AI drug discovery patents: molecule, method or model?

Yes, you can get a strong AI drug discovery patent, but usually not on the AI. A molecule your model found is claimed like any other new compound: it has to be new, non-obvious, useful and fully described. The model and the pipeline around it face a harder test, because software and math claims run into patent eligibility rules under 35 U.S.C. § 101.

This article is for founders and chief scientific officers of computational drug discovery companies. It compares the three claim types you will use (composition of matter, method and platform claims), shows what investors and pharma partners look for in each, and explains when a trade secret protects your model and training data better than a patent. It reflects USPTO guidance as of October 2026.

Key takeaways

• A compound found with AI is examined like any other compound. The statute says patentability “shall not be negated by the manner in which the invention was made.”
• Composition of matter claims on the molecule are the asset that pharma partners and investors value most. Back them with synthesis and assay data, not only model predictions.
• Claims to the AI platform itself must clear § 101. They are strongest when the specification and the claims show a specific improvement to how the model works.
• Model weights, training data and curation methods are often better kept as trade secrets, because a patent application publishes about 18 months after filing.
• Only natural persons can be inventors. Under current USPTO guidance, AI is a tool, and the usual conception test applies.

Can you get an AI drug discovery patent on a molecule your model found?

Yes. A new chemical compound is a “composition of matter,” one of the four categories that 35 U.S.C. § 101 says can be patented. How you found it does not matter. The compound must meet the same tests as one found by a medicinal chemist at a bench: novelty, non-obviousness, utility, and adequate written description and enablement.

35 U.S.C. § 101 covers “any new and useful process, machine, manufacture, or composition of matter,” and says nothing about how the invention was found.

The last sentence of 35 U.S.C. § 103, the obviousness statute, says patentability “shall not be negated by the manner in which the invention was made.” The examiner asks whether the claimed structure would have been obvious to a person of ordinary skill in the art, not whether your model made the search easy.

Where AI-found compounds actually run into trouble

The risks for an AI-found compound are scientific before they are legal. Three come up often.

• Prior art you didn’t look for. Generative models often propose close analogs of known actives. A structure already described in a printed publication before your filing date is not new under § 102. One that differs from a known active by a single added methyl group invites an obviousness argument.
• Predictions are not data. A docking score or predicted IC50 shows only that a model thinks a compound might work. Evidence of unexpected results over the closest prior art is the classic answer to an obviousness rejection, and that evidence comes from the lab.
• Genus claims that outrun the evidence. A model can enumerate thousands of structures. That does not mean you can claim them all (see enablement, below).

Composition of matter, method and platform claims: what is the difference?

A composition of matter claim covers the molecule itself, however it is made or used. A method claim covers a process, such as treating a disease with the compound or making it by a particular route. A platform claim covers your discovery engine: the model, the training method or the computational pipeline that finds candidates.

Composition of matter claims

A composition of matter claim reads like “A compound of Formula I, or a pharmaceutically acceptable salt thereof.” It is usually drafted as a Markush genus (a claim that lists alternative groups at defined positions on a core scaffold) plus narrower claims to specific compounds. It is infringed by anyone who makes, uses or sells the compound for any purpose. That breadth is why it is the most valuable claim in drug development. My small molecule patent practice page goes into salts, crystalline forms and other follow-on filings.

Method claims

A method of treatment claim covers giving the compound for a named condition. A method of manufacture covers a synthetic route. A patient-selection method covers using a biomarker, sometimes scored by a model, to choose who gets the drug. These claims are narrower than a compound claim, but they extend protection into new indications.

Platform claims

Platform claims cover your discovery engine: a model architecture, a training method, a molecular-graph encoding, or an active-learning loop that picks which compounds to synthesize next. They protect the engine rather than any one output, and they are the claims most likely to draw a § 101 rejection.

How the three claim types compare

Each claim type protects a different asset and faces a different main hurdle. The table sets them side by side.

Why do AI platform claims face 35 U.S.C. § 101 problems?

Mathematical concepts and mental processes are “abstract ideas” that cannot be patented by themselves. A machine learning model is built from math, so a claim to training or running one often recites an abstract idea. The claim survives if it integrates that idea into a practical application, such as a specific improvement to how the model works.

