AI Drug Discovery Hits Phase 3. Who Owns the Patents?

On September 14, 2026, Takeda said FDA accepted its New Drug Application for zasocitinib under Priority Review. The drug treats moderate-to-severe plaque psoriasis. FDA’s target action date is the first quarter of 2027. [1] Schrödinger’s physics-based software helped design the molecule. [2] Four days earlier, Insilico Medicine announced the first patient dosed in a 320-patient Phase 3 trial of rentosertib. Insilico says generative AI found both the target and the molecule. [3]

These two events move computational drug design from research papers into regulatory review and large clinical trials. They also raise patent questions. Who can be an inventor? Can a company patent the AI method? How long does exclusivity last after approval? This article answers those questions. We use primary sources where we could access them. We label every calculation as our analysis.

The Short Answer

As of October 10, 2026, FDA has not approved zasocitinib or rentosertib. Zasocitinib is under review. Rentosertib is in Phase 3. [1] [3]

Four points frame the rest of the article:

US patent law still requires a human inventor. The USPTO treats AI as a tool, like laboratory equipment or software. [4]
FDA’s January 2025 draft guidance on AI does not cover drug discovery. [5]
The standard FDA exclusivity clock starts at approval, not at discovery. A drug designed with AI gets the same five-year new chemical entity exclusivity as any other new molecule. [6]
The clinical evidence is early. The best-known analysis of AI-discovered molecules found high Phase 1 success and Phase 2 success near historic averages. [7]

What the Data Shows: Eight Findings That Matter

Each finding below stands alone. We mark calculated figures as analysis.

Boston Consulting Group authors found an 80% to 90% Phase 1 success rate for AI-discovered molecules. They found about 40% in Phase 2. [7]
In two Phase 3 trials, 61% and 52% of zasocitinib patients reached PASI 90 at week 16. Apremilast patients reached 17% and 16%. The apremilast comparison is descriptive only. Our analysis: the gaps are 44 and 36 percentage points. [1]
Takeda paid about $4 billion upfront for the zasocitinib program. It owes two $1 billion milestones if annual net sales reach $4 billion and $5 billion. [8] Schrödinger reported $147.3 million in distributions. Our analysis: that is about 3.7% of the upfront payment. [9]
In the rentosertib Phase 2a trial, the 60 mg once-daily group gained 98.4 mL of forced vital capacity at 12 weeks. The placebo group lost 20.3 mL. Our analysis: the difference is 118.7 mL. [10]
The rentosertib Phase 3 trial plans 320 patients for 52 weeks. Our analysis: that is 4.5 times the Phase 2a patient count and 4.3 times the treatment length. [3] [10]
The USPTO issued revised inventorship guidance on November 28, 2025. It rescinded the February 2024 guidance and applies the conception test to AI-assisted inventions. [11]
FDA’s draft guidance on AI states that it “does not address the use of AI models (1) in drug discovery.” [5]
Recursion’s cash fell from $603.0 million on December 31, 2024 to about $509 million on March 31, 2025. Our analysis: that is a drop of about $94 million, or 15.6%, in one quarter. [12] [13]

What Is an AI-Discovered Drug? Definitions and a Five-Tier Framework

No regulator defines the term “AI-discovered drug.” Companies use it in different ways. The definition you pick changes the success rate you calculate.

Definition: AI-Discovered, AI-Assisted and Computation-Designed Drugs

The BCG analysis looked at the clinical pipelines of AI-native biotech companies. [7] That is a company-level definition. It does not test how much AI touched each molecule.

FDA gives no definition for discovery work. Its draft guidance leaves discovery out of scope. [5]

We use “AI-discovered” only when the developer says AI found the target, generated the molecule, or both. We use “computation-designed” for programs where the design method is software-based but the developer does not call it AI. The Takeda release we reviewed gives trial and regulatory data. It does not describe the design method. [1] Schrödinger describes physics-based prediction and virtual screening. [2]

Original Framework: Five Tiers of AI Involvement

This framework is our own. It is not an industry standard. It sorts programs by how much of the discovery work the software did. The tiers help you read claims from developers and compare programs on equal terms.

