Biopharma companies debate whether to patent their algorithms for finding medicines.
Investors are pouring billions of dollars a year into artificial intelligence (AI)-enabled drug discovery. A recent study found more than 70 AI-derived small molecules, antibodies and vaccines in clinical trials. And deals in the space have been mounting. It’s possible none of the candidates will work. Or that none will make good business sense. AI-enabled drug discovery has yet to prove itself in the clinic. And on top of that, companies have to worry about intellectual property — regarding not only the drugs but also the discovery methods. “Everybody patents molecules,” says Alex Aliper, the president of Insilico Medicine, headquartered in New York City and Hong Kong. “For algorithms, it’s a choice.”
Deciding whether to apply for a patent, however, can be tricky. And whatever choice a startup makes could have huge repercussions.
AI can assist in multiple stages of drug discovery. Pharma startup Verseon, for instance, uses AI to optimize molecules and to understand the interactions between drugs and genes in aging and degeneration, says Ed Ratner, head of machine learning applications. Viswa Colluru, the CEO and founder at Enveda Biosciences, a startup that mines natural compounds, says they apply several types of AI algorithms to predict the structure and function of unknown compounds from mass spectrometry and high-throughput screening data. Some researchers use AI to design proteins from scratch. One of the most obvious advantages is speed. Aliper, of Insilico, says that although it can take years and tens of millions of dollars to nominate a preclinical drug candidate, his company used AI across several stages to nominate ISM001-055, their drug candidate for idiopathic pulmonary fibrosis, in 18 months for $2.6 million.
Given their potential advantages, patents are crucial. Aliper says Insilico has patented around 50 AI methods, on top of their patents for compounds. Every year, according to Bloomberg Law Analysis, there are close to 20,000 AI patents. And according to preliminary work by Mateo Aboy, the director of research in biomedical innovation, AI and law at the University of Cambridge, around 3,000 patents have been filed to date on AI methods related to drug development and delivery alone. Aboy says they’re distributed among universities, big pharma, big tech and startups.
But algorithms are hard to patent. Although the United States Patent and Trademark Office (USPTO) stipulates that patents can go to “any new and useful process,” among other things, an algorithm, which is a set of steps, is not necessarily considered a process. Instead, courts have often seen it as an abstract idea, and the USPTO doesn’t allow patents for “abstract ideas, laws of nature and natural phenomena (including products of nature).” To be eligible, a patent application should describe how the claimed invention as a whole integrates the algorithm to solve a specific practical application that improves a technology or technical field. And it is not enough to simply say, of the algorithm, “apply it with a computer.” A landmark 2014 Supreme Court case between Alice Corp., who had sued CLS Bank for infringing their patented financial transaction, highlighted this.
The eligibility waters have remained muddy, even after the USPTO revised its guidance in 2019. So in June, two US senators proposed the Patent Eligibility Restoration Act of 2023, not yet passed, which explicitly marks mathematical formulas and mental processes as ineligible. But it’s unclear how this revision goes beyond the Alice decision.
Matthew Chun, a technology specialist and patent agent at Fish & Richardson, offers applicants advice in a recent preprint. Inventors should talk about how they trained their machine-learning models, describe what part of a method can’t be done by human mind, frame the innovation as a technical improvement, tie it to a particular machine, and be specific.
Even when inventors think their AI systems might be patentable, they need to decide whether it’s worth patenting them. Shelby Newsad, an investor with the venture capital firm Compound, works with companies in their very early stages. For many, “algorithms themselves aren’t the sole proposition,” she says, “because there’s a new state-of-the-art method published every couple of weeks.” It’s also uncommon, she says, for companies to patent AI very early on, when they might not have a lab or even any code, just a plan. Because of the cost of applying for and defending patents, “in Silicon Valley, we talk about software patents being the sport of kings,” says Vijay Pande, a general partner and head of the a16z Bio + Health fund at the venture capital firm Andreessen Horowitz and a Stanford biologist with patents and startups of his own. (One patented system allows you to search for binding subgroups anywhere on a molecule, like looking for a cat anywhere in a photo.)
Wen Xie, a patent attorney and partner at Global IP Counselors who works with tech including autonomous vehicles, says that in the autonomous vehicle case, the AI is in the product, so it’s important to patent it first. Public deployment can preclude later patenting. A company like Novartis, she says, is more interested in patenting the drug that comes out of the AI than the AI itself. Aliper also says the guidelines are clearer for patenting drugs than methods.
When deciding whether to patent an AI system, to publish it (giving you only a year to patent it in the United States before it’s no longer new) or to keep it as a trade secret, inventors must weigh several factors. First, there’s the time and effort of applying. And it’s not clear whether the applicant will receive the patent — it depends greatly on the examiner and changing interpretations of patent law by the patent office, says Gregory Rabin, a patent attorney at Young Basile — nor whether it will hold up in court. And in some countries the application is published even if it fails. In the United States, you can request that the application not be published until the patent is granted, which may occur several years after the filing date. Otherwise, a patent application is published 18 months after filing.
