Key takeaways
Depending on the health innovation and jurisdiction, innovators in AI-enabled health care should consider a combination of IP protection mechanisms. They should evaluate strategic decisions to manage their IP.
Different innovation decisions are required at each development stage, from ideation through global deployment.
Innovation in health care calls for careful consideration of regulatory and compliance issues as well as clinical and safety considerations.
1.1 The foundations of AI-enabled health innovation
Jurisdictions recognize and understand the use of AI in health care through regulatory and legal frameworks. These vary across jurisdictions and may impact innovation strategies, including protection by IP. Health data governance and standards (see Chapter 3) and regulatory pathways
1.2 IP protection mechanisms for AI-enabled health innovation
AI-enabled health innovation typically requires considering the relevance of a combination of IP protection mechanisms as well as IP licensing models that may be used strategically to protect IP rights. Some IP rights relevant to AI-enabled health innovation include patents, trade secrets and copyright. Decisions on IP rights are not static. IP protection must be assessed throughout the innovation life cycle (see Section 1.4).
Patent rights are territorial in nature. Patents are granted based on an application by national or regional offices for inventions, such as products or processes, in all fields of technology. In order to obtain a patent, an invention must fall under the category of patentable subject matter, and meet the patentability criteria. As far as subject matter eligibility is concerned, the definition of an invention and permitted policy choices may exclude some creations from patentability, namely: scientific theories, aesthetic works, mathematical methods, plant or animal varieties, discoveries of natural substances, business methods, methods for medical treatment and computer programs, which are not patentable in some jurisdictions. The substantive criteria of novelty and inventive step require the invention to be a globally new technical solution to a problem that can be applied in industry, which is non-obvious for a person skilled in the art.
After successful registration, with associated costs, the patent owner secures exclusionary rights for a certain term, allowing the inventor to stop others from using, making or selling the invention without permission. A patent is granted for a limited period, usually not exceeding 20 years from filing the application. In return, the patent applicant must disclose the invention in the application, with such disclosure being published in the patent documents. Once the period of protection terminates, the invention becomes part of the public domain; anyone can commercially exploit it.
For AI-enabled health innovations, patents may protect technical solutions – such as model architectures; training pipelines; data-processing methods; medical devices, including sensors; and algorithms, encompassing those for drug design – as long as the claimed invention meets applicable patentability requirements (see Table 1). When seeking patent protection for an invention which discloses nucleotide or amino acid information, all applicants must file what is known as a sequence listing, compliant with WIPO Standard ST.26.
For AI-enabled health innovators, understanding comparative legal IP landscapes across jurisdictions is critical for both innovation strategy and market entry. Subject matter eligibility rules differ among countries; an AI-enabled invention that is patent-eligible in jurisdictions such as Japan or the Republic of Korea may require different claim drafting to satisfy patent eligibility requirements in the United States or before the European Patent Office. While an extensive analysis of the approach to subject matter eligibility for AI inventions is outside the scope of this chapter,
In cases where a patent examination is a requirement to grant a patent, some patent offices provide more detailed guidance on the treatment of inventions using AI and machine learning (ML) in their patent examination guidelines.
Trade secrets protect confidential information that has commercial value stemming from its confidentiality against unauthorized access, disclosure or use in a manner contrary to honest commercial practices (i.e. misappropriation) provided that reasonable steps have been taken to keep their secrecy. There is no registration process for trade secrets, and protection lasts as long as the information remains secret. Protection varies by jurisdiction – some protect trade secrets through dedicated statutes, while others rely on a combination of statute and case law, or primarily on common law principles such as breach of confidence. Successful approaches to trade secret protection may rely on clear written policies, strict access controls, physical and technical security measures, training, non-disclosure agreements (NDAs) and incident protocols. Trade secrets in AI-enabled health innovation may include large proprietary data sets, model weights, feature engineering methods, operational procedures and evaluation harnesses. The use of trade secret protection for AI-enabled health innovation algorithms involves navigating expectations around transparency, including explainability standards for clinical decision-making, auditability mechanisms for quality assurance and regulatory disclosure obligations for safety monitoring. However, regulatory expectations increasingly extend beyond explainability alone to include technical documentation, quality management, human oversight, traceability, and lifecycle monitoring, while preserving the protection of legitimate trade secrets.
