Some background: medicinal chemists are responsible for identifying drug candidates, usually by screening large (10^6) libraries of combinatorially generated molecules. Some of these hits turn out to be biologically active, and then it's up to the medicinal chemists to decide whether these hits are false positives or not, and further, to synthesize analogous compounds to see if they can tweak the biological activity of each compound.

It's in this 'synthesizing analogous compounds' step that subjective judgment comes in, with Lipinski's rule of five being the most 'basic' of the heuristics, and with most medicinal chemists adopting more and more complex heuristics. Or,  as this paper shows, perhaps they're just deluding themselves and their true metric is something very simple, and after making their decision, they dress it up with fancy post-hoc rationalizations.

New Comment
5 comments, sorted by Click to highlight new comments since: Today at 10:39 AM

The blog from which that's taken is frequently interesting and amusing. In particular, the author's occasional series of "Things I won't work with" is very nice. [EDIT: Where by "nice" I mean "amusing but intermittently disgusting and/or downright terrifying".]

I've found his blog a very useful source for a long time about systematic problems & biases in science.

"Things I won't work with" is wonderfully humorous! It continues to remind me of John D. Clark's Ignition!, which is a delightful survey of liquid rocket propellant chemistry, in all its hazardous glory circa 1960 (which is really most of it -- there hasn't been much chemistry work in liquids since then). Print copies are hard to find, but Google will probably get you the PDF that floats around.

Geoff Hinton's group recently had some impressive success using deep neural networks in a drug candidate discovery contest sponsored by Merck. This article sheds some light on the room for gains from such programs.

A comment suggests the paper misunderstood importance of the within and between chemist (non-)correlation because the target lists include multiple similar chemicals that everyone agrees should not all be selected.

It's still troubling that the stated decision process seemed to match so poorly with the revealed decision process of each chemist.

New to LessWrong?