gene Developer 1.7 build 763 [25]. A set composed
of 24 experimental data was used for building it (Supplementary Material Table 3). It was selected
to cover all of the polymer properties affecting the particle size and PDI of NPs. Ammonio
methacrylate copolymer and EC represented polymers with surface-active properties, while PLA,
different MWs of EC and PLGA were utilized to represent diverse viscosities and contact angles.
The set was randomly split into training (85%) and validation (15%) data. The testing data were
divided into three major groups; the first included polymers used in the training phase, the second
was represented by polymers new to the network but closely related to those used in first group
and finally the third group contained polymers with which the network was completely unfamiliar.
The model was trained using a feed forward neural network with a standard back propagation
learning algorithm [26] using hyperbolic tangent as the transfer function. The sum of error was
calculated using the following equation:
𝑁t 𝑁o

𝑆𝑢𝑚 𝑜𝑓 𝑒𝑟𝑟𝑜𝑟 = ∑ ∑(𝑌𝑎 − 𝑌p)2 (1)

where Nt is the total number of training data sets, No is the total number of output variables, Ya is
the actual value of the output variable and Yp is the predicted value of the output variable. To
avoid overtraining of the network which may lower its prediction capabilities, the target sum of
error was adjusted to 0.01. To test the network efficiency, the coefficient of determination (R2),
root mean square error (RMSE) and percentage bias [27] were calculated using the equations
mentioned in supplementary data.

The contribution of each factor to the particle size and PDI was calculated using the
connection weights method [28] using the following equation [29]:
𝐶ik = |∑ 𝑎ij × 𝑏jk| (2)
where aij and bjk represents the weights of the connections between ith input node and jth hidden
node and between jth hidden node and kth output node, respectively. Then the relative impact of
each input was determined by dividing the contribution of each factor by the sum of the
contribution of all factors.

2.8. Statistical analysis

Data were compared using a one-way analysis of variance (ANOVA) followed by Tukey–
Kramer multiple comparisons test using GraphPad Prism 5 (GraphPad Software, CA, USA).
Surface plots have been generated using a demo version of TableCurve 3D version 4 (Systat
Software, CA, USA).

5
Figure 1: A diagram of the artificial neural network design representing the components of the
input, hidden and output layers.

3. Results
3.1. Nanoparticle formulation parameters
The viscosities of all polymers dissolved in ethyl acetate was measured at a fixed
concentration of 1.81% (w/v) to obtain comparable values (Figure 2). Results revealed that there
was a simultaneous increase of viscosity, as a consequence of higher EC and PLGA molecular
weights. This was accompanied by a significant increase (P<0.05) in particle size from 145 and
187 nm to 201 and 210 nm (Figure 2A) together with a decrease in PDI (Figure 2B) from 0.133
and 0.088 to 0.098 and 0.054 for EC and PLGA, respectively.

Figure 2: Effect of increasing the polymer molecular weight (expressed as viscosity) of ethyl
cellulose and PLGA on the (A) particle size and (B) polydispersity index of the prepared
nanoparticles. Particles were prepared using 0.5% poly(vinyl alcohol) at a solvent to water ratio
of 0.3.

6
 Contact angles were measured between water and the solvent-cast films of 4 polymers with
comparable viscosities: PVAc (0.635 mPa.s), EC 4 (1.322 mPa.s), PLGA 503 H (0.666 mPa.s) and
PCL Mn 10000 (0.634 mPa.s). It is evident from the results that PCL is the most hydrophobic
polymer followed by PLGA, EC and finally PVAc (Figure 3). Correlating these findings with the
particle size of the NPs declared that more hydrophobic polymers with higher contact angles (with
the exception of EC) resulted in significantly larger particles (P<0.05). Accordingly, PCL produced
the largest NPs (214 nm) followed by PLGA 503 H (202 nm) and finally PVAc (192 nm), when
all other conditions were kept constant. On the contrary, increasing polymer hydrophobicity
decreased PDI progressively from 0.133 to 0.055. According to the measured θc of all the polymers
used in this study, they can be arranged in the following ascending order; PVAc (70.8°) < EC
(72.9°) < Eudragit RS (73.6°) < acid terminated PLGA (74.1°) < acid terminated PLA (76.2°) <
ester terminated PLGA (76.6°) < PCL (81.1°).