Copyright protects eligible works of authorship, including original literary and artistic works, which cover an enormous range of items, including books, journals, music, paintings, sculpture and films as well as computer programs, compilations of data,
Copyright may also be used to protect a database that meets relevant requirements, including displaying originality (the amount can vary by jurisdiction) in the selection or arrangement of its contents. In some jurisdictions, for databases meeting the relevant threshold, database rights may apply adjacent to the copyright protection. For example, sui generis database rights (a specialized protection for specific subject matter not covered by traditional IP laws) may protect against the extraction and reutilization of substantial parts of the database without permission. Sui generis database rights only exist in certain jurisdictions (principally the EU and UK).They are not international rights.
1.3 IP licensing mechanisms for AI-enabled health innovation
The IP framework creates the legal foundation for the control and use of protected innovations and creations. Within this framework, some approaches may allow permissions and uses depending on the choice of the owner of the IP, including the innovator or creator. These approaches may comprise assignment (the exercise of rights enabled through the permanent transfer of IP ownership) or licensing (the exercise of rights contractually enabled without the transfer of IP ownership). Deliberate and strategic licensing choices, along the spectrum from narrow to permissive, can complement proprietary assets and standards of participation (see Box 1).
Ultimately, AI-enabled health innovation that achieves the objectives of the IP owner may depend on a multidimensional approach. This requires strategic decisions about what to patent, what to keep secret, when to use open source and when to participate in standards. Licensing as a relevant access model has a significant role in the management and use of IP rights to deliver the specific objectives of rights holders. This may encompass for example, protection and control, collaboration and partnership, the leveraging of a strategic advantage to maximize commercial opportunity and/or support for access models (see chapter 2). Licensing approaches may differ according to the IP right, that is, the copyright or patent. Moreover, the role of IP rights in supporting access to technologies and their management will be impacted by other relevant factors, such as the disease burden, the size of the company, regulatory considerations, production sites, manufacturing skills and capacity, the expertise of the licensee and the nature of the product.
WHO's Regulatory Considerations on Artificial Intelligence for Health
IP licensing approaches relevant to AI-enabled health innovations exist on a spectrum, from permissive to restrictive, reflecting the diverse strategies that innovators employ to balance openness, collaboration and business sustainability. Exclusive voluntary licensing (a license granted to one party by the IP owner to permit particular uses of the IP-protected work) and non-exclusive voluntary licensing (a license granted by the IP owner to permit particular uses of the IP-protected work by multiple parties simultaneously) are two approaches. Open-source licensing is another way to license.
Open-source licensing approaches are numerous. They may apply to copyright-protected works, including software. Open-source licensing approaches may contribute to rapid iteration and trust within the IP framework. The Open-Source Initiative maintains criteria for licenses to be considered truly "open-source," including the freedom to use them for any purpose, access to source materials and the ability to create derivatives. In recent years, however, some AI developers have adopted source available or restricted licenses that allow inspection but impose limitations, such as restricting commercial use, competing services or use by certain entities. These licenses represent a middle ground between proprietary and open approaches.
1.4 The innovation journey: A life cycle IP matrix
AI-enabled health innovation follows a clear journey from the earliest spark of an idea to potentially large-scale, global deployment. At each stage, innovators face different decisions that shape not only the technology but also its long-term integration into health systems. These decisions are shaped by regulatory and compliance issues as well as clinical and safety considerations. Early choices around IP can lay the groundwork for later success.
This chapter introduces the innovation life cycle IP matrix, designed to help innovators in AI-enabled health care to manage IP strategically, from concept through commercial scale-up. Each phase of development presents distinct opportunities, risks and requirements across six dimensions: patents, trade secrets, copyright and data protection with due regard to sui generis database rights, open-source, regulatory and standards considerations, and contracts. While these dimensions apply across technology sectors broadly, by mapping them to the product life cycle, AI-enabled health innovators can anticipate challenges, avoid common pitfalls and unlock the full value of their IP. This, in turn, supports access to and the availability of technologies. Areas that may be particularly relevant for AI-enabled health innovation include trade secrets and regulatory and standards issues. While the matrix largely focuses on predictive AI, the rapid advancement of generative AI and multimodal models may require additional considerations, including the copyright implications of generated outputs.
