AI in Drug Discovery: IP, Regulatory Validation, and Licensing Considerations

Artificial intelligence is reshaping drug discovery at every stage — from target identification and lead optimization to candidate selection and clinical trial design. For biotechnology companies, pharmaceutical startups, and the investors and collaborators who work with them, this transformation creates a set of legal questions that did not exist a decade ago. Who owns a compound that an AI system identified? What does the FDA expect when a drug candidate was selected through machine learning? And how should licensing agreements be structured when the platform that generated the asset is as valuable as the asset itself?

This post works through those three questions systematically. The answers matter whether you are a startup building an AI drug discovery platform, a pharmaceutical company licensing AI-generated compounds, or a biotech spinning out assets that emerged from machine learning pipelines.

Part One: Patent Inventorship in AI-Assisted Drug Discovery

The Foundational Rule: Natural Persons Only

Under U.S. patent law, only a human being can be named as an inventor on a patent application. This principle was reaffirmed emphatically by the Federal Circuit in Thaler v. Vidal (2022), which held that the Patent Act’s use of “individual” unambiguously refers to natural persons. The Supreme Court declined to take the case. The door to AI inventorship is closed under current U.S. law.

But that framing asks the wrong question for most drug discovery programs. The operative question is not whether the AI is an inventor — it is not — but whether the humans working alongside the AI have made the kind of contribution that qualifies them as inventors under the legal standard.

The USPTO’s Revised Inventorship Guidance (November 2025)

In November 2025, the USPTO issued revised guidance on inventorship for AI-assisted inventions, expressly rescinding the February 2024 guidance in its entirety. The 2024 guidance had attempted to apply the Pannu factors — a framework originally designed for multi-human joint inventorship — to the question of whether a human’s contribution was sufficient when AI was also involved. That approach created confusion, and the 2025 revision removes it.

The revised guidance returns to a simpler framework. The critical inquiry is whether one or more natural persons made a significant contribution to the conception of the claimed invention. Conception remains “the touchstone of inventorship,” defined as the formation in the mind of the inventor of a definite and permanent idea of the complete and operative invention. Because conception is an act achievable only by a natural person, AI is treated as a sophisticated tool — analogous to other laboratory instruments — rather than as a co-inventor.

The practical implication: a researcher who merely pushes a button to run an AI model, accepts its output uncritically, and files a patent on the resulting structure may have a weak inventorship claim. A researcher who defines the design parameters, selects the chemical space for the model to explore, evaluates and iterates on candidate outputs using scientific judgment, and directs the optimization process is a much stronger candidate for inventorship under the revised framework.

What This Means for Drug Discovery Programs

Several patterns create inventorship risk in AI drug discovery:

Over-reliance on automated output. If your discovery process is described internally as “the AI found the compound and we confirmed it was synthesizable,” you have a documentation problem. The humans involved need to have exercised genuine scientific judgment — in problem formulation, parameter setting, output evaluation, or structural modification — not just passive acceptance.

Black-box AI platforms. Platforms that produce candidate compounds with minimal human input into the generative process increase inventorship uncertainty. This is not a reason to avoid these platforms, but it is a reason to document carefully what human scientists contributed at each stage.

Employee and contractor contributions. In AI-assisted programs, significant contributions often come from the data scientists and ML engineers who trained or fine-tuned the model, not only from the medicinal chemists who selected among its outputs. Companies should review whether these contributions rise to the level of inventorship and assign IP rights accordingly through employment agreements and contractor IP assignment clauses.

Trade secret as an alternative. Where human contributions to the inventive process are thin or difficult to document, some companies are choosing to protect AI-generated compounds through trade secret law rather than patents — at least until clinical validation makes a patent more defensible. This involves structuring programs so that manufacturing processes, formulation details, and discovery parameters remain confidential.

International Dimension

Inventorship rules vary by jurisdiction. The European Patent Office, for example, has taken a similar position to the USPTO — an AI system cannot be listed as an inventor — but the analysis of what constitutes a sufficient human contribution differs in detail from U.S. law. For companies seeking international patent protection on AI-discovered compounds, jurisdiction-specific patent counsel is essential.