Figure 3: Effect of polymer hydrophobicity (expressed as contact angle) on (A) the particle size
and (C) polydispersity index of the prepared nanoparticles. Particles were prepared using 0.5%
poly(vinyl alcohol) at a solvent to water ratio of 0.3. (B) Contact angle images between water and
polymer films of (a) poly(vinyl acetate), (b) ethyl cellulose 4, (c) PLGA 503 H and (d) poly(ε-
caprolactone) (PCL) Mn 10000.

The effect of increasing the concentration of various polymers in ethyl acetate on the IFT
was evaluated. In presence of PLGA, PVAc and PCL, no significant change in the IFT value
recorded between water and ethyl acetate (6.23 mN/m) was observed, indicating the absence of
any surface active properties (Figure 4). Similarly, no change in the droplet volume was observed
with the previously mentioned polymers (Supplementary Materials Figure 1). On the other hand,
EC displayed evident surfactant-like properties as demonstrated by droplet shrinkage and
significant decrease in the IFT (P<0.05) by increasing its concentration. This was reflected in its
surprisingly small NPs of size 145 nm (Figure 3).

7
Figure 4: Effect of increasing the concentration of different polymers dissolved in ethyl acetate on
the interfacial tension between water and ethyl acetate.

In addition, the effect of changing certain formulation parameters was also investigated
namely; PVA concentration and S:W. Decreasing the PVA concentration and increasing S:W
increased the particle size from 124 to 398 nm, 140 to 400 nm and 146 to 339 nm, as demonstrated
in the surface plots for EC, PLGA and PVAc NPs, respectively (Figure 5).

Figure 5: Surface plots of the effect of poly(vinyl alcohol) % and solvent to water ratio on the
particle size of nanoparticles made of (A) ethyl cellulose 4, (B) PLGA 502 H and (C) poly(vinyl
acetate).

8
 3.2. The artificial neural network model
The ANN model was successfully built and the predicted particle size and PDI were close
to the values recorded experimentally for the training, validation (Figure 6 and 7) and testing data
(Table 1). It successfully covered particle size and PDI range of 70-400 nm and 0.05-0.5
respectively. The R2, RMSE and % bias of the model are provided in Table 2.

Figure 6: Practical vs predicted particle size of nanoparticles of the training and validation data
with the percentage bias displayed above each bar.

Figure 7: Practical vs predicted polydispersity index of nanoparticles of the training and

validation data with the percentage bias displayed above each bar.
9
Table 1: Testing the predictive capabilities of the artificial neural network model with a cocktail of polymers
Viscosity IFT PVA Practical Predicted Practical Predicted Bias Bias
Category Polymer θc S:W
(mPa.s) (mN/m) (%) PS (nm) PS (nm) PDI (nm) PDI (nm) PS PDI
4.14 0.05 0.7 341 340 0.334 0.237 0% 29%
0.25 0.3 182 190 0.113 0.151 4% 34%
Ethyl Cellulose 4 1.322 72.9
2.46 0.5 0.3 144 162 0.133 0.090 13% 32%
First 1.5 0.3 133 116 0.075 0.067 13% 10%
Group 0.15 0.3 297 307 0.200 0.287 3% 43%
0.5 0.3 187 204 0.088 0.115 9% 30%
PLGA 502 H 0.514 74.1 6.23
0.5 0.5 228 244 0.150 0.152 7% 1%
1.5 0.2 135 127 0.108 0.105 6% 3%
Ethyl Cellulose 7 1.591 72.9 2.37 0.5 0.3 159 174 0.092 0.040 9% 56%
Second
PLGA 503 H 0.666 74.1 6.23 0.5 0.3 202 207 0.081 0.084 2% 4%
Group
PLGA 505 1.049 76.6 6.23 0.5 0.3 224 219 0.080 0.018 2% 77%
0.2 0.3 287 272 0.098 0.236 5% 141%
0.4 0.90 339 346 0.420 0.399 2% 5%
0.7 0.55 214 204 0.136 0.141 5% 4%
Poly(vinyl
0.635 70.8 6.23 0.7 0.9 247 229 0.113 0.248 7% 119%
acetate)
1.00 0.55 193 186 0.155 0.160 4% 3%
Third
1.00 0.20 146 127 0.130 0.095 13% 27%
Group
1.25 0.3 157 148 0.047 0.105 6% 123%
0.25 0.4 314 313 0.117 0.146 0% 25%
0.634 81.1
Poly(ε- 0.5 0.3 214 225 0.055 0.052 5% 6%
6.23
caprolactone) 0.5 0.3 219 228 0.060 0.013 4% 78%
0.893 81.1
0.75 0.7 252 243 0.110 0.059 3% 46%
IFT: Interfacial tension PS: Particle size PVA: Poly(vinyl alcohol) S:W: Solvent to water ratio θc: Contact Angle

