LLMs will undeniably have a major impact on the legal industry, but are any of these tools suitable for the pharma and biotech industry? PatKat has decided to take a look to see if any of currently available tools can assist with patent drafting for life sciences. As a caveat, these reviews will focus on the technical capabilities of the software and not the pricing or security features of the tools.
LLMs for patent drafting
We are now awash with companies claiming to provide LLM software capable of drafting patent applications. Whilst there are already a number of blogs out there reviewing the different companies, for this Kat, patent drafting is a subject-specific activity. Patents in the life sciences field are very different to those in the mechanical and software fields. The process of drafting patents in life sciences is correspondingly also very different. In life sciences the focus is very much on the data and how this supports the claims. The claims in life sciences require very domain-specific details about the drug modality, such as sequence information and chemical structure. So how do the existing AI software tools cope with biotech and pharma inventions? The first tool put through its paces by IPKat is the AI patent drafting software from Qatent (now part of Questel).
 |
| Mechanical and life sciences patent drafting: Worlds apart? |
Qatent's patent drafting software
- A "hybrid AI Model" that allows user input from the patent attorney into what the LLM is doing.
- A "paraphrasing tool" that promises to "give your patent application the correct scope by choosing among different variations of the same sentence/phrasing."
- A definitions tool that detects technical terms in the text and proposes alternative definitions, derived from Qatent's "extensive terminology database".
- "Relation tables" that "keep close tabs on how your claims express the relation between terms. Find typical errors, missing links, etc."
- A Diagrams generator, with the current offering producing "simple step diagrams".
The Qatent tool user interface is very simple and easy to use. The claim and description drafting and editing happens within a browser to which you can upload documents of various types. The finished claims and/or complete application can then be exported (clean or with mark-up) in PDF or docx format. This flexibility is appreciated. Overall the user experience is simple and pain-free (not always the case with new software!).
The Qatent tool provides two workflow options. The user can either upload an Invention Disclosure for drafting of the claims and description, or the user can upload their own claims. The optional workflow is a good feature, allowing the user to draft their own claims and use the tool for drafting the description.
Multiple types of document can be uploaded to the Invention Disclosure, including PDFs. It is also possible to copy and paste plain text into an Invention Disclosure dialogue box. If the user is uploading their own claims, these can also be pasted into a dialogue box or uploaded.
In a potentially nifty feature, the tool will automatically extract figures from the uploaded document. However, in this Kat's test example, in which the input was a scientific paper in PDF format, the figures appeared as blank black squares. Presumably, a simpler image format is needed. There is also the option of just uploading the figures separately.
The AI outputs can be edited within the tool itself, and then exported in PDF or docx format (with or without the notes provided by the tool).
Claim drafting based on data
The first test example this Kat tried was a dose invention for a monoclonal antibody. To test the AI claim drafting software, the data uploaded was a scientific publication describing in detail the identification of the optimal IV dose of a monoclonal antibody for a phase 3 clinical data based on data from a phase 2 clinical trial. The target claim was therefore a method of treatment claim for treating the relevant indication in the subject, comprising administering the antibody to the subject at the identified optimal dose.
In the first test, only the data document was uploaded without any additional description of the invention. Without the context of an invention disclosure the software clearly struggled to identify the invention. The first claims suggested by the AI software, for example, included a method for determining an optimal dose of the drug, as opposed to the more appropriate claim for method of treatment comprising administering the optimal dose of the drug.
The tool also hallucinated incorrect subject matter for claims. The software suggested, for example, claims for a random drug combination that was not mentioned in the data, a whole list of odd antigens for the antibody (despite the data being very clear on the identify of the antigen), and suggested some claims that assumed the monoclonal antibody was a small molecule (e.g. by suggesting the monoclonal antibody "or its salt"). The target and identity of the drug as a monoclonal antibody was very clear from the data, so it is unclear why the software chose to ignore this information. The software also failed to suggest any claims in different formats that are relevant to this type of invention, e.g. EPC 2000 second medical use.
Claim drafting based on data and the Invention Disclosure
In a second test, the same data was uploaded, but this time with a brief Invention Disclosure describing the invention as the optimal IV dose of the monoclonal antibody. The Invention Disclosure specified the name of the drug, the amount and frequency of the optimal dose, and the disease to be treated.
The software performed much better having been given this additional information. The first claim and the first 3 dependent claims were passable method of treatment claims, indicating the correct drug, indication and optimal dose, as identified in the invention disclosure. However, the subsequent dependent claims descended into random features that made little or no sense based on the data.
The claim dependencies also needed work. The developers have clearly tried to avoid incorrect dependencies in the claims, and lack of antecedent basis, by telling the tool to limit the use of multiple dependencies. Most of the claims thus depended only on the first independent claim. In terms of ensuring sufficient basis for all combinations of features, this is less than ideal. The structure of the claims was also odd, with claims dependent only on independent claim 1 alternating with claims dependent only on independent claim 2.