The innovation life cycle IP matrix in Table 2 serves as a quick reference tool, distilling actionable steps for each stage, from ideation through health system integration. It supports AI-enabled health innovators to better identify, protect, manage and leverage their assets according to their objectives. It is not a substitute for legal advice, but it can enable more effective, informed conversations with IP advisors, helping innovators to translate strategic IP planning into successful, scalable health care solutions. To use this innovation life cycle IP matrix:
Identify your current life cycle stage
Review all columns for that stage
Identify relevant IP actions and find more detail in section 1.5 of this publication
1.5 Details of the innovation life cycle IP matrix
Stage A: Ideation and concept development
Innovation begins with ideation, but this early stage is also where many IP missteps may occur. Capturing ideas promptly and accurately, clarifying ownership and making early strategic IP decisions establish the foundation for the rest of the innovation life cycle.
Invention disclosure form: An invention disclosure form is a confidential document that captures a technical innovation before a patent application is filed, given that patent filing may be premature at the ideation and concept development stage. The form is an important procedural step in internal decision-making. While this information may not be readily available, it may broadly describe the following aspects:
The invention itself, including its general technical field, the problem it addresses, and a clear explanation of how it works at a technical level, including its key features and distinguishing aspects, without limiting the description to a single implementation.
The novelty and technical contribution, describing what is new relative to existing solutions, why prior approaches are insufficient and what technical effects or improvements the invention achieves.
The state of the art, including what was known or used before the invention, relevant publications or patents if known, and how the invention differs from or improves upon those references.
The development context, such as how and when the invention was conceived and developed, what stage it has reached and what work has already been carried out versus what remains to be done.
The people involved, identifying who contributed to the invention, the nature of their contributions and any information necessary to later determine inventorship or authorship.
Disclosures and confidentiality, including whether the invention has been disclosed or shared in any form (e.g., publications, presentations, demonstrations, etc.), whether such disclosures were confidential and whether any further disclosures are planned.
Funding and third-party involvement, covering whether the work was supported by external funding, carried out under collaboration agreements or relied on third-party materials, data, software or ideas that could affect ownership or freedom to operate.
Software- or data-related aspects (where relevant), such as the role of code, models, data sets, tools or external components, and any licensing or access conditions that might be relevant to protection or exploitation.
Potential applications and use, describing how the invention could be used in practice, what kinds of products, services or processes it might enable, and in which fields.
Commercial and strategic considerations, at a high level, including why the invention may be valuable, how it compares to existing solutions in the market and what kinds of exploitation pathways (e.g., licensing, spin-out, collaboration or publication) might be used to further develop the invention.
Any constraints or risks, such as prior disclosures, contractual obligations, co-ownership issues or dependencies that could affect patentability, ownership or commercialization.
Access controls and NDAs for trade secrets: From the ideation phase, confidential information must be secured even if it may not necessarily qualify as a trade secret. Access should be limited, on a need-to-know basis, with secure storage in encrypted, password-protected repositories, logging of access events or breaches and periodic audits of permissions. Any external parties should sign robust NDAs before receiving confidential information. These agreements should define confidential information broadly enough to cover complex assets such as models or data sets, specify exclusions, set confidentiality periods (typically two to five years, or indefinitely for trade secrets), detail obligations upon termination (including material return or destruction) and outline remedies for a breach (injunctive relief or damages).
Ownership clarity of IP, including data: Employment, consultant and contractor agreements should address the ownership of work-related IP in accordance with the applicable jurisdiction. This may be achieved either through explicit assignment and work-for-hire clauses, where required by local law, or by referencing statutory provisions governing service inventions and employee inventions, where such legislation exists.
In research collaborations, IP terms should be negotiated upfront, including background IP brought into the collaboration, such as preexisting code, data and patents, and the handling of joint ownership or licensing rights.
