AI is deeply penetrating the pharmaceutical industry from drug target discovery to clinical trials, but the unconventional R&D model of AI pharmaceuticals also brings entirely new patent legal issues.

Defining ethical boundaries, avoiding diagnostic methods, the requirement for “full disclosure” in AI drug screening, and the path of “inventive” proof have become core propositions in patent layout in the AI+ biomedicine field.

This paper systematically outlines the key examination points for patent applications in the AI+ biopharmaceutical field around five dimensions: ethical boundaries, object identification, full disclosure, and practicality and inventiveness, and proposes targeted strategic recommendations.

1. Typesof AI-Related Patents From the perspective of patent applications, AI-related applications can be divided into four types:

• First, applications related to AI algorithms or models themselves, involving improvements in model structures and training methods such as deep learning and support vector machines, but not specific application scenarios.
• The second is applications based on the functions or domain applications of AI algorithms or models, covering functional layers such as natural language processing and computer vision, as well as application layers applying AI to specific scenarios like healthcare and transportation.
• Third, AI-assisted inventions, such as using AI to identify specific protein binding sites and ultimately obtain new drug compounds. AI acts similarly to an information processor, with human inventors still making creative contributions to substantial features.
• Fourth, inventions generated by artificial intelligence, meaning inventions without substantial human contribution and entirely generated by AI, according to the “Patent Examination Guidelines,” AI cannot act as the inventor—the inventor must be a natural person, and the request cannot include the name of the artificial intelligence.

For innovators in the AI+ biomedicine field, the core focus should be on two types of applications:
one is the specific application of AI in the pharmaceutical field. For example, compound screening, protein structure prediction, medical imaging analysis, and more. The key to these applications lies in what specific pharmaceutical technology problems AI solves, not just what the AI itself does.
The other category is AI-assisted inventions, which are inventions completed using AI as a tool during the R&D process, such as new drug compounds obtained by identifying specific protein binding sites using AI. In such applications, the AI’s role is similar to an information processor or drawing tool, and human inventors still make creative contributions to the substantive features of the invention, so they can be credited normally.
The common logic of both types of applications is: AI is the tool, and people are the main driver. The application documents must clearly state how AI solves specific pharmaceutical problems while ensuring that human creative contributions are fully realized—this is the underlying logic behind the patent layout in the AI+ biopharmaceutical field.

2. Ethical Issues: The red line

in Article 5 of the Patent Law is the most fundamental standard in patent examination.

Article 5 of the Patent Law stipulates that inventions that violate social ethics cannot be granted patent rights. The review guidelines further clarify that social morality refers to ethical and moral concepts and codes of conduct that are generally accepted and legitimate by the public. The key to determining whether an invention violates social morality lies in whether it conforms to widely accepted ethical and moral concepts.
In the ethical review of the AI + biomedical field, brain-computer interface solutions are the most controversial and require the most cautious approach. The Ministry of Science and Technology’s “Ethical Guidelines for Medical Research on Neurotechnology Involving Humans” and “Ethical Guidelines for Brain-Computer Interface Research” point out that neural regulation may have a significant impact on essential characteristics such as human free will, personal dignity, personal identity, and mental integrity, and should avoid replacing or weakening human judgment and decision-making abilities.
Below is a case study illustrating the specific application of this examination standard:
the application involves an “olfactory assisted enhancement system based on brain-computer-brain interface.” The technical solution is as follows: By implanting multi-channel electrodes in the animal’s olfactory channel, neural signals from the animal’s olfactory system are collected and recorded in real time; Using the decoding module to extract and convert odor information from animal olfactory nerve signals into decoding instructions; Then, by implanting an array of stimulation electrodes in the patient’s olfactory-related brain regions, electrical stimulation signals are output to precisely stimulate the olfactory-related brain regions, ultimately achieving olfactory perception in patients. The instructions state that the protocol is implemented with the patient’s informed consent, aiming to help patients with olfactory impairments rebuild their sense of smell.
This scheme was ultimately found to violate Article 5 of the Patent Law and could not be granted a patent right.
The core logic of the review is:

• First, the risk of personality identity. There is a fundamental difference in the sensitivity of humans and animals to olfaction. When humans are implanted with animal olfactory signals, their cognitive patterns, perceptual mechanisms, and emotional responses change, which may affect individual identity and self-continuity. The review guidelines explicitly identify “personal identity” as one of the core dimensions for measuring ethical risk.
• Second, the standards of social morality. Cross-species neural signal implantation directly affects animal perception signals in the human brain, altering the way humans perceive essential things. Currently, it is not widely accepted by the public as a legitimate medical method and does not meet the requirements of social ethics.
• Third, clear boundaries of ethical guidelines. The Ministry of Science and Technology’s ethical guidelines clearly state that brain-computer interfaces should avoid replacing or weakening human decision-making abilities.

