Deep Learning for Solid Form Engineering

: (Alternative Format Thesis)

Abstract

Novel pharmaceutical products are only valuable if they can be mass-produced effectively. Consequently, designing efficient, cost-effective and sustainable processes for drug manufacturing is crucial for the longevity and future of the pharmaceutical industry. Crystallization, a technique used for both purification and separation, is one of the most crucial steps during manufacturing. Crystallization produces solid-state substances which can be precisely engineered and controlled to exhibit desirable properties such as solubility, stability and flowability. The morphology of the solid form plays an important role in determining these physical and chemical properties of the drug; therefore, morphology is of utmost importance when developing manufacturing approaches. At present, trial-and-error laboratory experiments dominate the solid form engineering landscape. Despite sophisticated platforms continuing to increase the throughput of experimental screening efforts, the pharmaceutical industry has experienced a significant productivity decline as these methods continue to incur high costs and low success rates. For the approach to be considered efficient, the inverse scenario must be true, where there are lower costs and greater success rates. Although becoming less effective at producing novel chemical entities as pharmaceutical products, screening on these scales 10³-10⁶ molecules) generates a plethora of information that can be leveraged to make data-driven decisions in the future. In this thesis, the use of artificial intelligence is presented as a method of capturing the relationships uncovered by screening efforts and employing them to develop intelligent models capable of predicting solid form properties. This includes predicting solid form morphology and flowability. In addition to training the predictive models, this thesis incorporates a novel method of representing molecules in machine-readable form. Using this representation approach enables transfer learning training strategies which deliver superior model performance metrics. Although industrial screening is conceivable, few researchers have access to sophisticated platforms. To overcome these restrictions, a custom low-cost alternative is presented as a means of improving throughput while also creating reproducible labels that are less susceptible to human bias when labels are assigned by multiple researchers. The work presented in this thesis aligns with the technologies incorporated as part of the drive towards Industry 4.0 as applied to pharmaceutical solid form engineering. By using data-driven approaches, industrial manufacturing can become more effective by lowering costs, reducing resource consumption (thus making them more sustainable), and leveraging faster digital experimentation. In the end, this will not only increase the success rate of pharmaceutical discovery, bringing therapies to market faster, but it will also satisfy the industry-wide push towards sustainable manufacturing.

Date of Award	15 Nov 2023
Original language	English
Awarding Institution	University of Bath
Supervisor	Bernardo Castro Dominguez (Supervisor), Uriel Martinez Hernandez (Supervisor) & Chick Wilson (Supervisor)