Licensing approach: Deciding early on an IP licensing strategy is critical, as it dictates how you engage stakeholders and position yourself in the market. For example, full proprietary protection may be used for commercial medical device companies, keeping code and models confidential. The selective open-source release of non-core components may attract community input while safeguarding competitive advantages. Policies should be documented, and all team members trained on compliance, including to avoid improper code integration from incompatible licenses.
Regulatory and standards planning: Even at the ideation stage, regulatory pathways must be identified. Clarify whether the product will be classified as a regulated medical device and determine its risk classification. Identify relevant pathways, such as predicate devices for clearance, applicable quality and safety standards (e.g., International Organization for Standardization (ISO) 13485 Quality Management System for Medical Devices) and data privacy and protection laws. In some instances, regulatory pathway selection at the ideation stage is also influenced by whether the product will be classified under EU Medical Device Regulation/In Vitro Diagnostic Medical Device Regulation, the EU AI Act or both. These classifications may diverge, requiring parallel compliance tracks.
Background IP definition in contracts: Background IP exists before or outside a collaboration. It is brought into a project by the parties and may include preexisting patents, trade secrets, code or data owned by each party before collaboration begins. Foreground IP may include new IP generated during a collaboration or project. In collaborations, it is important to precisely distinguish between the two. The rights to use each should be defined, including ownership structures, licensing conditions and publication policies, including review and approval requirements.
Stage B: Data strategy and model training
During the training stage, data governance and trade secret protection become paramount. This is where the AI model is developed and the most valuable IP assets, such as data sets and trained model weights, are typically created.
Method/system patent claims for training pipelines: Novel aspects of model training, such as unique architectures, targeted data augmentation techniques, active learning methods, approaches to preserving privacy or optimized hardware and software integration, may be patentable. Giving due regard to relevant subject matter eligibility criteria, consider whether it may be helpful for claims to emphasize practical technical implementation rather than abstract concepts.
Securing data, weights and logs as trade secrets: Training data sets and model weights should be stored securely, using encryption and strict access controls. Logs of training runs containing hyperparameters and performance metrics should remain confidential. Version control systems should be private repositories and secured with multifactor authentication. NDAs should be executed with vendors and cloud providers, and contracts should ensure that data use is confined to agreed purposes.
Manage data sets as appropriate and license tracking: It is important to maintain provenance for every data set, track license terms and keep records on informed consent for patient data. Document de-identification methods and ensure compliance with relevant data privacy regulations. Execute agreements with data providers covering usage rights, retention and disposal.
Open-source and license management: Track and audit all open-source software (OSS) dependencies, verify compatibility with commercial goals, identify any licensed data that could create obligations to make relevant code open source, ensure attributions are properly maintained and conduct quarterly license audits as new dependencies are added.
Data transfer compliance with relevant regulations and standards: Cross-border data movement must align with local regulations, using contractual clauses, agreements or localization measures as required, and ensuring encryption in all transfers and storage.
Contracts and data use agreements: For each data source or service provider, define restrictions, allowed derivative works, publication rights, retention terms, breach obligations and compliance requirements, including with other applicable laws, particularly on data protection and data privacy (see chapter 3) and other underlying IP rights. For cloud service agreements, negotiate terms for protecting data security and confidentiality and prohibiting secondary data use. Include provisions requiring immediate notifications of any data breach or unauthorized access.
Stage C: Validation and evidence generation
Validation demonstrates that an AI model performs as intended in real-world conditions. This stage generates critical evidence for regulatory submissions, publications and commercial negotiations.
Refining patent claims based on technical effect: Validation data may allow the narrowing of patent claims to specify measurable performance improvements, clinical benefits or hardware integration advantages, and to add new claims through the continuation of application filings.
Protecting evaluation harnesses as trade secrets: These may include test set curation, evaluation metrics, validation protocols, negative results, internal records of failed experiments and model limitations, and post hoc analyses.
Copyright and data protection for model cards and documentation: Detailed technical documentation should be created, including reports, user manuals, training materials and summaries. These will be protected by copyright and may support both regulatory submissions and operational needs.
Open-source strategic release of non-core tools: Releasing generic evaluation frameworks, visualization utilities or benchmarking tools may build trust, while keeping core assets proprietary.
