Assessing Bio-AI/ML Convergent Threats: Using Historical Data for Future Trajectories

1.1 Bio-AI/ML Convergence and its National Security Relevance

Bio-AI/ML convergence is the integration of biotechnology with artificial intelligence and machine learning, producing capabilities that neither domain achieves in isolation. Current global trends show the United States, alongside both allied and adversarial nations, advancing in biotechnology, particularly in genetic engineering, through this integration (Pauwels & Vidyarthi et al., 2017). The 2025 National Security Commission on Emerging Biotechnology (NSCEB) report identifies China as an emerging global leader in this domain, supported by mass genomic data collection, intellectual-property acquisition, and supply-chain control. Roughly 80% of U.S. biomanufacturing companies depend on Chinese suppliers such as WuXi AppTec for components of their production pipeline (NSCEB, 2025). This structural dependency was especially evident during the COVID-19 pandemic which left the U.S. exposed to supply-chain disruption, drug shortages, and constrained medical resources.

Within the broader convergence domain, precision medicine occupies a central position. Precision medicine uses patient-level data such as genetics, lifestyle factors, and family history to personalize diagnostics, treatment, and prevention. Advances in gene-editing platforms such as CRISPR (You et al., 2019) have expanded the technical surface of precision medicine into gene therapy, drug development, and pharmacogenomics. AI/ML methods accelerate this further by improving predictive accuracy in gene-editing design and by enabling analysis of genomic datasets at a scale unavailable to manual annotation (Kim et al., 2018; Eraslan et al., 2019).

Taken together, these tooling and data trends raise a measurement question: how can capability development in this domain be tracked using observable, openly available data, in a way that supports the comparison of national positions over time?

1.2 Bio-AI/ML Convergent Advantage Framework

To structure the measurement question, this paper adopts a framework with four characteristics of bio-AI/ML convergent advantage (Figure 1). Each characteristic is observable, and each maps directly to one or more results presented in 3.

Figure 1. Bio-AI/ML Convergent Advantage — a conceptual diagram of the four characteristics that jointly produce convergent advantage.

The first characteristic, Platformization of Biology, refers to the transformation of biological tools — CRISPR, base and prime editors, synthetic-biology toolkits — into modular, programmable, and updatable platforms. Platformization allows a nation to pivot quickly across disease targets and emerging threats. In 3, this characteristic is operationalized through publication output in the gene-editing domain (Scopus B1–B2 searches).

The second characteristic, AI/ML-Driven Acceleration, refers to the use of AI/ML tools to compress timelines for therapeutic and synthetic output, optimize design pipelines, and increase the predictive accuracy of biological discovery. The Exscientia OCD-medication case, in which an AI-designed drug entered clinical trials in approximately one year against an industry baseline of a decade (UKRI, 2023), is illustrative. This characteristic is operationalized through AI+biotechnology and ML+biotechnology publication output (Scopus B5–B8).

The third characteristic, National Innovation System (NIS), refers to the network of academic institutions, private-sector companies, and government agencies that jointly drive innovation (Weerasinghe et al., 2024). A strong NIS provides the infrastructure, funding, and policy environment in which platformization and AI/ML-driven acceleration can scale. NIS is operationalized in 3 through the composite scoring of institutions across Activity, Impact, and Strategic Relevance dimensions, and through the comparative growth-rate signals observed over the 2015–2025 period.

The fourth characteristic, Geopolitical Biotech Rivalry, is the comparative dynamic between nations pursuing bio-AI/ML convergent capability. Rivalry intensifies both innovation pressure and the strategic stakes attached to capability concentration. This characteristic is operationalized in 3.2 through the security-designation analysis, specifically, the prevalence of U.S. government designations across the leading Chinese biotech institutions identified by Scopus output.

These four characteristics provide the conceptual foundation for the empirical framework developed in this study. Instead of measuring these characteristics directly, the lifecycle data-collection framework described in 2 operationalizes them through three dimensions:  Research Activity, Research Impact, and Strategic Relevance. Collectively, this framework enables a reproducible and systematic assessment of bio-AI/ML convergent capabilities across different institutions and countries.

1.3 Research question and contributions

This research made three key contributions. First, a descriptive research question was investigated: how do the bio-AI/ML convergent capabilities of the U.S. and China compare across research activity, research impact, and strategic relevance, and what do the historical trajectories imply for near-term projection? Second, a reproducible lifecycle data-collection and composite-scoring framework was developed (2). Third, an empirical baseline (3) against which future updates can be measured was established. Furthermore, the synthesis of what the historical trends imply for future trajectories (4); research limitations (5); and major conclusions (6) were presented.