10
 By calculating the contribution of each factor from the weights of the connecting links
between the input and output layers passing by the hidden layer as mentioned in Equation (2) it
was found that the most influencing factor on the particle size is the PVA concentration followed
by IFT, solvent to water ratio, viscosity and finally the contact angle. While for the PDI, PVA
concentration has the highest impact followed by viscosity > IFT > contact angle > S:W (Figure
8).

Figure 8: Pie chart showing the relative impact of each factor on the particle size and
polydispersity index of nanoparticles. PVA: Poly(vinyl alcohol), S:W: Solvent to water ratio, IFT:
Interfacial tension, θc: Contact angle and η: Viscosity.

Table 2: Statistical evaluation of the artificial neural network model for the particle size and
polydispersity index

Particle Size Polydispersity Index

Training Validation Testing All Training Validation Testing All

Data Data Data Data Data Data Data Data

R2 0.994 0.995 0.966 0.981 0.992 0.996 0.625 0.854

RMSE 6.770 15.964 11.939 10.282 0.012 0.025 0.058 0.042

% Bias 2 4 6 4 8 17 41 24

11
4. Discussion
Particle size is a critical parameter for the biological performance of NPs, where small
differences can have a significant impact. Moreover, high PDI values indicate a broad size
distribution presenting difficulty in concluding a relationship between biological effects and certain
particle size [30]. Therefore, developing a mathematical model capable of predicting particle size
and PDI from a pharmaceutical polymer of choice based on the respective properties, would reduce
the formulation efforts enormously during the design of NPs with a specific particle size.
A large majority of polymeric NPs are prepared by using water-insoluble hydrophobic
polymers. Accordingly, the commonly-used emulsification solvent evaporation technique was
chosen for NPs preparation because of its simplicity and low purification requirement compared to
other methods [31]. This limited the solvent choice to water-immiscible and partially miscible
solvents. Chlorinated solvents like dichloromethane and chloroform are among the most frequently
used solvents. However, according to the ICH guideline for residual solvents Q3C(R6), both
solvents belong to class 2 and should be limited in the pharmaceutical industry because of their
toxicological profile and health hazards. This shifted the focus in the past years towards ethyl
acetate as a better alternative. It is a class 3 solvent with lower toxicity and still is able to dissolve
many pharmaceutical polymers. Nonetheless, although this model was based on ethyl acetate, it
can be also extended to combine other solvents, where a similar behavior is expected, however,
this was beyond the scope of this study since the polymer properties was the main focus.
Viscosity of the organic phase directly affects the produced particle size of polymeric NPs,
where higher viscosity results in larger particles. This is attributed to droplet breakdown resistance
during the emulsification step, lowering the efficiency of particle size reduction by ultra-sonication
[32]. Factors controlling organic phase viscosity include solvent-polymer interactions, polymer
concentration and molecular weight. Good solvents for a certain polymer usually display higher
viscosity than poor ones. This is based on the fact that better solvent-polymer interaction unfolds
polymer chains and entangles solvent molecules [33]. Among the investigated polymers, EC
solutions in ethyl acetate were the most viscous due to its better solvation leading to chain
expansion more than the other polymers, which were comparatively coiled. Polymer chain
expansion creates gaps within the polymer structure, entrapping the solvent, hindering its flow and
subsequently increasing the viscosity [34]. Similar observations were noticed for PCL in
dichloromethane against PLGA and Eudragit [35]. Moreover, increasing either polymer
concentration [36,37] or molecular weight [34,38,39] produced larger NPs as a consequence of
higher viscosity [34,36] resulting from the increased number [38] and length of polymer chains,
respectively. This is in line with the results obtained here, as the MW of EC or PLGA increased,
solutions viscosity became higher and led to larger NPs with lower PDI.
Polymer hydrophobicity depends on its molecular structure. The ability of the polymer film
to interact with water, during contact angle measurements, depends on the presence of hydrophilic
groups on the polymer surface and their flexibility to orient themselves towards water molecules.
Polyester polymers commonly used for drug delivery include PCL, PLA and PLGA. From their
chemical structure, it can be postulated that PCL is more hydrophobic (as confirmed by measuring
θc) due to the lower oxygen:carbon ratio. For PLGA, hydrophobicity depends on two factors;
whether it is acid or ester terminated and ratio of the lactide moiety to the glycolide one. Ester
terminated PLGA 505 had a higher θc than acid terminated PLGA 502 H, 503 H and 504 H. Thus,
the carboxylic group enhances polymer-water interactions, lowering the contact angle and reducing
polymer hydrophobicity. On the other hand, polymer MW did not affect hydrophobicity, as all acid
terminated PLGA, although of different MW, had the same θc. This suggests that the terminal group
and not the molecular weight of PLGA is the dominant factor affecting polymer hydrophobicity.
12
Since lactic acid is more hydrophobic than glycolic acid due to its extra methyl group, therefore as
the lactide ratio increases PLGA hydrophobicity rises [40]. Applying the same principle PLA
should be more hydrophobic than PLGA, where acid terminated PLA had a larger θ c than acid
terminated PLGA. PVAc and methacrylate copolymers are characterized by the presence of the
ester group as a side chain and not direclty integrated in the backbone of the polymer like
polyesters. This provides the ester group more flexibility and potentially allows for more intense
