How to Patent an AI-Designed Drug: The Rules After November 2025

Copyright © DrugPatentWatch. Originally published at https://www.drugpatentwatch.com/blog/

On November 28, 2025, the USPTO published a notice that rescinded, “in its entirety,” its February 2024 guidance on AI-assisted inventions. [1] The new rule is short. AI is a tool. A human must conceive the invention. The test for a human inventor is the same test the Office uses for every invention. [1][2] This article explains what that rule means for a company that finds a drug with machine learning, and what it does not settle.

The change matters because AI-derived drugs are now in human trials. A drug that an AI platform helped to find has reported Phase IIa data in a peer-reviewed journal. [14] Other AI-discovery programs have been cut after clinical data did not support further spending. [16] Patent claims on these assets face four legal tests: inventorship, eligibility, obviousness, and enablement. Each test has recent authority. This article collects that authority in one place and then adds our own analysis.

AI-discovered molecules showed an 80-90% success rate in Phase I trials, compared with an average of 40-65% for human-discovered molecules. The Phase II success rate was about 40%, on a limited sample. [17]

Source: Boston Consulting Group research published in Drug Discovery Today, as reported by Health Tech World, May 10, 2024.

The Short Answer: Can You Patent a Drug That an AI Designed?

Yes, if a natural person conceived the claimed invention and the claims meet the normal tests for novelty, non-obviousness, and disclosure. An AI system cannot be an inventor under US law. [3][4] The USPTO now treats AI systems as “instruments used by human inventors,” in the same class as laboratory equipment, software, and research databases. [1]

The hard questions are about proof and scope. A team must show which person had a “specific, settled idea, a particular solution to the problem at hand.” [1] A team must also write a specification that supports every claim to the full scope claimed. The Supreme Court held in 2023 that a patent on a whole class of compounds must enable the whole class. [8] A generative model can propose far more molecules than a team can describe. That gap is the main drafting risk.

We use one original framework in this article. It splits the IP around an AI-derived drug into five layers: compound, use and formulation, target biology, model and training method, and data and know-how. The framework is our own. It is not an established industry classification. The section titled “A Five-Layer Map of AI-Drug IP” explains it.

What the Data Shows: Ten Findings That Matter

Each finding below stands alone and has a source. Findings 8 and 9 are calculations made for this article. They are labeled as analysis.

  1. The USPTO notice of November 28, 2025 rescinds the February 13, 2024 guidance on AI-assisted inventions in its entirety. [1][2]
  2. The USPTO now applies one conception standard to all inventions. It has “no separate or modified standard for AI-assisted innovations.” [2]
  3. The Pannu factors apply only when the Office decides whether several natural persons are joint inventors. [1]
  4. The Supreme Court denied review of Thaler v. Vidal on April 24, 2023, so the Federal Circuit rule that an inventor must be a natural person stands. [3]
  5. On April 18, 2025, the Federal Circuit held in Recentive Analytics v. Fox Corp. that claims to known machine learning methods applied in a new field are patent ineligible. [5]
  6. On November 4, 2025, the USPTO designated Ex parte Desjardins as precedential. The panel found claims to machine learning model training eligible because they improve the model itself. [6][7]
  7. In Amgen v. Sanofi, the patents disclosed 26 antibody examples and claimed a whole functional genus. The Supreme Court held the claims not enabled. [8]
  8. Analysis: In three hypothetical term calculations built from 35 U.S.C. 156, patent term extension ranged from 1 year to 5 years. The result depends on when approval comes after filing. [11][12]
  9. Analysis: A patent that claims only an AI model does not claim an active ingredient. Under Teva v. Amneal, it cannot be listed in the Orange Book. [9][10]
  10. Recursion ended or paused five programs in May 2025, including two clinical programs that met safety goals but showed no meaningful clinical benefit. [16]

Methodology: How Did We Build the Derived Figures?

What data did this analysis use?

We did not query a private patent database. We read primary legal documents where we could reach them: the Federal Register notice, the Federal Circuit opinion in Teva v. Amneal, the USPTO announcement on Desjardins, the text of 35 U.S.C. 154 and 156, and the FDA draft guidance page. [1][10][6][11][12][18] For other cases we used law firm analyses and trade press. We label these as secondary sources in the text.

Which period and records are included?

The legal events run from August 2022 to November 2025. The clinical and corporate events run from February 2023 to July 2026. We include a record only if a regulator, a court, a peer-reviewed paper, or a named publication states it. Status can change after publication. Check current dockets and regulator pages before you rely on any item.

How were the calculations done?

