CL Talk

Narrator 1 [00:00:06] This episode is presented in partnership with Bereskin & Parrr. 

Greg Hudson [00:[00:10] Hello, welcome to the CL talk podcast from Canadian lawyer. I'm Greg Hudson. It's not breaking new ground to say that generative artificial intelligence will change how law is practiced and understood, including and especially intellectual property law. One area where AI shows great promise is chemical research. Now to get a better understanding of how these two complicated fields intersect, today, we're speaking with two partners from Bereskin & Parr, Canada's leading full service intellectual property law firm. Reshika Dhir is a member of the firm's Electrical and Computer Technology Group, and CO lead of the firm's fin tech practice group. She's an electrical engineer, lawyer, and registered US and Canadian Patent Agent, and Patricia Folkins is a partner and leader of the firm's chemical practice group. She has a PhD from McGill and completed a postdoctoral fellowship at a little school called Yale. For our purposes today, Reshika and Patricia are both co leads of Bereskin & Parr new chemical AI practice group, Reshika and Patricia, welcome to the podcast. And thank you so much for being here today. 

Reshika Dhir [00:01:17] Thanks, Greg. Happy to be here. 

Patricia Folkins [00:01:19] Thanks Greg. It's our pleasure. 

Greg Hudson [00:01:21] Let's jump in with the first question. In what ways does AI assist with chemical research? 

Patricia Folkins [00:01:28] So there are many ways that AI can assist in chemical research. And that has assisted in chemical research. In the past, I started doing a lot of research and reading into the AI space, because many of us have recently heard about chat GPT. And I thought, well, this seems kind of cool. So I wonder what it can do. So in all the articles that I read, and some of the podcasts I've listened to, there are some scary things that I've heard. But in everything that I've read, one of the bright spots seems to be its potential application in chemical research, in particular in drug discovery, that hopefully it will improve the drug discovery process and lead to improve probability of success for researchers in this space. Other things that it can specifically be used for improving predictions of molecular properties, whether it be in molecules, or in general in materials to optimize reaction conditions or to develop synthesis synthetic roots for molecules from scratch can be used to develop laboratory automation procedures, and specifically in the drug discovery realm. It can be used to help select leads based on predictive binding to targets, whether it be a protein or an enzyme, an enzyme protein or receptor can also help predict pharmacokinetic properties such as solubility and half life. And yeah, as I said before, hopefully, like having that knowledge, you can hone in on more leads that are that are predicted to improve success rate. 

Greg Hudson [00:02:59] Interesting. Do you have any examples of things that have come out of AI assisted chemical research? 

Patricia Folkins [00:03:05] Yeah, I am aware of a couple of small molecule drugs that have come out of chemical AI research. There are a couple of companies in particular, I know of Exscientia. It's an AI based chemical drug research company. And they have several leads that I think there's three or four that are now in phase one phase two clinical trials. There's one in particular that was just recently licensed to Bristol Myers Squibb, which is a big pharma company for treating inflammatory and autoimmune conditions. I think their deal was reported to be around $1.2 billion. So definitely a valuable proposition for Exscientia. Insilico is another company that has a small molecule drug for treating fibrosis, it's in phase two. So it's still early. I don't know of any actual drug that's being like on the market and approved for sale that has been solely developed by AI, but they're working their way through the clinic right now. And hopefully soon there will be there's AlphaFold, which is actually an AI system that was developed by Deep Mind in combination with Emble, EBI, and it's a fully publicly open access. It's a publicly available database that predicts protein 3d structures and proteins are very important in drug research because many drugs modulate protein activity, whether it be a receptor or be an enzyme. So it has open access to over 200 million structures that were predicted using chemical AI. There's also Chemical AI, actual name of a company and China that provides several services like what I mentioned earlier, in terms of coming up with molecular syntheses. And lab automation. They say that more than 80% of the design routes to synthesize a molecule are comparable to or better than those designed by chemists with greater than 10 years experience. So I don't know maybe they're gonna put us chemists out of business, but certainly is facilitating the process and making research go quicker and hopefully more successfully. 