interaction with water molecules. This was reflected in the θc, where both had lower values than
polyester polymers. PVAc was even less hydrophobic because of its higher oxygen: carbon ratio.
EC exhibited a θc lower than polyesters which may be attributed to the ethoxy content of 48 - 49.5%
explaining its lower hydrophobicity. Moreover, more hydrophobic polymers usually produce larger
NPs. In accordance to our findings, previous studies showed that PCL produced larger NPs than
PLGA [35,41]. The same was also observed for the more hydrophobic ester-terminated PLGA
against its acid-terminated counterpart [42]. This is attributed to the polymer affinity to water,
where the more hydrophobic polymers have lower affinity to the aqueous medium and therefore
tend to lower the surface area in contact with water opting for larger particles. Therefore, higher
surfactant concentration will be needed to stabilize the surface and get the same particle size
obtained by the less hydrophobic ones.
Whether the polymer of choice possesses surface-active properties or not, is another
important character affecting the size of particles produced by different polymers. Smaller particles
were produced by EC due to its surface-active properties, where it acts as a polymeric emulsifier,
adsorbs at the organic/aqueous interface lowering the IFT and stabilizing the particle surface.
Moreover, EC was reported earlier to lower the IFT and stabilize ethyl acetate/water emulsions
[43]. Similarly, ammonio methacrylate copolymers are also surface-active and were successfully
used in preparing surfactant-free NPs [44]. Owing to their positively charged quaternary groups,
they were adsorbed at the interface and formed an electrical double layer stabilizing the interface
[44]. The higher the emulsifying properties of the polymer, as detected by the IFT reduction, the
smaller the obtained NPs. This is evident, where Eudragit RS PO produced smaller NPs than EC.
Increasing the PVA concentration decreased the particle size because of the availability of
more surfactant molecules, which stabilized the expanded polymer/water interface of smaller
emulsion droplets. Furthermore, increasing surfactant concentration lowers PDI to a certain limit,
after which PDI will eventually increase due to the formation of small and large NPs populations
[43]. On the other hand, increasing the S:W was accompanied by larger particles and wider
distribution in accordance with previous findings [37] and in contrast to others [45]. This may be
referred to the influence of the larger volume of organic solvent involved in the production of the
initial emulsion causing; net shear stress reduction [46], PVA dilution and larger organic/aqueous
interface formation.
The ANN mathematical model was very efficient as displayed from the R 2, RMSE and %
bias values. The established model demonstrated very good predicting abilities for the seen data
and test data. The polymers used to test the predictive capabilities of the network were carefully
selected to examine the network thoroughly. The first group of formulations represented polymers
(EC and PLGA) to which the network was already exposed in the training phase. For each polymer;
there were preparations within the PVA% range (0.1 to 1% for EC 4 and 0.3 to 1% for PLGA 502
H) introduced to the network and preparations outside it. The second group represented new
polymers but closely related to the ones used to train the network. These included different
molecular weight of EC (EC 7) and PLGA (PLGA 503 H, ester terminated PLGA 505). The third
group represented polymers with which the network was completely unfamiliar. These included;
PVAc and PCL as examples of less and more hydrophobic polymers respectively with contact
13
angle values located outside the range used for training the network. Against all testing
formulations, the model successfully predicted the particle size with minor error demonstrating its
efficiency. However, it was less successful for PDI where the error was relatively higher than
particle size. This is attributed to the smaller numeral values of PDI data compared to particle size.
Accordingly, minor deviations from the practical PDI values appear as high error values. Moreover,
PDI measurements usually suffer from less reproducibility than particle size making its predictions
much more difficult.
Furthermore, the connection weights method was used to calculate the contribution of each
input to the output [47] and arrange them according to their relative impact. Surface activity,
represented by both PVA% and surface-active properties of the polymer, was found to be the main
contributor to the particle size as stabilizing the interface between the hydrophobic polymer chains
and the hydrophilic water molecules, is the major driving force for NPs formation. S:W and
polymer viscosity had a comparable impact, while the influence of the contact angle was limited
which means that altering polymer hydrophobicity will cause only a minor shift in the particle size.
For PDI, polymer viscosity was the second most influential factor after surfactant concentration.
This indicates that increasing viscosity of polymer solutions will probably produce NPs with
narrower size distribution due to reduction of particle collisions avoiding aggregation and
coalescence.
Generally, the results described in this study indicate the extensive power of ANN in
finding relationships between various factors and building a well-performing predictive model.
This opens the door for further models capable of predicting the influence of other parameters on
the characteristics of NPs. Employing such machine-learning techniques for predicting the zeta
potential of NPs, the influence of different preparation methods on NPs physico-chemical
properties or the proper type and concentration of surfactant needed to achieve a certain particle
size will be definitely more complicated, but worth the investigations.