Ratner, of Verseon, says that when you patent an AI system you must worry about discoverability and enforceability. If your patented AI system is secretly adopted by another company for use in-house, it can be hard to discover the infringement. And because you can patent only an implementation of an idea, not the naked idea, someone else might find another way to implement the idea in your patent, and you wouldn’t be able to enforce the patent against them. Deciding to patent is “a business question — a cost–benefit tradeoff,” Ratner says.
“In our case, patenting and IP protection has kind of a dual role,” Aliper says, of Insilico. “It can be a shield and a sword.” They might patent an AI method and enforce the patent against others (although they haven’t done so) or patent it so that someone else doesn’t patent it and enforce it against them. Patents give them freedom and remove business uncertainty. Colluru, of Enveda, says they’ll patent a method if they worry someone else might independently invent and patent it. Previously, they would have been able to challenge the patent by showing old notebooks showing they were there first, but in 2013, the United States joined other countries and implemented a first-to-file rule. “Just because you have invented it first,” Colluru says, “doesn’t make you the rightful inventor.”
Dodging others’ patents is just as nuanced as deciding whether to file your own. “Since patents are valid for 20 years and this is a very fast-changing area of technology, you’re probably stepping on someone’s toes,” Rabin, of Young Basile, says. “There are patents that were filed in 2003 that are still valid,” he continues, “and that could be for foundational things in the industry that are now being used by everyone, and you can’t do your job without it.” A patent doesn’t need to mention an AI system’s application to drug discovery for it to be enforceable in that area.
Chun notes a couple of elements in US patent law that protect potential infringers. One states that if a patent owner waits too long to sue, the damages for infringement they can recover may be limited. Another states that if a patent owner gives the false impression that they don’t intend to enforce a patent, they may give up the right to do so later. Ratner says that Verseon will often use published algorithms for tasks common in the pharmaceutical industry because published algorithms can’t be patented more than a year later.
Chun says that patent lawyers sometimes perform a ‘freedom to operate’ analysis, searching for patents that may get in the way of a business. “In my experience, these FTO analyses are often initiated at the request of an investor,” he says, “or at least because a client knows that an investor might ask.” Colluru says Enveda keeps an audit of the algorithms they use so they don’t run afoul of software licenses. If they don’t like the license for a tool, they might develop an alternative themselves. Rabin notes that some companies actually avoid doing patent searches because knowingly infringing on a patent triples damages.
Aliper sympathizes with newcomers to drug discovery who may have a hard time avoiding others’ patented methods. Inventing methods that turn out to be useful is hard, he said. “It is not like ChatGPT, where you can immediately assess the quality of the response. If a model gives you a molecular structure, you need to synthesize it in the lab, you need to test it in the lab, and that takes time and resources.”
In general, there isn’t much suing over AI patents. In 2019, the World Intellectual Property Organization reported that there were more than 300,000 patent families related to AI and that, as of 2017, fewer than half a percent had been mentioned in litigation cases. Experts contacted for this piece couldn’t recall any litigation over AI in drug discovery. “People play pretty nice,” Pande says. In 2018, Google patented the ‘transformer’ neural-network architecture that’s at the heart of many of today’s generative AI models, including ChatGPT and some drug-discovery tools, but they have not sued anyone. Rabin suspects it’s because they patent defensively. Xie says it may be because it’s not worth their time to try to calculate damages. If an algorithm isn’t embedded in products sold, it’s hard to know how much it’s being used and what its commercial value is. “This is not to say that AI patents will never be litigated,” Chun says. “The whole AI-in-drug-discovery field is relatively new.”
When asked if there’s a lot of uncertainty in the area, Colluru says AI algorithms aren’t so different from other computational tools, such as those used in gene sequencing. But as they get trained on larger sets of data, uncertainty about copyright infringement and proprietary data will increase. OpenAI and Stability AI (maker of Stable Diffusion) have been sued over use of text and images to train their language and image models. And some datasets have become accessible accidentally.
Looking farther out, Colluru says success will depend on “your ability to build a wet-lab–dry-lab integration loop,” referring to active machine learning using test tubes and software tools. “There’s a fair amount of automation, robotics and software engineering of those data pipelines that goes unappreciated.”
Insilico will keep patenting AI, but also wants to share it with others, for free or fee. “We do have a software platform which we specifically designed to enable others to do what we can do,” Aliper says. Thousands of diseases have no adequate treatment. While humanitarian and business interests are sometimes at odds, they needn’t always be, he says. “The therapeutic space is enormous. No single company can cover it.”