Machine learning patents in biotech: getting past § 101

A machine learning patent in biotech gets past 35 U.S.C. § 101 when the claim ties the model to a concrete technical result: a better measurement, a lab instrument that acts on the model’s output, or a specific treatment. Claims that only apply a known type of model to biological data usually fail. Claims that show how the model changes what happens in the lab usually have a path.

This article is for founders and in-house teams at AI-bio companies who want a machine learning patent biotech investors can rely on, with claims on assay pipelines, lab automation or diagnostic workflows. It covers USPTO practice after Ex parte Desjardins, the Federal Circuit cases that set the limits, and a drafting checklist, as of October 2026. If your question is whether to claim the molecule, the method or the model, start with AI drug discovery patents: molecule, method or model.

Key takeaways

• Biotech ML claims can face two judicial exceptions at once: the math in the model (an abstract idea) and the biology it measures (a law of nature).
• The USPTO’s precedential Desjardins decision and the December 2025 MPEP changes ask two questions: does the specification describe a technical improvement, and does the claim reflect it?
• Federal courts are stricter. Training a generic model on new biological data is “incident to the very nature of machine learning” and is not enough.
• The strongest anchors are a physical step the model controls, a new measurement technique, or a particular treatment.
• USPTO memos in 2026 describe how to use eligibility declarations (SMEDs) as evidence. Plan the data for one before you file.

Why is § 101 harder for a machine learning patent in biotech?

Because two exceptions can apply to one claim. Courts treat mathematical concepts and mental processes as abstract ideas, and a trained model is math. Courts also treat natural correlations, such as a biomarker level that tracks disease, as laws of nature. A claim that uses a model to detect a biological correlation can trip both at once.

Under the Supreme Court’s Alice and Mayo framework, a claim “directed to” a judicial exception (an abstract idea, a law of nature or a natural phenomenon) is ineligible unless it adds something more. The USPTO applies that framework through MPEP § 2106, asking whether a claim integrates the exception into a “practical application.”

A diagnostics company that trains a classifier on cell-free DNA fragment patterns fights both exceptions. The model is an abstract idea, and the link between fragment patterns and disease is a natural phenomenon. Fixing one problem does not fix the other.

What changed at the USPTO after Ex parte Desjardins?

Examiners must now look for a technical improvement in the specification and then check whether the claim reflects it. An improvement to how a machine learning model itself operates can count. A bare statement that the invention is better does not. These rules came from a precedential 2025 appeal decision that the USPTO wrote into the MPEP in December 2025.

The Appeals Review Panel decided Ex parte Desjardins on September 26, 2025. The USPTO designated it precedential on November 4, 2025. The claims covered training a model on new tasks while protecting what it learned on earlier ones. The panel held that the claims, though they recited a mathematical concept, integrated it into a practical application, because the specification described improvements to “how the machine learning model itself operates,” including reduced storage and preserved performance on earlier tasks.

The USPTO’s December 5, 2025 advance notice of change to the MPEP revised §§ 2106.04(d), 2106.05(a) and 2106.05(f). Three points matter for drafting:

• The improvement must be apparent to a skilled person from the specification. A “bare assertion of an improvement” without supporting detail does not count.
• The claim must reflect the improvement, though it need not recite it in so many words.
• Examiners should not evaluate claims “at such a high level of generality” that the technical details disappear.

The new emphasis on evidence: SMEDs

A subject matter eligibility declaration (SMED) is a declaration under 37 C.F.R. § 1.132 that offers objective evidence on eligibility. After a December 2025 memo to examiners, the USPTO issued best-practice memos for applicants on April 30, 2026 and September 29, 2026. The September memo says a SMED is optional but can help in “clarifying the record or in overcoming a subject matter eligibility rejection.”

The memo recommends filing the SMED separately, filing it early, and stating which part of the analysis it supports. An applicant’s own declaration is “less persuasive than that of a disinterested person,” so neutral support helps, such as a published paper describing the technical problem as it stood at filing.