Clinical evidence protocols aligned with regulatory and standards expectations and guidelines: Validation protocols should align with regulatory expectations, follow standard reporting guidelines and include comparator studies against standard care. Where applicable, validation should also comprise subgroup performance reporting across relevant demographic groups (such as age, sex and ancestry) to address algorithmic bias and fairness obligations under emerging AI-health regulations.
Contracts and clinical trial data rights: Agreements with validation partners should clarify ownership, publication rights, regulatory use, future uses (including derivative works) and the confidentiality of trial data.
Stage D: Regulatory approval and market access
Regulatory approval is a critical milestone that reduces risk for commercial deployment of the technology. The alignment of an IP strategy and regulatory submissions will ensure consistency and enforceability. To note, regulatory authorization does not validate IP ownership or the freedom to operate. Regulatory clearance enables commercial deployment but is independent of IP ownership validation.
Patent claims aligned with intended use: Claims must match the approved clinical application, with continuation filings expanded as new indications gain approval.
Operational deployment as trade secrets: Integration procedures, monitoring protocols, incident handling and customer support processes should be protected as trade secrets, given that they involve significant operational know-how.
Copyright in labeling and instructions for use: Regulatory-compliant labeling, instructions for use (IFU), information to use with patients such as simplified explanations where applicable, marketing content and translations for international markets should be treated as copyright works.
Open-source compliance in distributed products: Ensure that any open-source components in the device meet license obligations, including source code provision, license notices and modification documentation.
Demonstrating conformance with standards requirements: Demonstrate compliance and conformance with relevant standards, including those relevant to quality management or interoperability, and maintain proof of compliance.
Field-of-use restrictions and updated rights in contracts: Structure customer agreements to manage IP and regulatory obligations, including by limiting customer use to approved indications, retaining rights to update software, meeting regulatory compliance obligations, applying data for improvement and termination rights.
Stage E: Deployment and post-market monitoring
Once IP is deployed, ongoing management, regulatory compliance and operational security are essential. Post-market surveillance generates real-world evidence and may identify new IP opportunities.
Filing a patent for continuous learning innovations: Continuation filings can protect methods for model updates, new clinical indications identified in real-world use and integration advances.
Protecting operational runbooks and monitoring procedures as trade secrets: Deployment playbooks, monitoring dashboards, incident protocols, scaling methods and disaster recovery plans should be confidential as they are valuable trade secrets.
User materials and user interface designs protected as copyright: Updated manuals, guides, videos, visual designs and integration documentation should be protected as copyright.
Community governance through open-source tools: Consider open sourcing certain monitoring, explainability, performance or audit tools to strengthen ecosystem engagement, as an option.
Monitoring obligations for regulations and standards: Many regulators require post-market surveillance for AI devices, including performance tracking, vigilance reporting and periodic safety updates. AI systems may also be subject to AI-specific obligations, such as monitoring for performance drift, defining retraining triggers and providing real-world performance reporting.
Service-level agreements and safety provisions in contracts: Customer agreements should set performance benchmarks, grant rights to deploy urgent safety updates, allocate obligations if model performance falls below specified thresholds and enable the use of deployment data for safety monitoring.
Licensing and Deployment Strategies: During deployment and post-market monitoring, innovators may implement licensing and distribution strategies aligned with the innovator or organization’s objectives. Licensing terms should also address software updates, post-market monitoring obligations, data governance and ongoing regulatory compliance.
Stage F: Scaling up and strategic partnerships
At the scale-up stage, strategic decisions about licensing, partnerships and potential exit strategies become central. An IP strategy may help to balance proprietary protection with ecosystem collaboration.
Patents portfolio expansion: Develop a diverse patent portfolio through divisional filings and continuations; foreign filings, including by pursuing the WIPO Patent Cooperation Treaty's national phase entry in key markets; portfolio reviews and freedom-to-operate analyses.
Protection of trade secrets in joint ventures or other strategic partnerships: Implement information firewalls, escrow arrangements and strict terms when working with partners, especially to safeguard crown jewel trade secrets such as proprietary data sets, unique model architectures, algorithms for data preprocessing, augmentation or feature engineering and methods for addressing AI-specific challenges.