This proposal directly involves cross-species neural signal implantation and touches all the above ethical red lines. This case demonstrates that technical feasibility cannot replace ethical review.

3. Object Issue: How to Avoid the “Diagnostic Method” Trap

Article 25, Paragraph 1, Item (3) of the Patent Law stipulates that “methods for diagnosis and treatment of diseases” are not granted patent rights. When AI is used in medicine, a key question arises: do AI-automated diagnostic or predictive methods really count as “diagnostic methods”?

The 2023 edition of the Patent Examination Guidelines made significant revisions to this issue, explicitly excluding “Inventions Not Considered Diagnostic Methods” in Section 4.3.1.2 of Chapter 1, Part II, excluding “information processing methods performed entirely by computers or similar devices” from diagnostic methods. The core logic of this revision is: traditional diagnostic methods are doctor-led and should not be monopolized by patents for humanitarian reasons; The information processing schemes performed by AI are essentially data processing rather than medical actions, and should not be subject to the same restrictions. Let’s look at a case below.
The application relates to a character-based classification method for arrhythmias. The technical solution is as follows: preprocessing ECG signals; Morphological and time-frequency features are extracted from the detected R wave positions to construct raw feature vectors; Feature weights are calculated using the ReliefF algorithm; Population initialization is guided by feature weights, and through genetic operations such as selection, crossover, and mutation, iterative optimization is carried out to output the most well-fitted individuals as preferred traits; Finally, a multi-classification strategy is used to combine multiple binary classifiers into an identification classifier, enabling the identification of various arrhythmias.
The core issue facing this program during the review was: Does arrhythmia identification count as a “diagnostic method for diseases”?
The review concluded that this protocol is not a diagnostic method, for the following reasons:

• First, all steps are carried out by computer. From ECG signal preprocessing, feature extraction, feature selection to classification and recognition, the entire process is completed by computer-executed algorithm programs, without involving manual operation or clinical judgment by doctors.
• Second, it is essentially an information processing method. This solution processes ECG data, performing mathematical transformations and feature extraction to produce a classification label for arrhythmias. The entire process reflects data processing rather than medical actions.
• Third, no need for a doctor’s involvement. From data input to result output, the entire protocol requires no physician intervention or judgment during the process. The computer automatically completes all steps.
• Fourth, the output results are only intermediate information. The computer-generated arrhythmia classification results are merely intermediate information serving as a reference for doctors’ diagnosis. The mechanisms of arrhythmia are complex and manifest in various types, and there is no single correspondence between arrhythmia and heart disease. Computer-generated classification labels such as “ventricular premature beats” and “atrial fibrillation” cannot directly equate to a diagnosis of “the patient has a certain heart disease.” The final diagnostic conclusion still needs to be comprehensively judged by the doctor based on the patient’s specific symptoms, other test data, and clinical experience.

This case establishes the core criterion for determining whether an AI medical method qualifies as a “diagnostic method”: the core of judgment is not whether the plan is related to a “disease” or “patient,” but whether the plan is directly aimed at “obtaining diagnostic results.” If all steps of a regimen are executed by a computer and output intermediate information for the doctor’s reference rather than directly drawing diagnostic conclusions, then the scheme is not considered a diagnostic method.