2.1 Lifecycle data-collection framework

Building on the conceptual framework introduced in 1.2, this study translates the four characteristics of bio-AI/ML convergent advantage into a reproducible methodological framework for empirical assessment. A data-collection lifecycle structures this analysis into four sequential stages: source selection, data retrieval, composite scoring, and validation (Figure 2). Each stage applies operational criteria documented in the project’s planning materials. Source selection identifies databases that provide comparable data across both countries and across the three dimensions to be scored. Data retrieval applies inclusion criteria (Appendix A) and reproducibility logging. Every Scopus query records its parameters, the AFFILCOUNTRY filter applied, and a SHA-256 hash of the returned result set in a search log maintained by the authors. Composite scoring (2.3) combines retrieved data into a single per-institution measure across three dimensions. Validation (2.4–2.5) assesses whether the scoring is robust to weighting choices and consistent with known leaders. The same four-stage structure is applied to both the bibliometric (Track B) and security-designation (Track C) data streams.

Figure 2. Lifecycle Data-Collection and Composite-Scoring Framework. The four-stage lifecycle — source selection, data retrieval, composite scoring, and validation — operationalizes the four characteristics of the convergent-advantage framework in 1.2 (Figure 1); the three scoring dimensions, Research Activity (A), Research Impact (I), and Strategic Relevance (S), are combined at the composite-scoring stage.

2.2 Source strategy

Two database sources were used for the bibliometric and security-designation analyses. Publication counts (Track B) were retrieved from Scopus using TITLE-ABS-KEY queries with the AFFILCOUNTRY filter and a 2015/1/1–2026/2/28 date range. Eight searches were executed (B1–B8): four domain keyword sets (gene editing, precision medicine, AI + biotechnology, ML + biotechnology) crossed with two country filters. Using the same database and query method for both countries ensures directly comparable publication counts. Security designations (Track C) were retrieved from Data Abyss, a national-security research platform, via direct institution-name lookup against fourteen U.S. government designation codes: Section 1260H, Section 1286, BIS Entity List, FCC Covered List, FARA, Military End User, Military-Civil Fusion, Treasury NS-CMIC, and others (full code reference is located in 2.3 and the project’s risk-tier definitions). Data Abyss was the only source examined that provides native U.S. government designation data; it was not used for publication counts because its title-only search and absence of a date filter make counts unsuitable for bilateral comparison. Inclusion criteria for the supporting literature synthesis are documented separately in Appendix A.

2.3 Three-dimensional composite scoring

Each institution receives a composite score derived from three independently measured dimensions: Research Activity (A), Research Impact (I), and Strategic Relevance (S). Each dimension maps to inputs that can be retrieved from the sources in 2.2; the full Python implementation is documented and available from the authors upon request.

Research Activity combines two inputs: per-domain publication counts from the eight Scopus searches and Nature Index fractional count. Publication counts dominate the dimension because they cover every institution in the dataset; Nature Index supplements because it covers fewer institutions but reflects higher journal selectivity.

Equation 1:

A_raw = 0.6 * (pub_gene_editing + pub_pm + pub_ai_bio + pub_ml_bio)

      + 0.4 * nature_index_fc

Research Impact is the mean citations per paper, computed from Scopus cited by-count on a 25-paper sample per institution — the 25 most recent papers by Scopus cover date — drawn from the 2015–2023 publication window. This is a proxy for citation impact; the field-weighted alternative (FWCI from SciVal) was unavailable under the project’s API entitlement, and its absence is acknowledged as a limitation (5). For institutions without a Scopus AF-ID, the value is imputed with the mean of the institution’s country group.

Equation 2:

I_raw = mean_citations_per_paper  # imputed with country-group mean if AF-ID missing

Strategic Relevance converts the fourteen Data Abyss designation codes into an ordinal point scale grouped by severity tier. Tier 1 — Extreme (direct PLA operational unit) — contributes 4 points. Tier 2 — Sections 1260H, 1286, MEU, MCF, and Treasury NS-CMIC — contributes 3 points each. Tier 3 — BIS Entity List, FCC Covered List, and FARA — contributes 2 points each. Tier 4 — Defense-affiliated universities, central SOEs, top-tier VC/PE firms, and PRC-listed public companies — contributes 1 point each. Multiple flags accumulate, so an institution carrying both a Section 1260H and a BIS Entity List designation receives 3 + 2 = 5 points. The maximum points possible is 30. U.S. institutions serve as the reference baseline and are assigned S = 0 by design.

