I’ve been in early-stage drug discovery my whole career. Am I a scientist, or not? That seems like a silly question, or one that should be answered “You’d better be,” but not everything I do falls under the heading of “science.” Some of it’s more like art, and some of it verges on voodoo. I’ve managed, just like my friends on Wall Street, to work myself into a profession where brains are necessary but nowhere near sufficient. But the one thing it certainly isn’t is a well-worked-out, flow-charted process.

Things would make a lot more sense if that were so, for sure, and there are a lot of researchers who occasionally long for things to be a bit more sensible. It would definitely be a lot easier to manage if that were so, and the number of managers who would like a more defined, predictable approach to discovery is legion. But the harder you chase that prize, the faster it runs and the better it hides. I will illustrate, with reference to two figures that are so often linked together, Christopher Lipinski and Monty Hall.

Go into any medicinal chemistry department and ask them what they think of Lipinski’s “Rule of Five” (Ro5) criteria, and odds are you will start an argument From the outside, you wouldn’t think that these would be too controversial — they were developed by looking at the physical properties of approved drugs, after all. And the great majority of drugs on the market really do have molecular weights below 500, fewer than five hydrogen-bond donors and acceptors, and all the rest. So what’s the problem?

The problem is that when you plot these variables out, the distributions tail out on both sides for quite a way. So everyone can always think of exceptions, great big profitable exceptions. A lot of people in this business are getting paid off the profits of molecules that violate one or more of the rules. To be fair, Lipinski himself didn’t intend for his observations to become dogma, but the spirit of, “Whatever can be measured can be managed,” is a hard one to overcome. So in many shops, people are strongly encouraged to submit only compounds that fit within the boundaries, and the screening collections are curated with them in mind.

On one level, I can’t disagree with those actions at all. If you handed me a collection of 100,000 compounds that meet the Rule of Five, and another 100,000 where every one of them violates it in some way, I would definitely choose to screen the first group, if I could only screen one. But if I could screen both, I certainly would, especially against a hard target. Leads are where you find them, and in some areas you have to look far and wide. Protein-protein interaction hits, when there are any, tend to violate Ro5 criteria, and marketed antibiotics don’t even seem to have heard of them. Better to start outside the rules, in such cases, then not to start at all.

Now, here’s where I want to reference the Monty Hall paradox (you were probably wondering when he was going to make his appearance). If you haven’t encountered that one, it goes like this: I have three identical-seeming doors to show you. Behind one of them is a new car, and behind the other two are goats. We will assume, for the sake of argument, that you are more interested in winning the car (I stipulate this because when I once explained this to my daughter, she told me that she would rather have a pet goat). At any rate, those are the rules; I have two goats and a car, guaranteed. You get to choose one of the doors at random. Once you’ve done that, the rule is that I have to open one of the other two doors to reveal one of the goats. Now we’re down to two closed doors — the one you chose at first, and the one that I didn’t open. And we’re also down to one goat and one car. At this point, I give you the choice of sticking with your original choice, or switching to the other door. Here’s the question: should you switch? Should you stay? Or does it even matter?

The majority of people, their first time through this, look at the situation and say, “Well, there are two doors. I chose at random, there’s one goat and one car, so it’s 50-50. Doesn’t matter if I switch or not.” But that’s completely wrong; as it turns out, you should switch. Think about it this way: if your first random choice was actually the car, then you should stay with that door all the way. If your first random choice turned out to be a goat, though, then if you switch at the two-door stage, you are guaranteed to get the car, because it has to be behind that other door. You had a two out of three chance of picking a goat at first, and therefore, if you switch doors later, you have a two out of three chance of getting the car.

It works out that way because the goat-revealing step alters the amount of information that you have. The situation has changed, and you have adjusted based on what you’ve learned. The problem trips people up because our brains aren’t so great at conditional probabilities, and we don’t even realize that we’ve been given crucial data when it happens. And so it is in drug discovery: there are moments that are the equivalent of the opening of that door to reveal a goat, and it’s our job to realize what we’ve been shown and what it means.

The reason many Ro5 violators don’t work out is their pharmacokinetics — they don’t get absorbed or they get cleared too quickly. If you can get PK data, the whole landscape changes. A funny-looking compound with good exposure data is not just a funny-looking compound anymore. It’s already cleared a high hurdle, and you may be looking at something that’s a lot closer to a car than to a goat. For a difficult or high-value target, you should take data like this very seriously. (And besides, what are you working on, under the current conditions, other than high-value targets, anyway?)

The problem with many screening cascades is that they’re based, naturally enough, on weeding things out at every step. Everything that makes this cutoff goes on; everything that doesn’t drops out. If you have plenty to start with, and have something left over afterwards, that’s fine. But you may not have either one with some targets, and that means that you’ll have to be more flexible. Just to be clear, that doesn’t mean being so flexible that you twist yourself into following junk in the hopes that everything will work out in the end. But it does mean following the data, which means that you have to get the data in the first place, which probably means giving some less-likely compounds a second chance. It makes things messy, and it even makes the whole process a bit less intellectually appealing, if you like things orderly. But would you rather have everything in its proper place, would you rather find a drug?