Knowledge transfer plans for copyright works: In scaling or acquisitions, plan for knowledge transfer, including documentation, the development of training programs, the negotiation of source code rights, the clarification of data rights (which may go beyond copyight) and the definition of support obligations. Knowledge transfer may also require consideration of relevant licensing approaches including whether additional licenses are required.
Bilateral Licensing Models: To facilitate targeted commercialization and technology deployment, consider structuring negotiated bilateral licensing agreements with clearly defined rights, territories, fields of use and revenue-sharing mechanisms tailored to each partner. Revenue models should reflect the scope and exclusivity of the license, while bilateral arrangements can support strategic partnerships, market expansion, local adaptation and long-term collaboration.
Open-source and dual licensing models: To balance open collaboration with commercial revenue, consider releasing a core version under an open-source license while offering premium features commercially, backed by contributor agreements. The revenue models should reflect the licensing approach. Open-source versions can be used to build community, drive adoption and create network effects.
Standards participation: Where technology becomes essential to a standard, assess and manage licensing commitments while retaining exclusivity where possible. Join standards bodies (ISO, International Electrotechnical Commission, Institute of Electrical and Electronics Engineers (IEEE) and Health Level Seven (HL7)) to influence standard development, identify which patents may be essential to implementing a standard, declare the willingness to license essential patents on RAND or FRAND terms, retain full exclusivity on non-essential patents and trade secrets, and negotiate royalty rates that balance access with fair compensation.
Exit readiness through contractual assignment : Ensure complete IP assignment to the company, resolve encumbrances that could complicate a transaction, obtain confirmatory assignments from all inventors and contributors, prepare standard agreement templates for efficient negotiation, and maintain organized IP due diligence materials. Separate licensing and deployment strategies, including tiered pricing, licensing approaches, local patent filing, and voluntary technology transfer, may also be considered when expanding into low- and middle-income countries.
Caption Health, founded in 2013 in California, pioneered AI-driven real-time guidance for ultrasound image acquisition. It has addressed a critical bottleneck in diagnostic imaging: the shortage of trained sonographers and technical expertise to capture high-quality images. Its flagship software, Caption AI, uses computer vision and deep-learning algorithms to analyze live ultrasound video streams in real time, evaluating the probe angle, cardiac chamber visualization and image clarity. It guides clinicians through voice prompts and visual overlays to optimize positioning, effectively transferring expert sonographic knowledge to the point of care and enabling non-specialist clinicians to capture diagnostic-quality cardiac images.
Caption Health's IP strategy combined patents, trade secrets and copyright protection. Its core United States patent (US 11,844,654 B2, “Ultrasound guidance dynamic progression method”) covers methods for dynamically guiding ultrasound workflows, real-time feedback presentation, adaptive sequence management and the integration of guidance logic with ultrasound diagnostic systems. Notably, the claim structure emphasizes practical guidance and system orchestration rather than abstract neural network structures, tying protection to specific types of technical implementation and thereby supporting enforceability under United States patent eligibility standards. Alongside this, the company's most competitively sensitive assets, including proprietary training data sets comprising thousands of expert annotated cardiac ultrasound studies, model weights optimized for real-time clinical inference and annotation methodologies developed with sonographer input, were maintained as trade secrets. The software, user interface designs and technical documentation were protected by copyright.
In 2020, Caption Guidance received FDA marketing authorization via the De Novo pathway, establishing a new Class II device category that subsequently served as a predicate for 510(k) clearance (K200755). This authorization required a demonstration of safety and effectiveness, clinical validation of image quality, human factors testing with nonexpert operators, software verification and cybersecurity risk management. The regulatory clearance proved strategically transformative: it validated clinical utility, created a reusable regulatory predicate for future submissions and materially increased the company's acquisition value. In February 2023, GE HealthCare acquired Caption Health, integrating Caption AI into its Vscan and LOGIQ ultrasound systems and leveraging GE's global distribution network and regulatory expertise to scale up internationally. This outcome illustrates how regulatory clearance can serve as a value multiplier on underlying IP assets.