Tier 1: Computational Optimization of a Human-Chosen Target

Scientists pick the target. Software ranks and designs molecules. Zasocitinib fits here. Schrödinger reports free-energy perturbation (FEP+) predictions of potency, selectivity and solubility. It also reports a virtual screen of 2.4 million compounds. [2]

Tier 2: AI-Generated Molecule Against a Known Target

Scientists choose a known target. A generative model proposes the molecules. Exscientia and Sumitomo Dainippon Pharma developed DSP-1181 for obsessive-compulsive disorder in this style of work. [14]

Tier 3: AI-Identified Target Plus AI-Generated Molecule

The software picks the target and then designs the molecule. Insilico calls rentosertib “the world’s first drug candidate discovered and developed using generative AI.” That is the company’s claim. We did not verify it independently. [3]

Tier 4: Platform Programs Built on a Data Engine

A company builds a large data and automation platform. It runs many programs on that platform. Recursion calls itself a TechBio company and describes its Exscientia merger as the largest TechBio merger in history. That is also a company claim. [12]

Tier 5: Structure Prediction and Protein Design Tools

These tools support discovery but do not name a drug. The 2024 Nobel Prize in Chemistry went to David Baker for computational protein design. Demis Hassabis and John Jumper shared the other half for protein structure prediction. [15]

Tier	What the software does	Example in this article	Source
1	Optimizes molecules for a human-chosen target	Zasocitinib (TAK-279)	[2]
2	Generates molecules for a known target	DSP-1181	[14]
3	Finds the target and generates the molecule	Rentosertib (ISM001-055)	[3]
4	Runs many programs on one data platform	Recursion pipeline	[13]
5	Predicts or designs protein structures	AlphaFold, Rosetta	[15]

Table: Original five-tier framework. Tier assignments are our analysis.

Do AI-Discovered Drugs Succeed More Often? Phase 1 and Phase 2 Success Rates

The only systematic analysis we found says Phase 1 success is high and Phase 2 success is about average. The authors work at Boston Consulting Group. They published in Drug Discovery Today in 2024. [7]

What the BCG Analysis Found

AI-discovered molecules showed an 80% to 90% success rate in Phase 1, which Jayatunga and colleagues describe as “substantially higher than historic industry averages.” [7]

HTWorld reported the same study on May 10, 2024. It gives a historic Phase 1 success rate of 40% to 65% for molecules found without AI. It also reports a projected end-to-end success probability of 5% to 10% without AI and 9% to 18% with AI. [16]

Metric	AI-discovered	Historic benchmark	Our derived figure	Source
Phase 1 success rate	80% to 90%	40% to 65%	Gap of 32.5 points between range midpoints	[7] [16]
Phase 2 success rate	About 40%	Described as in line with industry averages	No gap calculated	[7] [16]
End-to-end clinical success	9% to 18%	5% to 10%	Ratio of 1.8 at both ends of the range	[16]

Table: AI-discovered versus historic clinical success rates. Midpoint gap and ratio are our calculations.

What the Numbers Do Not Show

Three limits apply. Each one matters for a reader who plans to cite the 80% to 90% figure.

Sample Size and Selection

The authors say the Phase 2 sample was limited. [7] HTWorld does not report the number of molecules in each phase. [16] A small sample can swing a rate by many points.

What Phase 1 Measures

Phase 1 trials mainly test safety and dose. They do not test whether a drug works. The authors read the Phase 1 result as a sign that AI designs molecules with drug-like properties. [7] The Phase 2 result is the first efficacy test, and it sits near the historic average.

The commercial implication is direct. A high Phase 1 rate cuts early cost and time. It does not remove the Phase 2 risk. Patent value for an AI-designed molecule still depends on Phase 2 and Phase 3 results.

Zasocitinib (TAK-279): A Computation-Designed Drug Under FDA Priority Review

Zasocitinib is an oral TYK2 inhibitor. It is the most advanced program in this article. FDA accepted the NDA on Priority Review, and the target action date falls in the first quarter of 2027. The European Medicines Agency accepted a marketing authorization application. [1]

Zasocitinib Timeline: From Nimbus and Schrödinger to an NDA

Schrödinger became a collaborative partner on the program in 2016. [2] Nimbus first targeted the catalytic JH1 domain. It later moved to the regulatory JH2 domain after new literature from Bristol Myers Squibb. Schrödinger says the team named a development candidate within two years of that change. That is a vendor statement. [2]

Nimbus reported positive Phase 2b topline results in November 2022. [2] Takeda closed its acquisition of the Nimbus TYK2 subsidiary on February 8, 2023. At that date Takeda expected a Phase 3 psoriasis study to start in 2023. [8]

The Phase 3 program has three registered studies:

LATITUDE PsO 3001 (NCT06088043): 693 participants.
LATITUDE PsO 3002 (NCT06108544): 1,108 participants.
LATITUDE PsO 3003 (NCT06550076): an open-label long-term study with about 2,100 participants. It gives supportive data only. [1]