The term calculations apply the statute text. Patent term runs 20 years from the earliest US filing date. [12] Extension under section 156 includes half of the testing period and all of the approval period. It cannot exceed five years. The total remaining term after approval cannot exceed fourteen years. [11] Each scenario uses invented dates. We label every one as hypothetical. We did not apply every statutory adjustment, such as reductions for lack of due diligence.

What are the limits of this analysis?

This article has no patent count dataset. We did not review the patent filings of any company named here. Where we discuss claim types, we describe general risk, not the claims of a real patent. The AI success-rate figures come from a secondary report of a study. We did not retrieve the study itself. Courts have not decided many questions in this area. Our scenario table is analysis, not a prediction.

Definitions: Which Terms Matter Here?

What is an AI-derived drug?

An AI-derived drug is a drug candidate for which a machine learning system did part of the target selection or molecule design. The term has no legal meaning. Patent law asks one question: did a natural person conceive the claimed invention.

What is conception?

Conception is “the formation in the mind of the inventor, of a definite and permanent idea of the complete and operative invention.” [1] The USPTO says conception is complete when the inventor has “a specific, settled idea, a particular solution to the problem at hand.” [1] A “general goal or research plan” is not enough. [2]

What is a patent that can be listed in the Orange Book?

Three patent types must be listed for an approved small-molecule drug: drug substance patents, drug product patents, and method-of-use patents. [9] Process, packaging, metabolite, and intermediate patents must not be listed. [9]

What is patent term extension?

Patent term extension (PTE) adds time to one patent for a drug whose regulatory review used up part of the term. The extension “may not exceed five years.” [11] The owner must apply within 60 days after the product receives permission for commercial marketing. [13]

Legal Timeline: What Happened and When?

DateEventSource
August 5, 2022Federal Circuit affirms in Thaler v. Vidal that an inventor must be a natural person[4]
April 24, 2023Supreme Court denies review in Thaler (No. 22-919)[3]
May 18, 2023Supreme Court decides Amgen v. Sanofi (No. 21-757) for Sanofi, unanimously[8]
February 13, 2024USPTO issues Inventorship Guidance for AI-Assisted Inventions (89 Fed. Reg. 10043)[2]
December 20, 2024Federal Circuit decides Teva v. Amneal (No. 2024-1936) on Orange Book listing[10]
January 2025FDA publishes draft guidance on AI to support regulatory decisions (docket FDA-2024-D-4689)[18]
April 18, 2025Federal Circuit decides Recentive Analytics v. Fox Corp. (No. 2023-2437)[5]
June 3, 2025Nature Medicine publishes Phase IIa results for rentosertib[14]
September 26, 2025Appeals Review Panel decides Ex parte Desjardins (Appeal 2024-000567)[7]
November 4, 2025USPTO designates Desjardins precedential[6]
November 28, 2025Federal Register publishes revised AI inventorship guidance, which rescinds the 2024 guidance[1]

Who Can Be the Inventor of an AI-Derived Drug?

Can an AI system be named as an inventor on a US patent?

No. The Federal Circuit affirmed on August 5, 2022 that the Patent Act requires an inventor to be a natural person. [4] The Supreme Court denied review on April 24, 2023. [3] The question presented was whether the Patent Act “categorically restrict[s] the statutory term ‘inventor’ to human beings alone.” [3]

The case involved Stephen Thaler and his system DABUS. The application was No. 16/524,350. The USPTO denied it in May 2020 because the application data sheet did not name an inventor by legal name. [3] The Eastern District of Virginia granted summary judgment to the USPTO. [3] The Goodwin summary of the 2025 USPTO notice cites the Federal Circuit decision as 43 F.4th 1207. [2]

What did the USPTO change on November 28, 2025?

The USPTO replaced a special AI inventorship framework with the ordinary conception test. The notice appeared in the Federal Register, Volume 90, Number 227. [1] It carries docket number PTO-P-2025-0014. [2]

What did the USPTO rescind?

The notice states that the February 13, 2024 guidance “is rescinded in its entirety.” [1] That guidance appeared at 89 Fed. Reg. 10043. It had applied the Pannu joint-inventor factors to inventions made with AI. [2] It also had examples. The rescission removed those examples. [2]

What does the conception test require?

The inventor must have formed “a definite and permanent idea of the complete and operative invention.” [1] The notice says AI systems “are instruments used by human inventors.” [1] A single person who uses AI is judged under the traditional conception standard. [1]

What happens to a foreign priority claim that names an AI tool?

The notice states that “a priority claim to a foreign application that names an AI tool as the sole inventor will not be accepted.” [1] The US application must name the same natural person, or at least one common natural-person joint inventor, as the earlier application. [1] This matters for teams that file first outside the United States. Check the inventor field of every foreign filing before you claim priority.

