How Do You Characterize AI’s Contributions in Drug Discovery?

The Bottom Line

• As AI use becomes more frequent in the invention of new drugs, it’s important to determine how it affects who’s credited with patentable inventions.
• Using AI for drug discovery can require significant collaboration between AI and scientific experts.
• Companies using AI tools in drug discovery programs should keep detailed records to demonstrate how the inventor contributed to the invention’s conception.

Artificial intelligence-based drug discovery platforms are accelerating every phase of drug development, promising the rapid development of new chemical matter, the identification of new biological targets, and more predictive clinical trial algorithms. As the application of AI systems becomes intertwined in the invention of new drugs, there are still questions about inventorship of patentable subject matter obtained using AI systems.

For example, when must an applicant list a person responsible for training an AI model as an inventor? How would their contribution be characterized? Understanding the type of AI model used to discover the new drug can aid in the analysis.

AI-Assisted Inventions

All inventors who have contributed to the claimed invention must be listed on a patent application. In late 2025, the US Patent and Trademark Office issued revised inventorship guidance for AI-assisted inventions, rescinding the February 2024 guidance. Former guidance mandated that inventorship analysis of an AI-assisted invention was to be carried out under the framework of the US Court of Appeals for the Federal Circuit’s decision in Pannu v. Iolab Corp. to account for an inventive contribution by AI.

The newest USPTO guidance acknowledges that the Pannu factors remain applicable for joint inventorship analysis when multiple human inventors are involved. However, the Pannu factors no longer apply when a single person uses AI to develop an invention. Because an AI system isn’t human, there is no joint inventorship to analyze.

This means the USPTO effectively treats AI systems as a research tool. The 2025 guidance reiterates that the standard for determining inventorship—from Burroughs Wellcome Co. v. Barr Labs. Inc.—focuses on conception: “the formation in the mind of the inventor, of a definite and permanent idea of the complete and operative invention.”

Assessing Inventive Contributions

While this description of conception is theoretically straightforward, applying it to AI-assisted drug discovery can be challenging. The process of discovering drugs by relying in part on AI architectures can require significant collaboration between AI and scientific experts.

Conception in AI-assisted drug discovery may even depend on the type of AI model used in the drug discovery process. Given these variables, determining who played a role in the conception of the resulting compound is complicated. For example:

• Who could be listed on a patent application as an inventor?
• Should the person generating, troubleshooting, or even supervising the AI model be included?
• What if a compound was discovered by generative AI?

While the newest USPTO guidance supports a straightforward determination of inventorship when a single inventor uses AI as a tool, current guidance fails to consider differences in AI models or the collaborative nature of drug discovery. Following USPTO guidance, the Pannu factors remain applicable for joint inventorship analysis when multiple natural persons are involved. To be a joint inventor, they must:

• contribute significantly to the conception or reduction to practice of the invention
• contribute to the invention in a manner that isn’t insignificant in quality, when that contribution is measured against the full invention
• do more than merely explain to the real inventors well-known concepts

Assessing the conception of a chemical compound—like in Oka v. Youssefyeh—requires “the idea of the structure of the chemical compound” and “possession of an operative method of making it.” Assessing the type of AI model used to discover a compound can aid in applying the Pannu factors.

Predictive AI

Predictive AI can forecast properties of existing compounds, identify new drug targets, and reveal structure-activity relationships by implementing machine learning or non-agentic AI models. For example, the PandaOmics AI platform used for target identification relies on AI approaches such as natural language processing of scientific literature, network analysis, and matrix factorization combined with biological datasets to identify promising disease-associated molecular targets for existing compounds.

Predictive AI models take an input and return a prediction. The collaboration between a scientific expert and a person supervising the predictive AI model may have less impact on the conception of a compound.

For example, as in Oka, if a scientist has identified a structure of a chemical compound and a method for synthesizing it, that person has conceived of the compound. Conversely, if the scientist collaborates with an AI expert to predict properties or identify targets of the compound, for the AI expert to be a co-inventor, they must contribute significantly to the conception or reduction to practice (that is, the making) of the invention, as in Dana-Farber Institute, Inc. v. Ono Pharmaceutical Co., Ltd.

Still, as acknowledged in Ono, to qualify as a joint inventor, the AI expert can’t use a predictive AI model to provide the scientist with well-known principles or explain the state of the art. Such contributions likely amount to providing well-known information that can be found in textbooks, as in Hess v. Advanced Cardiovascular Sys., Inc.

Precedent is clear. As articulated by Eli Lilly and Co. v. Aradigm Corp., the AI expert can’t use a predictive AI model to “merely suggest[] an idea of a result to be accomplished, rather than means of accomplishing it.” If the output doesn’t amount to anything more than a result to be accomplished, the output is “too far removed from the real-world realization of an invention,” and the contribution to realizing the invention may not amount to a contribution to conception.