Key lessons for innovators
Regulatory clearance often significantly enhances acquisition value. FDA 510(k) clearance de-risks the technology for potential acquirers and demonstrates clinical validation.
A combination of patent and trade secret protection is powerful. Patents provide defensive protection and licensing leverage, while trade secrets protect the core competitive assets (such as data sets, models and clinical know-how).
Strategic positioning supports acquisition. Caption Health built exactly what a large medical device company would want to acquire – validated technology with regulatory approval, strong IP and a clear path to integration.
A focus on clinical workflow integration adds value. Rather than building a standalone product, Caption Health designed for integration with existing ultrasound equipment, making adoption frictionless.
Deep Genomics, founded in 2015 and headquartered in Toronto, Canada, applies AI to the design and development of genetic medicines. Its proprietary platform analyses large genomic data sets to predict how genetic mutations affect cellular processes and to identify RNA-based therapies for rare genetic diseases, including Wilson's disease and beta thalassemia. The process enables the design of oligonucleotide therapeutics that modulate RNA splicing and gene expression in a more systematic way than traditional trial-and-error drug discovery approaches.
The company's IP strategy combines a robust patent portfolio with extensive trade secret protection. Patents issued by the United States, including US 11,887,696 B2, US 11,636,920 B2 and US 11,568,960 B2, cover machine-learning systems for classifying and interpreting genetic variants, training convolutional neural networks on biological sequence data and scoring variant impacts on molecular phenotypes. Spanning patent classifications across genetic engineering, therapeutic preparations and bioinformatics, these patents are framed around systems and methods applied to real biological sequence data rather than abstract algorithms, strengthening their commercial enforceability. Complementing this patent portfolio, the company's most valuable asset is its proprietary database of genomic variant effects. It built this through years of curation and experimental validation and protected it as a trade secret, alongside neural network architectures, oligonucleotide design algorithms and experimental validation protocols.
Deep Genomics has raised over USD 180 million in private venture capital from investors including SoftBank Vision Fund 2, and has advanced several RNA-based drug candidates into preclinical and clinical testing. The company has benefited significantly from Toronto's world-class AI research ecosystem, drawing on the University of Toronto, the Vector Institute and a deep pool of AI and genomics talent produced through strong academic and government-supported programs. It has also accessed federal R&D tax credits and orphan drug incentives from the province of Ontario to help conserve capital during the early stages of commercialization.
Key lessons for innovators
Building a strategic patent portfolio provides broad protection and a competitive advantage: Deep Genomics filed foundational patents on RNA-targeting methods before clinical validation. Patent filing (preclinical data) can be advantageous when technology is novel and competitors are emerging.
De-risking through partnerships before major fundraising is critical. The SoftBank Vision Fund 2 investment (USD 180 million) was predicated on showing proof-of-concept pharma collaborations.
Geography matters for accessing specialized talent and government incentives. Toronto's AI talent pool (such as the Vector Institute and the University of Toronto) provided technical talent, while Canadian orphan drug incentives and R&D tax credits reduced early-stage risk.
United States market validation is essential. Despite its origin in Canada, Deep Genomics prioritized United States patent protection and FDA orphan drug designations, recognizing access to this market as critical for exit valuation and pharma partnerships.
InstaDeep was founded in 2014 with roots in Tunis and headquarters in London. It specializes in decision-making AI systems using advanced machine learning techniques, particularly reinforcement learning. The company gained international recognition for applying AI to drug discovery, vaccine design and protein structure prediction, with its work on COVID-19 variant tracking bringing it to global prominence. Its platform combines reinforcement learning with domain-specific biological knowledge to explore vast chemical and biological design spaces through iterative optimization. This approach is particularly well suited to problems where the optimal solution must be discovered rather than learned from labeled data.
InstaDeep's IP strategy balances patents, trade secrets and strategic open-source contributions. As a co-assignee with BioNTech SE, the company holds United States patent application 20250292868A1, covering machine learning systems for multimodal biological sequence analysis that integrate sequence representations with natural language interfaces. Rather than claiming reinforcement learning algorithms in the abstract, the patent is framed around practical system architectures, including biological sequence encoding, multimodal input handling, flexible inference pipelines and human-in-the-loop interaction models. This approach provides commercially meaningful coverage while mitigating subject matter eligibility risks. Core competitive assets, including proprietary data sets combining computational predictions with experimental validation data, optimized model architectures and reward-shaping techniques, are maintained as trade secrets. Contributions to open-source reinforcement learning frameworks focus on general-purpose tools rather than domain-specific implementation, building ecosystem credibility without exposing proprietary biotech capabilities.