Greg Hudson [00:05:08] Luckily, AI won't do anything for writers or journalists, it's, we're completely protected from that, unlike you chemists, just kidding. But let's talk about patents. Because we're finding drugs, we're obviously gonna want to patent that. So when it comes to patenting chemical AI, what actually can be protected? 

Reshika Dhir [00:05:29] That's a great question and not an easy one to answer. But short answer would be a lot can be protected. So a lot of the aspects of AI based inventions generally can be protected under the IP regime, one of the popular ways of protecting AI inventions, including chemical AI inventions, would be patents. So patents generally protect the functionality of inventions and for something to be patentable, there are certain requirements that the invention has to meet. So the invention has to be novel has to be new. And first of its kind, it has to be non-obvious or inventive, has have utility be useful. And then it has to be the right kind of subject matter to which patent can attach. One of the first considerations here is that not all inventions are eligible for patent protection. So in most jurisdictions of the world, mathematical models or algorithms are not considered inventions. And as you can imagine, that kind of complicates things for AI based inventions. But the good news is, if you have an application of a machine learning method, or an AI model, that solves a specific problem in a manner that goes beyond that underlying mathematical algorithm, then that machine learning method can be considered patentable. So what that means in the context of the intersection of AI algorithms and chemistry is if you're using AI models for designing new molecules, or materials, or doing predictive modeling, where you're predicting molecular properties of certain molecules, or compounds, or you're optimizing chemical reactions, so in those situations, you're kind of applying a general AI based model or machine learning models are an algorithm to a very specific problem. And that can help increase the success of getting a patent. Generally, in patent applications, highlighting all the technical advantages of the AI underlying AI algorithm, as well as the practical details of how it applies to those chemistry to chemistry research, and these chemistry examples, often time helps a lot in increasing the likelihood of success in obtaining a patent. Now, the other area that is very commonly explored when it comes to AI based inventions is basically the fact that AI tools can be embodied in a physical form. So you have medical devices that can incorporate some sort of, you know, innovative AI, you can have tools and products or even larger systems that have sort of at their core have AI based innovation. So generally, when you have a hardware integration with a AI model, that also helps you increase a chance of success. So, you know, for example, if you have a robotic system for some sort of chemical synthesis, or testing, or you have some novel spectrometer that is using certain novel AI models, or you have a kind of neat diagnostic tools that's relying on some cool AI algorithm, then that integration of hardware and software can also be patentable. 

Greg Hudson [00:08:52] What are some of the key issues or concerns? When it comes to patenting chemical AI? You kind of touched on some of them, but I wonder if you could go a little deeper? 

Reshika Dhir [00:09:01] Yeah, absolutely. So certainly one of the key challenges is the subject matter eligibility of AI based dimensions. So as I was saying before, this issue comes up in the context of AI mentions and to be honest, generally in the context of computer based dimensions, because of certain exclusions that are built into either the statute itself or comes out of case law. So general tendency of the patent offices to not issue patents for abstract ideas, or just mathematical models, etc, is one of the challenges when it comes to protecting AI based dimensions. So certainly emphasizing the technical advantages that the IV system provides can help demonstrate to the patent offices that you're not dealing with just some random abstract mathematical algorithm that you are in fact, solving a particular technical problem with a particular technical solution that includes that AI model. And it's being used in a specific context. So that's a challenge. Some of the other sources of challenges would be one of the requirements of getting a patent is that you need to have sufficiency of disclosure in the patent application. So what that means is that the applicant has to provide a detailed and a clear description of how the invention works, so as to enable a person who is skilled in that field to replicate that invention when the patent expires 20 years later. Now, sometimes, because of the complexity around the AI algorithms, describing such complex algorithms and how they apply to certain chemistry applications can be challenging. And sometimes you may have applicants who may not be able to articulate how this actually works, partly because they may not know themselves what's happening under the hood. So having a patent application where you haven't really clearly described how that invention works is could be grounds for invalidation of that application. And the other rather interesting issue that has come up in recent years has to do with the issue of inventorship. So every patent application requires the identification of inventors, and the question of who the mentors are for a particular invention that's captured in a patent application. It is specific sort of consideration of law and facts, right. So for somebody to be considered an inventor, they need to, they need to be the person who has conceived of the idea, or reduced that idea into practical application. But if you have the innovation happening by AI models themselves, in other words, if AI is doing the inventing, then technically you should be naming the AI as inventor. And we've had a series of test cases, they're called dabbas cases where an applicant tested whether patent offices will allow patents where AI is named an inventor. And in almost all notable countries, including the US and Canada, the decision was that AI cannot be named an inventor, this issue of inventorship can become a serious issue, especially when we have prevalence of reliance on AI models and the AI algorithms and these models are becoming smarter, day by day, is not going to be surprised to people to realize that there could come a time where a lot of the innovation is happening by these AI models themselves. And so this issue of who should be named an inventor and whether you know, an application with a name doesn't inventor shouldn't be issued into patent that's going to, I think, become an interesting issue that we're going to follow for the next few years to come. 

