(189ab) Application of Artificial Neural Network As a Predictive Tool for Continuous Liposome Processing

Liposomes are lipid bilayer vesicles which are used as drug delivery systems, especially as part of anti-cancer and anti-fungal therapies. At UConn, a continuous liposomal processing system has been developed and implemented 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 time consuming mechanistic computational modeling. In this work, an artificial neural network (ANN) was used as a predictive tool. We used experimental data results to train the ANN with input features such as the critical material attributes (CMA) and critical process parameters (CPP) that are known to affect the ANN output representing liposomal critical quality attributes (CQA). A major benefit of this work is that ANN has been shown to have predictive power by relating CMA and CPP to the CQAs of liposome formulation.

Methods:

Experiments have shown that the CMAs and CPPs that affect the liposome formation process are hydrocarbon tail length, percent cholesterol, temperature and flow rate of aqueous phase, type of buffer used. Natural and synthetic phospholipids were used for the hydrocarbon tail length studies. The liposomal formation process was modeled using ANN to predict CQAs (such as particle size, particle distribution index, and zeta potential). Multiple-input-multiple-output model (MIMO) and multiple-input-single-output (MISO) models were implemented using Levenberg- Marquardt (LM) algorithm. Initially 175 points were used to test the MIMO and MISO models. Later the natural phospholipid was added to make a total of 210 data points. Buffer was also added as one of the input features and the total data points were around 600

Results:

Mean Relative Errors (MREs) between the target values and predicted values were calculated for performance evaluation. In case of particle size, the training and testing MREs for MIMO were 32% and 33% and for MISO were 18% and 21%, respectively with 175 points. MISO model tested with 220 points showed the training and testing MREs as 12.01% and 12.35%, respectively. The incorporation of an additional input feature did not cause any significant change in the error as the training and testing MREs were found to be 11.55% and 12.65% respectively.

Conclusions:

Our findings indicate that the MISO model was superior than the MIMO model in terms of quantitative prediction of the CQAs. Our model is capable of predicting CQAs for natural as well as synthetic phospholipids. Ongoing work includes increasing the number of test points performing process optimization to further decrease the errors.

Acknowledgements:

FDA Grant# 1U01FD005773-01.

Disclaimer:

This article reflects the views of the authors and should not be construed to represent FDA’s views or policies.