Contreas, Leonardo
(2023)
Application of Machine Learning techniques for biomaterials design.
PhD thesis, University of Nottingham.
Abstract
Background: Over the years, the ever-growing development of computational systems allowed the cheminformatics and machine learning disciplines to flourish. With the former, chemical structures of any kind (small molecules, proteins, polymers) can be translated into machine language, and the latter can perform predictive analysis on those chemicals to find a pattern that correlates with a property or biological activity. Machine learning is an invaluable tool for performing predictive analyses on an endless number of tasks, with the healthcare sector being one of the most notable, but there is often a trade-off between predictive power and complexity, which ultimately leads to a poor interpretability. Healthcare-associated infections frequently arise in hospital environments, heavily affecting economy and society, and they are mainly caused by bacterial colonisation of the surface of biomedical devices used in healthcare. Having biomedical devices naturally able to resist bacterial attachment is considered the preferred solution. More and more bacterial species are becoming immune or less susceptible to therapies currently available, with P.aeruginosa, S.aureus and uropathogenic E.coli being among the most frequent culprits.
Several polymers with biological activity are known to be able to form nanoaggregates when copolymerised with hydrophilic species to form polymeric surfactants, which find use for a multitude of applications. The aggregation behaviour is driven by the Critical Micellar Concentration/Critical Aggregation Concentration, which is a function of the size and chemical nature of the hydrophobic part of the surfactant.
Material ability to adsorb viral particles can be modulated too. Environmental factors can affect the adsorption behaviour of SARS-CoV-2 viral particles onto a variety of surfaces. Positively charged surfaces are known to promote irreversible adhesion, and subsequent mechanical disruption of many types of viruses, including SARS-CoV-2: this is thought to be due to the negative charge that the external viral envelope exhibits at physiological pH range.
Aims and Objectives: 1) To apply linear, binary classification to predict whether P.aeruginosa, S.areus and uropathogenic E.coli will attach on a library of polyacrylates; to provide a mechanistic interpretation of the models obtained; to perform a virtual screening of an external library of monomers and propose new, promising polyacrylates with improved properties. (Chapter 3).
2) To predict the CAC of a set of polyacrylate-co-mPEGMA surfactants; to provide an interpretation for the descriptors used by the model (Chapter 4).
3) To predict the adsorption behaviour of SARS-CoV-2 viral particles on a set of poly-(acrylates/acrylamides); to give a mechanistic explanation for the model; to suggest new polymers with greater affinity for the viral particles via virtual screening of a database of commercially available compounds (Chapter 5).
Methods: Logistic regression and linear regression models were used to perform quantitative modelling, using molecular descriptors and mass-spectroscopy data, where applicable. For classification tasks, filters (collinearity, diversity) and wrappers (sequential forward selection) were used as feature selection methods. For regression, filters and a combinatorial approach were used as selection methods. Particular attention was paid to the interpretation of the results, with a polynomial regression between the features used by the models and a pool of easy and interpretable descriptors. The contribution of each simple descriptor to the final models was also investigated, and a virtual screening of an external monomer library was carried out where applicable.
Findings: 1) Hydrogen bond donors or acceptors are able to contrast attachment of P.aeruginosa and E.coli, but seem to promote attachment of S.aureus. Lipophilicity and molecular rigidity of repeated units are the most important factors contributing to the attachment resistance against all three bacterial species. The models generated using exclusively molecular descriptors were used for a virtual screening to identify new polymers with improved anti-attachment properties (Chapter 3)
2) The model confirms the correlation between self-aggregation behaviour and hydrophobicity and alkyl chain length of the hydrophobic monomer, as widely reported in the literature. A larger dataset would help corroborate such findings and would allow to perform a virtual screening with an even more robust model (Chapter 4).
3) Results suggest hydrophobic, nitrogen-bearing monomers are more likely to be proadhesion. A possible explanation is that once protonation at physiological pH occurs, the positive charge can be stabilised by the tertiary amine, especially when close to an aromatic cycle, via inductive effect and resonance. This leads to a strong adhesion of the negatively charged viral envelope protein and to its compression and mechanical disruption. A virtual screening identified new potential monomers with high proadhesion behaviour that will be eventually tested (Chapter 5).
| Item Type: |
Thesis (University of Nottingham only)
(PhD)
|
| Supervisors: |
Williams, Philip
Laughton, Charles
Alexander, Morgan |
| Keywords: |
Machine learning, biomaterials, bacterial attachment, viral adsorption, SARS-Cov-2, COVID-19, surfactants |
| Subjects: |
Q Science > QA Mathematics > QA 75 Electronic computers. Computer science
Q Science > QD Chemistry
R Medicine > R Medicine (General) |
| Faculties/Schools: |
UK Campuses > Faculty of Science > School of Pharmacy |
| Item ID: |
73104 |
| Depositing User: |
Contreas, Leonardo
|
| Date Deposited: |
22 Jul 2023 04:40 |
| Last Modified: |
22 Jul 2023 04:40 |
| URI: |
https://eprints.nottingham.ac.uk/id/eprint/73104 |
Actions (Archive Staff Only)
 |
Edit View |
Tools
Tools
![[img]](/style/images/fileicons/application_pdf.png)