Machine learning identification of adenine methylated PAM sequences inhibitory to SaCas9 activity

Cas9 nucleases can be used with single guide RNAs (sgRNAs) as antimicrobials and genome engineering tools in bacteria, yet applications are hindered by an incomplete understanding of Cas9-target interactions. Here, we generate large-scale SaCas9/sgRNA in vivo bacterial activity datasets and train a machine learning model (crisprHAL) to predict SaCas9 activity. The highest predictive performance was found when downstream sequence flanking the canonical NNGRRN PAM motif at positions [+1] and [+2] was included in model training, correlating with high in vivo activity on sites that included T-rich di-nucleotides in the [+1] and [+2] flanking positions. Strikingly, model predictions and experimentally determined activity in pooled sgRNA experiments in Escherichia coli and Citrobacter rodentium showed significantly reduced SaCas9 activity at sites with 5-NNGGAT[C]-3 PAM sequences. Adenine methylation (*A) at 5-NNGG*AT[C]-3 PAMs in-vitro reduced SaCas9 activity approximately 10 fold, whereas cytosine methylation (5-NNGGAT[*C]-3) had no impact on activity. Our results show that a general purpose machine learning architecture can provide biologically relevant insights into SaCas9-PAM interactions that can better inform activity predictions. Avoidance of adenine methylated PAM sites by SaCas9 may be a mechanism of self versus non-self discrimination or reflect an evolutionary adaptation to counter methylation as an anti-restriction strategy by phage or plasmids.