This appendix documents the data sources, definitions, sample-construction rules and limitations behind two empirical pillars of the Special theme. The first section describes the Dealroom–WIPO Deep Science Startup Tracker, the basis of Special theme section How large and important is deep science entrepreneurship?. The second section describes the procedure used to match the resulting universe of firms to global patent data, the evidence base for Special theme section What role does intellectual property play in deep science entrepreneurship? and for the SPARK: Patenting in Deep Science Startups and Spinouts (Global Innovation Index 2026 Special) (
1 The Dealroom–WIPO Deep Science Startup Tracker
1.1 Purpose and scope
The Dealroom–WIPO Deep Science Startup Tracker measures the extent and distribution of deep science entrepreneurship across economies and sectors. It is the empirical basis of Special theme section How large and important is deep science entrepreneurship? and is searchable through the publicly available Dealroom–WIPO Deep Science Startup Tracker released alongside this GII edition. The tracker covers startups and spinouts founded from the year 2000 onward. This window is long enough to capture firms that take a decade or more to validate their technology, while still reflecting recent innovation activity.
1.2 Underlying data
The tracker draws on Dealroom’s proprietary database. Dealroom aggregates company-level information from four sources: public information; contributions from more than 120 partner tech ecosystems and over 10,000 individual contributors (local accelerators and incubators, venture capital funds with in-depth ecosystem knowledge, angel associations); user-submitted records subsequently verified by Dealroom; and data engineering on regulatory filings and other structured sources. All records are verified and curated through an extensive manual process. Underlying data are accessible at app.dealroom.co.
1.3 Defining deep science startups and spinouts
In this chapter, deep science startups and spinouts are companies built on tangible engineering innovation or scientific discovery applied for the first time as a commercial product. They are typically capital-intensive, time-intensive and R&D-intensive. A firm qualifies under two primary criteria:
Technology complexity and time to market. The underlying technology requires extended development before market-ready maturity. Firms typically employ a high share of researchers and engineers from inception and often create proprietary hardware or other novel intellectual property.
Capital requirement. Research, development, testing and scaling require substantial up-front investment, well beyond the working-capital needs of typical software firms.
A secondary criterion – holding significant IP or being directly spun out of a research institution – raises the probability that a firm is classified as deep science but is not required for inclusion.
The scope used here covers nine fields in which firms act on the physical world: Life Sciences; Medical Devices and Digital Health; Food and Agritech; Semiconductors; Robotics and Autonomous Systems; Space; Energy; Transportation; and Advanced Materials and Manufacturing. The scope deliberately excludes digital-first deep tech such as large language model developers, blockchain infrastructure, novel cybersecurity protocols and other foundational software. Software firms qualify only when their primary application is on the physical world and they meet the complexity and capital criteria above. Examples of qualifying firms include AI drug discovery, autonomous driving and digital therapeutics.
Some sectors are nearly always classified as deep science (Robotics, Semiconductors, Space), with rare exceptions such as marketplaces for semiconductor IP. Others (Energy, Transportation, Food and Agritech) include both deep science segments and segments that are not. A biotechnology firm developing crop protection qualifies; a food-delivery firm or a standard farm-management platform does not. Final classification is made firm by firm, applying the criteria above to documented company information. To allow consistent historical comparison, firms remain classified as deep science even after the underlying technology becomes mainstream.
1.4 Sample construction
For the analyses presented in Special theme Section 3, the scope is further restricted by three filters applied jointly:
VC-backed financing. Firms have raised at least one venture capital round (Pre-Seed, Seed, Series A, B or C+). Venture funding excludes debt, non-equity grants and other lending capital. Exits (M&A and IPO transactions) are analyzed separately.
Sector. Firms operate in one of the nine deep science fields listed above.
Technology risk and capital intensiveness. Firms meet the firm-level criteria for complexity, time to market and capital requirements set out in Section 1.3.
The VC-backed filter trades coverage for comparability. Theoretical definitions of startups based on firm age, headcount or growth rate are difficult to apply consistently to private firms across economies, and fit deep science ventures particularly poorly (long development phases, lumpy growth, late commercial revenue). Conditioning on at least one VC round is observable and comparable across geographies, and captures the great majority of successful deep science ventures.