Analysis: the time from deal close (February 8, 2023) to NDA acceptance (September 14, 2026) is about 43 months. The time from the start of the Schrödinger collaboration in 2016 to NDA acceptance is about ten years. We calculated both from the dates in the cited sources. [8] [1] [2]

Zasocitinib vs. Apremilast: Phase 3 LATITUDE Results at Week 16

Studies 3001 and 3002 ran in 21 countries. The co-primary endpoints were sPGA 0/1 and PASI 75 versus placebo at week 16. Both studies met them. [1]

Week 16 endpoint	3001 zasocitinib	3001 placebo	3001 apremilast	3002 zasocitinib	3002 placebo	3002 apremilast
PASI 75	76%	12%	37%	71%	12%	33%
PASI 90	61%	5%	17%	52%	4%	16%
PASI 100	33%	0.7%	3%	25%	1%	4%
sPGA 0/1	71%	11%	32%	69%	13%	30%
sPGA 0	40%	0.7%	8%	34%	1%	7%

Table: Phase 3 week 16 results. Source: Takeda press release of September 14, 2026. [1]

Our Calculation: The Gap Versus Apremilast

We subtracted the apremilast rate from the zasocitinib rate in each study. We also divided the two rates.

Endpoint	3001 gap (points)	3001 ratio	3002 gap (points)	3002 ratio
PASI 75	39	2.1	38	2.2
PASI 90	44	3.6	36	3.3
PASI 100	30	11.0	21	6.3
sPGA 0/1	39	2.2	39	2.3
sPGA 0	32	5.0	27	4.9

Table: Derived gaps between zasocitinib and apremilast. All values are our analysis of the rates in the previous table.

Why the Apremilast Comparison Is Descriptive

Takeda states that all comparisons with placebo and apremilast had p-values below 0.001. It also states that the apremilast comparisons are descriptive and not controlled for multiple testing. [1] A reader should not call zasocitinib superior to apremilast on this basis. The trials did not test that claim in a formal way.

Analysis: the placebo-adjusted PASI 75 effect is 64 points in study 3001 and 59 points in study 3002. We subtracted the placebo rate from the zasocitinib rate. [1]

Who Got Paid: Takeda, Nimbus and Schrödinger

Takeda paid about $4 billion upfront. The release says the amount was adjusted for closing indebtedness and accounts payable. Two further payments of $1 billion each are due if products from the TAK-279 program reach annual net sales of $4 billion and $5 billion. [8]

Schrödinger held a 3.8% fully diluted stake in Nimbus on December 31, 2022. It reported a $111.3 million cash distribution on February 14, 2023. It expected a second distribution of about $36.0 million in the second quarter of 2023. The total is $147.3 million. [9]

Analysis: $147.3 million is about 3.7% of the $4.0 billion upfront payment. That is close to the 3.8% equity stake. The milestone payments, if earned, would add $2 billion. That equals 50% of the upfront amount. [8] [9]

Zasocitinib Patents and Orange Book Status

The Orange Book lists approved drug products. FDA has not approved zasocitinib. Our analysis: the Orange Book therefore holds no patent listing for it today.

We do not list zasocitinib patent numbers here. We could not confirm them against USPTO records while we wrote this article. You can check patent and exclusivity entries in a data service such as DrugPatentWatch after FDA acts on the NDA. Check the Orange Book entry against the patent record before you rely on it.

Rentosertib (ISM001-055): AI-Identified Target and AI-Generated Molecule in Phase 3

Rentosertib is an oral small molecule that inhibits TNIK. Insilico develops it for idiopathic pulmonary fibrosis (IPF). It is investigational and no regulator has approved it. [3]

GENESIS-IPF Phase 2a Results Published in Nature Medicine

Insilico announced the Nature Medicine publication on June 3, 2025. The GENESIS-IPF trial was double-blind and placebo-controlled. It ran for 12 weeks at 22 sites in China. It enrolled 71 patients. [10]

Item	Value	Source
Design	Phase 2a, double-blind, placebo-controlled, 12 weeks	[10]
Sites	22, all in China	[10]
Patients	71 with IPF	[10]
Arms	Placebo; 30 mg once daily; 30 mg twice daily; 60 mg once daily	[10]
FVC change, 60 mg once daily	+98.4 mL	[10]
FVC change, placebo	-20.3 mL	[10]
Difference (our analysis)	118.7 mL	Calculated from the two rows above

Table: GENESIS-IPF design and headline result. The release text we reviewed does not state the FVC change for the other two dose groups.

