An awful lot of drug discovery comes down (sooner or later) to screening compound collections. This has been true for a long time now, and it doesn’t look like it’s going away, either. So with that in mind, what’s in your collection? Did you buy a bunch of stuff from the vendors to fill it out? If so, your chemical matter is (1) biased towards compounds that are easier to synthesize, (2) probably branches out from a relatively limited number of scaffolds and (3) is duplicated in a number of other screening collections as well. Or did you assemble it from inside, on the other hand, building it up from years of your own med-chem efforts? In that case, then, you might have some interesting stuff in there, but it’s surely been shaped by the kinds of targets and projects you’ve worked on. Company A’s collection might be over-weighted towards kinase inhibitors, while Company B’s has relatively big heaps of PDE inhibitors and attempts at GABA ligands, from that big project that just seemed to go on forever. It’s a safe bet that no one thinks that their screening collection is all that it should be.
Or, as a colleague of mine put it recently when talking to a room full of biologists, “The compound library is not a treasure chest that we scoop jewels out of – it’s more like a huge basement, full of some things that could be useful and other stuff that we really should throw away”. The latter category, everyone would likely agree, includes the sorts of structures that seem to show up as hits in every other screen (rhodanines, polyphenols, etc.). But what about the opposite set of compounds? You’d imagine that some of that should-be-tossed material might be compounds that have been screened a number of times over the years and have never hit in anything at all. At what point do you give up on these things?
A new paper in Nature Chemical Biology takes a look at these, which the authors (from Novartis) are calling the “dark chemical matter” (DCM) of a screening collection. (Here’s a “News and Views” look at the paper as well). They’ve gone back over both the Novartis screening deck and the NIH’s Molecular Libraries collection, and picked out just those compound that have been through a wide variety of screening campaigns without ever lighting up an assay (234 Novartis assays and 429 PubChem assays linked to the NIH compound deck). Only 19 targets overlapped in the two sets, and there were (as you’d imagine) a wide variety of protein- and cell-based screens represented. The cutoff was that molecules had to have been screened in at least 100 assays without ever showing biological activity.
This is a really interesting idea. There do seem to be structural motifs that produce (or, in some cases, over-produce) biological activity, so what’s the other end of the scale like? And how many of these are around? The Novartis set had over 800,000 compounds that made the 100-assay cutoff, and about 14% of these were consistently inactive. The authors ran some simulations to see how many such compounds one would expect if activity were randomly distributed among the original set of compounds, and the numbers came out far lower. These aren’t just random losers, in other words – there’s something about them that makes them persistently less likely to show activity. Similarly, there were about 35,000 compounds in the 100-assay set for both databases (Novartis and NIH), and about 2700 of these had never hit, which is a much, much larger overlap than one could possibly expect by chance. (Statistically, a one-sided Fisher’s exact test gave a P value down around ten to the minus one-hundred-sixty-fifth, which no matter what you think of P values, is hard to argue with).
I would have thought that a good number of these were compounds that just plain didn’t work out in cell assays because of membrane permeability issues, although there certainly are cell assays that depend on surface proteins as well. As it turned out, splitting out the compounds based on biochemical versus cell-based assays shows a large overlap, so my concerns (and theirs, probably) were unfounded. Another thing that occurred to the authors was that these compounds probably hadn’t been QCed in quite a while, since they never hit, but a check of both collections showed that there was no difference in compound purity between active and inactive selections from either one. (I might have expected a slight skew to the other direction, actually – in my experience, nasty decomposed compound wells tend to hit even more often than usual, since there are more compounds in there, and some of them are pretty reactive and colorful).
So what do these compounds look like? Greasy bricks of aromatic rings? Nah, those are the compounds that tend to hit, unfortunately. DCM compounds were, on average, more hydrophilic and had fewer aromatic rings in them. The set of DCM compounds that overlapped between the two compound collections (and thus two sets of assays) were even more hydrophilic and lower molecular-weight than ever. There were some substructures that seemed to be enriched in the dark matter of both collections – diketopiperazines, for example, and some aminoalkyl N-methylpyrazoles, among others. None of these structures, I have to say, look odd at all; I don’t think any medicinal chemist would look at them and say “You know what, you could screen that stuff through a hundred assays and never see a damn thing”. Quite the opposite – they look fine. Clustering them in chemical space didn’t show any obvious “dark nebulae” – all the clusters with DCM compounds in them also have active compounds in them (and sometimes these are very similar structures indeed).
Interestingly, the authors were able to go back and look at these DCM compounds after another 34 assays had been run on them. What they found was that while these compounds did indeed have a lower hit rate than average, but these further assays did cut into the overall numbers. Dark-matter compounds, in other words, can hit, it’s just that they tend not to. 88% of the DCM compounds that did hit something in these further 34 assays, by the way, hit only in one of them (and one of them was a 12 nM primary hit). The same lower-than-usual hit rates were found when subsets of active-as-usual compounds were tested and compared to physiochemically similar DCM compounds in phenotypic-style cell assays (reporter-gene arrays, yeast viability and chemogenomics, and so on).
In fact, what these experiments suggested was that if you screen at (say) ten micromolar in such open-ended cell assays, a rather substantial part of the compound collection is going to react, and probably nonspecifically. DCM compounds, on the other hand, give you better signal-to-noise. So a big take-home from this is when you see a compound that’s been kicking around in the collection for a while suddenly hit for the first time, you should definitely pay attention to it – it might well be one of your better leads, especially if you’re running an assay that otherwise has a high false-positive rate:
From these experiments, we concluded that DCM is indeed less active than other compounds under normal high-throughput assay conditions but is not generally biologically inert. Indeed, our experiments supported the hypothesis that dark matter compounds have the potential to be potent hits with little or no target promiscuity and thus could present an opportunity for identifying new leads. Consequently, we recommend their identification and prioritization in screening libraries and in hit follow-up activities.
The authors also make the point that the assays used for this evaluation were almost entirely directed at mammalian targets – if you’re going for antifungals or antibacterials, you might see a different effect. I’d be interested in seeing someone do that sort of analysis, because there’s been speculation for some time that these targets are possibly selecting for types of chemical matter that screening collections don’t tend to have. I’d wonder the same thing about “undruggable” low-hit-rate mammalian targets (protein-protein interactions, transcription factors and so on). What does the DCM collection (as defined here) think of these?
This paper made me think about screening collections from an angle that I really never had before, and I really appreciate the substantial time and effort that went into it. As far as I know, it’s a unique look at these issues, but I hope to see some follow-up on its concepts from other organizations eventually. Good stuff!