1.5 Identifying spinouts
The tracker treats spinouts as a subset of deep science startups: firms originating from research carried out in a university or research center. Equity ownership by the parent institution is not required. Non-equity arrangements (royalty-only agreements, professor-privilege regimes such as Sweden’s) and informal but foundational links to an institute’s research are sufficient. Firms that received support, acceleration or peripheral licenses from a university but were not founded on its research are not classified as spinouts.
Spinout identification follows two complementary procedures:
Data-driven detection. Deep science VC-backed firms in the database are processed through automated review of company websites, press coverage and founder backgrounds, followed by human validation. Public information from university websites and national registers (for example, the HESA spinout register in the United Kingdom) feeds this process.
Ecosystem network collaboration. Technology transfer offices, valorization teams, national TTO associations, governmental institutions and spinout-focused funds supply firm identifiers and parent-institution links. Submissions are validated by Dealroom, which retains final responsibility for classification.
Firms emerging from joint research at two or more institutes are classified as spinouts of each institute. In economy- or city-level analyses, each firm is counted once per economy, though it may still appear in two distinct geographies. The spinout is counted in the economy of the institute of origin, not its current headquarters. A Brazilian spinout that was established in or relocated to the United States is shown as Brazilian in the spinout-specific statistics. When a spinout is acquired, the acquirer is not retroactively classified as a spinout, even if the parent institution retains equity in the acquirer.
Applying these rules, university spinouts represent over 30 percent of the deep science startup universe.
1.6 Data limitations
The data are global in scope, but several limitations should be borne in mind when interpreting the results.
Deep science firms without VC funding. A small share of deep science ventures achieve scale without raising venture capital, financed instead by grants, founder equity, family offices or early revenue. These firms fall outside the tracker by construction. Coverage of VC-backed status is also uneven across regions: emerging ecosystems with limited regulatory transparency and thinner press coverage are harder to track. Dealroom mitigates this through its global network of 120 ecosystem partners and bilateral data exchanges with regional providers.
Firm-level classification. Whether a firm meets the deep science criteria is determined by Dealroom’s curation team from gathered company information. The assessment is comparable across geographies, but very early-stage or stealth firms that disclose little about their technology are harder to classify with confidence.
Spinout transparency. Tracking of spinout creation by universities, research centers and public funders remains uneven. Some institutions publish complete lists; some maintain private lists shared on request; others do not track spinouts systematically. Transparency varies by region for structural and cultural reasons, and smaller institutions often lack tracking capacity. Data-driven detection is also weaker where founder profiles and company-founding histories are less documented online.
Failed ventures and the most recent year. Removing every failed deep science venture from the database, and capturing every firm founded in the most recent year (in this edition, 2025), is not feasible. Counts and trends in the most recent year should be read as provisional.
Country coverage. Coverage in low-income and certain middle-income economies is solid but perfectible. China is a particular case. Transparency on startup creation is limited, especially in sectors with national-security implications; many firms have no presence on global platforms or search engines; and venture financing involves direct participation by government and local development agencies in ways that do not map cleanly onto standard round labels (
To address these gaps over time, a notification system has been introduced as of the GII 2026 launch in September 2026. WIPO Member States, deep science firms and their representatives can flag omissions through the dedicated portal; submissions are reviewed by Dealroom and integrated into subsequent updates, with the aim of improving global coverage over future editions.
2 Matching deep science startups to global patent data
2.1 Rationale
Patenting patterns across deep science ventures have until now been studied mostly at the national or regional level (
2.2 Source data
Two underlying datasets are matched:
the Dealroom–WIPO Deep Science Startup Tracker described in Section 1, with firm-level identifiers including legal name, country of headquarters, founding year, founder names and parent institution where applicable; and
Patsnap’s global patent database (data extracted in March 2026)Patsnap (2026). Patsnap global patent database. Data extraction, March 2026. Singapore: Patsnap. , which integrates application and grant records from over 100 national and regional patent offices, including Patent Cooperation Treaty (PCT) applications. Records include applicant names, applicant addresses, inventor names, filing and publication dates, IPC and CPC classifications, and family-level identifiers.