Conclusion
Through artificial neural network, a mathematical model was successfully built capable of
predicting the particle size and polydispersity index of polymeric nanoparticles manufactured from
a large choice of pharmaceutical polymers. This was accomplished by identifying polymer
properties affecting the particle size namely; surface activity, viscosity and hydrophobicity. This
study opens the door for predicting the properties of nanoparticles in silico saving time and effort.

Acknowledgments
John Youshia would like to disclose financial support from the Egyptian Ministry of Higher
Education and the German Academic Exchange Service (Deutsche Akademische Austauschdienst,
DAAD) (91527629).

Declaration of Interest
The authors have no other relevant affiliations or financial involvement with any
organization or entity with a financial interest in or financial conflict with the subject matter or
materials discussed in the manuscript. This includes employment, consultancies, honoraria, stock
ownership or options, expert testimony, grants or patents received or pending, or royalties.

14
[1] W. Park, K. Na, Advances in the synthesis and application of nanoparticles for drug delivery,
Wiley.Interdiscip.Rev.Nanomed.Nanobiotechnol. 7 (2015) 494–508.
[2] C. Saraiva, C. Praça, R. Ferreira, T. Santos, L. Ferreira, L. Bernardino, Nanoparticle-mediated brain
drug delivery: Overcoming blood-brain barrier to treat neurodegenerative diseases, J. Control.
Release Off. J. Control. Release Soc. 235 (2016) 34–47. doi:10.1016/j.jconrel.2016.05.044.
[3] M.M. Abdel-Mottaleb, B. Moulari, A. Beduneau, Y. Pellequer, A. Lamprecht, Nanoparticles enhance
therapeutic outcome in inflamed skin therapy, Eur.J.Pharm.Biopharm. 82 (2012) 151–157.
[4] J. Youshia, A.O. Kamel, S.A. El, S. Mansour, Design of cationic nanostructured heterolipid matrices
for ocular delivery of methazolamide, Int.J.Nanomedicine. 7 (2012) 2483–2496.
[5] H. Maeda, The enhanced permeability and retention (EPR) effect in tumor vasculature: the key role
of tumor-selective macromolecular drug targeting, AdvEnzyme Regul. 41 (2001) 189–207.
[6] Y. Matsumura, H. Maeda, A new concept for macromolecular therapeutics in cancer
chemotherapy: mechanism of tumoritropic accumulation of proteins and the antitumor agent
smancs, Cancer Res. 46 (1986) 6387–6392.
[7] A. Lamprecht, IBD: selective nanoparticle adhesion can enhance colitis therapy,
Nat.Rev.Gastroenterol.Hepatol. 7 (2010) 311–312.
[8] P.I. Siafaka, N. Üstündağ Okur, E. Karavas, D.N. Bikiaris, Surface Modified Multifunctional and