Insilico reports that treatment-emergent adverse event rates were similar across groups. Most events were mild to moderate. Serious events were rare. All events resolved after the drug was stopped. [10]

What the Phase 2a Data Can Support

The data support a signal. A placebo group lost lung function and the high-dose group gained it over 12 weeks. The safety profile looked similar to placebo. This justified a larger trial, which Insilico has now started. [10] [3]

What the Phase 2a Data Cannot Support

Seventy-one patients across four arms is a small base. We do not have the patient count per arm from the sources we reviewed. The trial ran for 12 weeks, and IPF is a disease that progresses over years. All sites were in China. Insilico itself says the results “warrant further investigation in larger-scale clinical trials of longer duration.” That statement is from CEO Alex Zhavoronkov, as reported in the release. [10]

GENESIS-IPF-3: Phase 3 Design and Timeline

Insilico announced the first patient dosed in GENESIS-IPF-3 on September 10, 2026. The trial is prospective, multi-center, randomized, double-blind and placebo-controlled. Patients take rentosertib once daily for 52 weeks. About 320 participants are expected at 47 centers in China. [3]

The primary endpoint is the annual rate of decline in forced vital capacity over 52 weeks. The key secondary endpoint is time to first disease progression event. The registry numbers are NCT07687459 (ClinicalTrials.gov) and CTR20262475 (China). [3]

Metric	Phase 2a (GENESIS-IPF)	Phase 3 (GENESIS-IPF-3)	Ratio (our analysis)
Patients	71	About 320	4.5
Treatment length	12 weeks	52 weeks	4.3
Sites	22	47	2.1

Table: Phase 2a versus Phase 3 scale. Sources: [10] [3]. Ratios are our calculations.

Bio-IT World reported a Phase 3 launch on July 8, 2026, after a press briefing. It described a 320-patient, 52-week, placebo-controlled study. [17] Our analysis of the dates: the first patient was dosed about two months after that report. We treat the September 10 release as the date that trial activity began. We do not use the timeline Bio-IT World printed for later milestones, because it contains a date that precedes the article itself.

Zhavoronkov told Bio-IT World: “Our records are 9-18 months to go from zero to this developmental candidate.” [17] That is a company statement about discovery speed. It is not a result from an independent study.

Analysis: the first Phase 3 patient was dosed about 15 months after the Nature Medicine announcement. We calculated that from June 3, 2025 to September 9, 2026. [10] [3]

Orphan Drug Designation and Seven-Year Exclusivity for IPF

FDA granted orphan drug designation to ISM001-055 for IPF in early 2023. Insilico’s press release is dated February 8, 2023. A rare disease affects fewer than 200,000 people in the US. Benefits include tax credits for qualified clinical testing and fee exemptions. They also include seven years of market exclusivity if FDA approves the drug. [18]

Analysis: orphan exclusivity attaches to the approved indication. It can run alongside new chemical entity exclusivity. We explain the combined dates in the exclusivity section below.

Insilico’s Hong Kong IPO: What the Capital Raise Shows

Insilico listed on the Hong Kong Stock Exchange on December 30, 2025 under stock code 3696.HK. The offering raised HKD 2.277 billion. Fifteen cornerstone investors took part, including Lilly, Tencent, Temasek and Schroders. [19]

Insilico calls the offering the largest biotech IPO in Hong Kong that year by fundraising size. The release gives no independent data for that claim. [19] We repeat it only as a company statement.

When AI-Discovered Programs Fail: DSP-1181, Recursion Cuts and BenevolentAI

A fair view of AI drug discovery includes the programs that stopped. Three cases are documented in public sources.

DSP-1181: A Phase 1 Study That Missed Its Criteria

Exscientia and Sumitomo Dainippon Pharma developed DSP-1181 for obsessive-compulsive disorder. It entered a Phase 1 trial by the end of 2020. The partners dropped it about a year later. The Phase 1 study “did not reach its expected criteria,” and development in Japan stopped. [14]

The same source says two more compounds from the collaboration, DSP-0038 and EXS-21546, entered clinical trials in 2021. It gives no success rate for AI-found compounds. [14] This case shows that a Phase 1 failure happens even for a molecule that drew wide attention at entry.

Recursion’s Pipeline Cuts After the Exscientia Merger

The Exscientia business combination closed in November 2024. [13] On May 5, 2025, Recursion reported first-quarter results and cut part of its pipeline. [13]

Program	Indication	Action reported May 5, 2025	Source
REC-2282	Neurofibromatosis type 2	Deprioritized	[13]
REC-994	Cerebral cavernous malformation	Deprioritized	[13]
REC-3964	C. difficile infection	Deprioritized; out-licensing under review	[13]
REC-4539	Solid tumor	Paused	[13]
Undisclosed	Preclinical	Wound down	[13]

Table: Recursion program actions. Fierce Biotech reports that Recursion kept six active projects, four in cancer and two in rare disease.

