Artificial Intelligence (AI) is changing the way we live and engage with the world. Besides smart devices (phones, speakers), and self-driving cars, AI is also set to revolutionize biotechnology and the health sector. Deep learning algorithms are being used in areas such as genomic data analysis, cellular profiling and characterisation which find use, for example, in cancer diagnosis, drug design and pharmacology, genome sequencing, and protein structure prediction. With spending on AI in healthcare projected to grow exponentially in the coming years, there is likely to be a boom in biological innovations harnessing AI technology. In such a scenario, obtaining patent protection for an invention ensures a way of securing innovator’s rights in an already crowded market. This article provides a guide to the necessary requirements for obtaining a patent in AI in the area of biology.
Are AI inventions patentable?
Machine learning (ML) and deep learning (DL) are the branches of AI that are used most frequently in the pharma and biotech industry, and for the purposes of this article, these terms are used interchangeably. As such, AI inventions use a complex system of algorithms to achieve intended results. While the key to the invention may lie in the code, the algorithms themselves are not patentable in most jurisdictions. It becomes necessary, therefore, to link the use of the algorithms to some tangible result. This becomes clearer from the guidelines provided by various patent offices.
The United States Patent and Trademark Office (USPTO) is yet to provide any special guidelines regarding AI, and inventions using AI are treated in the same manner as those using mathematical models and algorithms. 35 U.S.C. § 101 lists mathematical models and algorithms as non-patentable subject-matter. But if such mathematical constructs are linked to a practical application, then they are deemed patentable. The European Patent Office (EPO) AI guidelines regards AI inventions as computer implemented inventions. The EPO requires that AI inventions must have a specific technical purpose, i.e., they must be linked to a specific use, and the technical effect of the use of the algorithm needs to be specified. For example, the use of a neural network in a heart-monitoring apparatus for identifying irregular heartbeats is considered a technical purpose. This criterion is usually met by most biological inventions using AI. However, this also requires that the step-by-step interaction between the claimed algorithm(s) and the computer hardware, network or the cloud, must be reflected in the claim language in order to make the subject matter patentable. Such an approach is also recommended to overcome objections under Section 3(k) of the Indian Patents Act, 1970. The Computer Related Inventions (CRI) guidelines issued by India’s Intellectual Property Office (IPO) recommend including the physical constructional features of carrying out the functions of the mathematical methods in the claims.
The caveat here is that the generic use of terms referring to computer hardware is not sufficient; the method must clarify the relationship between the math and the computer/network components so that the terminology does not appear superfluous.
Who is the inventor of an AI invention?
As heuristic algorithms become more refined, and lessen human intervention in the development process, the actual inventor of an AI invention becomes ambiguous. Identifying the true and first inventor(s) of an invention is necessary in the application process. In such cases, who should be named as the inventor or inventors of an AI invention – a human being or a machine? This question, although pertinent, has a more theoretical relevance which can be resolved by the clarification of the legal definition of an ‘inventor’, which until now has primarily included natural persons.
The USPTO issued a notification in August, 2019, in this context, which included the following questions: “What are the different ways that a natural person can contribute to conception of an AI invention and be eligible to be a named inventor?”; and, “Do current patent laws and regulations regarding inventorship need to be revised to take into account inventions where an entity or entities other than a natural person contributed to the conception of an invention?”. Such questions can help further clarify the definition of an inventor going forward.
The more relevant question, arguably, though, is, who should own the rights to an AI invention if there is ambiguity of inventorship? This question was recently resolved at the EPO, which denied the rights of two inventions to DABUS, an AI machine designated as the inventor in an application. The human applicants of the patent applications argued the machine should be recognized as the inventor and the applicant, and the owner of the machine, should be regarded as an assignee of any intellectual property rights created by the machine. The EPO however refused the applications with the machine designated as the inventor under Article 81 and Rule 19(1) of the European Patent Convention (EPC).
In its decision, the EPO concluded that the inventor designated in a European patent must be a natural person and that the understanding of the term inventor as referring to a natural person appears to be an internationally applicable standard. It stated that Article 81recites that if the inventor is not the same as the applicant, then the application should include a statement indicating the right to the European patent. Rule 19(1) requires that the designation of the inventor be given which includes the family name, given name, and address. The EPO noted names given to things may not be equated with the names of natural persons as things have no rights which a name would allow them to exercise i.e. AI systems have no legal personality. It stated that, AI systems or machines cannot have any legal title over their output which could be transferred by operation of law or agreement and that the question of ownership of an output is different from the question of the rights associated with inventorship.
