Patenting AI: Artificial Intelligence and Drug Discovery
The challenges involved in protecting innovations in the drug design process at the EPO

The application of artificial intelligence (“AI”) to drug discovery is expected to have a significant impact on future efforts to discover new and more effective therapies, and indeed at the time of writing, compounds that have been identified by AI-based techniques are already being tested in clinical trials[1]. The traditional approach to drug discovery (target identification and validation, compound screening, hit-to-lead and lead optimisation, etc.) can be expensive and time-consuming, and AI-based techniques may offer researchers in this field additional tools to expedite the process and bring potentially life-saving drugs to patients much more quickly.

AI-based techniques may assist at several stages of the drug discovery process. For example, AI-based techniques have been applied to protein structure prediction, which may assist with biological target identification and validation. Screening of existing compounds, which historically has been performed using large compound libraries, may be streamlined by AI-based techniques that can predict the properties of compounds. Alternatively, AI-based techniques with the capability to analyse large datasets may be applied to the results of high-throughput screening exercises to identify patterns in the data. Further, AI-based techniques may be used to generate wholly new compounds with desirable features based on the target binding site.

In this article, we explore some of the issues around patentability that may arise now and in the future at the EPO, as AI becomes further integrated into the drug discovery process.

Patentability of AI inventions and AI-created inventions at the EPO

As has been explained in previous articles in this series, there is no absolute statutory bar on patentability for AI inventions at the EPO. To succeed, claims must be drafted to include at least one technical feature (e.g., that the method is conducted on a computer). Providing that a claim includes at least one technical feature, any new features must make a technical contribution based on their technical effect in the context of the claim as a whole. During the assessment of the new features, non-technical features (e.g., an algorithm) may also be taken into account. In the field of drug discovery, we expect that most AI inventions would be deemed to produce a technical effect by the EPO. An important factor to consider is whether the technical contribution results in an objective outcome (i.e., one that is not dependent on subjective interpretation by a user). For example, providing novel drug targets, improving the bioavailability of drugs, and determining the affinity of a drug for a molecule, are all metrics that could fulfil the “technical contribution” requirement.

A further potential issue is that of enablement. While sufficiency requirements do not normally cause issues for computer-implemented inventions, in the field of AI it may be much harder to determine whether a method would produce the intended outcome. For example, it may not be self-evident that a series of functionally-defined method steps would always permit a user to identify therapeutically useful drug structures – the ability to do so may depend on the particular training data that was used, or the particular implementation of the method. Indeed, the Technical Boards of Appeal have confirmed in T 161/18[2] that the use of AI per se is not inventive unless the AI achieves a special technical effect in a reproducible manner. Applicants should therefore consider whether worked examples showing that the AI works in the manner claimed should be included in their patent applications, as discussed in a previous article.

When patenting inventions which have been created, at least in part, through the use of AI, applicants should ensure that applications are drafted to meet the usual enablement requirements for any biotechnology or pharmaceutical inventions at the EPO. In particular, applicants should strive to include in vitro or in vivo data supporting the asserted technical effect (such as a therapeutic effect). In silico data may also be used to support a technical effect, but applicants should avoid relying solely on computational data in view of the potential risks of the EPO finding that the technical effect is not plausibly demonstrated.

Will disclosure of AI methods affect drug patents?

Perhaps the most pressing question for applicants in the drug discovery industry is whether they should disclose any information about their AI in patent applications. It is possible that any patents directed to any drugs or new medical uses which are discovered using AI may be far more valuable than a patent for the AI itself, and applicants may wish to keep details of their proprietary processes secret to avoid any competitors gaining access to their methods.

As far as we are aware, the EPO has not objected to any drug-related patent applications on the basis that the drug structure or medical use would have been obvious to a skilled person in the field of AI drug discovery. However, as the use of AI becomes more widespread, it is possible that computational drug discovery or optimisation could become part of the skilled person’s arsenal of tools to be used in the course of routine work and experimentation. In this scenario, it is feasible that the EPO’s requirements for drug applications will become much more stringent, for example the discovery of an alternative, novel inhibitor for a known target may not be deemed inventive unless the inhibitor provides an additional surprising effect that could not have been foreseen. Indeed, we may see the application of similar criteria to those currently required of antibody inventions. In particular, if antibodies against a given target antigen are already known, a new antibody that binds to the same target must have a surprising technical feature (e.g., high affinity) in order to be deemed inventive[3]. This requirement has developed, in part, because the EPO considers that the mere discovery of a novel antibody cannot be inventive due to perceived ease with which antigen-specific antibodies can be raised and screened. It remains to be seen whether similar requirements are enforced for non-biologics, or whether such criteria could become even more stringent as computational methods become available which aid in improving affinity, bioavailability, specificity, etc.

A further consideration is whether the drug structures or medical uses suggested by an AI are even “inventive”. It is a common misconception that AI will always find the best solution to a problem, but simply applying AI to a problem does not guarantee that the output will be any better than if a team of skilled people were posed the same problem. In other words, the fact that AI was used to find a solution does not guarantee that the solution would not have been obvious to a skilled person or skilled team. It is therefore important for applicants to conduct the same patentability analysis before filing an application that they would conduct for drugs or medical uses which are discovered traditionally.

Other potential uses for AI and computation include target identification and affinity measurement, and the ability to substitute traditional assays for in silico simulations. Supporting data are usually required for applications for new drugs or new medical uses. These data can include binding assay data showing that a drug targets a particular molecule which is known to be implicated in a disease, affinity data showing that a drug binds to a particular molecule much more strongly than prior art drugs, epitope data, or data showing that a drug does not bind to molecules which are implicated in off-target effects. Clinical trial data is not normally required unless the technical problem to be solved is particularly ambitious, such as the provision of a synergistic combination. There is a possibility that, particularly as AI and computational methods improve and become more commonly used, the EPO may accept in silico data in place of traditional “wet lab” assays. This would enable the generation of such data much more quickly and easily. Moreover, this may be of particular benefit in cases where applicants are asked to provide comparative data showing that their claimed solution is better than a prior art solution – producing such data would be far less burdensome if in silico assays could be used. Of course, the EPO may also increase the pressure on applicants to provide supporting data in patent applications in response to the increasing ease with which such data could be generated. For example, the EPO may restrict the ability of applicants to extrapolate experimental results across a broader claim scope or may raise the threshold for plausibility.


In summary, applicants should continue to include as much data as possible in initial filings, preferably including in vitro and in vivo data wherever possible. Applicants should also ensure that any description of computational methods contain enough detail to enable a skilled person to recreate the method, and any training data should be described as thoroughly as possible. Before disclosing any aspect of a proprietary AI-based method, applicants should carefully consider whether the potential reward of a patent outweighs the risks entailed by disclosing your AI processes in a patent application, namely the usual risks associated with providing proprietary information to competitors and the risk that the disclosure of any computational processes may prejudice later applications.

AI may not yet have reached a point where the skilled person could simply use AI to solve a technical problem such as providing an improved drug structure or finding a new medical use for a known drug. However, there is a risk that we may begin to see objections at the EPO in the coming years that any AI-devised solutions to such technical problems are obvious. At the same time, the increasing use of AI may, in time, incline the EPO (and other patent offices) to accept in silico data, such as AI-generated data, as evidence of useful real-world effects. This will potentially ease the burden on applicants seeking to generate supporting data for new filings, though may encourage the EPO to require an increasing amount of supporting data in applications.