Where to even begin? A mold flew in through the window and, presto, penicillin. Hey, then, why not screen all the compounds we have on the shelves in those bacterial/assay/etc dishes? And so began screening for drug candidates.
But bigger quickly became assumed to be better: The more compounds screened, the better chance of finding an active one. And the race was on to have the biggest. Somewhere, some rational person whispered, “Wait, some compounds are just generically better than others, so isn’t bigger only better if there are more clean, diverse compounds? Did we look to see if that’s true of our library?” But like Michael Burry’s puzzlement over the lack of reaction to his postulate that not all home-mortgage-backed bonds were the same, everyone ignored the elephant in the room.
But in subprime markets, as in drug discovery, the chickens eventually come home to roost. First came just a horrific long term track record for high throughput screening leading to FDA approved drugs. Then came the PAIN(s), then the pivot to biologics or chemogenomics/chemical genetics phenotype screens.
That last one should be reassuring, but it turns out that saying something is true doesn’t mean that it is. The assumption of the chemogenomics library is that each compound is relatively target-specific. But, of course, this assumption is very likely to be laughably wrong. So the next best thing might be just to at least make sure that the ensemble of targets for each compound in a library overlaps minimally with those for the other compounds in the library. Indeed, with such an approach, one might imagine assembling the library entirely based on this minimal-overlap metric.
Hard work, but fortunately, Moret et al took the first step for the field. Restricting their analysis to the heavily annotated library of kinase inhibitors, they developed a polypharmacology vector for each compound and analyzed several commonly used chemical libraries. They found, as expected, that someone should have looked under the hood long ago: the libraries were terribly biased and unbalanced. Manipulating the same vectors, they were able to assemble a library that minimized overlap of targets. Along with this library, their methods serve as useful tools for customizing libraries in general.
Their work adds to the growing collection of reports that are starting to get a handle on polypharmacology of small molecule drugs/drug candidates. Genecentrix’ remains the only user-friendly, off the shelf tool for obtaining the polypharmacology ensemble of small molecule drugs, giving the predicted tissue activity of the drugs based on those targets as well.