Intellectual Property Considerations for Drug Discovery Via Artificial Intelligence


Michael B. Harlin

Michael B. Harlin

The use of artificial intelligence (AI) in various stages of drug discovery holds enormous potential for the pharmaceutical industry, and investment in AI drug discovery technology has doubled annually over the past five years.1 AI companies seeking customers, partners, and investors have sought to capitalize on this excitement by touting various proprietary platforms that use AI in some way. Pharmaceutical executives would be wise to consider what intellectual property (IP) issues may arise from the use of AI-based platforms for drug discovery.

Proprietary platforms are offered by many AI companies for various aspects of the drug discovery pipeline, and the use of such platforms can accelerate timelines and reduce expenditures. The various models applied by such platforms can be used for target discovery; identification of chemical structures likely to bind a target; design of de novo molecules; lead optimization to enhance effectiveness and diminish potential toxicity; and profiling a compound’s absorption, distribution, metabolism, and elimination (ADME). Artificial neural networks can be trained to perform virtual screening or to facilitate traditional high-throughput screening assays by the automation of data collection and analysis. For instance, AI can be used to conduct image analysis on cellular assays producing a visual signal with high speed and accuracy.

AI companies tout their proprietary technology to differentiate themselves, and this may include patents, which confer exclusive rights to new and useful processes, machines, manufactures, and compositions of matter. Methods of screening compounds in a laboratory for activity against a target have long been patentable, so at first glance, one might think screening methods performed in a virtual laboratory would also be eligible for patenting. However, US law has been especially hostile over the past decade to inventions based on data analysis, mathematical calculations, and statistical modeling, on the grounds that abstract ideas, mathematical operations, and natural phenomena are not “eligible” subject matter for a patent. For instance, computerized statistical methods for determining which parent passed on a gene were held ineligible for patenting in 2021.2 Accordingly, algorithms and models that underlie AI-based platforms are more difficult to patent in the US, compared to other countries.

A screening method that combines an AI program with one or more physical laboratory assays has a far easier path to patent eligibility. An algorithm does not necessarily become patent-eligible just because it includes a physical step to generate an input nor because it is run on a cloud computing system. However, a model that is integrated with biological or chemical tests can more readily meet the law’s current standards. As an example, consider a screening method that relies on a fluorescent signal to indicate a compound’s binding to a target within a cell. The use of AI to analyze and interpret the fluorescent signals from a library of millions of screened compounds could be patent-eligible as an improvement to laboratory technology and significantly more than merely an algorithm.

New chemical compounds discovered with the aid of AI-based platforms are certainly eligible for patenting, but one must consider how a molecule (or class of molecules) can be adequately described in a patent application. This requires the chemical name or structure, a way to synthesize the compounds, and a practical utility or immediate benefit to the public. One might be tempted to quickly file a patent application on a chemical compound predicted to have exceptional activity, relying on a model’s prediction rather than a traditional laboratory assay. Although the model might provide a compelling rationale as to how the compound would perform, US patent law requires that a patent identify a “specific and substantial utility” for a chemical compound3. Whether that utility can be established by an AI model—without actually making and testing the compound—has not been addressed. Moreover, if a chemical compound is entirely conceived by an AI program, it raises a question as to who invented it. The issue of whether AI can be an “inventor” under US law is unsettled at present, as a trial court recently agreed with the US Patent Office that only a natural person can be an inventor4. If a patent application for a chemical compound does not identify a natural person as its inventor, it will not be accepted.

Turning back to drug discovery platforms, an AI company can still have valuable IP even if it does not have a patent covering its platform. Models, software, datasets, and other know-how can constitute trade secrets, but only if they are kept confidential. The need for confidentiality may constrain how the AI company conducts its business and provides services to pharmaceutical companies who work with them. The AI company will likely use its models internally and provide data packages to its partners or customers that do not fully reveal how they were obtained. The need for confidentiality may constrain how the AI models can constitute trade secrets, assuming they are kept confidential. The AI company is likely to run its models internally and provide results to its partners or customers in a way that does not reveal how the models produced those results.

Deep learning with neural networks requires massive, high-quality datasets for training and for discovery efforts. The content of those datasets may be generated by the AI company itself, or they may be obtained from a public database, a consortium, or a vendor. If the model developer obtained those datasets from external sources, license or copyright issues might arise from their use. Data from external sources may come with conditions or restrictions, such as a royalty on inventions made possible by use of the data, or an obligation to share improvements.

AI companies in the field of drug discovery have employed different business models. Some have operated like traditional contract research organizations (CROs) who receive a fee for their service, while others have positioned themselves as development partners who share in the revenue of a drug that they help to develop. If a pharmaceutical company’s drug development project employs AI-based platforms from more than one AI company, it could create multiple royalty obligations on the resulting drug. If a first company’s platform identifies a series of hits for a target, and a second company’s program guides the optimization of the hit into a lead compound, one may owe royalties on the drug to both companies. As a result, a pharmaceutical company that engages an AI company should carefully consider the contractual terms and the expectations underlying the relationship.


  1. Margaret Ayers, et al, Boston Consulting Group, “Adopting AI in Drug Discovery,”, accessed July 14, 2022.
  2. In re Bd. of Trustees of the Leland Stanford Junior Univ., 991 F.3d 1245 (Fed. Cir. 2021).
  3. In re Kirk, 376 F.2d 936, 942 (C.C.P.A. 1967) (“It cannot be presumed that a steroid chemical compound is ‘useful’ under §101, or that one of skill in the art will know ‘how to use’ it, simply because the compound is closely related only in a structural sense to other steroid compounds known to be useful.”)
  4. Thaler v. Hirshfeld, 558 F.Supp.3d 238, (E.D. Va. 2021) (“Congress’ use of the term ‘individual’ in the Patent Act strengthens the conclusion that an ‘inventor’ must be a natural person.”)

About the Author

Michael B. Harlin is a partner at Neal Gerber Eisenberg, specializing in life sciences intellectual property.

Related Videos