(288f) Using Artificial Neural Network As a Predictive Tool for Critical Quality Attributes (CQAs) of Continuous Processing of Liposomes
Liposomes are bilayer vesicles of lipids which are used as drug delivery systems. A continuous liposome processing system has been developed and implemented at UConn, where lipids are dissolved in ethanol and mixed with an aqueous phase in a co-axial, co-flow arrangement. Predictive modeling of this process is essential, due to the tedious nature of laboratory experiments and the challenging computational simulations. The first aim of this research is to use experimental data to train and test artificial neural network (ANN) input features such as the critical material attributes (CMA) and critical process parameters (CPP) which affect the ANN output of the critical quality attributes (CQA). The second aim is to use molecular descriptors to train the ANN with more information. Molecular descriptors give information about the structural or functional features of molecules.
ANN is useful in case of non-linear data and can help in relating CMA and CPP to the CQAs of continuously processed liposomes. The input parameters such as hydrocarbon tail length, percent cholesterol, type of buffer (osmolarity), formation temperature of liposomes and flow rate of aqueous phase used were included as a part of the ANN model. These parameters were taken from the experimental data which corresponded to the CMAs and CPPs of the liposome formation process. Natural and synthetic phospholipids were used for the hydrocarbon tail length studies. ANN was used to predict the CQAs of liposomes such as particle size and particle distribution index. Multiple-input-multiple-output model (MIMO) and multiple-input-single-output (MISO) models for ANN were implemented using Levenberg- Marquardt (LM) algorithm in MATLAB and tested with an initial 175 data points. Later, additional data points were added and a total of 950 data points were used to build the ANN model. PaDel software was used to get the molecular descriptors which can help to obtain useful data and information. Decision Trees were able to predict a good measure of the relative importance of each descriptor which has higher probability of impact on the prediction result. MATLAB based Graphical User Interface (GUI) was implemented after the model development.
Mean Relative Errors (MREs) and Mean Squared Errors (MSEs) between the target values and predicted values were calculated for performance evaluation. For the MIMO and MISO models, the training MREs were 32% and 18% respectively with 175 points. Hence, MISO was used for the further model development. The incorporation of additional datapoints decreased the training error significantly. 1D and 2D descriptors from PaDel software were used to get ~800 descriptors. Decision tree was used to identify ~10 important descriptors from them. Molecular descriptors were added along with the other input parameters which dropped the training MRE for particle size to 4.2% and PDI to 2.5%. GUI was implemented and associated with the ANN based model to exhibit its predictions.
Our findings indicate that the MISO model was superior than the MIMO model in terms of quantitative prediction of the CQAs. Our model has shown capabilities of predicting CQAs such as particle size and PDI for natural as well as synthetic phospholipids. A GUI was successfully built for an end user to perform interactive risky analysis via data entry and visualization of the model predictions.
FDA Grant# 1U01FD005773-01.
This article reflects the views of the authors and should not be construed to represent FDAâs views or policies