Claim drafting: Conclusion
The major issue with claim drafting was that the Qatent software does not understand different categories of pharmaceutical invention or drug modalities. Overall, the Qatent tool is not fit for drafting claims for this type of invention. It is possible that crafting a highly detailed invention disclosure might have improved the situation. However, the data uploaded was sufficiently detailed such that the invention should have been easy to understand and identify. It is therefore not clear that writing a detailed invention disclosure would have helped. Also, by the time one has finished optimising the invention disclosure, one might as well have either just manually drafted the claims or used a LLM model directly.
Description drafting from the Claim
One of the more tedious tasks of patent drafting is converting and expanding the claims into appropriate basis. To this Kat, this process is definitely long-hanging fruit for automation with AI. So how did the Qatent tool perform?
To test this functionality, this Kat uploaded her own drafted claims for the dose invention used in the previous test, together with the data and the invention disclosure.
The basis section provided by the software repeated the text of each claim in the appropriate format for the description. However sometimes crucial information was lost in the transfer. For example, for the first claim, the embodiment was described nonsensically as being the antibody "administered at a dose of about mg every weeks". This unfortunately does not inspire confidence.
For each claim, the tool also suggested some text describing the advantages of the features in the claims. Some of these were helpful. However, despite the data source document providing detailed information on the advantages of the invention, the majority of these suggestions were very high-level and vague, providing little value. The extensive definitions that would be normally expected in a patent application relating monoclonal antibody for a therapeutic use were lacking.
Technical field and background information
In a patent specification, the description will often include a short paragraph on the technical field. The Technical Field paragraph provided by the tool was clearly taken from the user inputted International Classification number and was therefore too broad. For example, for the monoclonal antibody dose invention the technical field was said to include "chemical compounds or medicinal preparations and preparations for medical, dental or toiletry purposes". The "Background Art" section generated by the tool was also very short, around 1 paragraph. The Background also appeared to be less of an introduction, and more or a description of the invention.
Final document format
After you have finished editing the draft in the Qatent tool, you can export it in word of PDF format. The exported word document had line numbering headings, but no other word formatting, e.g. no bold headings, no use of word Styles for headings for easy navigation and formatting, and the claims were not a numbered list. This made making further edits to the document rather cumbersome.
Final thoughts
Qatent's AI patent drafting tool shows promise with its user-friendly interface and flexible workflow options. However, the current offering falls significantly short when handling specialised life sciences patents. While the tool performed better with explicit invention disclosure information, the software still struggled with understanding pharmaceutical invention categories and drug modalities, leading to hallucinated dependent claims and missing critical formats relevant to biotech applications. The description drafting functionality suggested only vague advantages for the claimed features and lacked the specialised definitions crucial for pharmaceutical patents. The final document formatting also requires substantial manual refinement. For patent professionals in the pharma and biotech sectors, the Qatent software currently appears to be a supplementary tool at best, requiring significant human oversight and expertise to produce acceptable patent applications in these specialised fields.
However, for this Kat, the inadequacies of the Qatent tools relate more to the particular requirements of life sciences patent drafting as opposed to the overall capabilities of the tool itself. Qatent was developed, like so many of the AI-patent software tools, by AI developers and patent attorneys with a background in software and mechanical drafting. It is therefore not surprising that the resulting software does not cope well with biotech or pharma drafting, which requires highly specialised field-specific expertise and an entirely different approach to mechanical or software patent drafting.
The search for AI patent drafting software for life sciences continues! Do readers have any suggestions or recommendations of AI tools for pharma and biotech patent drafting?
Further reading
Reviewed by Dr Rose Hughes
on
Tuesday, February 25, 2025
Rating:
Follow the IPKat on LinkedIn 
Posts
ReplyDelete"...The exported word document had line numbering headings, but no other word formatting, e.g. no bold headings, no use of word Styles for headings for easy navigation and formatting, and the claims were not a numbered list. This made making further edits to the document rather cumbersome..."
Not sure there are many attorneys who would actively choose to include word styles and numbered lists in their drafts. They are a recipe for disaster with broken ref codes and the like sneaking into filed docs. US-first drafts are the worst for this.
It actually highlights how and why these kinds of tools are doomed to fail on UI/UX. Just because you like to draft using certain styles and formatting, it does not mean everyone else does. This means tool providers end up trying to satisfy too many individuals' quirky demands and end up with a patchwork of features, most of which are useless for most users.
