Pseudomonas aeruginosa transcriptome during human infection

Abstract

Laboratory experiments have uncovered many basic aspects of bacterial physiology and behavior. After the past century of mostly in vitro experiments, we now have detailed knowledge of bacterial behavior in standard laboratory conditions, but only a superficial understanding of bacterial functions and behaviors during human infection. It is well-known that the growth and behavior of bacteria are largely dictated by their environment, but how bacterial physiology differs in laboratory models compared with human infections is not known. To address this question, we compared the transcriptome of Pseudomonas aeruginosa during human infection to that of P. aeruginosa in a variety of laboratory conditions. Several pathways, including the bacterium's primary quorum sensing system, had significantly lower expression in human infections than in many laboratory conditions. On the other hand, multiple genes known to confer antibiotic resistance had substantially higher expression in human infection than in laboratory conditions, potentially explaining why antibiotic resistance assays in the clinical laboratory frequently underestimate resistance in patients. Using a standard machine learning technique known as support vector machines, we identified a set of genes whose expression reliably distinguished in vitro conditions from human infections. Finally, we used these support vector machines with binary classification to force P. aeruginosa mouse infection transcriptomes to be classified as human or in vitro. Determining what differentiates our current models from clinical infections is important to better understand bacterial infections and will be necessary to create model systems that more accurately capture the biology of infection.

Keywords: Pseudomonas aeruginosa; chronic wounds; cystic fibrosis; human transcriptome; machine learning.

Fig. 1.

PCA of P. aeruginosa RNA-seq results. This includes human samples listed in Table 1, as well as mouse and in vitro experiments from our laboratory and others (Dataset S1). (A) Analysis was performed with 1,707 genes that were expressed (i.e., contained at least 1 RNA-seq read) in all samples. (B) To include two chronic wound samples from Denmark with low-read coverage (labeled in figure), analysis was performed with 761 genes that were expressed in all samples.

Fig. 2.

Gene categories that are significantly different in in vitro transcriptomes compared with human transcriptomes. Analysis was performed with 1,707 genes that were expressed (i.e., contained at least 1 RNA-seq read) in all samples. Categories and enrichment calculation were obtained from the BioCyc database using Grossmann’s parent–child-union variation of the Fisher’s exact test with a P value cut-off of 0.05 (48). Plotted are the genes with a P-adjusted value of <0.05. “Considered” genes indicates the number of genes within that category that were analyzed (i.e., included in the 1,707 genes).

Fig. 3.

The expression of genes in the las core QS regulon in human samples compared with in vitro samples (18). (A) Average relative expression of 42 core QS genes in vitro and in humans. Fold-change in expression was calculated as a ratio of the geometric mean of relative expressions for each gene among the in vitro transcriptomes to the geometric mean of relative expressions of each gene among the human transcriptomes. Values above 0 indicate higher gene expression in vitro. Samples with fewer than three reads for a gene were removed from the analysis. (B) Average relative expression of 42 core QS genes within in vitro biofilm/aggregate transcriptomes or in vitro planktonic transcriptomes, compared with human transcriptomes. Fold-change in expression of each gene was calculated as above. Bars represent SEM.