Do the Pannu factors still apply to AI-assisted drug discovery?

Only among human collaborators. The notice states that “the Pannu factors only apply when determining whether multiple natural persons qualify as joint inventors.” [1] Each human contributor must still contribute significantly to the conception of at least one claim. AI use “does not change the joint inventorship analysis among the human contributors.” [2]

The 2024 guidance had treated the human-AI relationship through Pannu. Sterne Kessler notes that Pannu v. Iolab Corp., 155 F.3d 1344 (Fed. Cir. 1998), no longer applies where the other participant is an AI. [19] The practical effect is that a team cannot argue that a model “contributed” to a claim in a way that dilutes or shifts inventorship.

How should a team document human conception in an AI workflow?

A team should record, for each claim, which person formed the specific idea and when. DLA Piper makes the same point. It advises teams to record which natural persons conceived each claimed concept, and notes that reviewing an AI output is unlikely to be enough. [4]

Which records should the team keep?

  • Dated notebook entries that state the hypothesis before the model run
  • The model version, the input data set, and the settings used in each run
  • The reasons a scientist chose one output over the others
  • Synthesis and assay records that link each tested compound to a named person

Who is likely to qualify as an inventor? Role-by-role analysis

The table below is our analysis. It applies the quoted conception standard to common roles in an AI discovery team. It is not a legal opinion, and the answer for any real project depends on the claims.

Role in the workflowLikely effect under the conception test (analysis)Basis
Sets a general goal, such as “find an inhibitor of target X”A general goal is not conception[2]
Runs a model with standard settings and passes on the outputThe person uses an instrument. Reviewing output is unlikely to be enough[1][4]
Forms a specific structural idea and chooses a compound to claim, with a settled reasonMay qualify, if the idea is specific and settled for a claimed compound[1]
Synthesizes and tests a compound that another person conceivedCarrying out another person’s conception does not qualify[4]
Designs a new model architecture or training methodMay qualify for claims to the model or method. Does not by itself qualify for claims to a compound the model later proposes[1][2]

Is an AI-Derived Drug Invention Patent Eligible Under Section 101?

What did Recentive Analytics v. Fox Corp. decide?

The Federal Circuit decided on April 18, 2025 that claims to machine learning applied in a new field were ineligible. [5] The case number is 2023-2437. The District of Delaware had dismissed the claims under section 101. The Federal Circuit affirmed. [5]

The court found the claims directed to the abstract ideas of producing event schedules and network maps with known machine learning techniques. [5] It framed the issue as whether the claims “apply established methods of machine learning to a new data environment.” [5] It found the claims lacked “a specific implementation of a solution to a problem in the software arts.” [5] Recentive conceded at oral argument that the patents did not claim “improving the mathematical algorithm or making machine learning better.” [5]

What did Ex parte Desjardins decide, and why does it help?

It decided that claims to a better machine learning training method can be eligible. The Appeals Review Panel decided Appeal 2024-000567 on September 26, 2025. [7] The USPTO designated the decision precedential on November 4, 2025. [6]

The claims covered training machine learning models with continual learning techniques. [7] The Patent Trial and Appeal Board had found them directed to an abstract idea. The panel reversed that finding. [7] It held that the claims reflected “improvements in artificial intelligence (AI) technology.” [6] The claims still stood rejected under section 103. [6] The USPTO page says sections 102, 103, and 112 remain the main tools that limit patent scope. [6]

Where does section 101 matter for a drug company?

Section 101 matters most for claims to software, models, and data methods. It matters less for a claim to a specific new compound. This is our analysis. The two cases above both involve software claims. [5][7] Neither case addresses a claim to a new chemical structure. We found no source in this review that applies Recentive to a compound claim.

For a model claim, the two cases give a clear drafting rule. Claim a specific improvement to the model or its training. Do not claim “use machine learning to find a drug” as a general idea. [5][7] Recentive and Desjardins point in the same direction on this rule, even though they reach opposite results.

Will AI Make Your Compound Obvious? Section 103 Risk

Can an AI-found molecule be “obvious to try”?

It can be. Sterne Kessler argues that AI can evaluate millions of molecular combinations quickly, which may make some discoveries obvious to try where success could be reasonably expected. [19] The firm cites KSR Int’l Co. v. Teleflex Inc., 550 U.S. 398 (2007), on this point. [19] This is a legal argument from a law firm. No court has adopted it for AI-found molecules in the sources we reviewed.

Should the skilled person be presumed to use AI tools?

The question is open. Sterne Kessler lists two open issues: whether prior art includes training data sets, and whether the person of ordinary skill should be presumed to have computational skill and access to AI tools. [19] If a court says yes to the second issue, the obviousness bar rises for every molecule found by a standard model.