Cash and Net Loss Numbers

Recursion reported $603.0 million in cash, cash equivalents and restricted cash on December 31, 2024. A year earlier the figure was $401.4 million. Full-year 2024 net loss was $463.7 million. Revenue was $58.8 million. [12]

Fierce Biotech reports cash of about $509 million at the end of March 2025. [13]

Calculation (our analysis)	Inputs	Result
Change in cash, Dec 31, 2024 to Mar 31, 2025	$603.0 million to about $509 million	About -$94 million (-15.6%)
2024 revenue as share of net loss	$58.8 million / $463.7 million	12.7%
Q4 2024 revenue as share of net loss	$4.5 million / $178.9 million	2.5%

Table: Recursion burn metrics. Inputs from [12] and [13].

What the Cuts Say About Platform Risk

Analysis: a platform company carries two kinds of risk. A single program can fail in the clinic. The whole platform can also fail to turn data into approved drugs fast enough to pay for itself. The numbers above show a company that spent far more than it earned in 2024. The program cuts show a company that narrowed its scope in response. [13] [12]

For patent teams, a cut program is also a patent question. A company that out-licenses or abandons a program must decide whether to keep paying maintenance fees on the related patent families. Public sources do not show what Recursion decided for each family.

BenevolentAI: Delisting Instead of Growth

BenevolentAI announced a proposed delisting from Euronext Amsterdam on February 5, 2025. It would merge into Osaka Holdings S.à r.l., a Luxembourg company. Each share converts one-to-one into a share of Osaka Holdings. The release gives this reason: “the financial and administrative costs of maintaining a public listing are not justified by the benefits.” [20]

The release does not discuss pipeline results. We make no claim about why the company chose that route beyond its own words.

Can AI Be a Patent Inventor? US Law After Thaler v. Vidal and the 2025 USPTO Guidance

No. A US patent needs a human inventor. An AI system can help, but a natural person must conceive the claimed invention. Two sources fix this rule: a 2022 Federal Circuit decision and a 2025 USPTO guidance notice.

Thaler v. Vidal (Federal Circuit, August 5, 2022)

Stephen Thaler listed his AI system, DABUS, as the sole inventor on patent applications. The Federal Circuit upheld the earlier rulings and held that only natural persons can be inventors. The court read 35 U.S.C. § 100(f), which defines the inventor as “the individual or … individuals collectively who invented,” to mean a human being. It found the statute unambiguous. [21]

The court did not decide whether a human who invents with AI help can get a patent. It left that question open. [21] The citation is Thaler v. Vidal, 43 F.4th 1207 (Fed. Cir. 2022).

USPTO Revised Inventorship Guidance (November 28, 2025)

The USPTO published revised guidance on November 28, 2025. It rescinded the February 2024 guidance. The Federal Register document number is 2025-21457. [4] The guidance implements Executive Order 14179 of January 23, 2025. [11]

The USPTO treats AI systems as tools. Jones Day quotes the guidance as calling them “analogous to laboratory equipment, computer software, research databases” and similar tools. [4] An application that names an AI system as inventor will be rejected under 35 U.S.C. §§ 101 and 115. [11]

Single Inventor Using AI

The conception test applies. A natural person must form a definite and permanent idea of the complete and operative invention. [4] The USPTO asks whether the person can describe the invention with particularity. [11] The Pannu factors do not apply to a single person who uses AI. [11]

Joint Inventors

The Pannu factors still apply when several natural persons share inventorship. Each joint inventor must make a significant contribution to conception. [11] The guidance drops the Pannu factors as a test of AI’s own contribution. [4]

What Should a Drug Discovery Team Document?

Mayer Brown does not give a formal checklist. It describes steps that follow from the guidance. [11] Our reading, as analysis, gives four records for a discovery team:

A dated record of who defined each claim limitation, such as the structure, the substituents and the dose range.
A record of how a scientist selected, changed and combined any structure or parameter that the software proposed.
For several inventors, a record of each person’s contribution to conception.
For priority claims, a check that the earlier and later filings share at least one natural-person inventor. [11]

The commercial reason is simple. An inventorship error can put patent validity at risk. The cost of a clean record is small by comparison. This is general information and not legal advice.

Can You Patent the AI Method Itself? Section 101 After Recentive v. Fox and Ex parte Desjardins

It depends on the claim. A claim that applies a generic machine learning model to a new data setting is not eligible. A claim that improves how the model trains or runs can be. Two 2025 decisions set this line.