2.3 Matching procedure
Matching proceeds in three stages. First, applicant names from the Patsnap patent database are harmonized through standard cleaning: removal of legal-form suffixes (Inc., Ltd., GmbH, KK, AG, SA), case and punctuation normalization, and enumeration of all possible name variants (including former names, operating names, registered names, etc.).
Second, each firm in the tracker is searched against the Patsnap patent database: all identified name variants of each firm are used as search queries, and all retrieved patent records are aggregated to form a comprehensive set of potentially relevant patents associated with that firm. The search covers both the original assignee name field and the current assignee name field, as the current assignee of a patent may differ from the original assignee due to corporate name changes, mergers, acquisitions or patent assignments. This captures all relevant patents, including those filed under a former name, acquired through corporate transactions or subsequently transferred to another entity.
Third, candidate matches are reviewed against firm metadata. Country of headquarters, founding year and business information of each firm are cross-checked against applicant addresses, the earliest priority date of the candidate’s patent portfolio and the technical information contained in patent abstracts. Matches that pass all three checks are accepted. For acquired firms, patent portfolios are attributed to the deep science startup up to the date of acquisition; subsequent filings under the acquirer’s name are not added to the startup’s count.
For comparisons against non-deep science VC-backed startups, a sampling approach is used given the size of the comparator universe. Sample sizes are set to yield a margin of error of approximately ±1 percent at the 95 percent confidence level for the overall non-deep science comparator, with wider margins at the income-group level (high income: ±1.1 percent; upper-middle income: ±2.3 percent; lower-middle income: ±2.9 percent; low income: ±5.3 percent).
2.4 Coverage
Matching covers the full universe of VC-backed deep science startups and spinouts in the tracker. Across this scale range, almost half of deep science startups (47 percent) are matched to at least one published patent family, against 16 percent for non-deep science VC-backed startups in the comparator sample. Coverage by sector, economy and stage is reported in full in the corresponding WIPO Technology SPARK report. Patent activity is most prevalent in Medical Devices and Digital Health, Life Sciences, Semiconductors and Advanced Materials and Manufacturing, and lower in segments where firms may rely more heavily on trade secrets and copyright.
2.5 Limitations
Filing lags. Patent applications are published 18 months after the priority filing date. Firm-level patent counts therefore understate filings made in the most recent 18 to 24 months, and trends in 2024 and 2025 should be read as provisional.
Naming inconsistencies. Earlier-stage firms occasionally file under founder names, holding companies or special-purpose vehicles whose link to the operating company is not always documented in public sources. Manual review recovers most of these cases, but residual under-counting cannot be ruled out, particularly in economies with less complete patent-applicant records.
Non-patent protection. Patent counts capture only one form of intellectual property. In several deep science segments, notably software-rich life sciences and AI drug discovery, firms may also rely on trade secrets, copyright over code and exclusive rights over training data. The absence of patents in these segments should not be read as the absence of IP.
Sampling for the comparator. The non-deep science comparator is sampled rather than enumerated, with the margins of error reported in Section 2.3.
2.6 Related work
The matching procedure builds on a small set of recent national exercises. The Royal Academy of
References
DeepTech Asia (2026). 2025: Renaissance of China’s deep tech VC. DeepTech Asia, January 7.
EUIPO and EPO (2024). Patents, Trade Marks and Startup Finance. Alicante and Munich: European Union Intellectual Property Office and European Patent Office.
Patsnap (2026). Patsnap global patent database. Data extraction, March 2026. Singapore: Patsnap.
Royal Academy of Engineering and Beauhurst (2024). Spotlight on Spinouts: UK Deep Tech Report. London: Royal Academy of Engineering.
WIPO (2026). Technology SPARK Report: Patenting in Deep Science Startups and Spinouts (Global Innovation Index 2026 Special). Geneva: WIPO, November.