Search

Abstract This study investigates the application of various neural network-based models for predicting temperature distribution in freeze drying process of biopharmaceuticals. For heat-sensitive biopharmaceutical products, freeze drying is preferred to prevent degradation of pharmaceutical compounds. The modeling framework is based on CFD (Computational Fluid Dynamics) and machine learning (ML). The ML models explored include the Single-Layer Perceptron (SLP), Multi-Layer Perceptron (MLP), Fully Connected Neural Network (FCNN), and Deep Neural Network (DNN). Model optimization is achieved through the Fireworks Algorithm (FWA). Results reveal promising performance across all models, with the MLP demonstrating the highest accuracy on both test and training datasets, achieving an R2 score of 0.99713 and 0.99717 respectively. The SLP also exhibits strong performance, with an R2 of 0.88903 on the test dataset. The FCNN and DNN models also perform admirably, achieving R2 scores of 0.99158 and 0.99639 on the test dataset respectively. These results highlight the efficiency of neural network-driven models, specifically the MLP, in precisely forecasting temperature values based on spatial coordinates. Additionally, the integration of the Fireworks Algorithm for model refinement yields advantages in improving the predictive performance of these models.

Preparation of small-molecule API (Active Pharmaceutical Ingredient) at submicron size would be of great benefit for pharmaceutical engineering, as the drug particles at submicron size possess higher solubility in water. Indeed, the drug bioavailability can be enhanced when the nanomedicine is prepared. In this study, the solubility of the drug desoxycorticosterone acetate (DA) is being examined to assess its viability of nanonization using supercritical operation. Two inputs are temperature and pressure which were considered for machine learning modeling in this study. The drug's solubility is the only output to be estimated by the optimized models. This dataset has 45 rows of data that were gathered at 5 different pressure and temperature levels. Support vector machine (SVM) is used as the core of the models built in this research. Epsilon-SVR and nu-SVR are models based on this concept, which together with two different polynomial and RBF kernels form the four models built in this research for estimation of DA drug solubility. The models are also optimized with the help of a new TLCO method. All four final models have an R2 score higher than 0.9, and among them, the Epsilon-SVR model with RBF kernel has the best performance with 0.967.

Production of solid-dosage drug nanoparticles was assessed by theoretical models to investigate the possibility of drug treatment via supercritical green processing. Nanonization can enhance drug solubility and consequently its bioavailability which is of great importance for pharmaceutical industry. This research presents a comparative study of three different regression models including Gaussian process regression, k-nearest neighbors, and multi-layer perceptron for predicting solvent density and solubility of Hyoscine drug. The models optimized using political optimizer (PO) algorithm. The results showed that all three optimized methods were able to predict density and solubility with high accuracy. PO-GPR achieved the highest R2 score for solubility (0.9984) and same for density (0.9999). The PO-MLP model achieved the high R2 score for density (0.9997) and the second-highest score for solubility (0.9945). PO-KNN also showed good performance for density (R2 = 0.9557) and solubility (R2 = 0.9783) but was outperformed by the other two models. In terms of RMSE and AARD%, PO-GPR and PO-MLP achieved lower error rates compared to PO-KNN. Overall, the results suggest that PO-GPR and PO-MLP are promising methods for predicting density and solubility of values. The models were useful for the application of drug nanonization and can be used to optimize the process.