Where do the courts draw the line on ML claims?

Federal courts are stricter than the USPTO. The Federal Circuit holds that applying generic machine learning to a new data environment, without disclosing an improvement to the model, is ineligible. In 2026 it applied that rule to imaging and language-processing patents. A patent that clears the USPTO still has to survive this test in litigation.

The key case is Recentive Analytics, Inc. v. Fox Corp. (Fed. Cir. 2025). The court 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 ineligible. It called iterative training “incident to the very nature of machine learning,” but added that machine learning “may lead to patent-eligible improvements in technology.”

Two 2026 decisions, both nonprecedential (not binding on future panels), show how the rule plays out:

• Dental Monitoring SAS v. Align Technology, Inc. (July 7, 2026). Claims used a “deep learning device” to assess orthodontic aligners from phone photos. One dependent claim required training on more than 1,000 dental arch images. The court held all the claims ineligible. Training on a specific dataset was “incident to the very nature of machine learning,” and the patents said the device could be chosen from available image-classification networks.
• Rensselaer Polytechnic Institute v. Amazon.com, Inc. (Feb. 24, 2026). Natural language processing claims were ineligible. The court held that “[a] conventional application of case-based reasoning, even to a novel environment, is abstract,” and that “attorney argument is no substitute for evidence.”

The lesson for biotech is direct. Swap “dental arch images” for “histology slides” or “mass spectra,” and the same claim structure fails. A bigger or more specialized training set does not, by itself, make a claim eligible.

What biotech cases say about lab methods and diagnostics

The Federal Circuit separates methods that produce or change something from methods that only observe a natural fact. A new lab technique that yields a better preparation is eligible. Detecting a natural correlation with conventional tools is not. ML claims in biotech inherit this line, so the question is which side of it your model’s output lands on.

In Rapid Litigation Management Ltd. v. CellzDirect, Inc. (Fed. Cir. 2016), claims to a method of producing hepatocyte preparations that survive repeated freeze-thaw cycles were eligible. The court called it a “new and improved technique, for producing a tangible and useful result.” The end result was not “simply an observation or detection” of a natural ability of the cells.

In CareDx, Inc. v. Natera, Inc. (Fed. Cir. 2022), claims to detecting transplant rejection by measuring donor cell-free DNA were ineligible. The specification described each step as conventional, and the court found that CareDx “does not actually claim any improvements in laboratory techniques.” The court distinguished cases involving “a method of preparation or a new measurement technique.”

In REGENXBIO Inc. v. Sarepta Therapeutics, Inc. (Fed. Cir. Feb. 20, 2026, precedential), the court also declined “to read out or ignore limitations in a claim” when reversing an ineligibility ruling on engineered cells. And the USPTO’s 2024 AI eligibility examples include Example 47, where an anomaly-detection claim became eligible once it automatically dropped malicious packets and blocked their source, rather than merely alerting an administrator. The biotech analog is a model whose output triggers a defined physical action.

Three ways to tie a model to a concrete technical result

Under current law, the most defensible biotech ML claims use one of three anchors: the model improves how a measurement is made, the model controls a physical lab process, or the model’s output selects a particular treatment. Each works best when the specification explains the technical problem and the claim recites the elements that solve it.

1. Assay improvements

Claim the measurement, not the prediction. Suppose your model corrects cross-talk between channels in a multiplex fluorescence immunoassay, so the instrument resolves analytes it could not separate before. The claim should recite acquiring the raw signal, the specific correction, and the corrected concentrations the assay reports. The specification should explain why existing compensation methods fail at low signal and show the improved limit of detection with data. A model output drafted as a “risk score” that only correlates with disease reads like CareDx; a change in what the instrument can measure reads more like a “new measurement technique.”

2. Lab automation and closed loops

Claim the action the model drives. A model that flags likely artifact wells in a high-throughput screen is abstract on its own. A claim in which the flag causes a liquid handler to re-dispense the flagged wells, or causes an imager to re-acquire them at a different focal plane, ends in a physical step. That mirrors Example 47’s automatic remediation.

