Designing processes for pharmaceutical product manufacturing is a complex and
resource-intensive task. With increasing research costs and quality standards, the
pharmaceutical industry seeks innovative technologies to enhance productivity and
maintain competitiveness. While a variety of tools exist in the process design domain for
optimizing conditions or selecting materials, options for guiding the selection of
manufacturing operations remain limited.
In this context, deep generative models (DGMs) emerge as a promising approach.
DGMs, known for learning the probability distribution of data, have gained popularity for
their ability to generate realistic examples, commonly applied in text and image
generation. In drug discovery, DGMs have successfully generated new active substances
with desirable properties. However, their application in the manufacturing space remains
unexplored. These models have the potential to assist in operation selection and
experimental targeting, thereby reducing development time.
This thesis aims to investigate the applicability of DGMs in pharmaceutical manufacturing
process design, developing DGMs capable of generating plausible sequences of
operations for product manufacturing, taking input information about the target product.
A significant challenge in developing DGMs is the requirement for large datasets. To
address this, two datasets were constructed using natural language processing (NLP)
applied to primary and secondary manufacturing data extracted from patents. The
primary processing dataset comprises over 385K manufacturing processes, while the
secondary processing dataset includes approximately 9K procedures for various dosage
forms and active ingredients.
The study involved training and comparing several architectures based on generative
adversarial networks (GAN) and variational autoencoder (VAE) using different metrics.
Real and generated sequences were contrasted manually to evaluate how closely the
model outputs resemble typical manufacturing sequences. This research contributes to
the exploration of DGMs’ application in pharmaceutical manufacturing, offering insights
into their potential for operation selection and product development. In the end, DGMs
were successfully trained and their potential for the generation of plausible sequences
was demonstrated. A survey assessed by a panel of experts yielded that the models generated sequences at least as good as the actual procedures in 38% of occasions for
the primary domain. While this shows the potential of generative modelling in this field, it
also remarks there is room for improvement to make it applicable in real-world scenarios.
| Date of Award | 6 Feb 2025 |
|---|
| Original language | English |
|---|
| Awarding Institution | - University Of Strathclyde
|
|---|
| Sponsors | University of Strathclyde |
|---|
| Supervisor | Cameron Brown (Supervisor) & Blair Johnston (Supervisor) |
|---|