A new study claims AI drug companies are recklessly patenting untested molecules, but the analysis reveals a fundamental misunderstanding of how pharmaceutical patents actually work. The study entitled “What patents on AI-derived drugs reveal” (Freilich & Rai) was recently published in the journal Science. The authors analysed granted drug compound patents from over 100 AI-drug companies, and concluded that there is a worrying trend of premature patenting by AI drug companies which will result in potentially therapeutic molecules being abandoned by the industry. The authors suggest that this trend should be addressed with policy and regulatory changes, including a higher bar for patentability.
We respectfully disagrees with the authors’ analysis. Put simply, AI drug companies are in the expensive drug discovery business, and have little to no incentive to pursue granted patents for molecules they do not believe can be commercialised.
The study methodology
There is no requirement for drug compound patents to disclose the discovery process that led to the identification of the compound. The key legal test is whether a skilled person could carry out the invention based on the disclosure of the application as filed. Small molecule drug compound patents instead generally focus on the properties and synthesis protocols for the claimed compounds. As such, even drug companies devoted exclusively to the use of AI to develop drugs usually do not give details of the AI methods used to initially identify the compounds of interest.
Given the lack of disclosure about the use of AI in drug discovery, the authors of the study could not simply look for evidence in patent applications that a drug was discovered using AI. Instead, the authors began by identifying companies marketing themselves as exclusive AI drug developers, and examining the granted compound patents filed by these companies. The authors examined 77 compound patents from over 100 AI-drug companies. Each AI-drug patent was then matched with a random small-molecule compound patent that satisfied the criteria of 1) being filed by the same law firm, 2) being from a company in the same USPTO size classification, 3) having a priority date within ±2 years of the AI patent and 4) not being assigned to an AI company. So far, so good.
Results: AI drug company patents contain less in vivo data
The results of the analysis showed that the drug compound patents from the AI companies contained less in vivo data than those from non-AI companies. Only 23% of AI-derived drug patents included any in vivo experiments, compared to 47% of control patents from traditional drug developers. Additionally, the AI companies tested a mean of only 0.8 compounds in vivo per patent (representing just 3.0% of disclosed compounds) compared to traditional developers testing 3.1 compounds per patent (6.4% of disclosed compounds). Additionally, only 12% of AI patents included ADMET (absorption, distribution, metabolism, excretion, toxicology) experiments compared to 26% of controls patents, whilst 1% of AI patents disclosed a specific formulation for their drug compared with 14% of control patents (P < 0.01).
A problem for the industry?
The authors suggest that these disparities suggest that AI-native companies are rushing to patent protection with minimal experimental validation, potentially creating what the authors term “compound patents on molecules that disclose little evidence of real-world testing”. The authors argue that the AI drug companies may have a negative effect on drug development, because the early filings will discourage other companies from working on the new drugs. Additionally, the authors suggest that the AI companies will find it difficult to develop the drugs themselves because licensing arrangements typically require more extensive validation than AI companies are currently providing. In other words, the authors argue that the patent applications may effectively stop anyone developing the drug.
However, in our view, the authors’ conclusions overreach the evidence:
In vivo data is not required for patent validity
The authors’ first mistake is to place too high an importance on in vivo data for patent applications. The filing of patents based on preliminary data is nothing new. It is not necessary for patentability to provide in vivo experimental data. It is notable that in the data presented in the study itself, less than 50% of the granted patents from the non-AI companies contained in vivo data. In many cases, in vitro data was deemed sufficient by the patent offices. What matters for patentability is whether the data is relevant and persuasive with respect to the claimed invention, not whether it was obtained in vivo, in vitro or even in silico. From the perspective of the patent office, what matters is that there is some evidence particularly related to the claimed invention provided. As explained our recent article on how to read a biotech patent, the data needed to support a claim for a new therapy, for example, is usually far lower than is required by regulatory authorities or even peer review publication.
The authors suggest that AI companies should be required to disclose their computational testing methods and results and provide alternative evidence of compound viability when in vivo data remains unavailable. However, the provision of sufficient evidence supporting the invention is already a requirement for patentability in most jurisdictions. There is just no requirement that this evidence should be obtained in vivo. In some cases, in silico or in vitro data in a human system may in fact be considered more persuasive and supportive of an invention than in vivo data from a mouse model that fails to represent key features of the human disease.
Will AI drug patents block the field?
The authors’ argument that AI drug company patent application will discourage others from working on the new drugs, is only relevant if the drug companies do not intend to develop the compounds themselves (with or without a partner), or are not able to do so.
The fact that the companies have filed patent applications for the compounds in question indicates that they at the very least intend to develop the drugs. If the AI companies just wished to block the field, in the manner of patent trolls, they could simply disclose the compounds and related data, without the rigmarole and expense of a patent filing and prosecution to grant. Interestingly, the data showed that the AI drug companies include approximately the same number of compounds in their patents as the non-AI drug companies. This suggests that AI and non-AI drug companies are being equally selective about the compounds they are choosing to disclose.