InstaDeep's Tunisian origins are significant, demonstrating that world-class AI innovation can emerge from markets outside traditional biotech hubs when supported by strong science, technology, engineering and mathematics education; multilingual talent; government support for tech entrepreneurship and well-developed diaspora networks. In January 2023, BioNTech acquired InstaDeep for a reported GBP 362 million upfront, one of the largest AI-biotech deals in Europe. This enabled BioNTech to integrate reinforcement learning capabilities into mRNA vaccine design, enhance computational immunology and accelerate personalized cancer therapy development. InstaDeep's investors gained substantial returns and access to resources to scale up globally.
Key lessons for innovators
Companies in low- and middle-income countries can compete globally. Tunisia's tech ecosystem produced a company that attracted a major pharmaceutical acquirer, demonstrating that geographic origin is not destiny.
Reinforcement learning offers unique value in biopharmaceuticals. The ability to explore design spaces and optimize complex objectives aligns well with drug discovery challenges.
Strategic partnerships accelerate growth. InstaDeep's early collaborations with pharmaceutical companies validated its technology and created acquisition interest.
Infervision, founded in 2015 and headquartered in China, develops AI-enabled medical imaging software with flagship products focused on lung nodule detection from chest computed tomography (CT) scans and stroke triage workflows. Its commercial positioning emphasizes workflow-integrated decision support for radiologists alongside an expanding global regulatory footprint.
Infervision's IP strategy combines patent protection in key commercial markets with confidential know-how for proprietary training data and rapid iteration capabilities. Despite its origins in China, the company has pursued United States patent protection, recognizing that it provides global validation and supports commercial positioning for international partnerships. Its portfolio includes both narrow, application-specific patents (such as US 10,937,157 B2, covering convolutional neural network integration with CT image preprocessing pipelines for pulmonary nodule detection) and broader platform-level patents (such as US 11,200,982 B2, covering medical data analysis methods driven by machine learning and applicable across imaging modalities). This dual approach creates defensive coverage that protects current products while preserving the freedom to expand into new clinical indications. Alongside its patent portfolio, the company likely maintains confidential treatment of curated CT imaging data sets, neural network architectures, training protocols and operational processes, enabling rapid model updates and hospital integration.
Infervision has secured regulatory clearances in both China and the United States, strengthening its position across key markets. In the United States, it obtained FDA 510(k) clearance for its InferRead Lung CT AI system in 2020 (K192880), followed by a second clearance in 2025 (K240554) for an enhanced version with expanded nodule characterization capabilities. In China, it received NMPA Class III approval in 2022 for its InferRead CT Stroke solution. While these clearances do not provide statutory exclusivity in the manner of pharmaceutical data protection, they create meaningful first-mover advantages through established clinical evidence and predicate device status that later entrants must meet or exceed through their own validation processes.
Key lessons for innovators
A dual market regulatory strategy can reduce single-market dependence and strengthen international commercialization pathways. Infervision's approach entailed dual-track regulatory strategy – pursuing both FDA clearances (market credibility in the United States) and NMPA approvals (domestic market protection). This approach is particularly relevant for companies that are based in markets with evolving regulatory frameworks and are seeking to establish a global presence.
Benchmarking tools may support understanding of the regulatory landscape. For innovators targeting markets beyond China, the United States and the European Union, where national regulatory authorities may lack AI-specific frameworks, WHO's Global Benchmarking Tool
(8)Available at https://www.who.int/tools/global-benchmarking-tools. for regulatory system maturity offers a practical reference. It helps in understanding the regulatory landscape and engaging constructively with national authorities with varying levels of capacity. The tool assesses national regulatory systems across six functions, including market authorization, post-market surveillance and laboratory access, providing innovators with a structured basis for evaluating the regulatory environment in target markets.