Stimuli Responsive Nanoparticles for Drug Targeting: Current Status and Uses, Int. J. Mol. Sci. 17
(2016) 1440. doi:10.3390/ijms17091440.
[9] C. He, L. Yin, C. Tang, C. Yin, Size-dependent absorption mechanism of polymeric nanoparticles for
oral delivery of protein drugs, Biomaterials. 33 (2012) 8569–8578.
[10] D. Peer, J.M. Karp, S. Hong, O.C. Farokhzad, R. Margalit, R. Langer, Nanocarriers as an emerging
platform for cancer therapy, Nat.Nanotechnol. 2 (2007) 751–760.
[11] J. Youshia, A. Lamprecht, Size-dependent nanoparticulate drug delivery in inflammatory bowel
diseases, Expert. Deliv. 13 (2016) 281–294.
[12] L. Wu, J. Zhang, W. Watanabe, Physical and chemical stability of drug nanoparticles, Adv. Drug
Deliv. Rev. 63 (2011) 456–469. doi:10.1016/j.addr.2011.02.001.
[13] S. Honary, P. Ebrahimi, R. Hadianamrei, Optimization of particle size and encapsulation efficiency of
vancomycin nanoparticles by response surface methodology, Pharm. Dev. Technol. 19 (2014) 987–
998. doi:10.3109/10837450.2013.846375.
[14] D. Neumann, C. Merkwirth, A. Lamprecht, Nanoparticle design characterized by in silico
preparation parameter prediction using ensemble models, J.Pharm.Sci. 99 (2010) 1982–1996.
[15] A.A. Metwally, R.M. Hathout, Computer-Assisted Drug Formulation Design: Novel Approach in Drug
Delivery, Mol.Pharm. 12 (2015) 2800–2810.
[16] E. Esmaeilzadeh-Gharedaghi, M.A. Faramarzi, M.A. Amini, N.A. Rouholamini, S.M. Rezayat, A.
Amani, Effects of processing parameters on particle size of ultrasound prepared chitosan
nanoparticles: an Artificial Neural Networks Study, Pharm.Dev.Technol. 17 (2012) 638–647.
[17] R.A. Hashad, R.A.H. Ishak, S. Fahmy, S. Mansour, A.S. Geneidi, Chitosan-tripolyphosphate
nanoparticles: Optimization of formulation parameters for improving process yield at a novel pH
using artificial neural networks, Int. J. Biol. Macromol. 86 (2016) 50–58.
doi:10.1016/j.ijbiomac.2016.01.042.
[18] K.S. Shalaby, M.E. Soliman, L. Casettari, G. Bonacucina, M. Cespi, G.F. Palmieri, O.A. Sammour, A.A.
El Shamy, Determination of factors controlling the particle size and entrapment efficiency of
noscapine in PEG/PLA nanoparticles using artificial neural networks, Int. J. Nanomedicine. 9 (2014)
4953–4964. doi:10.2147/IJN.S68737.
[19] D. Bas, I.H. Boyaci, Modeling and optimization II: Comparison of estimation capabilities of response
surface methodology with artificial neural networks in a biochemical reaction, J. Food Eng. 78
(2007) 846–854.