It's similar to what happens in most IP management systems. Ever wonder what 99% of the features are in your IPMS that no one on your team uses? The answer is that it was one corporate client asking for feature X, and the the tool provider implementing it. Then another asking for feature Y, and the tool provider implementing it, and so on. Neither uses the features built for the other. The features are sort of overlapping but it's not clear why and that has long since been forgotten. The tool provider doesn't want to combine them because they were explicitly requested by big corporate clients.
It turns out that the profession is too small for there to be a universal, generalizable set of "best" features in these kinds of tools. Given that it is easier than ever to build cheap, custom, in-house solutions that do everything you'd want according to your team's style and requirements, while using the exact same models that every over-priced 3rd party tool is selling, there is no excuse for most firms not to in-house their drafting tools.
"Not sure there are many attorneys who would actively choose to include word styles and numbered lists in their drafts"
ReplyDeleteSo do you manually write out and correct all the claim numbers whenever you change the claim? What do you do when there are a hundred of claims? Do you also manually write the page numbers? For an industry that definitionally works at the forefront of technology, it seems that some patent attorneys are remarkably slow at adopting even the most basic of automation tools and in realising that automation in most cases decreases, not increases, the risk of error.
In my experience, if you are doing anything more complicated than adding a dependent claim to an independent claim, automatic numbering fails more often than it works. An important client of mine forbids automatic claim numbering for this reason.
DeleteThere are never hundreds of claims to renumber because that would cost a huge amount of money and I have never paid an excess claims fee. If your clients can afford to pay hundreds of claims fees or allow you to charge them for writing hundreds of claims, you do not need to worry about how efficiently you can renumber claims because your clients clearly have more money than sense. It takes seconds to renumber a set of 15 claims.
Thanks you anons for your comments. Possibly this is one of those examples where field of invention matters? It would be rare I believe for priority application in the pharma field to be filed with only 15 claims, for a variety of reasons. Amendments to avoid claim fees can be made at the national/regional phase where necessary.
DeleteRESPONSE BY QATENT CO-FOUNDER 1/2
ReplyDeleteDear Rose,
Thank you for testing qatent and sharing your thoughts.
I am one of the two co-founders of qatent (my cofounder is Kim Gerdes, Professor in Computational Linguistics at Paris-Saclay University, and former head of the master track in AI). As a qualified French patent attorney, I worked for patent law firms, IBM and Roche Diagnostics: I know the very high expectations of my former colleagues in pharma.
Please find below some thoughts:
- qatent did its computations on all IPC/CPC classes, including C and D. Qatent is workable for chembio but is surely not optimized for this huge and diverse field (at least as of today). But again, it is workable. We focus on the "fundamental" layer, mechanics, software, medical devices, etc. We make steady progress (new release every month in average, so please come back and test again!) but there are tons of challenges for this "very basics" layer (e.g. jurisdiction-dependent drafting styles, patentability diagnostics at claims' drafting, handling of consistency, handling of sufficiency of disclosure partly, machine vision to speed-up the description of figures, reference numerals management, etc).
- with respect to the generation of claims: they are not fit-for-purpose "out of the box". It lacks the incorporation of business insights (single or multiple providers, etc) and regulatory insights (e.g. mandatory redundancy of systems in medical devices), for example. Maybe this will change in the future. We are working hard to sophisticate this generation of claims, and our next release will likely improve the situation. But please note that today, you can already use the chatbot and reprocess claim sets (e.g. rephrase method claims into system claims, handle product-by-process, handle Markush claims, etc).
- the proficiency in the use of chatbots is highly variable amongst our users. A chatbot is like a virtual partner that knows a lot of things (like the skilled person, it knows all patents on earth), but is not proactive (not yet), it will respond only. If asked the right questions, you can get amazing value. It is your role as a patent attorney to dive deeper and deeper. The deeper the question, the better value you extract from the AIs. At qatent, we try to facilitate this extraction but we cannot (and do not want) to automate 100%. We encode the "canonical" best drafting practices, which may be considered as "non-debatable", to guide the AI models. This minimal set of rules is not trivial and we probably can discuss subtleties endlessly. As an aparté, for example, in my opinion, the problem-solution approach shall be encoded by the IP community as a whole, including patent offices (yes!), not by proprietary systems ... Code was law, now law is code, but who is writing the code ? For chembio, you know better than me that clinical data cannot be faked, that plausibility / credibility of technical effects is at stake, etc. Decision G2/21 is super interesting in this view.
RESPONSE BY QATENT CO-FOUNDER 2/2
ReplyDeleteBeyond clients, we are looking for partners in this field. Qatent can hyper-specialize AI on top of the base brick for specific verticals (e.g. cosmetics, etc).
So as a temporary conclusion: please continue considering this market as an emerging market, and please come back to test again. Meanwhile, if you have feature requests, from baby steps to big challenges, come find us and discuss !
All the best,
François Veltz
Qatent Cofounder