Microbes have disproportionate impacts on the macroscopic world. This is in part due to their ability to grow to large groups and cooperatively secrete massive amounts of secondary metabolites that impact their environment. Yet, the conditions enabling secondary metabolism without compromising primary needs remain unclear. Here we investigated the biosynthesis of thamnolipids, a secondary metabolite that Pseudomonas aeruginosa makes to decrease the surface tension of surrounding liquid. Using a combination of genomics, metabolomics, transcriptomics, and mathematical modeling we show that biosynthesis of rhamnolipids from glycerol varies inconsistently across the phylogenetic tree; instead, non-producer lineages are also those worse at reducing the oxidative stress of primary glycerol metabolism. The link to oxidative stress explains the inconsistent distribution across the P. aeruginosa tree, adding a new layer to the regulation of rhamnolipids—a microbial secondary metabolite important for fitness in natural and clinical settings.

Metabolite extraction. All P. aeruginosa strains were grown until the end of exponential phase of growth in glycerol minimal medium. Bacteria was then loaded into 0.25 μm nylon membranes (Millipore) using vacuum, transferred to pre-warmed hard agar plates with the same medium composition and incubated at 37ºC during 2.5 h. The filters were then passed to 35 mm polystyrene dishes (Falcon) with 1 mL of 2:2:1 methanol:acetonitrile:H₂O quenching buffer and incubated there during 15 minutes on dry ice. Cells were removed by scraping and the lysate containing quenching buffer was transferred to 1.5 mL tubes and centrifuged at 16000 rpm for 10 minutes at 4ºC. Supernatant transferred to fresh tubes and stored at -80ºC. 

Metabolomic data preprocessing. The extracts were profiled using liquid-chromatography coupled to mass spectrometry (LC-MS), identifying a total of 92 compounds (Supplementary Fig. 5). Some compounds contained missing values. These missing values in metabolite abundance can be (1) truly missing; (2) present in a sample but its level is below detection limit; (3) present in a sample at a level above the detection limit but missing due to failure of algorithms in data processing. Here we assume that a metabolite with missing values in all three replicates is truly missing in the sample and removed from our analysis (Supplementary Fig. 5). However, if the missing values were only found in one or two replicates, the missing values were imputed by the average of the non-missing values. After that imputation all compounds with missing values were removed (Supplementary Fig. 5).

The peak areas were normalized using Cross-Contribution Compensating Multiple Standard Normalization (CCMN) (80) with NormalizeMets R package (81). This method relies on the use of multiple internal standards. Since LC-MS lacks such internal standards, we used instead a set of metabolites assumed to be constant across all the strains. They were selected with a Kuskal-Wallis test, adjusting the p-value with Benjamini-Hochberg method. The ones with a p-value above 0.05 were considered constant (pyruvate, methylglyoxal, (S)-2-Acetolactate, Tyramine, D-Glucose, (S)-Lactate, N-acetyl-L-glutamate 5-semialdehyde, 4-Aminobutyraldehyde and Glycine), therefore after the normalization step they were removed (indicated in red, Supplementary Fig. 5A).  The processed area peaks for all metabolites are included in Supplementary Table 5.