Recentive Analytics v. Fox Corp. (Federal Circuit, April 18, 2025)

The Federal Circuit decided a question of first impression. It held that “patents that do no more than claim the application of generic machine learning to new data environments” are ineligible when they do not disclose an improvement to the models. The patents concerned event scheduling and network mapping for broadcasts. The court found them directed to an abstract idea with no inventive concept. [22]

Ex parte Desjardins (USPTO Appeals Review Panel, 2025)

The Appeals Review Panel decided Ex parte Desjardins on September 26, 2025. The USPTO designated it precedential on November 4, 2025. The panel held that claims to training a machine learning model can be eligible under 35 U.S.C. § 101. This holds when the claims integrate a mathematical concept into a practical application that improves how the model operates. The claims were still rejected on obviousness grounds under § 103. [23]

A December 5, 2025 advance notice of MPEP changes told examiners to look for a detailed disclosure of a technological improvement and to confirm the claim as a whole reflects it. [23]

What This Means for AI Drug Discovery Platform Companies

Analysis: the two decisions point the same way. A claim to “use machine learning to find a kinase inhibitor” resembles the generic application that Recentive rejected. A claim to a specific training method that improves the model resembles the claims in Desjardins. We state this as our reading of the sources, not as a legal opinion. [22] [23]

The molecule itself is a separate matter. A patent on a compound or on a method of treatment does not depend on a software-method claim. That is why the patent layers below matter more for the drug than for the platform.

The Patent Layer Cake: A Taxonomy of Protection for an AI-Designed Drug

This taxonomy is our own. The first three layers are common categories in pharmaceutical patenting. The grouping and the AI-specific notes are our analysis.

Layer	What it protects	AI-specific issue	Evidence in this article
1. Composition of matter	The molecule	A natural person must have conceived the claimed structure	[4]
2. Method of use	Treatment of a named disease with the molecule	The AI-identified target (rentosertib) supports a use claim, and inventorship rules still apply	[3] [11]
3. Formulation and dosing	The product form and regimen, such as an inhaled version	Insilico runs a parallel Phase 1 of an inhaled formulation	[3]
4. AI platform and method	The training or generation method	Eligible only if the claim improves the technology	[22] [23]
5. Regulatory exclusivity	Market entry, with no patent needed	The clock starts at approval, not at discovery	[6] [18]

Table: Five layers of protection for an AI-designed drug. Layer definitions are our framework.

Exclusivity Math: NCE, NCE-1, Paragraph IV and the 30-Month Stay

The use of AI in discovery does not change the length of FDA exclusivity. The dates run from approval. This section shows the arithmetic.

What Are the FDA Exclusivity Rules for a New Chemical Entity?

An FDA Office of Generic Drugs presentation states three rules:

During the five-year NCE period, no applicant can submit an ANDA or 505(b)(2) application for the same active moiety.
If the NDA lists patents in the Orange Book, an ANDA or 505(b)(2) application can be submitted with a Paragraph IV certification at the “NCE-1” date, which is year four.
If the brand company sues within the one-year window that begins four years after approval, the 30-month stay extends so that 7.5 years pass from NDA approval. [6]

Scenario Analysis: Zasocitinib Exclusivity Dates If Approved in Q1 2027

Assumptions: FDA approves zasocitinib on a date in the first quarter of 2027, which is the target action window. [1] FDA treats it as a new chemical entity. The NDA lists patents in the Orange Book. The brand company sues within the one-year window. These are our assumptions. FDA may act earlier, later or differently.

Milestone	Rule	If approved Jan 1, 2027	If approved Mar 31, 2027
Earliest ANDA with Paragraph IV (NCE-1)	Approval + 4 years	Jan 1, 2031	Mar 31, 2031
NCE exclusivity ends	Approval + 5 years	Jan 1, 2032	Mar 31, 2032
End of extended 30-month stay	Approval + 7.5 years	Jul 1, 2034	Sep 30, 2034

Table: Scenario dates. Rules from [6]. Dates are our calculations and not predictions.

Scenario Analysis: Rentosertib Exclusivity Formula for IPF

Insilico has not announced an approval date in the sources we reviewed. We therefore give a formula. Let A be the approval date. If rentosertib qualifies as a new chemical entity, the NCE-1 date is A plus four years and NCE exclusivity ends at A plus five years. [6] The orphan designation adds seven years of exclusivity for the IPF indication, so that protection ends at A plus seven years. [18]

Analysis: for IPF, the orphan period outlasts the NCE period by two years. The combined effect depends on the final approved labeling and on FDA’s orphan exclusivity determination at approval.