Part Two: FDA Regulatory Validation of AI-Generated Candidates

The Scope Question

Before analyzing FDA requirements for AI in drug discovery, it is important to note a scope limitation in the FDA’s own guidance. The FDA’s January 2025 draft guidance, “Considerations for the Use of Artificial Intelligence to Support Regulatory Decision Making for Drug and Biological Products,” explicitly states that it does not address AI when used in drug discovery — meaning AI for target identification, lead optimization, virtual screening, and other pre-development activities.

What the guidance does cover is AI and machine learning used in regulatory submissions — the models used to analyze clinical trial data, predict safety profiles, process biomarker data, or support other decisions that are directly referenced in an IND, NDA, or BLA. This is an important distinction for drug discovery companies to understand.

The Credibility Framework

For AI and ML that is used in regulatory submissions, the FDA’s 2025 draft guidance adopts a risk-based credibility assessment framework with four core steps:

1. Define the context of use — what decision is the AI model supporting, and how significant is that decision to regulatory review?
2. Assess model risk — higher-risk uses require more rigorous validation than lower-risk uses.
3. Plan and execute verification and validation — training and validation datasets must be clearly delineated; models must be validated on independent data.
4. Document results — technical documentation including algorithms, code, and performance metrics must be maintained as part of the regulatory submission.

The guidance distinguishes three risk tiers. Low-risk applications (AI for hypothesis generation with significant human oversight before any regulatory reliance) require minimal documentation. Medium-risk applications require model descriptions, validation data, and data governance records. High-risk applications — where AI outputs directly influence regulatory decisions — require full transparency, prospective validation, and ongoing monitoring.

Where Drug Discovery AI Intersects With Regulatory Obligations

Even though the FDA’s 2025 guidance expressly carves out drug discovery AI, that carve-out has practical limits. Consider the following scenarios:

Target identification to clinical hypothesis. A company uses AI to identify a biological target and uses the AI-generated hypothesis as the scientific rationale in its IND filing. At that point, the AI-generated content is being used to support a regulatory decision, even if it originated in discovery. Companies should be prepared for FDA questions about the basis for the hypothesis and the confidence level of the AI system that generated it.

AI-generated pharmacokinetic predictions. If PK/PD modeling using AI is referenced in a regulatory submission, it needs to meet the credibility standards in the 2025 guidance. This includes training-validation data separation and documentation of model performance.

In January 2026, the FDA published “Guiding Principles of Good AI Practice in Drug Development.” This document addresses broader principles of AI use across the drug development lifecycle and signals that the FDA expects sponsors to think systematically about AI governance, documentation, and quality even in stages the formal guidance has not yet fully addressed.

Practical Steps for IND Preparation

For companies preparing IND submissions that reference AI-generated data or analysis:

• Prepare a model card for each AI tool that generated data referenced in the submission. A model card documents the model architecture, training data, validation approach, known limitations, and intended use.
• Maintain a clear record of which AI tools were used at which stages of the discovery and development process.
• Ensure that training and validation datasets do not overlap — this is a consistent FDA expectation for any predictive model used in regulatory submissions.
• For novel AI systems, consider pre-submission meetings with FDA to discuss the agency’s expectations before submitting data generated by those systems.

Part Three: Structuring Licensing Agreements for AI-Discovered Compounds

The Landscape

The AI drug discovery licensing market is significant and growing. In 2025, AI-ML drug discovery and licensing transactions totaled approximately $12.3 billion across 99 deals, up from 84 deals totaling $11.8 billion in 2024. Major transactions include the Eli Lilly-Insilico Medicine research and licensing collaboration announced in November 2025, in which Insilico is eligible to receive over $100 million including upfront, milestone, and tiered royalty payments.

These deals reflect a structural split that is fundamental to understanding the licensing landscape: platform IP and compound IP are distinct assets with distinct legal and economic characteristics.

Platform IP vs. Compound IP

Compound IP covers specific molecules or clusters of related structures. It is defined by the claims in a patent application and derives its value from the therapeutic and commercial potential of that compound. Standard biotech licensing terms — upfront fees, development milestones, and royalties on net sales — apply.