The Federal Circuit’s warning

In Recentive Analytics, Inc. v. Fox Corp., decided April 18, 2025, the Federal Circuit held that patents that “do no more than claim the application of generic machine learning to new data environments, without disclosing improvements to the machine learning models to be applied, are patent ineligible under § 101.” The case involved TV scheduling, not chemistry, but the reasoning is not limited to that field. A claim to “training a graph neural network on binding data and using it to rank compounds” invites the same rejection.

The USPTO’s current position

At the USPTO, the picture improved for well-drafted AI claims in 2025. An August 4, 2025 memo to examiners in the software arts reminded them that the “mental process grouping is not without limits” and that a rejection should not be made just because an examiner is uncertain about eligibility.

Then the Appeals Review Panel decided Ex parte Desjardins on September 26, 2025, and the Office designated it precedential on November 4, 2025. The claims covered training a machine learning model on new tasks while protecting what it learned on earlier ones. The panel found that the specification identified “improvements as to how the machine learning model itself operates,” and that the claims were eligible. A December 5, 2025 notice revising MPEP § 2106 wrote that reasoning into the examiner’s manual. Examiners first check whether the specification describes an improvement, then whether the claim reflects it. A bare statement that the invention is “better” is not enough.

What this means for drafting a platform claim

Write the specification around a technical problem and its technical fix. Suppose your team built a training method that keeps a binding-affinity model accurate on a new target class without retraining from scratch. Explain the problem (lost accuracy on earlier targets, excess compute) and how your architecture solves it, then put the elements that deliver the fix into the claim. The computational chemistry and AI drug discovery practice page describes the kinds of platform inventions I work on.

Method of treatment claims that use AI

The USPTO’s 2024 AI subject matter eligibility examples include Example 49, which is close to drug development. A claim that used an AI model to score a patient’s risk and then gave an “appropriate treatment” was ineligible. A dependent claim that named a specific treatment, “Compound X eye drops,” was eligible, because it applied the risk score in a particular treatment. If your model selects patients, claim the specific drug and the step that acts on the result.

How much data does a compound claim need?

Enough to show that you made and tested representative compounds across the scope you claim. Under 35 U.S.C. § 112(a), the specification must describe the invention and teach a skilled person how to make and use it. For broad genus claims the Supreme Court has set a demanding standard, and in my view a list of model-generated structures will rarely meet it alone.

In Amgen Inc. v. Sanofi (2023), a unanimous Court held that the specification “must enable the full scope of the invention as defined by its claims,” and that “the more one claims, the more one must enable.” The case concerned antibodies defined by their function. The same logic applies to a Markush genus defined by a model’s output space.

Here is a hypothetical. Your generative model proposes several hundred analogs around a new kinase-inhibitor core. You synthesize a few dozen and test them. A genus claim drawn around the substituents you actually made and tested, with structure-activity data showing which positions tolerate change, is defensible. A genus drawn around everything the model could generate, most of which no one has made, is a target for an enablement challenge. Draft the genus to match your data, and file new applications as the data grows.

What do investors and pharma partners expect for each claim type?

Pharma partners licensing a candidate usually want issued or pending composition of matter claims on the lead compound and its close analogs, with a clean chain of title. Investors in a platform company also want to know how the engine is protected, and whether that protection comes from patents, trade secrets or both.

• For an asset deal: expect diligence on the compound claims first. Who invented the lead? Was it in a poster or preprint before filing? Do the claims cover the clinical candidate and its salts?
• For a platform collaboration: expect questions about who owns compounds generated for the partner, and improvements to a model trained on the partner’s data. These are contract terms as much as patent questions.
• For a financing: expect investors to ask whether your platform claims have survived or are likely to survive § 101, and how you protect the parts you didn’t patent. I wrote about these diligence questions in what investors ask about your patents.

A portfolio that protects only the platform is hard to enforce, because a private model’s use is hard to detect. One that protects only compounds leaves the engine exposed. Most companies need a deliberate mix.

When are trade secrets better for the AI model and data?

When the asset can stay hidden and a competitor cannot learn it from your products. Model weights, training datasets, data-cleaning rules and hyperparameter choices usually meet that test. A compound does not, because its structure becomes public once it enters clinical development and appears in regulatory and patent filings.

Under the federal Defend Trade Secrets Act definition, 18 U.S.C. § 1839(3), information is a trade secret only if the owner has taken “reasonable measures to keep such information secret” and the information derives independent economic value from not being generally known. Patents work differently. An application generally publishes about 18 months after its earliest filing date, and the patent term generally ends 20 years from filing. You trade disclosure for a limited monopoly.