What Happens If a Generic Files a Paragraph IV Certification at NCE-1?

A generic company can file at year four. It must certify that a listed patent is invalid, unenforceable or not infringed. If the brand company sues in the one-year window, the stay extends to 7.5 years from approval. [6]

Analysis: this rule gives the brand company about 2.5 years of stay after the five-year NCE period ends. For a drug with $1 billion in annual sales, that is a large amount of revenue at stake in the litigation. We do not forecast sales for either drug here.

Does FDA Regulate AI in Drug Discovery? Draft Guidance and the FDA-EMA Principles

Not by a binding rule. FDA’s draft AI guidance excludes drug discovery. A joint FDA and EMA set of principles covers early research but is voluntary. [5] [24]

FDA Draft Guidance of January 2025: Scope and the Seven-Step Framework

FDA published the draft on January 7, 2025 under docket FDA-2024-D-4689. It is a Level 1 draft with non-binding recommendations. [25] It covers AI used to produce data that support decisions on drug safety, effectiveness or quality. It states that it “does not address the use of AI models (1) in drug discovery.” It also excludes some uses for operational efficiency. [5]

The guidance sets a risk-based credibility framework with seven steps:

Define the question of interest.
Define the context of use for the AI model.
Assess the AI model risk.
Develop a plan to establish AI model credibility within the context of use.
Execute the plan.
Document the results and discuss deviations from the plan.
Determine the adequacy of the AI model for the context of use. [5]

Model risk combines model influence and decision consequence. FDA defines decision consequence as “the significance of an adverse outcome resulting from an incorrect decision concerning the question of interest.” [5]

FDA and EMA Good AI Practice Principles (January 14, 2026)

The EMA and FDA published joint guiding principles on January 14, 2026. They apply across the medicines life cycle, from early research and clinical development to manufacturing and post-marketing safety. Adoption is voluntary. [24] Jones Day lists eight concepts, such as a risk-based approach, a clear context of use and robust data governance. [24]

Regulatory Gap Analysis: Where Discovery-Stage AI Sits

Analysis: a discovery team that uses AI faces no binding FDA credibility rule for that work. The credibility framework applies when AI output supports a regulatory decision. A company that submits AI-derived evidence in an NDA should expect FDA to ask about the model’s context of use. We cannot confirm from the sources how often FDA has asked so far.

Capital Flows: Isomorphic Labs, Insilico and Recursion

Investors keep funding the field. The three data points below come from company releases.

Company	Event	Amount	Date	Source
Isomorphic Labs	Series B led by Thrive Capital	$2.1 billion	May 12, 2026	[26]
Insilico Medicine	Hong Kong IPO (3696.HK)	HKD 2.277 billion	Dec 30, 2025	[19]
Recursion	Year-end cash, cash equivalents and restricted cash	$603.0 million	Dec 31, 2024	[12]

Table: Selected capital events. The rows measure different things and are not additive.

Isomorphic’s other Series B investors include Alphabet, GV, MGX, Temasek, CapitalG and the UK Sovereign AI Fund. The company says it will use the money for its AI drug design engine, IsoDDE, and for moving its pipeline toward the clinic. It lists partnerships with Novartis, Lilly and Johnson & Johnson. The release gives no candidate names or trial dates. [26]

AI Drug Discovery Timeline: Key Events From 2016 to 2026

Date	Event	Source
2016	Schrödinger becomes a collaborative partner on the Nimbus TYK2 program	[2]
End of 2020	DSP-1181 enters Phase 1; dropped about a year later	[14]
Nov 2022	Nimbus reports positive Phase 2b topline results	[2]
Feb 8, 2023	Takeda closes the Nimbus TYK2 subsidiary acquisition	[8]
Feb 2023	FDA orphan drug designation for ISM001-055 in IPF	[18]
Oct 9, 2024	Nobel Prize in Chemistry for protein design and structure prediction	[15]
Nov 2024	Recursion and Exscientia complete their merger	[13]
Jan 7, 2025	FDA publishes the draft AI guidance	[25]
Feb 5, 2025	BenevolentAI proposes delisting	[20]
Apr 18, 2025	Recentive v. Fox decided	[22]
May 5, 2025	Recursion cuts programs	[13]
Jun 3, 2025	Rentosertib Phase 2a published in Nature Medicine	[10]
Nov 28, 2025	USPTO issues revised inventorship guidance	[11]
Dec 30, 2025	Insilico lists in Hong Kong	[19]
Jan 14, 2026	FDA and EMA publish joint AI principles	[24]
May 12, 2026	Isomorphic Labs announces $2.1 billion Series B	[26]
Sep 10, 2026	First patient dosed in rentosertib Phase 3	[3]
Sep 14, 2026	FDA accepts zasocitinib NDA under Priority Review	[1]
Q1 2027 (projected)	FDA target action date for zasocitinib	[1]

Table: Timeline assembled from the cited sources. The Q1 2027 row is a company-reported target, not a confirmed event.

