One of the tricky parts about trying to find new antibiotics is that many screening modes will just discover things that have been discovered before. You’d think that if you’re looking for “bug killers” that you could just run through the compound collection looking for stuff that, well, kills bugs, but the problem is that everyone has done that before. For decades. So unless your compound collection is pretty unusual, odds are that you’re going to hit things that others have already hit. (Odds are as well that you’re going to hit things that will kill off mammalian cells at the same rate as bacterial ones, too, but that’s another counterscreen!) There may well be some interesting compounds in there, but how do you know which ones those are?
High on the order of business, then, is figuring out how your hits are working so you can prioritize the more interesting (new) stuff over the well-trodden pathways. That is not so easy. This new paper, though, suggests a high-throughput way to sort such things out in the case of Mycobacteria. The authors are a multinational team from the ETH-Zürich, the Hungarian Academy of Sciences, the Institut Pasteur, the Swiss Tropical and Public Health Institute, the University of Basel, and GSK’s site in Spain, and there are certainly plenty of people on that list who know anti-infectives research. What they’re illustrating is a comprehensive metabolic profiling technique that allows potential antibiotic compounds of unknown mechanism to be narrowed down.
The first step was to build a reference set of metabolic data in M. smegmatis, using 62 known compounds that cover 17 different mechanisms of action. The bacterial broth medium was extracted directly at various time points and treatment concentrations, and shoveled right into a mass spec instrument. This generated a pretty ferocious pile of data, which was cleaned up by eliminating drug-specific ions and the like, and standardized across different plates and extractions to try to get a workable baseline. About 5% of the detected metabolites in the broths showed real changes on drug treatment, in the end. The time courses of these changes made sense – drug targeting metabolic processes directly showed up quickly, while things like protein synthesis targets took longer to alter the profiles.
Correlating the pathways, enzymes involved, and mechanisms of action (no small feat) suggested that MoAs could indeed be binned by looking at the metabolite profiles (even if they weren’t directly targeting a metabolic pathway). Some of them were firmer or less variable across different compounds (such as the quinolones) than others (cell wall synthesis inhibitors), but they all did end up with distinguishable signatures. The next test was to assay a list of 212 compounds that had been identified by the GSK team as active against Mycobacteria but whose mechanisms were unknown.
This picked out a number of inhibitors of DNA replication and folate biosynthesis (both of which mechanisms are broadly covered by existing antibiotics, although specific enzymes in these pathways can presumably still be new targets). One compound, GSK2623870A, appears to work in the fatty acid biosynthesis area (which Mycobacteria spend a good amount of metabolic energy on), and a total of six similar compounds were identified that nonetheless still showed somewhat varying metabolome profiles.
The belief is that this sort of profiling should translate to other varieties of bacteria – if it does, it also has ten-to-hundredfold higher throughput than the existing profiling methods. No libraries of mutant bacteria are needed (a known way of sorting such questions out). The culture conditions used for this sort of assay, though, will probably have to be investigated further, since there are surely some genes (and whole pathways) that are activated under some conditions and not others. Finding protocols that mimic the natural infection process as closely as possible will improve the chances of success. The authors propose moving on to those different bacterial families and expanding the list of reference compounds, since this study seems to have worked out, and trying to turn this into a general mechanism-clarifying platform for antibiotic discovery.
That’s going to generate mighty piles of data, but (as is often the case), this is a feature, not a bug. As the authors mention, it will be interesting to turn various machine-learning algorithms loose on these data sets to see what they can make of it. If things work out, you could also end up predicting new combination therapies as well, if you can match up complementary pathways, which could open the door to using more bacterial-nonlethal compounds that never would have been picked up in solo testing. Here’s hoping it works!