4. Practicality Issues: Why Are AI Solutions Lacking Reproducibility?

A practical invention must meet reproducibility requirements, meaning that those skilled in the art can repeatedly implement the scheme based on the disclosed technical content, and the results of each implementation should be the same, without relying on random factors. This requirement is especially significant in the AI + biomedicine field—many AI solutions rely on manual annotation or preprocessing, and the degree of standardization in manual operations directly affects the reproducibility of these solutions. Let’s illustrate this with a case study.
This application relates to a deep learning brain tumor prognosis analysis modeling method that combines attention mechanisms and multi-scale feature mining. The protocol requires “two clinical radiologists to independently outline areas of interest using ITKSNAP or 3D Slicer software and jointly confirm their effectiveness; When the two outline results are inconsistent, two doctors re-examine the brain tumor medical imaging sample together to jointly confirm the valid result of the area of interest.”
The review concluded that the plan was not practical.
The reviewer’s logic is that this plan heavily relies on the doctor’s clinical experience and subjective judgment. Different doctors’ outline results may be inconsistent, and when the outline results are inconsistent, both doctors are required to confirm together. However, this confirmation process also relies on subjective judgment rather than objective standards. The implementation results of the plan vary from person to person and are not reproducible.
If the doctor acts as a general operator in the protocol, using software to annotate according to a preset standardized process, and the annotation results meet objective standards, reproducibility issues are not involved. However, if doctors need to use professional knowledge and clinical experience for subjective judgment, and different doctors’ judgments may differ, the protocol is not reproducible.

5. Full disclosure of issues: Review standards

for AI-assisted drug screening AI-assisted drug screening is the most promising application scenario in the AI+ biomedicine field, but it is also the area with the most problems in patent examination. The core confusion faced by examiners was: the applicant claimed, “I used AI to screen the medicine,” but the instructions did not clearly state “how it was screened” or “on what basis did it say the screening was correct?” Let’s look at two cases: one and the other, to see what kind of manual is considered fully public.
Negative case: A virtual drug screening system
application involves a drug virtual screening system for a crystal complex. The claims describe multiple functional modules such as a visualization subsystem and an AI model management subsystem. Manual description: “Users only need to describe the mode of action of the drug on proteins and the requirements of the drug to generate a batch of compounds that meet expectations in this system.” “As for what architecture AI models use, where training data comes from, how input data is preprocessed, what the model outputs, and how to verify model reliability, none of these are recorded.
The review concluded that disclosure was insufficient.
The reason is: the biological activity of chemical drugs is widely recognized as highly unpredictable, and the instructions do not provide any technical details. Experts in the field cannot be certain that this “large model virtual screening subsystem” truly has reliable screening capabilities for specific targets.
Positive Case: Screening Method for EGFR-Targeted Anti-Tutorious Active Compounds This application relates to a deep learning-based screening method for EGFR-targeted anti-tumor active compounds
. The claims clearly state three core elements:

• Clear model architecture: using a graph-attention network architecture, the input is the graph structure of compound molecules, and the output is the predicted activity value of the compound for EGFR targets.
• The training process is clear: compounds known to have activity confirmation data for EGFR targets are collected as training sets, with measured activity data as annotations.
• The validation method is clear: 1,000 random compounds mixed with 8 known active compounds for testing, achieving a hit rate of 87.5%.

Thorough disclosure of review and recognition.
Verifying the effectiveness of AI models with known compounds is sufficient for those skilled in the field to reasonably expect the method to achieve its intended purpose without the need to provide additional wet experimental data for new compounds screened by AI.
Boundaries that require special attention: If the AI designs a completely new chemical entity, simply recording its molecular structure is not enough; characterization data and preparation methods of the compound must be provided. The protected object of chemical products is the “physical entity” rather than the “design,” and only the structural formula is stated, without providing experimental data, so examiners cannot confirm whether the compound truly exists.

