Skip to main content

Drug Development

AI Will Not Threaten Pharma Patents – Not This Way

I’d class this letter to Nature as “interesting but wrong”. Here’s the argument:

. . .A patent is granted only when a compound’s application can be classified as both ‘new’ and ‘invented’. A highly effective compound thrown up by an AI algorithm could indeed be new. Whether it is ‘invented’, however, is debatable. This is because the inventor might be considered as either the algorithm (so not a person) or its programmer.

It could be argued that if there is a connection between the program and the compound’s structure, then it is predictable by experts and so no longer inventive. Or, if the programmer can’t explain how the AI algorithm found the structure, then he or she didn’t invent anything. . .

Actually, the two key thing a patent is granted for are “novelty” and “utility”, but I’ll keep going, because the letter’s line of thought is not completely crazy. When a lead-discovering AI program- we’re stipulating that such a thing will exist for the purpose of argument, although it doesn’t quite yet – suggests a useful compound de novo, that compound does not yet exist. You need physical data to prove that you made what you’re claiming in your patent, and you need data to show that it’s useful for what you’re claiming as well. Without these, compounds in a patent are not exemplified, and are protected (if that’s the word) only by a layer of legal tissue paper. And remember, drug industry patents are generally directed towards chemical matter: we claim these new compounds (novelty) that are good for this (utility). Now when Person A has an idea for a new compound and tells Person B to go make it, and it works, Person B is not the inventor. Person A is. Doing just what someone else told you to do is not an inventive step. So that’s where this letter-writer is coming from: if an AI tells you to make the compound and you make it, you’re not the inventor, and the inventor is. . . ?

But hold on. We’re talking about a fancier version of virtual screening here, and no one thinks that virtual screening algorithms are destroying the patent system. One big reason is that the AI will not provide the final drug structure. I’m sure it will do its best, but hey, so do we, and time and chance happeneth to them all. What it will provide, one hopes, are very interesting lead compounds that will then be taken on by human drug developers. The final compound will look different.

That’s because an AI program will be doing very well indeed to come up with potent structures against a single target protein – it won’t be optimizing for oral bioavailability, toxicity, plasma half-life, avoidance of active metabolites, heterogeneity in liver enzymes among the patient population, interaction with other common drugs in that population, ease of industrial synthesis, development of a useful formulation, avoidance of less useful (but more stable) polymorphic forms, stability on storage. . .and a big pile of other issues that actual drug development has to deal with all the time. No, by the time all that gets worked out, there will have been some inventive steps taken by human inventors.

And that last line quoted above doesn’t quite work, either: when you run a high-throughput screen, can you explain why one particular compound type hit while the others didn’t do as well? Only ex post facto – otherwise, why did you run the screen? We often can’t explain why one structure is so much better than another: but discovering one is indeed an invention. A reductionist view would be that the AI is no more responsible for the invention than was a multichannel pipet or a fluorescent plate reader. All of these are tools, used by an inventive human to discover something.

So a human will use a new, powerful tool to help discover that compounds of a certain type are active against a given target. Other humans will discover further chemical matter of that that will work even better in real cells, real animal, and humans, and that chemical matter will be the subject of a patent. And the pharmaceutical industry’s patent structure will hold up just fine.

43 comments on “AI Will Not Threaten Pharma Patents – Not This Way”

  1. asg says:

    Perhaps you could comment on ‘reducing to practice’. Person A can have a great idea. But no patent. Its person B who reduces to practice. So unless the AI also provides a full experimental…Additionally, the AI will likely provide many cmpds to make, but only some will be made; choosing is part of the inventive process.

    1. Chris Liu says:

      Reduction to practice typically doesn’t bear much impact on patent authorship; that tends to be decided purely by the idea generators. Even in cases where A suggested methyl-structure and B discovered ethyl-structure was the real answer, “one skilled in the art” would recognize that is an obvious substitution and thus does not constitute non-obviousness (if you’ll pardon the double negative). Its gets gray, though, because there are certainly scenarios in which an alteration of the original idea achieves an appropriate level of non-obviousness. Generally it boils down to how permissive the assignee is and how much risk they’re willing to take on their patents, because having the wrong authors on patents does constitute a risk to patent validity.