Greg Hudson [00:12:47] That's so interesting to have a computer be an inventor, I want to talk about the chemical aspect as well. How do we think about patenting chemical processes? 

Patricia Folkins [00:12:59] Well, I guess that's the chemist. I'll take that one. Certainly, as you mentioned, we are developing, perhaps new chemical molecules, new chemical entities, new materials, new processes, these are all standard and common traditional subject matter for patent applications and don't necessarily run into the issue of you know, whether or not they fall under patentable subject matter or not because they are classic people subject matter, because they are classical things that are protected using patent applications. We don't fall into those issues of whether or not they represent patentable subject matter. But the issue is if you only have data that is generated by AI, so Insilico data, and you haven't actually reduced some of these predictions to actual practice actually made the chemistry chemicals and proven that they do bind to that receptor or inhibit that enzyme, then you could run into problems for not having sufficient disclosure, as Reshika, has mentioned as one of the criteria for patentability, you need to have enough data to enable what the computer is predicting will work actually does work. So that's another caution is that you? It should researchers should make sure that you know, once they have sort of these predictions in hand that they actually go into the lab and try them out and make sure that they work the way that the AI predicts that they're going to work at least a representative exact number of examples, all the molecules or the processes, etc. Interesting. 

Greg Hudson [00:14:31] 

Now, I know that patenting is one way to protect IP, but there are other ways to protect IP as well. So how does those relate to chemical research? Chemical AI research? 

Reshika Dhir [00:14:43] Yeah, so I can take that one. So one of the other commonly used ways of going about protecting certain aspects of the IBS chemical inventions would be trade secrets, and trade secret kind of, as the name suggests, is the commercial value you get by keeping your valuable information a secret. One of the aspects of chemically AIBs dimensions that can be kept a secret would be your proprietary training data, right. So because data itself can be very valuable for many companies, especially if it was carefully curated and labeled, and set out to make for effective train di models. Another aspect of chemically, these dimensions that can be protected under trade secrets would be the AI algorithms themselves. Now, there could be a couple of reasons for choosing to protect them via trade secrets versus patterns. Sometimes you have AI models that may be just too mathematical, and they may not be considered the right kind of subject matter for patent protection. But many other times companies may choose to keep such algorithms trade secret, as opposed to protecting them via patents is for strategic reasons. So for a patent application, one of the requirements is, as we kind of alluded to before, is that the invention has to be disclosed in a patent application. And once it's filed in a patent is issued, the patent is only valid for 20 years from the date of first filing, which by no means is a short period of time, it's a very good chunk of time for companies to recoup the investment they made into their research and development, and to generate some revenue out of their inventions and products. However, if companies can successfully protect that algorithm through trade secret means, then the protection can extend way beyond the 20 years. Of course, trade secret protection or reliance on trade secret can be tricky, because if the algorithm or the How it works is released to the public, where it gets out there, then the protection is over, then there was really the value is gone. So that's risky. So obviously, companies are relying on trade secret mechanism to protect the data or algorithms themselves, they have to be very careful and have a curated plan on how they're going to implement their that trade secret regime. Sometimes even output data of the AI system can be protected under trade secret, provided it can be sufficiently kept a secret. And it's something that cannot be easily reverse engineered. Another bucket of IP that can be used in this context of chemical AI inventions would be copyright. Now copyright can again, protect the software, or the code itself, including the source code and the object code. But copyright protection, it only extends to the expression of that code, and not to the underlying functionality, which means that your code can be easily or potentially easily reworked, or at least the functionality of it can be reworked and used by others. So there has to be again, certain caution to be warranted, if copyright protection is solely being considered for software protection. And also sometimes, you know, if you have your raw data or training data that has, or it can have copyright attached to it, especially if say you're relying on photographs as training data, or some content that is generated as training data, you have to be kind of careful if there is already copyright attached to that data. And if you have the right with the permissions and the rights to use that data. And likewise, if you are exercising any skill and judgment in preparing the data and labeling it and putting in databases and things like that, you will maybe be able to claim copyright in those creations as well. 

