Personalized medical intervention is in a transformative phase as artificial intelligence (AI) algorithms are increasingly deployed to tailor treatments for individual patients based on their unique characteristics. Developers face the task of adequately protecting AI-based technologies under current US patent jurisprudence. Enablement and necessary disclosures to obtain patents related to AI are one such challenge.
Roughly a year has passed since the US Supreme Court delivered its enablement ruling in Amgen Inc. v. Sanofi, where the court refined the boundaries of undue experimentation. In Amgen, the challenged patent included broad functional claims to any antibody that bound to a protein and functioned a certain way.
The Supreme Court found that 26 examples of antibodies in the Amgen specification were insufficient to support claims to a broad genus of functionally defined antibodies absent some enabling guidance. This ruling raises questions for the patentability of AI-related inventions in precision medicine.
Black Box Problem
AI models that employ deep learning or complex neural networks are difficult to describe. These models may be based on a small amount of actual software code but are instead trained using voluminous datasets—such as genetic profiles, clinical histories, and treatment outcomes—to identify patterns that inform tailored therapeutic approaches.
While the resulting outputs can be reliable, the model’s operations frequently occur within a “black box” that can be difficult to describe in retrospect. Providing an enabling disclosure sufficient to support claims to the AI system or model, particularly those with functional terminology, can be challenging.
Because these models are often built through trial and error, the iterative process used to create the model may be as relevant to its function as the underlying training data. Disclosing training data alone may not provide the “general quality” thought to run “through the class” unless accompanied by detailed accounts of each specific trial and error during the development.
- Will training data need to be disclosed in patent specifications?
- If aspects of the training data cannot be disclosed, are there other solutions for properly enabling AI claims?
- Is it plausible to shift the enablement inquiry in AI applications to overall functionality rather than requiring a detailed disclosure of the inner workings of the model?
Though the US Patent and Trademark Office (USPTO) is actively exploring solutions and seeking public comment on how “patent applications for AI inventions [can] best comply with the enablement requirement,” no affirmative guidance exists on the subject.
AI Models’ Evolution
AI models used in personalized medicine typically aim to be dynamic and accommodate evolution over time with new data, improved algorithms, and enhanced computational resources. This continuous development is crucial for increasing the models’ accuracy, efficiency, and versatility.
Narrow AI claims allowed today may not be relevant to the models being used six months or six years later. Providing an enabling disclosure that supports broad claims covering the underlying concept of the model may not be trivial.
- What claiming strategies are sufficient to allow AI patents to achieve commercial relevance over the life of the patent and the product?
- What types of data and description might be necessary to support a model that evolves?
- What types of description would satisfy the Amgen requirement of a “general quality” that gives the class its “peculiar fitness” for a particular purpose?
These questions remain largely abstract until the USPTO provides affirmative guidance, and the Federal Circuit applies Amgen to AI claims.
How Section 112 law applies to AI claims in view of the Amgen v. Sanofi holding remains uncertain. Patent disclosures will need to provide detailed and clear descriptions of the development process, including the iterative training of models, to navigate the post-Amgen landscape. Practitioners should also consider whether their AI model technology is more suitable for trade secret protection.
The case is Amgen Inc. v. Sanofi, U.S., No. 21-757.