Is it worth patenting your AI tools for drug discovery?

About the author

Parminder Lally is a senior associate at Appleyard Lees IP LLP. Parminder specialises in drafting and prosecuting patent applications for computer-implemented inventions. She has built a substantial reputation working with high-growth start-ups, spin-outs and SMEs in Cambridge, and has in-house experience.

 

After a year of working from home during the global pandemic, many companies have reconsidered their flexible working policies, and employees may now have more freedom to decide where they work.  This may extend beyond simply working from home some of the time to permitting employees to move to, and work in, another country entirely.  Even during the past year, people temporarily moved to be nearer to family abroad or to take advantage of better weather overseas.  But what are the implications for tech companies of having employees based anywhere?  In this article, we take a look at one possible issue. 

 

In the last few years, there has been a lot of interest in how artificial intelligence (AI) can be used to improve the drug discovery and development processes.  During the global Covid-19 pandemic, many people will have become aware of just how long it takes and how much it costs to develop a drug and release it onto the market.  AI is being used by many companies to try to speed-up and increase the efficiency of the process to identify new drugs for new and existing diseases.  Sometimes, the AI is used to discover existing drugs that could be used for a different purpose.  Often, when the AI identifies a drug, and experiments confirm it is useful for the intended purpose, the drug can be patented.  But what about the AI itself – can the AI be protected by patents, and should you try to?  In this article, I will share some thoughts on whether it is possible to patent AI tools for drug discovery and drug design.

AI Tool Types

Many AI tools used for drug discovery are based around natural language processing (NLP).  This is because huge amounts of useful biomedical data is contained within documents aimed at human researchers: academic journal articles, patents, clinical records, and so on.  One reason traditional, human-powered drug discovery takes time is that humans cannot possibly trawl through all the existing material and spot patterns or possible drug candidates.  In contrast, NLP algorithms can be designed to do exactly this, because they can ‘read’ these documents and identify biologically relevant pieces of information, such as the names of genes, proteins, drugs, diseases, and so on.  The NLP algorithms can also identify patterns and relationships between these words, which may suggest, for example, that a particular gene is related to a particular disease pathology or etiology.  These patterns and relationships may prompt scientists to consider whether a particular drug or type of drug might be suitable for treating a disease.

The patterns and relationships identified by the NLP algorithms can be combined with other data, such as data from other fields, to create a “knowledge graph”.  For example, the knowledge graph can combine the insights from the biomedical data with chemical data, drug data, and disease data.  The knowledge graph is used to store links between such data, and the links may enable better predictions to be made on which therapeutic products could be used to safely and effectively target a particular disease or condition. 

Another type of AI tool that may be used for drug discovery is computer vision.  Again, there is lots of information in medical images that are collected of patients.  Computer vision algorithms may be used to identify features in the medical images which are present only in patients with particular diseases.  They can also be used to identify features in the images that are early indicators of disease onset or disease progression, which can then be used to determine suitable drugs for the various stages of the disease.   

AI tools may also be used to design new drug molecules.  Algorithms could be used to tweak existing chemical compounds or antibody drugs so that, for instance, their efficacy is improved or their toxicity is reduced, Similarly, new molecules may be designed from scratch, which have various desired structural, chemical or biochemical features that are known to be useful for treating particular diseases.

Can these AI tools be patented?

Generally speaking, AI methods can be patented at the European Patent Office (EPO), as long as the AI methods provide a technical solution to a technical problem.  However, as with most aspects of law, the devil is in the details!  Let’s take a look at each of the tools mentioned above separately.

(a) Computer Vision

AI and machine learning (ML) algorithms which are directed to image processing are patentable at the EPO. The EPO’s Guidelines for Examination explicitly state that the classification of digital images and videos by, for example, performing object recognition or feature detection, is considered technical because the algorithms typically involve analysing pixels of images.  Similarly, techniques to enhance the images so that they can be more easily interpreted are considered technical. 

Therefore, if you have developed an AI tool that involves analysing or classifying medical images, the AI tool may be patentable (subject to what is already in the public domain, of course). 

(b) Design

The use of AI to modify existing molecules or generate new molecules from scratch may be patentable at the EPO.  The computer-aided design of drug products is patentable if the design is based on technical considerations.  In this case, it is likely that parameters needed to obtain a suitable drug product, such as the structure, function, toxicity level, synthesis method, and so on, are based on technical considerations. 