Use this checklist for each component of your platform.

1. Can a competitor see it? If the component is visible in your product, outputs or publications, a trade secret will not hold. Consider a patent.
2. Can you detect infringement? If a competitor could use your method privately and you would never know, a patent may be hard to enforce. Lean toward secrecy.
3. Will the advantage last 20 years? Model architectures change fast. A trade secret has no fixed end date, but it ends the day the secret gets out.
4. Does it clear § 101? If you can’t describe a concrete technical improvement in how the model works, a platform patent is a weak bet.
5. Do you have the controls? Secrecy depends on access limits, confidentiality agreements, and employee and contractor assignment terms.
6. Do partners need to see it? If a collaboration requires sharing the method, a patent filed before disclosure may protect you better than an NDA alone.

A common pattern: patent the training method, keep weights and curated data secret, and patent every development candidate. For a broader view of early-stage choices, see my notes for startups and university spin-outs.

Who is the inventor of an AI-discovered drug?

A human. Only natural persons can be named as inventors on a U.S. patent. The USPTO’s current guidance treats AI systems as tools, like lab equipment or software, and applies the same conception test to AI-assisted inventions as to any other invention.

The USPTO published its revised inventorship guidance for AI-assisted inventions on November 28, 2025. It rescinded the February 2024 guidance and states there is no separate standard for AI-assisted inventions; conception is “the touchstone of inventorship.” It cites Thaler v. Vidal (Fed. Cir. 2022) for the rule that only natural persons can be inventors. In practice, record who chose the target, set the design constraints and picked compounds from the model’s output.

A filing sequence that works for most AI drug discovery companies

File early on a platform improvement that clears § 101, keep the rest of the engine secret, and file on compounds once you have synthesis and assay data for a coherent series. Add method of treatment, salt and crystalline form filings as the program matures, and time every filing before public disclosure.

1. Inventory the platform and sort each component into “patent,” “trade secret” or “publish,” using the checklist above.
2. File a provisional application on any platform improvement you can describe in concrete technical terms, before a talk, a preprint or a partner pitch.
3. Run prior art searches on the model’s top candidates before you invest in synthesis.
4. File on the compound series once you have made and tested representative members, with a genus drawn to match the data.
5. Within 12 months of the provisional, file the nonprovisional application, and a PCT (international) application if you want foreign coverage, adding the new data.
6. Add method of treatment and patient-selection claims as indications and biomarkers become clear.

The U.S. one-year grace period under § 102(b)(1) covers only certain disclosures, and patent offices outside the U.S. apply their own rules, so treat a conference abstract as a filing deadline. I explain the timing traps in when to file: the clock starts earlier than you think.

Frequently asked questions

Can you patent a drug discovered by AI?

Yes. A compound found with AI is patentable on the same terms as any other compound. It must be new, non-obvious and useful, and the application must describe and enable what is claimed. 35 U.S.C. § 103 says patentability is not negated by the manner in which the invention was made. A human who conceived the claimed compound must be named as the inventor.

Can an AI model be patented?

Sometimes. Claims to a model or training method must pass 35 U.S.C. § 101. The Federal Circuit has held that applying generic machine learning to a new data environment is not eligible. Claims that describe and recite a specific improvement to how the model itself operates, as in the USPTO’s precedential Ex parte Desjardins decision, have a much better chance.

Should I patent my training data or keep it a trade secret?

Usually keep it secret. A patent application generally publishes about 18 months after filing, which would disclose your curation methods to competitors. Curated datasets, data-cleaning rules and model weights are usually better protected as trade secrets, provided you take reasonable measures to keep them secret. Patent the methods that competitors could learn from your products or publications.

How much experimental data do I need before filing on an AI-designed compound?

Enough to support the scope you claim. A narrow claim to a specific compound can rest on its synthesis and some activity data. A broad genus needs representative examples across the genus, because the Supreme Court held in Amgen v. Sanofi that the specification must enable the full scope of the claims. Model predictions alone are a weak foundation.

If you’re sorting out which parts of your platform to patent, which to keep secret, and when to file on a compound series, I’m happy to talk it through.

This article is general information, not legal advice, and reading it does not create an attorney-client relationship. Patent law changes and every situation differs. Please talk to a patent attorney about yours.

Message Alexander on LinkedIn