Kyle Millar Thesis - A Random Forest Approach to Understanding CRISPR-Cas Associations in Bacteria

CRISPR-Cas systems are a crucial and intriguing defence mechanism foundin bacteria and archaea. This defence mechanism is able to adapt and defendagainst attacks from Mobile Genetic Elements (MGEs). This mechanism also hasmany uses outside of genomic defence. For example, certain types of CRISPRCas proteins allow for modification of eukaryotic genomes in vivo. Currently,there are applications and algorithms capable of finding CRISPR-Cas types andthe associated arrays, however, the idea of predicting whether a genome mightcontain a CRISPR-Cas locus, based purely on the background genome contentoffers a faster query time. To test whether the presence of CRISPR-Cas systemscould be predicted from the background genome, a Random Forest algorithmwas employed using a large data set - a bacterial pangenome containing 9,689genomes. To annotate this pangenome with ’CRISPR’ identifiers, the annotationtool Bakta was used, allowing for the use of custom scripts to find the relevantinformation needed from the annotated genomes. The algorithm was shown tohave an accuracy of 0.89, and an AUC-ROC score of 0.96. These results imply astrong ability to classify the predictions correctly, based on background genomecontent. The algorithm calculated the ’feature importance’ of all genes that werepresent in the pangenome; the gene of highest importance was ’pbp4b’ followedclosely by ’csy3’ (a positive control variable). The ten genes that had the highestfeature importance all had a statistically significant association with CRISPR-Cassystems when evaluated using chi-squared tests. The algorithm was capableof predicting CRISPR-Cas systems in γ-proteobacteria and offers potential forresearch candidates when investigating CRISPR-Cas associations. This approachcould be used to predict CRISPR-Cas more broadly across prokaryotic life, upondata availability.