Deciphering the relationship between AMR phenotype and genotype utilising machine learning models.

We hope to then expand this to predict genes important to MDR. In Chapter 2, we investigate how well AMR gene identification tools can predict AMR phenotype and how the AMR databases vary. Several of the techniques were applied to investigate the resistome of hospital pipelines and the spread of AMR genes, which are now published in two papers of the Journal of Hospital Infection [1, 2].

We then used biologically interpretable machine learning models in Chapter 3 to try to improve AMR phenotype prediction accuracy and further understand what is involved in AMR phenotype mechanisms. The work completed in Chapter 3 is now published in the Journal of Life Science Alliance, DOI: 10.26508/lsa.202302420 [3]. We tried to make more general models which may have higher prediction accuracy for unseen data in Chapter 4.

We opted to build neural network models, which may not offer the same biological insight as the decision tree models but may have a higher accuracy. In Chapter 5 we use the models developed in Chapters 3 and 4 to develop a user-friendly tool and validate the models using an external dataset of 451 Pseudomonas aeruginosa genomes. In Chapter 6, we did two pangenome analyses on P. aeruginosa and Escherichia coli genomes and then investigated the distribution of AMR genes and eggNOG gene families across the pangenomes, respectively.

The pangenomes were also analysed using Coinfinder, which identifies genes within the pangenome that always co-occur together (associated genes) and genes which are never found together (disassociated genes). Machine learning models were built to predict MDR (for pairs of antibiotics) using decision trees and CNN models. We then investigated how MDR is related to the pangenome by investigating the decision tree models further.

Using a combination of machine learning and pangenomic techniques, we explored the routes to resistance and whether different routes can be mutually exclusive.<br><br>We found that AMR gene identification tools may not accurately predict the AMR phenotype. The AMR phenotype is very complex, involving multiple gene interactions (including gene absence) and sometimes can be species-specific. For the cross-validation of the training data and external validation, convolutional neural networks were found to have the highest accuracy across all models for AMR phenotype prediction.

Our pangenome study revealed AMR genes were present in the core genome of P. aeruginosa and E. coli. Some of these AMR genes conferred resistance to multiple antibiotics, suggesting that MDR genes are present in the core genome of both species. Combining the pangenomics and machine learning methods revealed that several genes associated with the MDR phenotypes in the same decision tree were disassociated in the Coinfinder network, suggesting possible routes to MDR may be mutually exclusive.