Intellectual Property Challenges in AI-Driven Drug Discovery

The adoption of artificial intelligence-based drug discovery platforms within the biopharmaceutical industry presents new challenges to intellectual property rights that may require reexamining existing patent doctrines.

Although machine learning systems autonomously produce patentable subject matter, practitioners must now consider unprecedented questions regarding inventorship, patentability standards and infringement liability.

Inventorship determination is the most immediate challenge. Under Title 35 of the U.S. Code, Section 101, and corresponding provisions in most other jurisdictions, human conception is an essential statutory obligation.

Yet when AI networks autonomously discover novel chemical entities or large molecule therapeutics through the analysis of structure activity relationships in a wide array of datasets, the classical framework of conception and reduction to practice falls short.

On Nov. 26, the U.S. Patent and Trademark Office issued revised inventorship guidance for AI-assisted inventions with the hope of streamlining its approach to determine inventorship in AI-assisted inventions.[1]

The new guidance rescinds the February 2024 framework, which applied the Pannu joint-inventorship factors, as established in the U.S. Court of Appeals for the Federal Circuit’s 1998 decision in Pannu v. Iolab Corp., to evaluate AI-assisted inventions.[2]

Pannu is still a doctrine to determine joint inventorship among multiple people, but it does not apply where the other participant is an AI system, which by definition is not a person.

The key change in the revised guidance is that the USPTO now treats AI systems as tools that are analogous to laboratory equipment, rather than joint inventors.

The guidance emphasizes that there is no separate or modified standard for AI-assisted inventions — the same legal standard for determining inventorship applies to all inventions, whether AI was used or not.

Conception — which a 1929 Court of Customs and Patent Appeals case, Townsend v. Smith, defined as “the formation in the mind of the inventor of a definite and permanent idea of the complete and operative invention” — remains the touchstone of inventorship.

However, if conception requires a specific solution to a problem to show possession, such that only ordinary skill is needed to reduce it to practice, this remains problematic in drug discovery even with the revised framework.

AI facilitates the identification of disease-related molecular patterns and relationships by constructing multiomics data networks.[3]

For example, natural language processing techniques have been used to map gene functions into high-dimensional space, enhancing the sensitivity of target identification without a great deal of gene function overlap.[4]

If the AI network is identifying new chemical entities, or perhaps old chemical entities that are useful for new indications, is a human’s selection of a compound to move forward in drug development the inventive contribution? What happens when the AI system only identifies one such useful compound?

Even the training data poses serious issues regarding derivative rights and infringement. Biopharmaceutical companies can scrape proprietary and public databases, published patents and clinical trial data, to name a few, to train their algorithms. But when they produce outputs, liability under direct and contributory infringement of the intellectual property rights that are built into the training data itself presents a potential risk.

Similarly, when these systems scrub real-life data, such as medical records, electronic health records and insurance claims, concerns about privacy are most certainly present.

The nonobviousness standard may need to be revisited as well. If AI can identify therapeutic compounds through computational screening and predictive modeling orders of magnitude faster than any human researcher, the most pertinent question becomes what it means for AI to constitute a nonobvious discovery.

AI complicates this analysis because machine learning algorithms can rapidly evaluate millions of molecular combinations, which can render certain discoveries “obvious to try” when success could be reasonably expected.[5]

The courts, though, have held that inventions are obvious to try when there is a finite number of identified predictable solutions and a reasonable expectation of success.

In drug discovery, AI can be used for target discovery, de novo drug design, biomarker discovery, predicting pharmacometrics properties, drug repurposing and improving clinical trial efficiency, further confounding a one-size-fits-all analysis.

The result-effective variable doctrine is also likely to be problematic in nonobviousness analyses.[6]

How these variables are defined in the context of drug discovery and development will require flexibility as the variables can include factors from many discovery categories, including data quality; molecular representation; model architecture; the biological context; novelty and synthesizability; absorption, distribution, metabolism, excretion and toxicity; and safety.

As AI-driven optimization becomes common, the courts may increasingly view parameter tuning as predictable, reinforcing the doctrine’s relevance.

Therefore, AI-driven discoveries will need to be treated differently. In drug discovery itself, the computational capability to filter large amounts of data in the chemical space does not necessarily result in an obvious compound selection.

Key factors that must be considered include the size and diversity of the chemical space explored, whether AI revealed an unexpected structure-function relationship, the degree of human intervention in training and interpreting the model, and whether the discovery required innovative AI architecture and novel training approaches.

While the Graham factors, which were established by the U.S. Supreme Court’s 1966 decision in Graham v. John Deere Co., remain relevant, they will also likely require adaptation.[7]

Will the scope and content of the prior art now include training datasets? Should a person having ordinary skill in the art now include computational expertise and be presumed to have access to similar AI tools?

As AI and machine learning become standard in biopharmaceutical research, what was once innovative may migrate into the realm of routine optimization.

Patent practitioners will need to develop prosecution strategies that emphasize unexpected results, synergistic effects or superior properties that were not predictable from the training data alone.

For example, experimental data demonstrating that the AI-identified compound significantly outperforms computationally similar candidates becomes crucial evidence of nonobviousness.

Moreover, it will be important to document any human insight beyond routine computational tuning.

Lastly, practitioners should work with discovery teams to gather evidence of objective indicia of nonobviousness, such as a long-felt need or failure of others that are linked to the claimed invention, ensuring there is a clear nexus between these factors and the claimed features of the invention.

Trade secret protection offers strategic alternatives but raises enforceability concerns, especially regarding misappropriation claims when competitors develop similar AI architectures.

However, trade secrets can readily be used to protect model architectures and predictive insights.

In addition, many AI components, such as raw data, intermediate outputs and heuristic methods that may not meet subject matter eligibility requirements under Title 35 of the U.S. Code, Section 101, can be protected as trade secrets.

In addition to updating their approach toward patentability, practitioners must also now carefully structure licensing agreements, collaborative research arrangements and employment contracts to address AI-generated inventions explicitly, anticipating evolving judicial interpretations and potential legislative reforms in this rapidly developing area.

[1] Fed. Reg. 90:227 November 28, 2025.

[2] Pannu v. Iolab Corp., 155 F.3d 1344 (Fed. Cir. 1998).

[3] Zhang et al., Nat. Med. 31:45-59 (2025).

[4] Chen et al., Nat. Commun. 15:1853 (2024).

[5] KSR International Co. v. Teleflex Inc., 550 U.S. 398 (2007); In re: O’Farrell, 853 F.2d 894 (Fed. Cir. 1988).

[6] In re: Aller, 220 F.2d 454 (CCPA 1955); In re: Peterson, 315 F.3d 1325 (Fed. Cir. 2003).

[7] Graham et al. v. John Deere Co., 383 U.S. 1 (1966).