(c) NLP

The EPO has long-held that techniques for classifying text documents or data records solely in respect of their textual content is not considered a technical purpose but a linguistic one. 

In order for the documents mentioned above to be used by AI algorithms, they may need to be classified or otherwise augmented first before they can be used to train an algorithm to identify patterns.  However, the EPO’s stance means that methods for classifying data records – such as academic papers and clinical data – so that they can be used to generate a training data set are not patentable. 

Furthermore, the EPO does not allow new NLP algorithms to be patented if the only difference between the new NLP algorithms and existing ones is the underlying linguistic model.  This is because the linguistic model is not considered by the EPO to make a technical contribution. 

It may seem strange that AI for image processing is acceptable but AI for text processing is not, but unfortunately the EPO has held this view for several decades.  In the decision in T1316/09 from December 2012, the EPO considered a method for suggesting automated responses to an incoming electronic message based on content analysis and categorization (i.e. a predictive text method) to be unpatentable. This was on the basis that the methods of text classification were not considered to produce any relevant technical effect or provide a technical solution to a technical problem.

Similarly, in T1784/06 from September 2012, the EPO considered an algorithm for classifying data records more efficiently to lack an inventive step because the algorithm was not used for a technical purpose.  Essentially, the EPO considered the classifying of the data records to be non-technical because the classified records were used to improve a billing procedure, which is considered to be an administrative or commercial purpose rather than technical.

The precedent for these decisions comes at least in part from the EPO’s decision in T0052/85.  In 1989, the EPO decided that a method of generating a list of expressions semantically related to an input linguistic expression using a programmable data processing system was unpatentable because the method lies in the field of linguistics.  The semantical relationships were not considered technical in nature – the relationships were determined based on the (linguistic) meaning of those expressions rather than anything technical.  The EPO considered the invention to provide no contribution outside of the field of linguistics or the field of conventional computer performance. 

It is important to note that if a linguistic model or NLP algorithm serves a technical purpose (such as the discovery of a new drug target for treating a disease), the steps of generating the training data set and training the model/algorithm may contribute to the technical character of the invention if they are linked to achieving that technical purpose. 

The EPO will usually divide the features of a patent claim that is directed to using linguistic models or NLP algorithms to mine medical datasets or documents for the purpose of drug discovery into technical features and non-technical features.  For example, a computer-implemented model may be used to trawl through academic papers, determine relationships between diseases and potential drugs, and provide a score for each relationship which indicates how suitable the drug may be for treating that particular disease.  The computer is a technical feature, and any steps to input the papers into the model may be technical.  However, the relationships and scores are non-technical features.  These non-technical features need to be linked in the claim to the technical purpose of the claim in order to contribute to the technical character of the invention.

Therefore, it may be possible to patent an NLP algorithm for drug discovery if the claim explicitly recites how the algorithm is used for that specific technical purpose, and in particular, how the steps performed by the algorithm contribute to discovering a new drug. 

Should you go down the patent route?

There are many reasons why companies may want to file patent applications for their AI tools, even if the grant of a patent is unlikely.  One of those reasons is obviously because they want to obtain a granted patent for their AI tool.  This may be because the company is considering commercialising the tool itself, and wishes to stop others from, at least, offering the same tool for use to others. 

Another reason for starting the patent process is for defensive purposes.  The company may simply want to patent the AI tool to stop others from doing so.  Once the patent application is published, it will be in the public domain and form prior art against any later filed patent applications for the same or similar AI tools. 

Another reason is that it demonstrates to the public what the company is working on, which may create some buzz and interest in the company (which is useful for acquiring funding, commercial partners or interest from potential acquirers).  Even if the patent application may ultimately not be successful, it is likely that the patent application will be pending for several years, during which it contributes to the size of the company’s patent portfolio.

Some companies working in the AI drug discovery space have not made any attempts to patent their AI tools, and have instead focused on protecting the drugs that are identified by their algorithms.  Protecting the AI tools as trade secrets or know how may be useful for tools that may not be patentable in the first place, but which provide the company with its competitive edge or USP. 

Conclusion

If you have developed AI tools for drug discovery or drug design and would like to discuss how you should protect it, get in touch with us. At Appleyard Lees, we have a dedicated team of patent attorneys who specialise in protecting AI inventions, and we can advise you on whether the patent route is suitable for your AI invention and your business goals.

Author

Related posts