The question then becomes whether the AI companies have inadvertently shot themselves in the foot and will not be able to develop the compounds. The authors’ argument is that the patent filings will themselves be a barrier to development because they will scupper the AI companies licensing opportunities. But can this argument be followed?
Is a lack of in vivo data in the patent a barrier to licence discussions?
The authors argument that the AI companies will not be able to find a licence deal, because the patent application does not contain sufficient in vivo data, is flawed. In the field of pharmaceuticals and biotechnology, it is true that the usual strategy of large pharmaceutical companies is to file compound patents as late as possible, e.g. just before entering the clinic, in order to preserve patent term and maximise return of investment. By contrast, smaller biotechs, universities and platform companies tend to file their patent applications earlier, with less data, as an aid to marketing and securing investment. The disconnect in IP strategy between large pharmaceutical companies and early stage companies is therefore something the industry was grappling with well before the emergence of AI in drug discovery. AI companies that have thought about their IP strategy will be aware of the need to balance an attractive offering for licensing with the need to secure investment and market their AI drug discovery platform.
Furthermore, as the authors of the article themselves admit, the data does not necessarily indicate that AI companies are filing earlier than non-AI companies. The data presented in the paper merely show that the AI companies are filing patents with different data. It is possible that this is because the AI companies do not have the expertise or funding to carry out in vivo studies and/or they are able to provide sufficient alternative data in the application. Indeed, if the promise of AI acceleration of the drug development pathway is fulfilled, AI companies may actually have the advantage when it comes to licencing deals. After all, if the time it takes for a drug to reach the market is decreased, then the available patent term for securing return of investment will be longer.
The authors’ argument that prospective licensees may require more in vivo validation than is provided in the patent applications of the AI companies, is more persuasive. However, this argument does not really pertain to patent strategy. The data evaluated in due diligence may or may not be included in the patent. Potential licensees will also carry out separate IP due diligence (e.g. does the IP cover the molecule, when is loss of exclusivity, who owns the IP, etc.) and scientific and commercial due diligence (e.g. proof of mechanism of action, toxicology risks, likelihood of efficacy, manufacturing costs, etc.). Whether the licensee is attracted by the company’s offering will thus be determined by their view, not of the IP alone, but of the data as a whole, most of which may not be included in the patent application. The amount of in vivo data in the patent is therefore immaterial from the perspective of the licensee if the patent is considered valid based on the data that is provided.
Final thoughts
The authors’ argument that the limited in vivo data represents a worrying barrier to the development of the AI-generated drugs, can therefore not be followed. Currently, AI-drug companies appear to be adopting an IP strategy very similar to that of non-AI drug companies of a similar size and development. There is even an argument that not including in vivo, toxicology and formulation data represents a more savvy IP strategy, given that these data may contain further inventions (e.g. dose, formulation, method of treatment), which could be used to bolster the overall patent portfolio later down the line.
As the authors of the article themselves admit, An important limitation of this study was its inability to assess how much non-AI companies may also be using AI behind the scenes, given that there is no requirement for drug companies to disclose this in their patent filings. Many large pharmaceutical companies have entered into collaborations with AI companies, and are also developing their own internal AI R&D tools. However, it is impossible to know from the patent for a new compound how much the initial drug discovery process relied on the use of AI. Interestingly, regulatory authorities have hinted that they may require more information about the use of AI in drug development, which may impact the IP strategy for drug development inventions involving AI.
Finally, the authors note that the data in the study relates to patents filed before the rise of LLMs and the proliferation of tools for patent drafting. The authors go on to suggest that AI drafting tools may make the problem they perceive worse, by enabling AI companies to file earlier with disclosure of even more untested compounds. However, this argument again misses the point that the AI companies have no incentive to file patent applications full of AI slop. AI-drug companies are in the expensive business of drug discovery and development. An IP strategy involving simply disclosing random AI-generated lists of untested compounds is quite simply not in their interests.
The rapid advancements in AI-assisted drug design demand an adaptable IP strategy. At Evolve, we specialize in crafting robust IP strategies that align with your scientific breakthroughs and commercial goals. Our team, with extensive experience in both AI-assisted drug discovery platforms and the broader pharmaceutical industry, understands the unique challenges you face. We’ll help you consider the full scope of possible IP protection, from trade secrets and platform patents to drug substance patents, ensuring your innovations are positioned for maximum value. Don’t delay protecting your valuable AI research, the sooner you develop a robust IP strategy, the stronger your position will be. Visit our dedicated AI drug discovery page or contact us today for tailored IP advice that drives real business value and keeps you ahead in this fast-moving field.
Subscribe to our newsletter to stay up to date with the latest pharma IP.
This post is based on a previous article on IPKat.