Which facts help a compound claim survive an obviousness challenge?

Sterne Kessler suggests four factors: the size and variety of the chemical space explored, unexpected structure-function relationships, the degree of human intervention, and novel AI architectures or training methods. [19] It also advises that teams collect objective evidence such as long-felt need and the failure of others, with a clear link to the claimed features. [19]

Why do unexpected results matter more now?

Sterne Kessler advises that prosecutors stress unexpected results or superior properties that the training data would not predict. Its example is a showing that the AI-identified compound clearly outperforms computationally similar candidates. [19] In our analysis, this evidence also helps in a later Paragraph IV case. The brand company can use the same data to answer an obviousness attack.

Enablement and Written Description: What Does Amgen v. Sanofi Mean for AI Claims?

What did the Supreme Court hold in Amgen v. Sanofi?

The Court held, unanimously, that Amgen’s patents on a genus of antibodies were not enabled. [8] Justice Gorsuch wrote the opinion. [8] The two patents claimed a whole genus defined by the function of binding to specific amino acids. They identified 26 examples. [8] A patent on a class must enable others “to make the whole class without undue experimentation.” [8] The Court rejected Amgen’s argument that its disclosure was a “roadmap.” Mayer Brown describes the disclosed method as “open-ended trial and error.” [8]

Why do generative models create a disclosure problem?

A model can propose a very large set of molecules that share a property. A team may be tempted to claim the whole set by function. This is our analysis, and it follows from the Amgen holding. [8] If the specification gives a few examples and leaves the rest to trial and error, the claim is at risk.

DLA Piper raises a second disclosure issue. It notes that patents claiming “machine learning” or “neural networks” without specifics risk section 112 challenges. [4] It also notes that each generative system is unique and depends on its inputs and training data. How much of the system must be disclosed is not settled. [4]

Which claim types carry which disclosure risk? A comparison

The table below is our analysis. It ranks claim types by how closely they match the facts of Amgen. It does not predict outcomes.

Claim typeDisclosure risk (analysis)Reason
A specific compound with a defined structure, with synthesis and assay dataLowerThe claim names one structure and the specification can support it
A defined structural family, with many worked examplesMediumThe claim is wider than the examples. The examples must support the full scope [8]
Any compound that inhibits target X for use in disease Y, defined by functionHigherIt matches the functional genus claimed in Amgen [8]
A method that uses “a machine learning model” to find such compounds, with no model detailHigherSection 101 and section 112 risk together [4][5]

A Five-Layer Map of AI-Drug IP: An Original Framework

We propose a five-layer map to organize the IP around an AI-derived drug. The map is our own framework. It is not an established industry classification. It helps a team decide which tool protects which asset, and which listing or disclosure rule applies. Each row is drawn from the sources cited in the table.

LayerWhat it coversMain protectionMain legal riskOrange Book listable?Source
1. CompoundThe active ingredient and its salts or formsPatent (composition of matter)Inventorship proof, obviousnessYes, as a drug substance patent[1][9][19]
2. Use and formulationApproved indication, dosage form, deliveryPatent (method of use, drug product)Obviousness, label matchYes, as method-of-use or drug product patent[9]
3. Target biologyThe link between a target and a diseasePatent, if claimed with specific compoundsEnablement for functional claimsOnly if the patent claims the approved drug[8][10]
4. Model and training methodThe algorithm, architecture, and training processPatent (if a specific technical improvement) or trade secretSection 101, section 112No (analysis)[5][7][4]
5. Data and know-howTraining data, intermediate outputs, assay recordsTrade secret and contractMisappropriation proof, enforceabilityNo[19]

The map shows one point that many term sheets miss. Layers 1 and 2 earn exclusivity at the pharmacy. Layers 3, 4, and 5 protect the discovery engine. A generic company challenges layers 1 and 2. A competing AI company challenges layers 3, 4, and 5. The two groups use different legal tools.

Case Studies: What Do Real AI-Derived Programs Show?

Rentosertib (ISM001-055): What did the Phase IIa trial show?

Rentosertib is a TNIK inhibitor for idiopathic pulmonary fibrosis (IPF). Its target was found with a generative AI approach, and Insilico Medicine designed the molecule with its Pharma.AI platform. [14] Nature Medicine published the Phase IIa results on June 3, 2025. [14]

What were the trial results?