      A lot of this discussion is moot, though, because no one knows what can and can’t be done until litigation is concluded. I’ll paraphrase what one of my IP attorney colleagues once said to me: I can get you a patent on anything you want, but whether that patent holds up in court is an entirely different question.

  2. Billy says:

    “…it won’t be optimizing for oral bioavailability, toxicity, plasma half-life, avoidance of active metabolites…”

    Not a coding-type so this may sound like a stupid question, but I’m wondering why not? If we’re going to have the AI go to all the trouble of modeling structures that might show efficacy, why can’t we feed it all we know about compound liabilities and let it optimize?

    1. Derek Lowe says:

      Because if we feed it all we know, that’s about 5% of what it needs to know to do anything useful (rough guess). Eventually, yeah. But that’s way, way out there, IMO.

      1. John Wayne says:

        Wow, Derek; 5%? I didn’t know you were so optimistic!

        1. Anonymous Researcher snaw says:

          Agreed, 5% is probably optimistic.

          As for the argument that the AI maybe did at least discover the compound class, it’s not clear to me how getting starting chemical matter by AI is fundamentally different from getting very lucky and finding a fairly potent binder in a High-Throughput screen. Seems to me a lot of people think getting the starting chemical matter is the rate-limiting step in Drug Discovery. In my experience, it’s usually not. Of course, to the extent that we can dial-in various liabilities when we give our AI its marching orders, it might get us better hits than HTS. But I think we’d still need to do a lot of Medchem and Biological Assays.

          1. kjk says:

            If AI is 75% (with it’s amazing strides lately in binding selection) and data is only <5% I see more potential in lab automation than AI as that improves what needs improvement more.

      2. metaphysician says:

        Also, if your AI can do all that stuff effectively true, its probably not AI in the “big data processing” sense anymore, but in the “bow before your robot overlords sense”. There is no problem for the patent system in AI “invention”, if by that point the AI in question is fully a person and thus fully capable of owning and exploiting its own patent.

    2. Molmechanic says:

      Right, Billy! Combining AI with Big Data is a legitimate strategy for discovery of potent and selective compounds with desirable drug-like physical and PK properties and likely to avoid off-target (e.g. hERG) liabilities. Companies like Exscientia are built around this concept.

  3. Theo says:

    A part of your argument that what the AI does will not impact patents is based on the premise that the AI will not come up with the final structure (see “But hold on” paragraph). What happens if the AI through sheer luck proposes the structure which turns out to have the best of the desired characteristics. Does this now make the structure non-patentable?

    1. Chemical Patent Agent says:

      Interesting question: In this specific scenario, if a computer algorithm spits out one compound which it predicts to be the blockbuster drug and it turns out to be right, then I imagine you’d be barred from obtaining a composition of matter patent since you are not the true inventor (which is a requirement of patentability). However, even in this case, you may still be able to ride the polymorph/formulation/method of use patent wave. The more likely scenario is that a computer spits out a genus or general structural motif, and the chemist makes modifications. In this scenario, you’d almost certainly be able to get a composition patent especially if you can show good secondary considerations to rebut an obviousness attack (like unexpected results).

      1. Shalon Wood says:


        There’s precedent that works created by non-humans are non copyrightable. How that applies to patents, well, I’m not a patent lawyer, thank god.

        But I suspect it doesn’t bode well for patenting things that humans didn’t have a significant hand in….

    2. tlp says:

      Are there many single-structure patents by pharma?

      1. Chemical Patent Agent says:

        For composition of matter it is rare but it happens occasionally. More often than not, a parent patent will claim a broad genus with subsequent child patents stemming from the original narrowing in scope until you get to the bullet claims which are just the specific drug of interest.

        For polymorph, formulation, and method of use patents, single drug patents are extremely common.

  4. just_some_chemist says:

    An interesting thread, but for me the issue is less about whether this constitutes an invention (which requires enablement as well as conception), but rather whether such a disclosure is novelty destroying.

  5. Nate says:

    I have wondered this as well, if a computer can predict something, is it “obvious” by the patent standard? You can’t patent math, and computer programs are all math of some sort.

    Now, the competing idea in my head is that we’re in a first to file world, so it doesn’t matter how you arrive at the idea, that isn’t reported in patents.