What This Means for Generic Manufacturers, Brand Companies and Licensing Teams

What This Means for Generic Entry

Generic teams should track NCE-1 dates, not discovery stories. If zasocitinib wins approval in the first quarter of 2027, the earliest Paragraph IV filing date falls in the first quarter of 2031 (our scenario). [6] DrugPatentWatch and the Orange Book are the places to check which patents the brand lists after approval.

What This Means for Brand Manufacturers

Brand IP teams should build the inventorship record at the time of discovery. The USPTO test is conception by a natural person. [4] Teams should also keep platform claims separate from compound claims, since the two face different eligibility tests. [22] [23]

What This Means for Business Development and Licensing

The Takeda and Nimbus deal shows how value is priced: a large upfront payment and sales-based milestones. [8] The Recursion cuts show the other side. [13] Diligence should test clinical data and patent position separately from the AI claim.

Methodology and Limitations

What Data We Used

We used company press releases, a regulator’s web pages and presentation, law-firm summaries of court and agency documents, and one peer-reviewed analysis. The period is 2016 to October 10, 2026. We included an event only when a cited source states it and gives a date.

How We Calculated

We subtracted and divided reported percentages and dollar amounts. We counted months between dated events. We added years to hypothetical approval dates using FDA’s stated rules. Every calculated figure is marked as our analysis. No duplicates arose because each event came from one dated source.

Limits

Several primary documents were not available to us during research, including the Nature Medicine paper, the Federal Register notice and Recursion’s SEC filings. For those we relied on named company releases or law-firm summaries. The reference list names the source for each claim. We did not query patent records, so we list no patent numbers. Company statements about being first or largest are labeled as company claims. Readers who need legal advice should consult counsel.

Key Takeaways

Zasocitinib has an NDA under Priority Review with a target action date in the first quarter of 2027. [1]
Rentosertib began Phase 3 dosing in September 2026 in a 320-patient, 52-week trial. [3]
BCG authors report 80% to 90% Phase 1 success and about 40% Phase 2 success for AI-discovered molecules. [7]
US patents need a human inventor. The USPTO treats AI as a tool. [4]
FDA’s draft AI guidance excludes drug discovery. [5]
Exclusivity runs from approval. Five-year NCE exclusivity and the year-four Paragraph IV filing date apply to AI-designed drugs as to others. [6]
Generic machine learning claims face eligibility problems. Claims that improve the model do better. [22] [23]
Failures are documented: DSP-1181, Recursion’s cuts and BenevolentAI’s delisting. [14] [13] [20]

FAQ

Has FDA approved an AI-discovered drug?

Not either of the two lead cases here. Zasocitinib is under review with a first-quarter 2027 target date. Rentosertib is in Phase 3. [1] [3]

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

No. The Federal Circuit held in 2022 that an inventor must be a natural person. The USPTO repeated this in its November 2025 guidance. [21] [11]

Who owns a patent on a drug that AI helped design?

The USPTO guidance addresses inventorship, not ownership. It names natural persons as inventors. Ownership usually follows assignment agreements, which the guidance does not cover. [11]

Does FDA require companies to disclose AI used in discovery?

The January 2025 draft guidance does not address drug discovery. It covers AI used to support regulatory decisions. [5]

How long is new chemical entity exclusivity?

Five years from approval. A generic applicant may file with a Paragraph IV certification at year four. [6]

What is the projected end-to-end success rate for AI-discovered drugs?

HTWorld reports 9% to 18% with AI versus 5% to 10% without it, from BCG research. Treat this as a projection. [16]

What did Takeda pay for the zasocitinib program?

About $4 billion upfront, plus two $1 billion milestones tied to annual net sales of $4 billion and $5 billion. [8]

When will FDA decide on zasocitinib?

Takeda gives a target action date in the first quarter of 2027. That is a target, not a decision. [1]

Can machine learning methods be patented in the US?

Sometimes. Generic use of machine learning in a new setting is ineligible. Claims that improve how a model trains can qualify. [22] [23]

How much money did AI drug discovery companies raise recently?

Isomorphic Labs announced a $2.1 billion Series B on May 12, 2026. Insilico raised HKD 2.277 billion in its December 2025 IPO. These are two examples, not an industry total. [26] [19]