15
[20] K.M. Desai, S.A. Survase, P.S. Saudagar, S.S. Lele, R.S. Singhal, Comparison of artificial neural
network (ANN) and response surface methodology (RSM) in fermentation media optimization: Case
study of fermentative production of scleroglucan, Biochem. Eng. J. 41 (2008) 266–273.
doi:10.1016/j.bej.2008.05.009.
[21] Y. Li, M.R. Abbaspour, P.V. Grootendorst, A.M. Rauth, X.Y. Wu, Optimization of controlled release
nanoparticle formulation of verapamil hydrochloride using artificial neural networks with genetic
algorithm and response surface methodology, Eur. J. Pharm. Biopharm. 94 (2015) 170–179.
doi:10.1016/j.ejpb.2015.04.028.
[22] M.R. Zaki, J. Varshosaz, M. Fathi, Preparation of agar nanospheres: Comparison of response surface
and artificial neural network modeling by a genetic algorithm approach, Carbohydr. Polym. 122
(2015) 314–320. doi:10.1016/j.carbpol.2014.12.031.
[23] A. Lamprecht, Y. Bouligand, J.P. Benoit, New lipid nanocapsules exhibit sustained release properties
for amiodarone, JControl Release. 84 (2002) 59–68.
[24] Z.G. Tang, R.A. Black, J.M. Curran, J.A. Hunt, N.P. Rhodes, D.F. Williams, Surface properties and
biocompatibility of solvent-cast poly[ε-caprolactone] films, Biomaterials. 25 (2004) 4741–4748.
doi:10.1016/j.biomaterials.2003.12.003.
[25] S.-K. Jung, K. McDonald, Visual gene developer: a fully programmable bioinformatics software for
synthetic gene optimization, BMC Bioinformatics. 12 (2011) 340. doi:10.1186/1471-2105-12-340.
[26] D.E. Rumelhart, G.E. Hinton, R.J. Williams, Learning representations by back-propagating errors,
Nature. 323 (1986) 533–536. doi:10.1038/323533a0.
[27] S.M. Abdel-Hafez, R.M. Hathout, O.A. Sammour, Towards better modeling of chitosan
nanoparticles production: screening different factors and comparing two experimental designs, Int.
J. Biol. Macromol. 64 (2014) 334–340. doi:10.1016/j.ijbiomac.2013.11.041.
[28] J.D. Olden, D.A. Jackson, Illuminating the black box: a randomization approach for understanding
variable contributions in artificial neural networks, Ecol. Model. 154 (2002) 135–150.
doi:10.1016/S0304-3800(02)00064-9.
[29] T. Hattori, S. Kito, Analysis of factors controlling catalytic activity by neural network, Adv. Catal.
Catal. Mater. Energy Environ. Prot. 10th Jpn.-Korea Symp. Catal. Held Shimane Prefect. Assem. Hall
Matsue Jpn. 10--12 May 2005. 111 (2006) 328–332. doi:10.1016/j.cattod.2005.10.044.
[30] M. Gaumet, A. Vargas, R. Gurny, F. Delie, Nanoparticles for drug delivery: The need for precision in
reporting particle size parameters, Eur. J. Pharm. Biopharm. 69 (2008) 1–9.
doi:10.1016/j.ejpb.2007.08.001.
[31] C. Pinto Reis, R.J. Neufeld, Ribeiro António J., F. Veiga, Nanoencapsulation I. Methods for
preparation of drug-loaded polymeric nanoparticles, Nanomedicine Nanotechnol. Biol. Med. 2
(2006) 8–21. doi:10.1016/j.nano.2005.12.003.
[32] S. Galindo-Rodriguez, E. Allémann, H. Fessi, E. Doelker, Physicochemical parameters associated
with nanoparticle formation in the salting-out, emulsification-diffusion, and nanoprecipitation
methods, Pharm. Res. 21 (2004) 1428–1439.
[33] K.S. Gandhi, M.C. Williams, Solvent effects on the viscosity of moderately concentrated polymer
solutions, J. Polym. Sci. Part C Polym. Symp. 35 (1971) 211–234. doi:10.1002/polc.5070350117.