Recite the control step in enough detail that it is not a token “and then act on it.” Some of the Dental Monitoring claims ended by sending a message to guide a new image capture, and they were still ineligible.

3. Diagnostic workflows that end in treatment

Claim the particular treatment. In USPTO Example 49, a claim that used a model to identify high-risk patients and then gave “an appropriate treatment” was ineligible. A dependent claim naming a specific drug was eligible. For a diagnostic company, that means adding claims in which the model’s classification selects a named therapy or a defined procedure, alongside the broader detection claims you want for licensing.

How weak and strong framings compare

A drafting checklist for machine learning patent claims in biotech

Write the specification around a technical problem in the lab and the mechanism that solves it, then make sure the claims recite that mechanism. Collect the evidence you would need for a declaration before filing, because comparison data is easiest to gather while the experiments are running.

1. Name the technical problem. Describe what the instrument, assay or workflow could not do before: drift between reagent lots, channel cross-talk, low-abundance analytes lost in noise.
2. Explain the mechanism, not just the result. Say why your architecture, input representation or training method fixes the problem. Recentive and Dental Monitoring turned on the absence of this explanation.
3. Don’t describe your model as off-the-shelf. If you used a standard architecture, say what you changed. A specification that says any available network will do can be used against you.
4. End at least one claim set in a physical step or specific output. Instrument control, sample handling, a corrected measurement, or a named treatment.
5. Put the improvement-delivering elements into the claims. Under the December 2025 MPEP changes, the examiner checks whether the claim reflects the improvement described in the specification.
6. Keep comparison data. Benchmarks against the prior method, run on the same samples, support both a SMED and an obviousness response. Note published papers that describe the problem as it stood before your filing date.
7. Check disclosure timing. File before posters, preprints and partner decks, as I explain in why the patent clock starts earlier than you think.

The computational chemistry and AI drug discovery practice page describes the types of platform and pipeline inventions I work on.

What should investors and partners expect?

Expect questions about whether your ML claims would survive a § 101 challenge in court, not only at the USPTO. A patent allowed under the post-Desjardins guidance can still be challenged in litigation under Recentive. Claims that recite a measurable lab result or hardware action are easier to defend in diligence.

Be ready to show which claims recite a physical anchor and what you kept as a trade secret. I cover broader diligence questions in what investors ask about your patents.

Frequently asked questions

Can you patent a machine learning model used in biotech?

Yes, but not as a bare model. The claim must integrate the model into a practical application, such as a specific improvement to how the model operates or how an assay or instrument works. The USPTO’s precedential Ex parte Desjardins decision supports such claims. Courts reject claims that only apply generic machine learning to new data.

Are AI-based diagnostic method claims patent eligible?

They can be, but detection claims alone are risky. The Federal Circuit held in CareDx v. Natera that detecting a natural correlation with conventional techniques is ineligible. Claims that recite a new measurement technique, a method of preparing a sample, or a particular treatment based on the model’s result have a stronger basis under current law and USPTO examples.

What is a subject matter eligibility declaration?

A SMED is a declaration under 37 C.F.R. § 1.132 that provides objective evidence on eligibility, such as a skilled scientist’s explanation of the technical improvement your specification describes. The USPTO’s September 2026 memo recommends filing it separately and early, stating its purpose, and supporting it with neutral sources where possible.

Do the USPTO’s Desjardins changes bind the courts?

No. The MPEP and USPTO precedential decisions guide examiners, not federal judges. A court reviewing an issued patent applies Supreme Court and Federal Circuit precedent, including Recentive Analytics. Draft claims that satisfy both the USPTO’s improvement test and the courts’ stricter view.

If you’re working out how to claim an assay pipeline, an automated lab workflow or an AI-driven diagnostic before a filing deadline or a financing, 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