6. Creativity Issues: The Boundaries

of Algorithm Conversion In the application of general AI algorithms in the biopharmaceutical field, there is a core principle for determining creativity: simply applying existing algorithms to new scenarios without improving the algorithm itself does not constitute a creative contribution. Only when algorithmic features and technical features are closely integrated to form a technical means to solve a technical problem can algorithmic features contribute to creativity. Let’s look at a case below.
(1) Lack of creativity:
The SVM-based abnormal body temperature sequence detection device is a typical case. This solution uses a support vector machine to analyze body temperature data and determine whether the user is infected with COVID-19: by regularly measuring body temperature with a temperature detection device worn on the body, time series and higher-order feature values are extracted, input to the support vector machine model for classification, and output tags such as normal population, COVID-19 immune population, and COVID-19 population.
The comparison document has publicly disclosed a method for analyzing temperature data using a support vector machine to assess breast cancer risk, which also includes collecting temperature sequences, calculating mean/variance/approximate entropy measures, and classifying using support vector machines. The two schemes are highly similar in data samples, computational models, and screening mechanisms; the only difference is that the test target is changed from breast cancer to COVID-19, and there is no improvement to the support vector machine algorithm itself.
The review concluded that it lacked creativity.
The reasoning is that the support vector machine is a mature algorithm. Technicians in this field know that fever is one of the clinical manifestations of COVID-19, and are motivated to apply a breast cancer classification model based on body temperature sequences to COVID-19 screening. This conversion directly applies existing algorithm application methods without overcoming technical difficulties.
(2) Creative Cases:
If the applicant adjusts the information transmission method of the graph neural network to better predict the activity of that target based on the molecular characteristics of a specific target; or to process specific types of medical image data, the hierarchical structure and receptive field of convolutional neural networks have been specifically optimized, thereby improving the accuracy of lesion identification. These improvements are substantive adjustments made to AI models to address specific technical problems in the biopharmaceutical field, reflecting the deep integration of algorithms with specific application scenarios, and considering the creative contribution of algorithmic features to technical solutions.

7. Practical Points

for AI + Biomedical Patent Layout Based on the aforementioned examination standards and case analysis, patent applications in the AI + biomedicine field should pay attention to the following six core points:
1. Shift from “Functional Description” to “Technical Implementation” 。 The application document should not only state “what AI can do,” but also clearly state “how AI does it”—model architecture, input-output settings, training data sources and annotation methods, training process, and verification methods, all of which are indispensable. If it only states “using AI to screen drugs” without “how to screen,” the examiner will directly consider disclosure insufficient.
2. Avoid diagnostic method pitfalls. The claims emphasize “computer execution,” “intermediate results,” and “for physicians’ reference,” avoiding terms like “diagnosis” and “treatment.” Topic names often include “Information Processing Methods,” “Data Analysis Methods,” and “Model Training Methods.”
Third, focus on algorithm improvements addressing biomedical issues. Simply applying general AI algorithms to new scenarios without improving the algorithm itself does not constitute a creative contribution. It is essential to clearly state what substantive adjustments were made to the algorithm model to solve specific technical problems in the biopharmaceutical field—what was adjusted, why the adjustment was made, and what problems were solved after the adjustment.
4. Emphasize data compliance and ethical review. Plans involving patient data or genetic data must specify in the application documents that the data source is legal and complies with ethical requirements. Protocols involving brain-computer interfaces and neural regulation must confirm that there are no ethical risks that alter personality identity or autonomous cognition. Technical feasibility cannot replace ethical review.
5. Reasonably grasp the threshold for full disclosure. Verifying the effectiveness of AI models with known compounds is sufficient to meet the requirements for full disclosure, without requiring the full experimental data of the new compounds screened by AI. However, if the application for protection is a completely new chemical entity, characterization data and preparation methods must be provided, with different standards for the two.
6. Accurate identification of the inventor. AI cannot be considered an inventor; the inventor must be a natural person. During AI-assisted R&D, companies should establish a complete contribution record system, identify natural persons who have made creative contributions to technical solutions, and ensure the accuracy of inventors’ credits.

8. Conclusion
AI is reshaping every aspect of biopharmaceutical R&D. The core logic of intellectual property protection has always remained unchanged—exchanging publicly available technical content for exclusive patent rights. The non-traditional path of AI-assisted R&D requires applicants to clearly and thoroughly explain in patent documents how AI solves specific problems. Only in this way can a solid intellectual property moat be built alongside technological innovation.

About Us

MeritsIP: Your Global Intellectual Property Partner

At MeritsIP, we provide comprehensive IP services in biomedicine, medical devices, manufacturing, semiconductors, and AI. We support 200+ top clients globally with strategic, high-quality IP services.

Stay Connected

• • Website: https://www.meritsipglobal.com

• • LinkedIn: Follow us on LinkedIn

• • Email: meritsip@meritsandtree.com

© 2025 MeritsIP. All Rights Reserved.