[34] S. Desgouilles, C. Vauthier, D. Bazile, J. Vacus, J.-L. Grossiord, M. Veillard, P. Couvreur, The Design of
Nanoparticles Obtained by Solvent Evaporation:  A Comprehensive Study, Langmuir. 19 (2003)
9504–9510. doi:10.1021/la034999q.
[35] Y.V. Chernysheva, V.G. Babak, N.R. Kildeeva, F. Boury, J.P. Benoit, N. Ubrich, P. Maincent, Effect of
the type of hydrophobic polymers on the size of nanoparticles obtained by emulsification–solvent
evaporation, Mendeleev Commun. 13 (2003) 65–67. doi:10.1070/MC2003v013n02ABEH001690.
[36] S. Bohrey, V. Chourasiya, A. Pandey, Polymeric nanoparticles containing diazepam: preparation,
optimization, characterization, in-vitro drug release and release kinetic study, Nano Converg. 3
(2016) 1–7.
16
 [37] N. Sharma, P. Madan, S. Lin, Effect of process and formulation variables on the preparation of
parenteral paclitaxel-loaded biodegradable polymeric nanoparticles: A co-surfactant study, Asian J.
Pharm. Sci. 11 (2016) 404–416.
[38] X. Song, Y. Zhao, S. Hou, F. Xu, R. Zhao, J. He, Z. Cai, Y. Li, Q. Chen, Dual agents loaded PLGA
nanoparticles: Systematic study of particle size and drug entrapment efficiency, Eur. J. Pharm.
Biopharm. 69 (2008) 445–453. doi:10.1016/j.ejpb.2008.01.013.
[39] G. Mittal, D.K. Sahana, V. Bhardwaj, M.N.V. Ravi Kumar, Estradiol loaded PLGA nanoparticles for
oral administration: Effect of polymer molecular weight and copolymer composition on release
behavior in vitro and in vivo, J. Controlled Release. 119 (2007) 77–85.
doi:10.1016/j.jconrel.2007.01.016.
[40] H.K. Makadia, S.J. Siegel, Poly Lactic-co-Glycolic Acid (PLGA) as Biodegradable Controlled Drug
Delivery Carrier, Polymers. 3 (2011) 1377–1397. doi:10.3390/polym3031377.
[41] M. Snehalatha, K. Venugopal, R.N. Saha, Etoposide-Loaded PLGA and PCL Nanoparticles I:
Preparation and Effect of Formulation Variables, Drug Deliv. 15 (2008) 267–275.
doi:10.1080/10717540802174662.
[42] Y.N. Konan, R. Gurny, E. Allémann, Preparation and characterization of sterile and freeze-dried sub-
200 nm nanoparticles, Int. J. Pharm. 233 (2002) 239–252. doi:10.1016/S0378-5173(01)00944-9.
[43] P. Wachsmann, A. Lamprecht, Ethylcellulose nanoparticles with bimodal size distribution as
precursors for the production of very small nanoparticles, Drug Dev. Ind. Pharm. 41 (2015) 1165–
1171. doi:10.3109/03639045.2014.935393.
[44] C.A. Nguyen, Y.N. Konan-Kouakou, E. Allemann, E. Doelker, D. Quintanar-Guerrero, H. Fessi, R.
Gurny, Preparation of surfactant-free nanoparticles of methacrylic acid copolymers used for film
coating, AAPS.PharmSciTech. 7 (2006) 63.
[45] R.M. Mainardes, R.C. Evangelista, PLGA nanoparticles containing praziquantel: effect of
formulation variables on size distribution, Int. J. Pharm. 290 (2005) 137–144.
doi:10.1016/j.ijpharm.2004.11.027.
[46] X. Song, Y. Zhao, W. Wu, Y. Bi, Z. Cai, Q. Chen, Y. Li, S. Hou, PLGA nanoparticles simultaneously
loaded with vincristine sulfate and verapamil hydrochloride: Systematic study of particle size and
drug entrapment efficiency, Int. J. Pharm. 350 (2008) 320–329. doi:10.1016/j.ijpharm.2007.08.034.
[47] J.D. Olden, M.K. Joy, R.G. Death, An accurate comparison of methods for quantifying variable
importance in artificial neural networks using simulated data, Ecol. Model. 178 (2004) 389–397.

17

View publication stats