Are AI-Predicted Compounds Patentable in Japan? What JPO Example 52 Tells Foreign Applicants
Japan Patent Office examination practice for AI-driven inventions · Enablement & support requirements · By Takuya Ohsugi, Japanese Patent Attorney (Benrishi)
If your R&D uses machine learning to predict new molecules, materials, or drug candidates, one question decides whether you can patent them in Japan: how much experimental data must the specification actually contain? The JPO has answered this directly in its AI case examples. Here is what Example 52 means for foreign applicants entering Japan.
Why this matters for foreign applicants
AI-driven discovery — in drug discovery, diagnostics, materials, and bioinformatics — increasingly produces inventions where a model proposes a compound before anyone makes it in the lab. Applicants naturally want to file early and claim broadly, sometimes before full experimental validation.
Japan, like the EPO, scrutinises whether the specification truly supports and enables the claimed invention. A specification that would pass in one jurisdiction can fail Japan’s written-description rules if the data is thin. Understanding the JPO’s stance before drafting (or before national-phase entry) is where a Japanese patent attorney adds the most value.
The relevant Japanese requirements
Two provisions of the Japan Patent Act govern this:
| Requirement | Provision | Plain-English meaning |
|---|---|---|
| Enablement | Art. 36(4)(i) | The specification must let a skilled person make and use the invention without undue trial and error. |
| Support | Art. 36(6)(i) | The claims must be supported by — not broader than — what the specification actually discloses. |
For AI-predicted subject matter, the recurring question is whether a prediction (rather than a measured result) is enough to satisfy these.
JPO Example 52: the fluorescent-compound case
The fact pattern
A machine-learning model is trained on public data pairing known fluorescent compounds’ chemical structures with their emission properties. The model then predicts novel structures expected to show a target profile — a peak emission wavelength of 540–560 nm and a fluorescence lifetime of 5–20 µs. It outputs two novel compounds, A and B. Compound A is then actually synthesised and measured: peak wavelength 545 nm, fluorescence lifetime 12 µs — within the target range. Compound B is only predicted, not made.
The JPO Examination Handbook uses this example to illustrate when such a specification meets the enablement and support requirements. The key takeaway is that the specification is treated as sufficient when it contains at least one of the following:
Three routes to a sufficient specification (JPO)
- Experimental results — actual measured data for a manufactured embodiment (e.g., synthesised Compound A measured at 545 nm / 12 µs).
- Verification of the model’s prediction accuracy — evidence that the AI’s predictions are reliable for the property at issue.
- Common technical knowledge in the field — established knowledge that makes the predicted property credible for a skilled person.
The practical reality (as of December 2025)
The three routes above are how the Handbook frames the test. In day-to-day prosecution, however, the routes are not equally proven. Based on granted-case review to date, one practical observation stands out:
Only Route 1 (actual experimental results of a manufactured compound) has been confirmed in granted Japanese patents. As of December 2025, no granted patent has been identified that relied on Route 2 or Route 3 alone — i.e., prediction-accuracy verification or common technical knowledge without measured data — to carry the enablement/support burden.
This does not mean Routes 2 and 3 are unavailable; the Handbook expressly contemplates them. It means the safe, tested path today is to include real measured data for at least a representative embodiment, and to treat purely predicted embodiments (like Compound B) as more vulnerable to a written-description objection.
What this means when you file in Japan
- Anchor broad claims with at least one worked, measured example. A single synthesised-and-measured embodiment materially strengthens enablement and support for the surrounding claim scope.
- Do not rely on the model’s output alone. If the specification presents only predicted compounds, expect scrutiny under Art. 36 — and plan how you would respond.
- If measured data is limited, build the record for Routes 2 and 3. Documenting prediction-accuracy validation, or the relevant common technical knowledge, gives you arguments even if they are less battle-tested.
- Mind the timing. Japan does not allow new matter to be added after filing. Data you leave out of the priority/PCT specification generally cannot be inserted later, so the enablement strategy must be set before filing — or, for PCT cases, reviewed carefully before national-phase entry.
How this compares abroad
Applicants from the US and Europe will recognise the tension: the JPO’s enablement/support analysis is closer in spirit to the EPO’s sufficiency and support practice than to a more permissive approach. If your global filing strategy assumes prediction-heavy specifications will be accepted everywhere, Japan is a jurisdiction where that assumption should be tested early with local counsel.
Planning to patent an AI-driven invention in Japan?
I am a Japanese patent attorney working at the intersection of biotechnology, AI, and IP strategy, communicating directly with you in English. I can review whether your specification is ready for JPO examination — before you file or enter the national phase. Book a free initial consultation →
Reference: JPO, “Patent Examination Case Examples pertinent to AI-related technologies” (Patent Examination Handbook, Annex A), Example 52. The JPO publishes an official English translation of these case examples on its website.
While every effort is made to ensure the accuracy of the information on this site, its completeness and accuracy are not guaranteed. The content reflects the personal views of the author and does not represent the official position of the affiliated patent office. This article is general information, not legal advice. © 2026 Takuya Ohsugi. All Rights Reserved.