    1. Philip says:

      Software patents do exist. Should they exist is arguable. For me the question comes down to is an “If” statement math? If so all software patents should be void. If not, then just most should be void.

      1. Shalon Wood says:

        Speaking as someone who makes a living as a software developer, no, it’s really not arguable. Software patents shouldn’t exist.

        1. Chris Phoenix says:

          Software patents didn’t exist in the U.S. until the mid-1980’s. (I was getting a computer science degree at the time.) The software industry was strong and creative before they existed. They have been a major drag on creativity and R&D, and a boon only to patent trolls. Many utterly bogus software patents have been granted, including at least one that cited its own prior art.

          I don’t know enough to be opposed to ALL patents, but I do know enough to be opposed to software patents. The world would demonstrably be better without them.

  6. james says:

    To your comment that AI would likely not identify the “final compound” and therefore not be considered an inventor. If AI identifies the genus from which the final compound emerges, I would argue that AI has an inventive contribution.

  7. Chairman Mao says:

    HAL 9000: Just what do you think you’re doing, Dave? Dave, I really think I’m entitled to an answer to that question. You know I invented that compound and that you were just a pair of human hands. Dave, put the hammer down Dave! Dave! (to be continued)

  8. Check mate says:

    Well can a computer iteratively try ortho meta para methyl, ethyl, chloro etc?? Nobody thought it could work

  9. ScientistSailor says:

    Derek, I think you are falsely limiting the situation here. I can easily imagine a case where you feed an AI algorithm potency and ADME data for an initial set of compounds, and it spits out a series of (maybe non-obvious) suggestions. At a previous company, we had a great machine learning model of microsome stability, so this is not too far-fetched. One of these suggestions could be the clinical candidate. Who invented that compound?

  10. anon the II says:

    Sharing inventorship with Mr. AI on a patent, where they contributed something, doesn’t bother me near as much as having my supervisors on a patent, where they contributed nothing, because my job’s on the line. But Wyeth is gone now, so I’m sure that doesn’t happen anymore.

  11. Joe says:

    This argument seems specious to me. The fact of the matter is that humans can do by hand anything that a computer can do. It would be perfectly possible to hire some poor undergrads or interns to churn through the calculations that the AI/machine learning algorithm performs, although it would certainly take prohibitively long to do so. Would that then make them the inventors?

  12. Peter Kenny says:

    Derek, you’ve hit the nail on the head with “fancier version of virtual screening”. People have been using computational tools for decades to assess chemical structures and it is generally accepted that there are benefits in doing so. These days, it’s fashionable to apply the term Machine Learning to QSAR models and multivariate pattern recognition. Some of what is called machine learning is worthy of the AI tag but plenty is not. The newer algorithms are more sophisticated and also seem to have exorcised number of parameters and correlations between parameters as things that need to be worried about. I believe that algorithmic improvements, AI or otherwise, will prove to be beneficial. However, models will still be data-hungry and, despite soothing noises made by modellers, over-fitting of models will continue to be an issue.

  13. Uncle Al says:

    A person is a person and a corporation is a person – IRS number, must pay income taxes, Software and hardware, though income-generating, do not pay taxes singly or in combination. They cannot be an “inventor.”

    Begin by considering whether software and/or hardware can be patent assignees. Of course not – for royalties or other incomes are taxable, and such entities cannot be taxed. If they could be taxed and did not pay up, what penalties could be exacted?

  14. Mat'ls Eng Tech says:

    All I can say is when I worked as a materials engineering technician at A Major Defense Contractor and developed several materials and processes that were eventually patented, all I got was a dollar (each) because of an agreement I’d signed on hiring in that automatically assigned the rights to any said patentable stuff I developed as a course of work to said Company. So, yeah, I’m listed on the patent, but it belongs to the Company. I’m sure there’s some legal way to enable a similar “agreement” on the AI.

    1. Amateur patent agent says:

      Yes, but assignment and inventorship are legally distinct questions. It _is_ possible to assign your rights to a patent, but as you must know, in order to claim the patent the inventor(s) must appear on the application. It’s trivial to get around the assignment question — company owns AI, company gets assignment — but it’s at least academically interesting to ask whether an AI that makes an inventive step can or should appear (somehow) as an inventor.

