Abdalla, Youssef; (2025) Applying Machine Learning Methods to Pharmaceutical Datasets for Developing Personalised Medicinal Products. Doctoral thesis (Ph.D), UCL (University College London).

Preview

Text Abdalla_10207411_thesis.pdf Download (33MB) | Preview

Abstract

Current one-size-fits-all medicines fail to account for significant patient heterogeneity, rendering them ineffective in up to 70% of patients. Recognising this, there is a growing shift towards personalised medicine. This thesis proposes a machine learning (ML)-driven personalised medicine pathway where a patient’s optimal dose is predicted, used to generate a new drug formulation, and three-dimensional (3D)-printed into a personalised medicine. However, a key challenge in ML using pharmaceutical data is the lack of high-quality, large datasets, which limits model generalisation. To address this, VECT-GAN (Variational Encoded Conditional Tabular Generative Adversarial Network) was developed to generate high-quality synthetic data to augment small pharmaceutical datasets. This model was applied to multiple pharmaceutical case studies, significantly outperforming state-of-the-art generative models and halving model prediction errors. VECT-GAN also improved performance in predicting 3D printing outcomes. Building on this, ML’s ability to optimise the drug formulation and printing processes using Selective Laser Sintering was explored. To achieve this, an ensemble of neural networks (NNs) was trained to predict 3D-printability with 90% accuracy, and the optimal laser scanning speed and printing temperature with 92% accuracy. By integrating these NNs with a Differential Evolution algorithm, a pipeline was developed to generate 3D-printable medicines along with the required printing parameters. Experimental validation showed that 80% of the generated formulations were successfully printed. Having established a personalised medicine manufacturing method, this research further explores dose personalisation using tacrolimus as a candidate drug. Data from two hospitals was used to train a novel long short-term memory (LSTM) network to predict tacrolimus blood levels with a mean absolute error of less than 5%. These predictions facilitate therapeutic drug monitoring and can be used to determine an optimal tacrolimus dose, providing a practical example of personalised dosing. Together, these approaches demonstrate the role of ML in enabling the production of personalised medications.

Download activity - last month

Export as XMLJSONCSV

Download activity - last 12 months

Export as JSONCSVXML

Downloads by country - last 12 months

Export as XMLCSVJSON

Archive Staff Only

View Item