The trial was double-blind and placebo-controlled. It ran for 12 weeks at 22 sites in China and enrolled 71 patients. [14] The arms were placebo, 30 mg once daily, 30 mg twice daily, and 60 mg once daily. [14] In the 60 mg once-daily group, mean forced vital capacity (FVC) rose by 98.4 mL. In the placebo group it fell by 20.3 mL. FVC was a secondary endpoint. The primary endpoint was safety and tolerability, and the sponsor reported that it was met. [14]

Analysis: The difference between the 60 mg group and placebo is 118.7 mL (98.4 minus negative 20.3). This is a calculation from the reported means. It is not a reported statistic. The trial was small and short, and FVC was not the primary endpoint.

What does the program show about speed?

The sponsor reports that 22 candidates nominated from 2021 to 2024 took an average of 12-18 months from project start to preclinical candidate nomination. The typical time is 2.5-4 years. Each project needed about 60-200 synthesized and tested molecules. [14] Analysis: The midpoints are 15 months and 39 months. The reported AI-assisted midpoint is 61.5% shorter. This figure comes from the sponsor’s own numbers and has no independent check.

What are the IP questions?

FDA granted Orphan Drug Designation to the asset in February 2023. [15] A July 2026 report says Insilico is advancing rentosertib toward Phase III. It gives no design details and says nothing about patents. [15] We did not review any rentosertib patent filing. We cannot say which claim types protect it. The case still shows the general point: the AI step saved time in discovery, but the value of the asset now depends on clinical data. Exclusivity comes from layers 1 and 2 of the map.

Recursion: What happens when clinical data do not support a program?

On about May 5, 2025, Recursion ended or paused five programs after its merger with Exscientia. [16] It terminated REC-2282 for neurofibromatosis type 2 and REC-994 for cerebral cavernous malformation. [16] It paused REC-4539, a clinical oncology program, and explored out-licensing REC-3964 for C. difficile infection. [16] It also stopped one preclinical program. [16]

The REC-994 Phase II trial enrolled 62 patients. It met its safety and primary endpoint but showed no meaningful clinical benefit. Follow-up data did not confirm earlier signals. [16] The company ended the quarter with $509 million in cash and equivalents. It said its runway extended into mid-2027. [16]

The lesson for IP is plain. A valid patent on a molecule has little value if the molecule fails in the clinic. The source does not discuss the patents on these programs. We make no claim about them.

Do AI-discovered drugs succeed more often in clinical trials?

The early data say they do in Phase I and do not in Phase II. Research reported in Drug Discovery Today found an 80-90% Phase I success rate for AI-discovered molecules, against 40-65% for human-discovered molecules. [17] It found about 40% in Phase II, which the report calls in line with industry averages. The Phase II sample was limited. [17] The same report states that the overall probability of success rises from 5-10% to 9-18% with AI. [17] We did not retrieve the study itself, and the sources we read do not list the molecule counts.

What Is the Patent Clock on an AI-Derived Drug?

How long does a patent on an AI-designed drug last?

A utility patent ends “20 years from the date on which the application for the patent was filed in the United States.” [12] If the application claims priority to an earlier US application, the 20 years run from the earliest such filing date. [12] AI-assisted discovery tends to push the filing date early, because teams file on a compound soon after the model proposes it. An early filing date uses up term before approval.

How does patent term extension work?

Extension adds back part of the time lost to regulatory review. The statute includes only one-half of the testing period and all of the approval period. The total extension “may not exceed five years.” [11] The statute also reduces the extension so that the remaining term after approval plus the extension “does not exceed fourteen years.” [11] The owner must apply within 60 days after approval. [13]

Three hypothetical scenarios: how much extension would a patent earn?

The table is hypothetical analysis. All dates are invented. Each case assumes a patent filed on January 1, 2020, so the base expiry is January 1, 2040. [12] The testing period starts January 1, 2022. We applied the one-half rule, the five-year cap, and the fourteen-year cap. [11] We ignored other statutory adjustments.

ScenarioApproval dateTerm left at approvalHalf of testing period + approval period14-year cap allowsExtension earnedNew expiry
A: Approval 11 years after filingJan 1, 20319 years3.5 + 2 = 5.5 years5 years5 years (five-year cap)Jan 1, 2045
B: Approval 14 years after filingJan 1, 20346 years5 + 2 = 7 years8 years5 years (five-year cap)Jan 1, 2045
C: Approval 7 years after filingJan 1, 202713 years1.5 + 2 = 3.5 years1 year1 year (fourteen-year cap)Jan 1, 2041

The table gives one clear result. A fast program loses extension to the fourteen-year cap. In scenario C, a drug that reaches approval 7 years after filing earns 1 year. In scenarios A and B, the five-year cap decides the result, and the expiry date is the same. Speed in discovery does not add to patent life. It can add to the time on the market only if the filing date stays fixed.

Which patents on an AI-derived drug go in the Orange Book?

