As regular readers will know, I am not yet of the opinion that artificial intelligence has led to any new drugs – yet. But I’m also of the opinion that there is no reason that it can’t, or that it shouldn’t, given the advances in the field. Now, AI (and machine learning as well) are not the amazing back doors to biopharma riches that you might imagine from reading some press releases. Applying these things usefully to drug discovery is actually quite difficult, and what’s becoming more clear is that in many cases we’re not going to be able to rely on the existing data as much as we might have hoped. Rather, we might have to generate new piles of the stuff under cleaner and more controlled conditions to give the software a more healthy diet, one that will generate conclusions that are more reliable. That’s going to take plenty of money and effort.

But it could easily be worth it. The difficulties of drug discovery and development are so severe (and the rewards for success so great) that anything that increases its chances could be of great value. I have a talk I give on this subject that comes across a good deal more pessimistic, but that’s because I’m sometimes brought in as the Designated Bucket of Cold Water, a role that my son and daughter will tell you that I was born for. There I point out that the most lucrative problems to solve are also the most intractable, such as (just to pick a few) “What are the chances that this drug candidate will actually benefit the patients it’s being developed for?”, and the related questions of “What are the chances for unexpected toxicity that could derail my clinical trials?” and the overarching “Should I be working on this target at all, and if not, what’s a better one? How about the whole disease, while we’re at it?” These are of great importance and they’re very, very hard to answer, and I don’t think that AI is going to provide answers to them any time soon – largely because the difficulty is lack of knowledge in the first place. You can’t expect the software to take a bunch of used pizza boxes and find a route to convert them into gold jewelry, and unfortunately a lot of our knowledge (when you look at it on the absolute scale) is at the used pizza box stage.

That said, there are (somewhat) easier problems that I would expect AI to help us with before we get around to those whoppers, such as “What molecule should I make next? What’s the best way to make it? Which of these molecules on this list over here has the best chance of being a potent ligand?” God knows those are hard enough, but they do provide a bit more traction for the algorithmic tires. And now we finally get to the subject of today’s blog, after that excruciating windup. It’s a question that has come up here before, and it’s getting more and more attention: if you do use an AI system to come up with something useful and novel (ie patentable), what do you do about inventorship? In that earlier post just linked, I didn’t have too many worries, because I thought well, we already do virtual screening and molecular modeling and docking, and no one thinks that those programs are “inventors” if something gets discovered, not any more than an NMR machine or a cell sorter (or their software) is an inventor. I still believe that, but I also have to admit that there are some legal issues that this viewpoint does not quite resolve.

This comment at Nature is a good look at the subject. The authors do start out with what I think is an unrealistic doomsday idea: what if courts decide that AI-assisted inventions aren’t patentable at all? I don’t see that happening, but the article is correct that while past challenges to the patent system turned around what an invention is (such as the disputes about patenting gene sequences), this one is about what an inventor might be. The most troubling line of argument that I see here is the one about the “skilled in the art” standard, for obviousness. Some big IP decisions have turned on this issue: does this new application detail something that is obvious to one skilled in the art, and is therefore not patentable? Was this already-patented invention over here in fact obvious to one skilled in the art at the time, and therefore its patent is invalid? You can just see the billing hours racking up, like the meter on a gas pump.

But here’s the problem: the reason that we’re putting money into AI/ML systems is for them to tell us things that are not obvious to us humans. Think of the moves that emerged in Go when AlphaGo (and AlphaGoZero) started rooting through the rules of the game looking for useful strategies. They weren’t things that human Go players had ever really tried, not in centuries of messing around with the game. No one has been thinking about patenting a Go gambit, but what if your AI system spits out a structure for a potential lead compound that seems weird and unlikely, not something you’d have made yourself and not something that anyone else seems to have made, either, but works anyway? What does “obvious” mean at that point? Obvious to a human? Obvious to an AI system? And if the latter, what the hell does that mean?

I should mention here that I very much do not think we are at this point yet. To me, the AI/ML compounds (and even drug candidates) that have been touted seem like perfectly normal stuff, unexciting extensions or rearrangements of what was already known. Some of them probably skate pretty close to an obviousness problem already, but that certainly doesn’t have to be the case for all time. I don’t see any reason why we can’t end up in “Dang, you’re telling me that works?” territory at some point, and that’s when we’re going to have to wrestle with this problem of patentability. As the article notes, right now, courts are not currently in a mood to deal with this. There are patent applications out there from an AI firm called Imagination Engines that explicitly name the AI software as the inventor of (for example) a new type of food container, and these have now been rejected by the UK, the US, Germany, the European Patent Office as a whole, South Korea, Taiwan, New Zealand, and Australia, on the ground that inventors have to be human. Apparently the German court was willing to have a human inventor listed as the person who prompted the AI system to create the inventions, but that’s about as far as things have gone. My own take is that Imagination Engines is pushing this as a means of publicity (as with their talk of having an “arguably sentient machine”, which I don’t buy, either), but it doesn’t mean that these issues are going to go away.

The authors here call for a comprehensive approach to all this, with a new “AI-IP” framework. This might incorporate all sorts of changes – different patent terms, different definitions of inventorship and patentability, requirements for the inventors of the AI system itself (and/or those who modified it for a specific use) to be listed on the patent, and so on. They’re also hoping for a global dispute resolution system, which would really be something to see (as it is now, you fight out patents country by country, at great expense). My guess is that even if this idea gains support that the centrall dispute system will be one of the first things to be jettisoned  –  it’ll be hard enough just to get an AI framework together without that change on top of it, is my guess. Europe is rolling out a unified patent court, which is truly a step in the right direction, but getting totally different parts of the world to sign on to something like this would be quite a lift.

Over the next ten or twenty years it looks like we’re going to have to wrestle with these issues, one way or another. It might well be a good idea to (as the authors of this new commentary recommend) get out in front of the issues and start thinking about them now. For better or worse, though, I think the more likely outcome is that this gets haggled out in fits and starts and dead ends and reversals in jurisdiction after jurisdiction. What will electrify things is the first really lucrative out-of-the-blue invention from an AI system, and that just hasn’t happened yet. But you’d have to think it will. . .right?