Computer- aided formulation development - Computer-Aided Applications in Pharmaceutical Technology - page 24

considering only the molecule features and gave an idea for more
general networks, trained with data on systems involving other oils,
surfactants, and surfactant-to-cosurfactant ratios. In a related study,
Agatonovic-Kustrin and Alany (2001) estimated the infl uence of
the cosurfactant on phase behavior of the fi ve-component systems
(ethyl oleate (oil)/a mixture of sorbitan monolaurate and polyoxyethylene
20 sorbitan monooleate (surfactant+cosurfactant)/deionized water/
n- alcohols (1-propanol, 1-butanol, 1-hexanol, and 1-octanol) or
1,2-alkanediols (1,2-propandiol, 1,2-pentanediol, 1,2-hexanediol,
and 1,2-octanediol)(cosurfactant)). A supervised network with a
multilayer perceptron (MLP) architecture with a BP learning rule (Neural
Networks ® , StatSoft Inc, Tulsa, USA), was used to correlate phase
behavior of the investigated systems with cosurfactant descriptors
(inputs), which were preselected by a genetic algorithm (GA) (Pallas 2.1,
Compu Drug Int., San Francisco, USA and ChemSketch 3.5 freeware,
ACD Inc., Toronto, Canada).
The most successful MLP ANN model, with two hidden layers
comprising 14 and 9 neurons, predicted the phase behavior for a new set
of cosurfactants with 82.2% accuracy for the microemulsion region.
Alany et al. (1999) presented the fi rst report describing the utility of
ANNs in predicting phase behavior of the four component system
(ethyl oleate (oil)/sorbitan monolaurate (primary surfactant)/
polyoxyethylene 20 sorbitan monooleate (secondary surfactant)/
deionized water) regarding the components ratio. The BP training
algorithm was selected. The training and testing data were extracted
from several pseudo-ternary triangles, which represented the cuts through
the phase tetrahedron. The inputs were percentages of oil and water and
HLB values of the surfactants blend. The outputs were the corresponding
systems (o/w emulsion, w/o emulsion, microemulsion, and liquid
crystals). The trained MLP (ANNs simulator software, NNMODEL
Version 1.404, Neural Fusion), with 1 hidden neuron, was tested on
validation data and an accuracy of 85.2 to 92.9% was estimated,
depending on the output critical values used for the classifi cation. The
low error rate demonstrated the success in employing ANNs to predict
phase behavior of quaternary systems.
The fundamental goal in SMEDDS development is to optimize the
surfactant/cosurfactant/oil mixture, in order to achieve suffi cient drug
solubility and infi nite dilutability with water phase. However, there is a
risk of disturbing the thermodynamic stability on dilution with the
subsequent drug precipitation (Kyatanwar et al., 2010). The study of
Mendyik and Jachowicz (2006) describes the development of the system

￿

￿ ￿