The idea of phenotypic screening is not new, but once you bring up the topic, you find that different people have different ideas about what a phenotypic screen really is. As has been discussed around here before, not everyone buys into the concept of a cell-based screen as “phenotypic”, but I’m willing to believe in the category. But cells aren’t animals. And “animal” itself is a concept that needs some clarification as well. Fruit flies and nematodes are noble creatures, in their way, and have provided many an insight into living systems, but for human drug discovery you need to move a bit closer to humans.
But doing such a screen in a whole vertebrate animal is not easy to realize. You’re going to need thousands of them, many thousands, in the same way that a screen against a cloned protein target uses many thousands of wells. So that rules out anything larger than. . .what? Xenopus frogs and zebrafish? Doing those on scale is pretty nontrivial, too. I remember back at the Wonder Drug Factory, when I was part of a task force (a good ten or twelve years ago now) trying to think of new ways to generate projects and leads, and one of the ideas we kicked around was the feasibility of just that sort of screening. “So which building are we going to turn into an aquarium?” was one response.
I wondered about this just a year or so ago, now a new paper has come along detailing the largest zebrafish screen I have yet heard of. The authors are looking at diabetes pathways, and have generated transgenic fish embryos with fluorescent reporter genes tied to various pancreatic cells – one wavelength specifically for beta-cell growth, and another for general pancreatic cell differentiation (as measured by delta-cells). The appearance and timing of these tissues in zebrafish development is pretty well worked out, so compounds that enhance or disrupt aspects of it will be apparent on inspection.
On inspection . . .there’s the rub. How do you “inspect” the thousands and thousands of zebrafish embryos that you’re going to need. This paper details an automated system that seems to do a pretty good job (but I’m mindful that people have been working on automated imaging tools for such things for many years now – it’s not easy, and you run a significant risk of generating, at significant cost, a terabyte of so of noise). These authors, actually, had previously tried a less-automated pancreatic growth assay in zebrafish, and seem to have been spending some time trying to avoid doing it that way again. The result is what they’re calling “the first truly high-throughput whole-organism drug screen in a vertebrate model”, and they may well be right about that, although there are a lot of press releases over the years that would appear to disagree.
So, how many compound are we talking about here? Probably not as many as you’re thinking. Their collection was 3348 (mostly approved) drugs, but you need 16 zebrafish embryos per concentration (six levels) across the dose-response, which added up, in the whole screen, to over 500,000 fish embryos. That gives you one drug per 96-well plate. This data-point buildup happens, naturally, because the data are noisy in whole animals, and if you don’t work on this scale you run the risk of that terabytes-of-noise effect, especially in what is, in the end, a (sorry) fishing expedition. (And if this makes you wonder if many rodent studies are underpowered, you may be onto something there, although keep in mind that many of these are narrowed down to very specific readouts).
225 compounds showed signs of activity at first, but 48 of these were false positives (fluorescent interference or general effects on embryo viability). 23 of the remaining compounds really looked as if they were really affecting islet cell formation, and 23 more might have been, so these compounds were taken into orthogonal secondary assays. 11 of the compounds (“Lead I”) from the best set also performed in an endocrine reporter assay, but they also went back to look for compounds that might have been increasing the beta-cell mass without such effects, and 15 of these (“Lead II”) were identified by further screening.
One possible commonality in the Lead I set was an affect on NF-kB function, and in the Lead II set, serotinergic activity looked like a possible common trait. Serotinergic compounds (like fluoxetine or serotonin itself) that weren’t in the original screening set also showed the desired effects on beta-cells, which leads some weight to the idea. Another recent (although smaller and less automated) zebrafish screen looking for beta-cell effects also identified serotonin as a hit, interestingly, and the two screens seem to generally overlap in activity classes, which must be reassuring.
Overall, I have to say that I’m impressed by efforts like these. There are still plenty of chances to go off the rails, but this screen seems to have been designed to take as many of these into account as possible. The next question is how many other screens of this sort are achievable – that is, how many different types of tissue can be usefully labeled? And how many non-embryonic-development endpoints can it be extended to? That zebrafish aquarium building may have merely been ahead of its time. . .