Platform IP covers the AI system, algorithms, training data, and methods that generated the compound. Platform IP can be licensed to generate royalty streams across an essentially unlimited number of programs. A platform licensor may seek to retain rights in the platform while granting a licensee only a limited right to use the platform’s output for a defined program or indication.

The distinction creates complex IP ownership questions that must be resolved explicitly in licensing agreements.

Key Provisions in AI Drug Discovery License Agreements

Ownership of discovered compounds. Who owns a compound that the AI platform identifies? This must be addressed explicitly. Common approaches include: the licensee owns the compound IP while the licensor retains platform IP; joint ownership with a defined agreement governing prosecution and enforcement; or a royalty-bearing license from the platform owner to the compound developer.

Background IP and foreground IP. The agreement should clearly define what IP each party brings to the collaboration (background IP) and who owns newly developed IP arising from the collaboration (foreground IP). For AI collaborations, foreground IP may include improvements to the AI model itself — which the platform company will typically insist on retaining.

Regulatory data rights. If the collaboration generates preclinical or clinical data, the parties need to agree on who can use that data in regulatory submissions and for which compounds or indications.

Milestone triggers. Milestones in AI drug discovery deals typically track regulatory events: IND filing, Phase I completion, Phase II initiation, FDA approval, and commercial launch. The agreement should also address what happens if the AI platform generates multiple candidates simultaneously — does each candidate generate independent milestone obligations, or is there a per-program cap?

Audit and verification rights. When milestone payments are tied to AI performance (e.g., compound advancement), the agreement should give the paying party audit rights over the platform outputs and validation data. This is less common in traditional pharma licensing but is becoming increasingly standard in AI collaborations.

Exclusivity terms. A licensee funding a drug discovery collaboration will typically want exclusivity over the compounds discovered for a defined indication or target class. The platform licensor will want to limit exclusivity as much as possible — the value of the platform depends on its ability to serve multiple programs. Negotiating the scope of exclusivity is one of the most contested points in AI discovery deals.

Human oversight representations. Some recent agreements have included representations by the platform company that all generative chemistry outputs subject to patent prosecution have undergone documented human oversight review. This is a direct response to USPTO inventorship requirements and protects both parties from inventorship challenges downstream.

Indemnification for IP disputes. The agreement should address who bears the cost if a third party challenges a compound patent on inventorship grounds. Platform companies may push for limitations on indemnification for claims arising from the nature of AI-assisted inventorship; pharmaceutical partners will push back.

Cross-Border Considerations

Many AI drug discovery platforms are domiciled outside the United States. Agreements with companies in the EU, UK, Israel, Canada, and increasingly China and India require careful attention to governing law, dispute resolution, export control (particularly where AI models may be controlled under EAR), and data protection requirements for any training data that includes patient-derived information.

Practical Takeaways

For any company operating at the intersection of AI and drug discovery, these are the questions to address now:

1. Document human inventive contributions at every stage of your AI-assisted discovery process. The documentation should be contemporaneous — generated at the time of the work, not reconstructed for a patent application.

2. Audit your employment agreements and contractor IP assignment provisions to ensure that all contributors to AI-assisted inventions — data scientists, ML engineers, and chemists alike — have assigned their IP rights to the company.

3. Review your regulatory documentation practices. Even where the FDA’s 2025 guidance does not technically apply to discovery-stage AI, building good documentation habits now — model cards, dataset logs, validation records — will reduce friction when those activities move into the regulatory submission context.

4. Negotiate platform vs. compound IP ownership explicitly in any collaboration or licensing agreement. Do not rely on default rules or boilerplate; they were written before AI drug discovery existed.

5. Consider trade secret protection for discovery processes and AI models alongside or instead of patent protection where inventorship is uncertain.

The legal infrastructure for AI drug discovery is still being built, but the foundational rules are clear enough to act on. Companies that invest in documentation, IP assignment, and thoughtful licensing structures now will be better positioned as the regulatory and enforcement landscape matures.

This post is for general informational purposes only and does not constitute legal advice. Reading this post does not create an attorney-client relationship. If you have questions about your specific situation, consult a qualified attorney.