Investigating the Transcriptional Responses of Escherichia Coli under Chemical Challenge

Abstract

Bacteria live in diverse and dynamic environments that necessitate adaptation to external stimuli. Here, we probe Escherichia coli K-12 with a wide array of chemical stressors spanning several drug classes, and gauge the transcriptional responses using a promoter-GFP fusion library. Assayed with PFIboxes, the output fluorescence images are temporally resolved and data rich. When quantified as gene expression, these transcriptional responses are seemingly unique to each molecule tested, yet clear differences exist between drug classes. Promoters showing large magnitudes of differential regulation in selective conditions, such as DNA damage, oxidative stress, and cell wall stress, can be used as diagnostic reporters for primary screening assays. The transcriptional signatures generated by these experiments were used to train a 10-layer convolutional neural network in Keras for mechanism of action (MOA) predictions. This model was used to predict the mechanism of action of cefmetazole, polymyxin B, as well as cinoxacin, a compound excluded from the training set. The model predicted the identity of cefmetazole and polymyxin B with 95-100% accuracy. Cinoxacin was predicted to be enoxacin, another fluoroquinolone antibiotic, with ~80% confidence, illustrating the power of prediction MOA of unknown molecules with a large training dataset. This deep learning model predicted the MOA of an unknown compound, MAC168425, as trimethoprim. Further characterization of the compound suggests that its inhibitory activity is involved in folate-binding and utilization, and glycine cleavage. In all, this work illustrates that microbial reporter arrays generate unique patterns which can be used to make hypotheses on the MOA of unknown molecules.