Drug substance, drug product, and method-of-use patents go in the Orange Book. [9] Process, packaging, metabolite, and intermediate patents do not. [9] A listed patent can trigger a 30-month stay. If the patent owner sues within 45 days after notice of a Paragraph IV certification, FDA cannot approve the generic application for 30 months. [9]

On December 20, 2024, the Federal Circuit decided in Teva v. Amneal that “a patent must claim an invention containing the active ingredient.” [10] It held that patents claiming “just the device components of the product approved in an NDA do not meet the listing requirement.” [10] The court affirmed the order that required Teva to delist five patents. [10]

Analysis: The same reasoning points away from listing a patent that claims only a discovery model or training method. Such a patent does not claim an active ingredient. We found no court decision that applies Teva to AI-model patents, so this is an inference from the holding and the listing rules. [9][10] The practical result is that a model patent gives no stay and no Orange Book signal to generic filers.

How can a team monitor patents and generic entry for these assets?

Use primary records first: the Orange Book, USPTO records, and court dockets. A commercial database can speed the work. DrugPatentWatch says it draws on the USPTO, the FDA Orange Book, ANDA filings, litigation records, and patent registers in 134 countries, with daily refresh. [20] It also tracks drugs with expiring patents and generic entry opportunities. [20] We treat its output as a starting point. Confirm any date or listing against the underlying government record before you rely on it.

Trade Secret or Patent: How Should a Team Protect the Model and the Data?

When does a trade secret protect better than a patent?

A trade secret can protect parts of the discovery engine that may not meet patent eligibility rules. Sterne Kessler lists model architectures, predictive insights, raw data, and intermediate outputs. [19] A patent requires disclosure. A trade secret does not. This matters because Recentive requires a specific technical improvement for a model claim to survive, and DLA Piper notes that disclosure of the system may be needed for section 112. [5][4]

What are the limits of trade secret protection?

Sterne Kessler notes enforceability concerns. A company must prove misappropriation, and competitors can build similar AI architectures on their own. [19] The firm also advises that licenses, research collaborations, and employment contracts address AI-generated inventions directly. [19] In our analysis, a team that mixes patents and trade secrets should decide early which model details go into the specification and which stay private. A later change is hard, because published text cannot be taken back.

What This Means for AI Biotech Companies, Pharma Licensors, and Generic Challengers

What this means for AI-native biotech companies

Build the human conception record from the first model run. Draft compound claims with worked examples that match their scope. [8] Draft model claims around a named technical improvement. [5][7] Check every foreign filing for an inventor field that names an AI tool. [1]

What this means for brand manufacturers that license AI-derived assets

Ask the licensor for the inventorship records, the list of named inventors by claim, and the prosecution history. Ask which patents are Orange Book listable. [9][10] Ask how the licensor treats the model and data: as patents, as trade secrets, or both. [19] Price the asset on clinical data first. The Recursion cuts show that a program can end even when a trial meets its safety endpoint. [16]

What this means for generic and biosimilar challengers

The grounds of attack are the old grounds. The USPTO page on Desjardins says sections 102, 103, and 112 remain the main tools that limit patent scope. [6] A challenger can test inventorship, obviousness, and enablement. It can use the argument that AI makes a molecule obvious to try. [19] It can use Amgen against functional genus claims. [8] No source we reviewed shows a court accepting the first argument for an AI-found molecule. Treat it as untested.

A challenger should also note the stay rule. A listed patent with a timely suit gives the brand a 30-month stay. [9] A patent that cannot be listed gives no stay. [9][10]

Scenario Analysis: How Could Courts Treat AI-Derived Drug Claims?

The table is our analysis. Each scenario rests on a stated assumption. We assign no probabilities, because we found no court decision that decides these points for AI-found molecules.

ScenarioAssumptionEffect on compound patents (analysis)Related source
1. AI treated as a lab instrumentCourts follow the USPTO position that AI is an instrument used by human inventorsInventorship fights shift to proof of human conception. The compound patent stands or falls on novelty, obviousness, and disclosure[1]
2. Skilled person presumed to use AICourts presume that the person of ordinary skill has AI toolsMore molecules look obvious to try. Unexpected results and objective evidence carry more weight[19]
3. Higher disclosure for AI systemsCourts require more detail on the model and data to support broad claimsBroad functional claims weaken. Specifications need model and data detail. Trade secret strategy narrows[4][8]

What Happens Next: FDA, USPTO, and the Courts

Does the FDA draft guidance on AI affect patents?