Unless an AI expert using a predictive AI model can prove a significant contribution to the conception of a compound, they won’t be able to satisfy the Pannu factors. Their contribution may not amount to co-inventorship.

Generative AI

Generative AI chemistry models represent a fundamentally different approach to molecular discovery compared with traditional computational methods. Rather than modifying known compounds of screening existing libraries, these models can design entirely novel molecular structures de novo. Generative chemistry models enable “inverse design,” which allows scientists to specify desired properties and have the model generate candidate structures to that satisfy those requirements.

A key distinction between predictive and generative AI lies in their fundamental objectives. While predictive AI can answer the question, “What will happen with this molecule?” the objective of generative models is to answer, “What molecules could have these properties?”

Once a user specifies desired chemical features, a generative model can design compounds that satisfy those criteria. Generated compounds can be evaluated and selected for further testing like traditionally discovered compounds.

With generative chemistry models, the collaboration between a scientific expert and a person supervising the generative AI model may affect the conception of a compound. For example, when a scientist has a chemical scaffold and can make it, they have conception of that scaffold as suggested by Oka.

If the scientist submits their chemical scaffold to a generative chemistry model and generates derivatives of that scaffold, the person chaperoning the model may contribute to the conception of the derivatives. Presumably, the derivatives generated wouldn’t have been identified without the contribution of the AI expert.

If the AI expert is familiar with the model processes and parameters used to generate the derivatives, their contribution would likely satisfy the two-pronged test for chemical conception. Because the AI expert made a significant contribution, it would be able to satisfy the Pannu factors and likely would amount to co-inventorship.

However, one must possess a way of making a compound identified by generative AI in addition to knowing its structure for conception of a chemical compound. As laid out in Nartron Corp. v. Shukra U.S.A., possessing parameters and inputting them into a generative chemistry model could be viewed as “suggest[ing] an idea of a result to be accomplished, rather than means of accomplishing it.” If the original scientist and the AI expert could only make the compound with the assistance of a chemist, then it would be likely that all three people could satisfy the Pannu factors, amounting to co-inventorship. Without a means of obtaining an AI generated compound, conception is incomplete.

Indeed, the Federal Circuit has held that “asking someone to produce something without saying just what it is to be [done] or how to do it is not what the patent law recognizes as inventing” in Morgan v. Hirsch. Aiming to obtain a compound as in Amgen, Inc. v. Chugai Pharm. Co., pose a problem as in Morgan, or validate an inventor’s expectation that a compound will work for its intended purpose as in Burroughs doesn’t amount to conception. Until the scientist possesses the structure of the chemical compound and an operative method of making it, they won’t have conception as in Oka. Given the highly collaborative nature of drug discovery, it’s likely that multiple people contributed to the conception of an AI generated compound.

Looking Ahead

As AI tools become more integrated into drug discovery programs, companies using these tools should keep detailed records to demonstrate how the inventor contributed to the conception of the invention. While scientists naturally document physical lab work and data analysis, they are not generally trained to keep track of mental processes, computer workflows, and ideas that would be imputed into AI. Moreover, because AI assisted drug discovery is an iterative process and since AI discovered compounds will undergo biological testing and chemical optimization, there will be several places where multiple people can make inventive contributions. Without robust documentation which scientist and what information was fed into an AI system to obtain a patentable compound, assessing the inventive contribution would be challenging if not impossible.

When multiple named inventors are involved, the records should demonstrate how each named inventor has satisfied the Pannu factors.

Companies should also evaluate whether an invention was the fruit of a predictive or generative AI model. The type of model, as well as how it was used, can affect both conception and co-inventorship inquiries. These inquiries can easily fit into an overall IP governance strategy.

When conducting IP due diligence of compounds derived from AI drug discovery, parties must assess whether a patent filing contains a defensible record of human conception or if it relies on undocumented AI output. Such a distinction could mean the difference between a defensible commercial asset and an unenforceable research result.

This article does not necessarily reflect the opinion of Bloomberg Industry Group, Inc., the publisher of Bloomberg Law, Bloomberg Tax, and Bloomberg Government, or its owners.

Author Information

Ben Bourke, of counsel in Womble Bond Dickinson’s San Francisco office, is a registered patent attorney who primarily represents life sciences patent holders in Abbreviated New Drug Application and related patent litigation.

Daniel Kremer is a patent agent in Womble Bond Dickinson’s Irvine, Calif., office, concentrating on patent prosecution concerning oncology, cell therapies, and immunotherapy matters.

Susan Krumplitsch, partner in Womble Bond Dickinson’s Silicon Valley office, is an intellectual property litigator representing life sciences companies across the biotechnology, pharmaceutical, and medical device sectors.

Write for Us: Author Guidelines