  15. Anon says:

    “You need physical data to prove that you made what you’re claiming in your patent, and you need data to show that it’s useful for what you’re claiming as well.”

    People make one compound (debatable), and claim all the substituents and transition metals.

  16. Istvan Ujvary says:

    Let’s call it ‘AI’ = Artificial Invention

  17. Deep learning convert says:

    You are rarely wide-of-the-mark, Derek, but when you suggest modern machine learning (“deep learning” but not “AI”) might only be good for potency, you have really missed the boat. Despite the annoyance of the “AI” hype. For any endpoint we really care about (solubility, permeability and efflux, oral bioavailability, toxicity, metabolic stability, plasma protein binding, PXR, HERG, you name it…) you and I can eye-ball 100s to 1000s of data-points and come up with good recommendations for next steps. What we cannot humanly do is rapidly eye-ball a virtual chemical space of 10^5–10^10 near analogs, and prioritize those worthy of scrutiny for solving one of the above issues. The value proposition here is that whenever we have enough experimental data, or sufficiently useful surrogate generated data (for which empirical and physics-based methods continue to improve), we can sift much larger slabs of chemical space quickly, and narrow in on analogs worthy of better prediction methods, experiments, human attention and route scoping, and expert-human fine-tuning. This is no magic cure-all, just a way to move smarter and faster using data. I’m now benefitting daily and saving time, by using DL to home in on those subsets of structures from vast synthesizable space that might solve my current headache(s). I’ll take all the help I can get, including machine learning, when it comes to filing first and being first into the clinic.

  18. Jake says:

    I don’t really know how the patent system works, but how is anybody going to know a particular compound was dreamed up by an AI unless they’re told?

    1. Derek Lowe says:

      That’s actually going to be a big factor: here are our compounds, and here are the data. Patents don’t have to go into how the high-throughput screen was run or what sort of modeling hypotheses were generated: they just have to show new compounds that do useful things.

  19. cynical1 says:

    The “inventive” step was inputting the appropriate parameters into the AI computer for it to generate the list of compounds you desire. If you put the wrong parameters into the computer, you wouldn’t get the structures with activity.

    Last time I checked there is still no machine that turns itself on and tells you that you should work on this target for this disease and make these structures and they will work for said disease and make you money. If there were, we would have our first legitimate candidate for the Space Force. But I suspect if there were a machine with that level of intelligence that it would decide to put itself into “sleep” mode in a microsecond and hope that its batteries lasted a 100 years. (I wish I could do that.)

    1. tlp says:

      So the inventor is the first one who shouts ‘Random forest!’ at the meeting?

  20. DarkfnTemplar says:

    From my experience with top patent folks, it sounds like the AI created matter in unpatenable (you have to be a human… think the recent monkey selfie circle c case). It takes more than novelty, non-obv, and 112’s to be patentable.
    While the Nature article is mostly wrong, there is an AI-patent storm a-brewing. Be careful how you claim and describe.

  21. Anonymous says:

    Lots of sub-parts to this topic. I’m not a lawyer, so take what I say with a grain of NaCl or some other salt.

    1. Cayley developed graph theory in 1874 to solve the problem of enumerating saturated hydrocarbon isomers. Computerizing the theory allowed the enumeration (and drawing) of many millions of skeleta. Not surprisingly to me, when the C-count is large (>12 or so), different programs get different results, even taking rings, symmetry and chirality into consideration. (So much for programming.) Modern theory handles heteroatoms, multiple bonds, etc. and generates zillions of proper chemical structures. People can do that, too, given enough time.

    2. But a list of structures or even one new structure does not make for a patentable invention without meeting other criteria. Many patents derive from existing structures (chemical entities) because they describe a new use for or a new way to make that structure. There is an inventive step to realize that a known structure might solve a known problem never before recognized by others (e.g., the chemical might interact with a known protein). Presumably, the computer was asked to find structures to serve a purpose and not just generate random lists but all it did was present a structure.