Thus, the EPO is unlikely to accept applications with a machine as an inventor, for now. Such applications may also be denied in the future as the EPO cited several documents demonstrating that the legal intent of the EPC was define an inventor as a natural person, i.e., a human being. Whether other jurisdictions will follow the same route remains to be seen.
Drafting a patent application for an AI invention
Navigating this framework through the careful drafting of claims is the first step to ensure that an AI patent application does not fall into the category of non-patentable inventions. However, two other hurdles that are likely to arise are (i) the inventive step requirement, and (ii) objections related to sufficiency of disclosure.
The inventive step requirement is similar in most jurisdictions, and is primarily determined by the closest prior art documents. In general, the patent practitioner needs to convince the examiner that the invention is not obvious from a reading of the cited prior art documents. AI inventions face the hurdle that if a problem was solved previously by simple mathematical analyses, using a DL or ML system in their place will not overcome an inventive step objection. The Japanese Patent Office (JPO) AI guidelines help in understanding this aspect through examples. For instance, if cancer levels were previously detected by measuring the levels of two markers, A and B in cells, using an ML system that uses training data related to the levels of marker A and B to do the same would not count as an inventive step. However, if an additional component, such as cell size, were added to the training data, this would count as an inventive step as the previous measurement criteria did not include this parameter. Similarly, a minor modification of the algorithm to achieve better results is likely to be viewed as routine optimization unless it can be shown that it leads to a superior technical results and that such a method has not been applied before.
In addition to inventive step challenges, AI inventions are also likely to receive objections related to insufficient disclosure of the invention in the specification. The claims must describe the essential features of the inventions in as much detail as necessary in the specification of a patent application. For instance, the EPO AI guidelines specify that terms such as ‘neural network’, ‘reasoning engine’, and ‘support vector machine’ may refer to abstract models or algorithms. Therefore, it is necessary to define these terms in the specification, in relation to their specific technical use in the invention. Most jurisdictions require that an invention should be described in a manner that a person skilled in the art is able to replicate the invention. This becomes particularly necessary in the case of AI inventions as they are inherently complex and there is a greater chance of relevant pieces of information being left out. In some instances, an explanation of the algorithm in pseudo code may help explain how the algorithm initiates the intended result. This is may be particularly useful when differentiating an invention from the prior art or for explaining the link between the mathematics and the technical effect.
Another useful practice is to include the details of the training data used to train the algorithm. The JPO requires that the correlation between multiple types of data be specified in the specification, unless the relations are common general technical knowledge. For example, if the object of the algorithm is to determine the Body Mass Index (BMI) of an individual from his face, it would not be sufficient to state that facial details are used to derive this result. Specific correlations such as measurement of the face-outline angle etc. should be described in relation to their use to determine the BMI.
Finally, it is also recommended to validate the results obtained though the ML or DL system. The JPO, for instance recommends that products obtained through AI must be evaluated to confirm that the invention is reproducible. This is clarified in from an example relating to an anaerobic adhesive composition in the JPO’s AI guidelines. In the example, a trained model is used to formulate an adhesive composition and to provide the estimated curing strength of the composition. However, the specification does not disclose any experiment where the adhesive composition is actually produced to test the prediction. Without a ‘real-world’ validation of the predicted results, the JPO would likely reject such applications for insufficient disclosure. The JPO further specifies that post-filing data would not be considered to tackle such an objection. However, in cases where it is not possible to experimentally validate AI predictions, the models can be tested on already available data/products to validate the results.
In summary, AI inventions can be patented, and although many of the major patent offices across the world are yet to provide specific guidelines, the general trend points to the broad acceptability of such inventions. In any case, the applications for such inventions must be drafted carefully, ensuring that the claims clarify the specific relationship between the computer infrastructure and the abstract mathematics. A lack of detail is likely to result in the death of an application. Most jurisdictions specify the need to describe the ML and DL methods in sufficient detail to carry out the invention. This is also a useful practice when the nuances of the invention may help to overcome inventive step issues. Most importantly, for inventions in biology, experimental validation of the results obtained through AI is recommended, and in the event that this is not possible, testing the system on previously published results or data may be sufficient.
JPO AI Guidelines:
Recent decision at the EPO refusing patent applications with an AI inventor:
CRI Guidelines, Indian Patent Office: