01. Introduction and Research Background
The explosive development of artificial intelligence (AI) is rapidly reshaping traditional industries, and its impact on R&D in the innovative drug field is particularly direct and intense. The 2024 Nobel Prize in Chemistry was partly awarded to researchers who used AI for protein structure prediction, showing that AI has entered an important stage of basic research in the life sciences. The U.S. FDA has also noted that the number of drug applications containing AI elements has increased significantly in recent years, with related applications covering all stages including preclinical research, clinical trials, drug marketing and manufacturing, and post-marketing surveillance. This fully demonstrates that AI is accelerating into the core processes of innovative drug R&D, and that the combination of AI and life sciences has moved from proof of concept into broader scientific research and industrial practice; AI drug discovery has become a common industry practice model. By way of definition, "AI drug discovery" mainly refers to the use of machine learning, deep learning, generative models and other technologies to analyze targets, compounds, experimental data and literature, and to assist in completing R&D activities such as target discovery, candidate compound screening, molecule generation and optimization, drug use prediction, and preparation route recommendation. At the current stage, the value of AI drug discovery is not to completely replace drug R&D personnel, but to improve the efficiency of information processing and candidate scheme screening and to shorten the cycle of some early R&D stages.
Facing the global AI innovation trend, on the one hand, large pharmaceutical companies (especially multinational MNC pharmaceutical companies) are investing huge amounts of funds to internalize AI capabilities as their own R&D infrastructure. Take Eli Lilly as an example: in 2025 it launched the TuneLab platform, opening some AI/machine learning models trained on long-term R&D data to biotechnology companies for property prediction of small molecule and antibody drugs; it later cooperated with NVIDIA to build AI computing infrastructure for drug discovery. Other MNC pharmaceutical companies are also actively investing in and embracing AI drug discovery. On the other hand, specialized AI drug discovery service providers are constantly emerging and growing, for example, China's leading AI drug discovery companies Insilico Medicine and XtalPi completed Hong Kong IPOs in 2025 and 2024 respectively, and achieved eye-catching AI drug discovery BD transaction results with large MNCs. Traditional pharmaceutical R&D contract outsourcing companies (CROs) are also entering the AI drug discovery track on a large scale.
However, the non-traditional R&D model of AI drug discovery also brings new patent legal issues. Specifically, in the scenario where pharmaceutical companies conduct AI drug discovery R&D independently, companies need to resolve issues such as whether AI-assisted results can obtain patent protection, who can be identified as the inventor, and how to prove the substantive contribution of natural person R&D personnel; in the AI drug discovery outsourcing R&D scenario, the focus of risk further shifts to whether the service provider's deliverables may fall within the scope of third-party patent protection, and how the sponsor and service provider allocate risk through freedom-to-operate (FTO) analysis, insurance and contract terms. In addition, related patent risks exist not only in the formation and implementation stages of R&D results, but have also extended to the process of preparing patent application documents. For example, on April 1, 2026, the China National Intellectual Property Administration issued a risk warning regarding the use of intelligent agents such as "Xiaolongxia" to draft patent application documents, pointing out that such applications may cause problems such as leakage of technical information, substantive defects in application documents, and dishonest applications. In summary, the patent issues in AI drug discovery are not simply whether AI drug discovery can apply for patents, but continuous legal risks running through all stages of R&D result formation, application, implementation and commercialization.
In the practice of providing legal services to pharmaceutical enterprise clients, the authors of this article have also directly faced the patent issues brought by AI drug discovery R&D to both parties of cooperation. This article attempts to analyze the patent law issues of AI drug discovery in light of our business practice experience.
02. Independent R&D Scenario: Patent Application Risks of AI-Assisted Results
(1) AI Cannot Be an Inventor
AI can participate in candidate molecule generation, screening, structural optimization and R&D route recommendation, but the current patent system still bases inventor identification on the creative contributions of natural persons. The 2025 revised Guidelines for Patent Examination further clarify that an inventor should be an individual, i.e., a natural person, and the name of artificial intelligence must not be entered in the request. Therefore, AI's participation itself does not exclude the relevant results from patent protection; the key is whether the natural person who made a creative contribution to the specific technical solution can be identified.
An enterprise can act as the patent applicant or patentee, but it should still identify the contributions of specific natural persons in the formation of the technical solution. In projects jointly completed by multiple teams such as algorithms, medicinal chemistry, biology and pharmacology/toxicology, inventors should be examined in light of the claims to be applied for, to see whether the relevant technical personnel made substantive judgments on the raising of technical problems, parameter setting, result screening, structural optimization and experimental verification. Personnel engaged only in project management, providing general data support, building general platforms or performing routine experiments usually do not automatically become inventors by participating in a project.
Conversely, if R&D personnel do not mechanically accept model outputs, but set unconventional constraints around specific targets, efficacy or druggability issues, screen and adjust candidate schemes, or finalize the technical solution through experimental results, their contributions are more likely to constitute creative contributions in the sense of inventorship. If algorithm personnel design special objective functions, screening mechanisms or model constraints for specific R&D problems and cause the relevant technical features to enter the claims, they may also constitute joint inventors.
(2) AI-Generated Results Must Still Undergo Examination of Patent Grant Conditions
Candidate compounds, use schemes and preparation routes formed with AI assistance still need to satisfy the requirements of novelty, inventiveness and utility. AI generation does not automatically mean inventiveness; if the relevant results are merely combinations, substitutions or optimizations of existing compounds, public databases or conventional technical paths, and do not produce unexpected technical effects, they may still fail to obtain authorization due to lack of inventiveness.
Patent applications in the pharmaceutical field also need to pay attention to sufficient disclosure and support by experimental data. Model predictions or computational conclusions alone are usually not sufficient to automatically support a relatively broad scope of compound or use protection. Enterprises should, in light of data such as efficacy, stability, toxicology and preparation feasibility, explain that the technical solution can achieve the expected effects, and highlight in the application documents the judgments, optimization and verification made by human R&D personnel on AI outputs.
In particular, when AI optimizes based on known parent nuclei, public activity data or existing synthetic routes, if the final scheme only reflects conventional substituent replacement, limited-range screening or combinations easily conceived by those skilled in the art, the fact that it was generated by AI cannot itself make up for the lack of inventiveness. Conversely, if the enterprise can explain why a specific candidate was retained from a large number of outputs, how R&D personnel continued to adjust the structure accordingly, and what unexpected effects the experimental results produced in terms of activity, selectivity, pharmacokinetics, safety or stability, it will be more conducive to forming a complete inventiveness argument.
(3) R&D Records Are an Important Basis for Inventor Identification and Ownership Management
Enterprises should incorporate the AI R&D process into a recordable and traceable R&D management system. R&D traces should not be limited to the final candidate molecules and successful experimental results, but should also cover model versions, sources of input data, parameter settings, candidate screening criteria, reasons for human rejection, structural optimization, failed experiments and subsequent adjustments prompted by failed results.
The above records help reconstruct the process by which model outputs were transformed into specific technical solutions and prove the substantive contributions made by R&D personnel, thereby providing a basis for inventor confirmation, ownership of service inventions, patent applications and invalidation proceedings. At the same time, this is also a necessary means to respond to CNIPA examiners' requests to disclose the "formation process of the technical solution" and to identify whether an application is an honest application, and can reduce and prevent compliance risks arising from AI-generated content being identified as an abnormal application. For enterprises planning financing, licensing or cross-border cooperation, clear R&D records, personnel contribution statements and ownership chains also help reduce the risk of determining inventors solely by after-the-fact recollection and enhance the credibility of core patent assets in due diligence and value assessment.
(4) Global Layout Should Pay Attention to Differences in Rules Across Jurisdictions
Major jurisdictions such as China, the United States, Europe and the United Kingdom generally do not accept listing an AI system as an inventor, but the judgments on human contribution in AI-assisted inventions and the requirements for application documents are not entirely the same. The USPTO's 2025 revised guidance emphasizes that AI-assisted inventions apply the same inventorship criteria as other inventions; the European Patent Office and the UK Supreme Court also confirmed in the DABUS case that artificial intelligence cannot be an inventor in a patent application.
However, practice in various jurisdictions does not consist only of diametrically opposed results. The German Federal Court of Justice in the 2024 DABUS case likewise insisted that the inventor must be a natural person, but held that even if artificial intelligence made an important contribution to the discovery of the technical solution, as long as the participation of a natural person had a causal influence on the overall result, that natural person could still be listed as the inventor; supplementing the natural person's name with an explanation that he or she prompted the specific AI to generate the invention would not automatically lead to rejection of the application. South Africa once granted a patent in 2021 listing DABUS as the inventor, but South Africa adopts a patent registration system focused on formal examination, and that result should generally not be directly understood as general recognition after substantive examination that AI has inventor qualification. The Federal Court of Australia at first instance once accepted an interpretation that AI could be listed as an inventor, but the Full Court subsequently overturned that judgment, and current practice still requires the inventor to be a natural person.[18] These differences show that when laying out cross-border, enterprises cannot merely count whether a jurisdiction has ever seen an "AI inventor patent"; they should further distinguish whether the result comes from substantive law interpretation, a stage judgment in judicial proceedings, or merely from a formal examination system.
For innovative drug enterprises planning a global patent layout, it is advisable to, at an early stage of the project and in light of target markets, uniformly sort out inventors, applicants, R&D records and experimental data, and adjust application strategies according to the requirements of each jurisdiction, so as to avoid affecting the protection of core results due to inconsistent inventor identification or document preparation.
03. Outsourcing R&D Scenario: Infringement Risks and Liability Allocation for AI Drug Discovery Service Results
(1) Typical Transaction Structure of AI Drug Discovery Outsourcing Services
Outsourcing R&D as referred to in this article mainly means that a pharmaceutical company or Biotech company, through commissioned R&D, technical services or other means, commissions an AI drug discovery platform, CRO or other professional service provider to use its algorithm models, databases, computing power and technical team to complete target analysis, candidate compound screening, molecular structure generation and optimization, use prediction or preparation route recommendation, and the entrusted party delivers phased results. In such transactions, the pharmaceutical company usually proposes R&D goals, provides some project materials and decides on subsequent development of candidate results, while the outsourcing service provider provides professional R&D services within the agreed scope.
The interim deliverables delivered by the outsourcing service provider may continue to be used in preclinical research, IND filing, financing, out-licensing or commercial development. Therefore, pharmaceutical companies are concerned not only with whether the outsourced R&D services are completed on time, but also with whether the deliverables can be safely advanced in subsequent R&D and in the target markets; and whether the outsourcing service provider can give an absolute non-infringement guarantee covering the whole world, all R&D stages and all third-party patents. The parties need to establish an executable allocation mechanism between the usability of the results and controllable risk. The parties need to establish a reasonable allocation mechanism between the usability of the results and controllable risk through FTO, insurance and contractual arrangements. It should be noted that the fact that a deliverable has patent application value does not automatically mean that it can be implemented without restriction by third-party patents.
(II) Main patent infringement risks
First, the candidate compound may fall within the scope of protection of a third-party compound patent. Pharmaceutical compound patents often form broad protection through general formulas, Markush structures, substituent ranges, salts, esters or prodrugs, and even if the AI output does not directly copy the patent examples, it may still be covered by the claims.
Second, the risk is not limited to the core compound. Peripheral patents such as indications, dosing regimens, crystal forms, salt forms, formulations, compositions, preparation methods and intermediates may also restrict subsequent R&D and commercialization. Risks that are not obvious at the early screening stage may be further exposed when the candidate is determined, at clinical filing or at the time of a licensing transaction.
Third, patent rights are territorial. The absence of an obvious obstacle to a result in China does not mean that it can be freely implemented worldwide. When conducting searches, the outsourcing service provider should determine the scope in light of the countries and regions where the client intends to implement or commercialize.
Fourth, any search carries residual risk. Patent applications that have been filed but not yet published, subsequent amendments to claims and changes in the legal status of patents may all cause risks not found at the time of delivery to emerge later. Therefore, it is usually difficult for an outsourcing service provider to promise "absolute non-infringement".
(III) Phased risk identification through FTO
FTO means Freedom to Operate, usually translated as freedom-to-operate analysis, and is used to determine whether the implementation of a technical solution or product in a specific territory may infringe a valid third-party patent. In AI pharmaceutical projects, FTO can be carried out around candidate compounds, uses, crystal forms, salt forms, formulations and preparation routes.
FTO should match the R&D stage. Specifically, at the project initiation or target selection stage, major competitors and obvious blocking patents can first be identified; after the candidate scope is narrowed, searches can then be conducted around general structures, similar compounds, uses and key intermediates; before the preclinical candidate is determined, IND filing, major financing or a licensing transaction, a more complete claim comparison and legal status verification should be carried out in light of the target jurisdictions; after the commercialization plan is determined, the analysis also needs to be updated according to the final formulation, dosage, manufacturing process and sales regions. Through such a phased arrangement, costs can be controlled while avoiding deferring all patent risks to the end of the project. In addition, the contract should specify the search targets, territories, timing, cost bearing and the boundaries of the report.
It should be emphasized that FTO can only identify and reduce risk, not guarantee zero risk. Its conclusions are affected by factors such as the search timing, publicly available information, claim interpretation and subsequent changes in the solution, and should not be directly equated with a non-infringement guarantee.
(IV) Insurance can buffer losses but cannot replace contractual arrangements
For projects with broader target markets, higher transaction amounts or greater commercialization value, the parties may evaluate products such as patent infringement liability insurance and overseas intellectual property infringement liability insurance to cover part of the defense, settlement or compensation costs. It is recommended that AI pharmaceutical outsourcing service providers negotiate with their own insurers on the specific matters and scope that can be covered.
However, insurance usually has compensation limits, deductibles and exclusions, and intentional infringement, known risks, pre-insurance disputes or liabilities additionally assumed under contract may not be covered. Insurance is more suitable as a risk buffer at the compensation funding level and cannot replace the parties' agreement in the service agreement on review obligations and the boundaries of liability.
(V) Service agreements should allocate risk around reasonable obligations rather than absolute guarantees
The commissioning party/sponsor may require the outsourcing service provider to conduct reasonable searches and risk warnings within the agreed scope, and to promptly notify and assist in handling when obvious patent obstacles are found or third-party claims are received. Where the outsourcing service provider delivers results despite knowing of major risks or conceals known problems, corresponding compensation liability may be agreed.
AI pharmaceutical outsourcing service providers should avoid absolute non-infringement commitments that are not limited by territory, time or mode of implementation. The contract may limit their obligations to taking reasonable efforts within the agreed scope based on publicly available information, and exclude liability arising from factors such as materials provided by the client, R&D directions designated by the client, subsequent modification of results, implementation beyond the agreed territory or use, and continued advancement without adopting risk warnings.
The parties should also agree on notification, defense and settlement mechanisms for third-party claims, reasonably set caps on compensation liability, and clarify the bearing of FTO, insurance and external expert costs. Overall, FTO is responsible for front-end identification, insurance provides a loss buffer, and the contract completes the final allocation of liability (including the maximum compensation liability limit); the three should be used in coordination.
In addition to infringement liability, outsourcing contracts should also avoid using "project results belong to the client" to summarize all intellectual property arrangements. The parties should separately define background intellectual property such as platforms, models, databases and trade secrets that existed before the cooperation, project inputs provided by the client such as targets, compounds and experimental data, foreground results formed during project implementation such as candidate molecules, uses, routes, experimental data and patentable solutions, and derivative results such as model improvements or subsequent new uses formed based on project experience, and separately agree on ownership, application rights, use licenses, sublicenses, regulatory filings, due diligence disclosure and post-termination use rights.
Conclusion
AI is improving the efficiency of information processing and candidate screening in innovative drug R&D, but improved technical efficiency may bring higher patent infringement risks. In independent R&D, enterprises should focus on inventor identification, patent grant conditions and R&D records; in outsourced R&D, they should further focus on third-party patent obstacles and the allocation of liability between the service provider and the sponsor.
The legal analysis in this article mainly provides general risk identification ideas; specific projects still need case-by-case judgment on relevant patent issues in light of the candidate compounds, R&D stage, target markets and transaction structure. For innovative drug enterprises and AI pharmaceutical outsourcing service providers, a more prudent path is to establish traceable records of result formation in the early stage of R&D, conduct FTO at key nodes, and form risk management arrangements matching the project value through insurance and contract terms.
Only by advancing AI R&D capabilities in parallel with patent applications, freedom-to-operate analysis and transaction risk management can the commercial value of AI-enabled innovative drug R&D be realized more robustly.
[Note: The first draft of this article was completed by our team intern Ye Xinmeng (undergraduate degree in pharmacy from Fudan University, currently a graduate student at Sichuan University Law School) under the guidance of the partners and team lawyers.]
Questions and answers
Can an AI system be named as an inventor in a patent application?
No. The 2025 revised Guidelines for Patent Examination clarify that an inventor must be a natural person, and AI's name must not be entered. Courts in Europe, the UK and Germany also require a natural person inventor. AI participation does not exclude patent protection if a natural person made a creative contribution.
What patentability requirements must AI-assisted drug discoveries meet?
AI-assisted results must still satisfy novelty, inventiveness and utility. AI generation does not automatically confer inventiveness; if results are mere combinations or optimizations of known compounds without unexpected effects, they may lack inventiveness. Applications also need sufficient disclosure and experimental data to support the claimed scope.
Why are R&D records important for AI drug discovery patents?
R&D records help reconstruct how model outputs became specific technical solutions, prove human contributions for inventor identification, and support ownership and patent prosecution. They also respond to CNIPA requests on the formation process and reduce compliance risks from AI-generated content being deemed abnormal applications.
What is freedom-to-operate (FTO) analysis in AI drug discovery outsourcing?
FTO analysis determines whether implementing a technical solution may infringe valid third-party patents in a specific territory. It should match the R&D stage, starting with competitor identification and progressing to claim comparison and legal status verification before IND filing or licensing. FTO reduces but does not guarantee zero risk.
How should service agreements allocate patent infringement risk in AI drug discovery outsourcing?
Service agreements should avoid absolute non-infringement guarantees and instead limit obligations to reasonable searches within an agreed scope. Contracts should address notification and defense of third-party claims, liability caps, and allocation of FTO and insurance costs. Insurance can buffer losses but cannot replace contractual arrangements.
Laws and cases cited
Regulations
- Guidelines for Patent Examination
Policy documents
- USPTO's 2025 revised guidance
Cases
- DABUS case
- German Federal Court of Justice 2024 DABUS case
- Federal Court of Australia judgment
References
- Standing Committee of the National People's Congress: Patent Law of the People's Republic of China (2020 amendment), effective June 1, 2021.
- State Council: Implementing Regulations of the Patent Law of the People's Republic of China (2023 revision), State Council Decree No. 769, effective January 20, 2024.
- China National Intellectual Property Administration: On Using Intelligent Agents Such as "Xiaolongxia" to Draft Patent Application Documents or Inducing Multiple Risks, April 3, 2026.
- China National Intellectual Property Administration: Decision of the China National Intellectual Property Administration on Amending the Guidelines for Patent Examination, China National Intellectual Property Administration Order No. 84, effective January 1, 2026.
- China National Intellectual Property Administration: Interpretation of the 2025 Amendments to the Guidelines for Patent Examination, December 4, 2025.
- United States Patent and Trademark Office, Revised Inventorship Guidance for AI-Assisted Inventions, 90 Fed. Reg. 54636–54637 (2025).
- European Patent Office, Legal Board of Appeal, J 8/20, Designation of Inventor/DABUS, 21 December 2021.
- UK Supreme Court, Thaler v Comptroller-General of Patents, Designs and Trade Marks [2023] UKSC 49.
- Nobel Prize Outreach AB, The Nobel Prize in Chemistry 2024: Computational Protein Design and Protein Structure Prediction.
- U.S. Food and Drug Administration, Artificial Intelligence for Drug Development.
- QuantumPharm Inc. (XtalPi): Global Offering Prospectus, Hong Kong Stock Exchange, June 4, 2024.
- VAMATHEVAN J, CLARK D, CZODROWSKI P, et al. Applications of Machine Learning in Drug Discovery and Development[J]. Nature Reviews Drug Discovery, 2019, 18: 463-477.
- ZHAVORONKOV A, IVANENKOV Y A, ALIPER A, et al. Deep Learning Enables Rapid Identification of Potent DDR1 Kinase Inhibitors[J]. Nature Biotechnology, 2019, 37: 1038-1040.
- HASSELGREN C, OPREA T I. Artificial Intelligence for Drug Discovery: Are We There Yet?[J]. Annual Review of Pharmacology and Toxicology, 2024, 64: 527-550.
- Eli Lilly and Company, Lilly Launches TuneLab Platform to Give Biotechnology Companies Access to AI-Enabled Drug Discovery Models, 9 September 2025.
- Eli Lilly and Company, Lilly Partners with NVIDIA to Build the Industry’s Most Powerful AI Supercomputer, 28 October 2025.
- Bundesgerichtshof, Beschluss vom 11. Juni 2024, X ZB 5/22 (DABUS).
- World Intellectual Property Organization, Getting the Innovation Ecosystem Ready for AI: An IP Policy Toolkit, 2024.
- IP Australia, Patent Manual of Practice and Procedure, 5.4.3 Artificial Intelligence - Inventorship and Entitlement, updated 8 January 2025.