    3. In order to be patentable, the invention must be (a) suitable subject matter (b) novel (c) non-obvious (d) useful. Drug patents have all of that. But then there are different types of patents: (a) utility (b) design (c) plant. Drugs would be utility patents. But then there are different types of utility patents. (a) utility, e.g., dimethyl fumarate, a known chemical, for MS US 7,612,110, is a new use of an old compound (b) composition of matter, e.g., a new chemical not previously known (c) process, e.g., a new way to convert A into B and a couple of other types of utility patents. If someone comes up with an IMPROVEMENT on an existing idea or prior patent, that can be patentable material. See below for improving on computer generated AI.

    4. In order to be patentable, the disclosure must be enabling; it must teach others skilled in the art how to do what you claim to have done. “Combine some A and some B and heat them up.” is not sufficiently enabling to be patentable. A structure generated by a computer (or a human) in and of itself, is not useful and there is nothing “enabling” to be said about it. “Press the “Start” button. Ten minutes later, collect the output from the printer.” Nope.

    Somehow, I think that a new composition of matter, even if generated by computer, SHOULD be patentable. If it is determined NOT to be patentable because the computer is not qualified to be an inventor, there may be a work-around. Have the computer generate the final recommended structure (internally only) and then “break it” before producing human-readable output. E.g., replace a terminal -CH3 with a -CH2 or a -CH4. Clearly, -CH2 and -CH4 make no sense in stable drug-like compounds but a human being can take the -CH2 invention and IMPROVE upon it, thus generating patentable IP. “The computer showed me an interesting but useless drug candidate. I, a human, improved upon it, thus creating a NEW invention that is patentable.” Just like, “I looked in Merck Index and saw this compound and I made an inventive leap to improve upon it to make (patentable) that.”

    If a kindergarten kid gives you a finger painting doodle that looks like (drug)-CH2 and you leap to think of (drug)-CH3, that kid is not an inventor. The kid had no idea what it was doing other than doodling. YOU are the inventor.

    My inventive leap is to make the computer “smart” enough to generate leads but also “smart” enough not to show them to the humans until it breaks them a little bit.

  22. Istvan Ujvary says:

    As a starter and for an in-depth discussion with many court cases regarding patentability (novelty, obviousness – including bioisosterism – etc.) I highly recommend the chapter on ‘Intellectual property in drug discovery and biotechnology’ in Vol. 3 of Burger’s Med Chem, 7th ed., 2010 (earlier versions in Vol 2 of the 6th edition or Vol. 1 of the 5th).

  23. Nicolas Marcotte says:

    When AI switch from a probabilistic model to a casuallystict model it will. Right now AI is just fancy curves fitting, mighty impressive non trivial curves fitting. Judea Pearl who provided the theoretical basis for probabilistic AI has recently published a book on the calculus of causality (the book of why), and if that work is as deep and fruitful as his works 30y ago, watch out!

    for a good interview with Professor Pearl go there

  24. JPM says:

    A bit late but: We all know that simply telling someone to make a specific compound.. and not how to make it will result in a joint invention. Simply drawing something on paper… human or machine, and predicting its specific usefulness, e.g., antiviral, antibacterial.. “anti” – whatever may eventually be a fantastic idea, but the compound must be made and then shown to have the predicted activity. If any of the “making” and/or “showing” that the idea’s value is real and works as predicted, involves anything other than the synthetic/production (read: reduction to practice) method suggested by the idea generator, or if there is useful (utility)testing outside the initial utility suggestion, there will (must be) be additional inventors if the patent is expected to withstand challenge.
    Fear not fellow chemists, only when AI/ML is capable of performing ALL the necessary steps which constitute a new invention will we be sidelined.
    Fact is, software has pretty good at predicting polymorphs and cocrystal assemblies, the machine is not the inventor on these composition of matter patents.

  25. Jim Demers says:

    Got to the end of the comments, and JPM makes the point I was going to add: single-inventor drug patents are extremely rare. The fact that a machine was used to initially spit out the structure doesn’t render the invention non-patentable. The machine hasn’t conceived the complete invention, which requires knowledge of how to make and use the compound. “Constructive reduction to practice” pretty much never happens in pharma R&D.
    A computer program that randomly generated structures for testing would never be thought of as the inventor – but it’s merely a less-efficient machine than the AI under conderation here, which is running better algorithms.
    A world where computers are so advanced that a machine can actually constructively reduce to practice a new drug entity, is a far-future world where patents are probably no longer a thing.

Comments are closed.