A generative model proposes a molecule. A chemist selects it, modifies it, and eventually synthesizes it. Somewhere in that sequence, a question that used to have an obvious answer stops being obvious: who actually owns the result? IP disputes in AI drug discovery exist precisely because artificial intelligence (AI) has inserted a genuinely new kind of contributor into a legal framework, patent law, that was built entirely around human inventors. The commercial question, who owns the compound, is often contractually straightforward. The legal question, who counts as its inventor, is not, and the two do not always point in the same direction.

This connects to the regulatory guide to AI in drug discovery for how FDA, EMA and ICH are approaching AI validation more broadly, and to the guide to AI across drug discovery for the wider technical picture. DDN has previously noted that intellectual property disputes remain one of the field's more unresolved questions even as adoption accelerates; this works through why, and what is actually being done about it.

Who owns an AI-generated drug molecule?

In practice, ownership of the molecule itself is usually the easier half of this problem. Most AI-driven drug discovery (AIDD) companies operate under well-established commercial models, service agreements, licensing deals, or in-house development, in which intellectual property for whatever gets discovered is contractually assigned, typically to the pharmaceutical partner funding the work. That clean default exists precisely because it makes deals possible: a licensee or partner needs confidence that the IP they are paying for is not encumbered by a separate claim from the AI vendor whose tools helped find it.

The complication is that contractual ownership and patent inventorship are legally separate questions. A company can own a patent outright while still being required, under patent law, to correctly name the individual humans who made a significant intellectual contribution to conceiving the invention. Getting that second question wrong, even when ownership is not in dispute, can expose a patent to a later validity challenge. That is the narrower, thornier problem the rest of this piece is really about.

The inventorship problem

The foundational rule, settled after years of test litigation over an AI system called DABUS, is that only a natural person can be named as a patent inventor; no jurisdiction that has substantively considered the question has concluded otherwise. That still leaves the harder question open: when a human and an AI model collaborate closely on a discovery, which human, if any, actually qualifies as the inventor?

The USPTO's February 2024 Inventorship Guidance is the most detailed attempt yet to answer that for AI-assisted inventions generally. It applies the long-standing Pannu factors, requiring a significant, non-trivial intellectual contribution to an invention's conception, and translates them into a handful of guiding principles specific to AI. Paraphrased, the core ideas are: using an AI system does not by itself disqualify a person from being an inventor; simply posing a research goal or running a pre-built model does not qualify as a significant contribution; but a person who builds an essential piece of the system that produces the invention, or who identifies a specific problem and designs an AI tool to solve it, may qualify even without being present for every step that followed.

Insilico Medicine's own General Counsel has written that these principles, while a genuine step forward, were illustrated with simplified, linear examples that do not fully capture how AIDD actually works: a real discovery program typically layers multiple AI platforms, target identification, generative chemistry, property prediction, across iterative rounds involving chemists, biologists and data scientists, none of whom individually resembles the guidance's tidy hypothetical inventor. The practical response many companies have adopted is procedural rather than legal: documenting, at each stage of a discovery program, who made which decision and why, so that inventorship can be reconstructed and defended later if a patent is challenged.

Proprietary data and training data IP

Inventorship is only one of three separable ownership questions running through an AI drug discovery program simultaneously: who owns the trained model, who owns the data that trained it, and who owns whatever molecule the model helped identify. These do not automatically travel together. A model trained on one company's proprietary assay data, then licensed for use on a second company's target, can easily end up with disputed ownership over improvements the model makes during that second engagement, even when the resulting drug candidate's ownership is contractually clear.

This is now a design consideration for informatics platforms themselves, not just a legal afterthought. Revvity Signals has described building its AI features specifically to keep a client's underlying data and intellectual property from becoming exposed to or absorbed by the third-party AI models integrated into its software, a direct acknowledgment that the software layer connecting a company's proprietary data to an outside AI tool is itself a point of IP risk that has to be engineered around, not just contracted around.

Data-sharing consortia: MELLODDY, OpenADMET

Two consortia illustrate genuinely different answers to the same underlying puzzle: how can competitors share enough data to build better models without giving away the proprietary information that makes their data valuable in the first place?

MELLODDY (Machine Learning Ledger Orchestration for Drug Discovery), an EU-funded effort launched in 2019 spanning ten pharmaceutical companies alongside technology and academic partners, completed a three-year federated learning experiment in 2022 built around 2.6 billion confidential experimental data points covering more than 21 million molecules. No partner ever saw another partner's raw data; each trained a shared model locally and only encrypted model updates, not underlying compounds or assay results, were combined into a global model. Every one of the ten partners measurably improved their own predictive models as a result, without any partner's proprietary data ever leaving its own servers.

OpenADMET takes the opposite approach: fully open data and fully open models, rather than private data feeding a shared model. Governed by the Open Molecular Software Foundation and funded by ARPA-H, the Gates Foundation and other backers, it focuses specifically on ADMET and toxicity properties, the pharmacokinetic and safety characteristics responsible for a large share of clinical failures, publishing open datasets, open-source models and public benchmark challenges rather than keeping any of it proprietary to a single sponsor.

Competitive advantage with shared models

The apparent paradox, why competitors would cooperate at all, resolves once the specific thing being shared is separated from the thing being protected. MELLODDY's partners never shared their compounds, targets or assay results; they shared only the statistical signal extracted from training on that data, aggregated in a way no partner could reverse-engineer to recover another's underlying molecules. The competitive asset, the proprietary dataset itself, stayed private throughout; only its generalizable predictive value was pooled.

OpenADMET's model rests on a related but distinct idea: pre-competitive data. ADMET liabilities, the properties that cause a compound to fail for reasons unrelated to how well it hits its target, are widely treated across the industry as a shared cost of doing business rather than a source of competitive edge. A company's actual target selection and lead compounds remain closely guarded; the underlying toxicology and metabolism data used to screen out bad actors increasingly is not, on the logic that better shared ADMET models reduce costly late-stage failures for everyone without giving any single competitor an edge over another.

Emerging legal frameworks

The DABUS litigation, filed across roughly a dozen jurisdictions starting in 2018 by an inventor seeking to name his own AI system on patent applications, has by now run its course in most of them, and the results are strikingly consistent:

What this settles is narrower than it sounds: it closes off naming an AI system itself as inventor, but it does not resolve the much more common, much murkier question of which human collaborator in a genuinely AI-assisted discovery program qualifies as one. That question is being worked out through patent office guidance and internal documentation practice at a time, not by a single sweeping ruling, and is likely to remain unsettled longer than the DABUS question ever was.

This article was produced in accordance with Drug Discovery News’ AI Editorial Policies.