Not directly. FDA published its draft guidance, “Considerations for the Use of Artificial Intelligence To Support Regulatory Decision-Making for Drug and Biological Products,” in January 2025. It is a Level 1 draft with non-binding recommendations. The docket number is FDA-2024-D-4689. [18] The page describes a risk-based framework to assess an AI model’s credibility for a stated context of use. It covers AI used to produce data about drug safety, effectiveness, or quality. [18] The page we reviewed does not address patents or inventorship, and it does not say whether drug discovery uses are in or out of scope.

What should a team watch over the next year?

Watch four items. First, any court decision that applies the conception test to an AI-assisted compound claim. Second, any decision that addresses the skilled person and AI tools in an obviousness case. Third, later USPTO examples or decisions that apply Desjardins to life sciences claims. Fourth, any final version of the FDA guidance. We found no source for these events as of November 16, 2026, but our search was not exhaustive.

Pre-Filing Checklist for AI-Derived Drug Patents

StepActionBasis
1Name only natural persons as inventors, and check them against foreign priority filings[1][3]
2Record each person’s specific idea, with dates, before and after each model run[1][4]
3Match claim scope to worked examples and assay data[8]
4Describe the model and data in the specification to the level the claims need[4]
5Claim a model only through a specific technical improvement[5][7]
6Collect comparative data that show unexpected results[19]
7Decide which assets are patents and which are trade secrets[19]
8Plan the Orange Book listing and the 60-day PTE deadline[9][13]

Key Takeaways

  • The USPTO rescinded its 2024 AI inventorship guidance in full on November 28, 2025, and now applies the ordinary conception test. [1]
  • An AI system cannot be an inventor. A human must form a specific, settled idea. A general goal does not count. [1][2][3]
  • The Pannu factors apply only among human collaborators. [1]
  • Claims to known machine learning in a new field are ineligible. Claims that improve the model itself can be eligible. [5][7]
  • Claims to a whole class of compounds must enable the whole class. [8]
  • Analysis: Extension on an AI-derived drug patent ranges from 1 to 5 years in our three hypothetical cases, and a fast approval can lower it. [11]
  • Analysis: A model patent is not Orange Book listable, because it does not claim an active ingredient. [9][10]
  • Clinical data decide asset value. Recursion cut programs that met safety goals but showed no meaningful benefit. [16]

FAQ: AI-Derived Drug Patents

Can an AI be named as an inventor in the United States?

No. The Federal Circuit held that an inventor must be a natural person, and the Supreme Court denied review on April 24, 2023. [3][4]

Does the USPTO apply a different test to AI-assisted inventions now?

No. The November 2025 guidance applies the conception standard to all inventions and states that there is no separate or modified standard for AI-assisted innovations. [2]

Can I claim priority to a foreign application that names an AI as inventor?

Not if the AI tool is the sole named inventor. The USPTO “will not” accept that priority claim. The US application must share at least one natural-person inventor with the foreign filing. [1]

Can I patent a machine learning model that I use for drug discovery?

You can, if the claim states a specific technical improvement to the model or its training. A claim that applies known machine learning to a new field failed in Recentive. A claim to a training improvement succeeded at the USPTO in Desjardins. [5][7]

Can I list a model patent in the Orange Book?

Probably not. Orange Book listing covers drug substance, drug product, and method-of-use patents. [9] Teva v. Amneal requires that a listed patent claim at least the active ingredient. [10] This answer is our inference from those two sources.

How long does protection last for an AI-designed drug?

The base term is 20 years from the earliest US filing date. Extension can add up to 5 years, but the total term after approval cannot exceed 14 years. [11][12]

What is the deadline to apply for patent term extension?

The application is due within the 60-day period that begins on the date the product receives permission for commercial marketing or use. [13]

Should I keep my training data as a trade secret?

Often yes. Sterne Kessler lists raw data, model architectures, predictive insights, and intermediate outputs as candidates for trade secret protection. The firm also warns about enforcement limits. [19]

Do AI-discovered drugs succeed more often in trials?

In Phase I, the reported rate is 80-90%, against 40-65% for human-discovered molecules. In Phase II, it is about 40%, on a limited sample. [17]

Can a generic company invalidate a patent because an AI designed the molecule?

Not on that fact alone. The USPTO says sections 102, 103, and 112 are the main tools that limit patent scope. [6] A challenger can still argue obviousness or lack of enablement. [8][19]

