We spend a lot of time in drug discovery thinking about ratios. As we accumulate data about our compounds, we start ranking them by how selective they are - "This one's 10x versus the other receptor subtype and that one's 50x," you'll hear someone say, or "We've got to get compounds at least 100-fold over that other enzyme or side effects are going to kill us." Generally you have several secondary assays that the compounds have to jump through along the way, and the ratios are what everyone looks at.

And when compounds start to get dosed in animals, you try to look for cutoffs that can tell you which compounds are worth trying in longer assays. Maybe they only work, for example, when the ratio of peak blood concentration, Cmax (or time-averaged total exposure, AUC) to the binding potency is 100x or more. (Other things being equal, that means you could get the desired effect with a really potent compound that doesn't get into the system all that well, or a weaker compound that hangs around a long time.)

So, how good are these numbers? There's the problem - not as good as we tend to think. Even experienced medicinal chemists can get caught over-interpreting data when it's expressed in ratio form. The problem is, on a graph we all expect to see error bars (and we get pretty antsy if they aren't there - it means someone didn't run the experiment enough times, or they're sloppy about making their graphs, or they're trying to pull a fast one.) And for single data points from an assay, we try to remember the variability - looking at the various runs that went into the number you see is always recommended.

But when things get expressed as ratios, all that disappears. We throw around "40-fold" as if it's different from "20-fold," and it takes a conscious effort to remember that it almost certainly isn't. The variability of biological assays would completely curl the hair of a physicist or physical chemist - at times it curls ours in med-chem, and we're supposed to be used to it. Plus or minus 100% is considered a nice, tight assay for many systems - really, it is. They get worse as the system gets more realistic, too - cloned proteins are usually tighter than isolated ones, which are invariably tighter than cell assays, which are certainly tighter than tissue preps, and anything's less variable than some of the animal assays. If you have one of the jumpier ones in the denominator of your ratio, well, prepare to get all sorts of crazy results.

This is why no one, and I mean no one of any competence at all, really trusts "N of 1" data, especially if it's saying something interesting or unusual. If you haven't run the assay again, you're often better off not telling anyone about your numbers until you have. I have seen many people fall flat on their faces because they couldn't resist trumpeting some startling result that later turned out to be junk. It's tough, because we live for startling results. But we die by error bars, and they rule the drug discovery world in the end.