Greg Hudson [00:18:43] This feels like something that will probably be tested in court, it sounds like are all these rules sort of defined perfectly. 

Patricia Folkins [00:18:50] I agree. And so not everything is sort of black and white, there's a lot of gray in here, especially when it comes to sort of copyright protection, what it extends to how much of the skill and judgment needs to be exercised before copyright can attach to data. That's a work in progress. Likewise, when we talked about patents, the discussion on what is the right kind of subject matter that can attach patent protection to it. That's also work in progress. There's, we see case law in Canada once in a while. And but all the case law just will suggest to us that we're just getting started in resolving a lot of the complicated legal questions that are part of those considerations. 

Greg Hudson [00:19:34] Oh, that makes sense. Now you, Patricia and Rashida, you're the CO leads of the chemical AI practice group by um, why was that group formed? 

Patricia Folkins [00:19:43] We were starting to see you know, as a result of all of these companies and all these researchers working in this space, I think chemical research has been using computer modeling for many, many years, but the AI aspect has really accelerated the innovations that are taking place in this space. And I'm particular was getting quite a few or several disclosures from our clients in Canada. I know the University of Toronto has, I think, in 2021, set up this acceleration Consortium, which is a global coalition of scientists design to design discover materials that don't exist using AI. So there's a lot of exciting things happening in Canada and starting, some of it is starting to trickle into our firm. And I can handle the chemical side of it, but I can't do the AI side of it. So I said, this is going to need an interdisciplinary approach. So I needed to reach out to someone in our ECT group, Electrical and Computer Technology Group, and Reshika and I've worked together on other matters, and she's a great person to work with, and very knowledgeable on that side on the computer side. So we thought, well, let's team up and make a practice group and focus our efforts on some of this research.  

Reshika Dhir [00:21:01] Interesting to collaborate on a topic like chemistry, because I think the last chemistry course I ever took was in grade 12. But we have in Patricia and other members of our chemistry group, great teachers who know how to tell an electrical engineer what's going on the chemistry side of things, which really helps when I'm trying to talk about AI models and the data that's being used to train those models and the output that's being generated. So we have a great team here. And it's such a pleasure to work with Patricia and the team. Awesome. 

Greg Hudson [00:21:33] Patricia, Reshika. Thank you so much for stopping by the podcast today. And explaining patenting chemical AI, which as it continues to get better with drug discovery, it's going to become more and more important to patent those drugs. 

Patrica & Reshika [00:21:46] Thanks. Thanks. Our pleasure. Thanks, Greg.  

Greg Hudson [00:21:50] 

This has been Greg Hudson with Canadian lawyer talk podcast. Thanks so much. 

Narrator 1 [00:21:55] Thank you for listening to this episode of CL Talk. For more from Patricia, Reshika and the team at Bereskin & Parr visit them at bereskinparr.com. For the latest episodes, be sure to follow us on all major listening channels.