References

  1. U.S. Patent and Trademark Office. (2025, November 28). Revised inventorship guidance for AI-assisted inventions. Federal Register, 90(227). https://www.govinfo.gov/content/pkg/FR-2025-11-28/html/2025-21457.htm
  2. Goodwin Procter LLP. (2025, December 1). USPTO issues revised inventorship guidance for AI-assisted inventions: Key takeaways for AI/ML innovators. https://www.goodwinlaw.com/en/insights/publications/2025/12/alerts-lifesciences-uspto-issues-revised-inventorship-guidance
  3. IPWatchdog. (2023, April 24). Supreme Court dodges AI inventor question with denial of DABUS case. https://ipwatchdog.com/2023/04/24/supreme-court-dodges-ai-inventor-question-denial-dabus-case/
  4. DLA Piper. (2024, April 15). Generative AI, drug discovery, and US patent law. https://www.dlapiper.com/en-us/insights/publications/synthesis/2024/generative-ai-drug-discovery-and-us-patent-law
  5. JD Supra. (n.d.). Federal Circuit refines Section 101 eligibility of machine learning inventions [Analysis of Recentive Analytics, Inc. v. Fox Corp., No. 2023-2437 (Fed. Cir. Apr. 18, 2025)]. https://www.jdsupra.com/legalnews/federal-circuit-refines-section-101-7961235
  6. U.S. Patent and Trademark Office. (2025, November 4). PTAB designates as precedential an Appeals Review Panel decision. https://www.uspto.gov/subscription-center/2025/ptab-designates-precedential-appeals-review-panel-decision
  7. JD Supra. (n.d.). Precedential shift: USPTO clarifies patentability of AI training methods [Analysis of Ex parte Desjardins, Appeal No. 2024-000567]. https://www.jdsupra.com/legalnews/precedential-shift-uspto-clarifies-6859750/
  8. Mayer Brown LLP. (2023, May 18). Decision alert: Amgen Inc. v. Sanofi, No. 21-757. https://www.mayerbrown.com/en/insights/publications/2023/05/decision-alert-amgen-inc-v-sanofi-no-21-757
  9. Finnegan, Henderson, Farabow, Garrett & Dunner, LLP. (2018, July 3). Requirements, benefits, and possible consequences of listing patents in FDA’s Orange Book. https://www.finnegan.com/en/insights/requirements-benefits-and-possible-consequences-of-listing-patents-in-fdas-orange-book.html
  10. U.S. Court of Appeals for the Federal Circuit. (2024, December 20). Teva Branded Pharmaceutical Products R&D, Inc. v. Amneal Pharmaceuticals of New York, LLC, No. 2024-1936. https://cafc.uscourts.gov/opinions-orders/24-1936.OPINION.12-20-2024_2439730.pdf
  11. BitLaw. (n.d.). 35 U.S.C. 156: Extension of patent term. https://www.bitlaw.com/source/35usc/156.html
  12. BitLaw. (n.d.). 35 U.S.C. 154: Contents and term of patent; provisional rights. https://www.bitlaw.com/source/35usc/154.html
  13. U.S. Patent and Trademark Office. (n.d.). MPEP 2750: Patent term extension (35 U.S.C. 156). https://www.uspto.gov/web/offices/pac/mpep/s2750.html
  14. News release on the Nature Medicine publication (June 3, 2025) of Phase IIa results for rentosertib in idiopathic pulmonary fibrosis, DOI 10.1038/s41591-025-03743-2. EurekAlert. https://www.eurekalert.org/news-releases/1086096
  15. Artificial Intelligence News. (2026, July 7). Insilico Medicine advances AI drug for IPF to Phase III trials. https://www.artificialintelligence-news.com/news/insilico-medicine-advances-ai-drug-for-ipf-to-phase-iii-trials/
  16. Pharmaceutical Technology. (2025, May 6). Recursion axes drug programmes to streamline pipeline. https://www.pharmaceutical-technology.com/news/recursion-axes-drug-programmes-to-streamline-pipeline/
  17. Health Tech World. (2024, May 10). AI-discovered drugs achieved higher success rate than those by humans: Study [Reporting on research published in Drug Discovery Today]. https://www.htworld.co.uk/news/ai/ai-discovered-drugs-achieved-higher-success-rate-than-those-by-humans-study/
  18. U.S. Food and Drug Administration. (2025, January). Considerations for the use of artificial intelligence to support regulatory decision-making for drug and biological products (Draft guidance; Docket FDA-2024-D-4689). https://www.fda.gov/regulatory-information/search-fda-guidance-documents/considerations-use-artificial-intelligence-support-regulatory-decision-making-drug-and-biological
  19. Sterne, Kessler, Goldstein & Fox P.L.L.C. (n.d.). Intellectual property challenges in AI-driven drug discovery. https://www.sternekessler.com/news-insights/insights/intellectual-property-challenges-in-ai-driven-drug-discovery/
  20. DrugPatentWatch. (n.d.). Home page. https://www.drugpatentwatch.com

Make Better Decisions with DrugPatentWatch

» Start Your Free Trial Today «

Copyright © DrugPatentWatch. Originally published at
DrugPatentWatch - Transform Data into Market Domination