Equation 3:

S_raw = 4*flag_extreme + 3*(flag_1260+flag_1286+flag_meu+flag_mcf+flag_cmic)

      + 2*(flag_el+flag_cl+flag_fara)

      + 1*(flag_defense_univ+flag_soe+flag_vc+flag_pe+flag_public_company)

# Max possible: 30. US institutions: S_raw = 0 by design.

All three raw scores are normalized to the [0, 1] interval using min-max normalization, computed across the combined U.S.–China dataset. Min-max normalization is preferred over z-score standardization because the resulting composite is bounded on [0, 1] and interpretable as the fraction of the distance between the lowest- and highest-scoring institutions on each dimension. The edge case in which a metric has zero variance is handled by assigning the midpoint, so that an uninformative dimension does not penalize any institution.

Equation 4:

X_norm = (X_raw – X_min) / (X_max – X_min)    if X_max > X_min

X_norm = 0.5                                   if X_max = X_min

The composite score is a weighted sum. Chinese institutions use all three dimensions; U.S. institutions use a two-dimensional re-normalization in which the security weight is redistributed proportionally between activity and impact, so that the baseline country is not penalized for having S = 0.

Equation 5:

C_china = 0.40*A_norm + 0.35*I_norm + 0.25*S_norm

C_us    = 0.533*A_norm + 0.467*I_norm

The default Chinese weights (w_A = 0.40, w_I = 0.35, w_S = 0.25) place research quality at 75% of the composite and strategic relevance at 25%, reflecting that institutions enter the table primarily because they perform biotechnology research, not because they hold security designations. The activity weight slightly exceeds the impact weight because publication count is available for all institutions while citation impact is sample-based and noisier.

Because Strategic Relevance is defined from U.S.-government designations of foreign entities, U.S. institutions carry S = 0 by construction, and the redistribution in Equation 5 reallocates the security weight to Activity and Impact so that the baseline country is not penalized for a dimension it cannot, by definition, register. One consequence is that the composite is a strategic-relevance-adjusted score rather than a pure research-capability score: a Chinese institution whose normalized security score exceeds its research composite receives a higher composite than a U.S. institution with identical Activity and Impact. This is intended — Strategic Relevance is treated as a first-class dimension of the assessment, not a research-quality proxy — and its effect is made transparent in 3.3, where the research-only (Activity + Impact) ranking is reported alongside the full composite so that security-driven placements can be distinguished from research-driven ones.

2.4 Stability analysis

The default weights are assumptions and warrant a sensitivity test. Nine weight-perturbation scenarios are evaluated (Table 1), spanning equal weighting, single-dimension dominance, and pairwise dominance. For each pair of scenarios (i, j), the Spearman rank correlation ρ(i, j) is computed between the institution rankings the two scenarios produce. The resulting 9 × 9 matrix yields ρ_min, the worst-case ranking agreement across all scenario pairs.

Table 1. Weight-perturbation Scenarios for Stability Analysis.

Classification follows three thresholds: stable if ρ_min > 0.85 (worst-case ranking agreement at least 85%), marginal if 0.70 < ρ_min ≤ 0.85, and unstable if ρ_min ≤ 0.70. These thresholds were fixed in advance of computing the correlation matrix, so the marginal classification is not a post-hoc choice. For the dataset analyzed here, ρ_min = 0.752 and ρ_mean = 0.919, placing the result in the marginal range. The worst-case scenario pair is the activity-only vs security-only comparison (S8 vs S9), as expected when contrasting two extreme single-dimension views. The discrete four-tier classification is nevertheless fully stable: no institution changes tier across any of the nine scenarios (3.3).

2.5 Tier assignment and external validity

Tier boundaries are assigned by k-means clustering on the composite score (k = 4, deterministic seed), so that breakpoints reflect the score distribution rather than arbitrary percentile cutoffs. The four clusters are mapped to Tiers 1–4 in descending order of cluster mean, and institutions in the bottom cluster are excluded unless they qualify on security relevance grounds. External validity is checked against a pre-specified ground-truth list of eight institutions: Chinese Academy of Sciences, Tsinghua, Peking, and the University of Science and Technology of China (USTC) for China; Harvard, the Massachusetts Institute of Technology (MIT), Stanford, and UC Berkeley for the U.S. The scoring is considered well-calibrated if at least three of these eight appear in Tier 1 or Tier 2.

3.1 Research activity and research impact

Across the 2015/1/1–2026/2/28 window, the Scopus searches retrieved 31,026 U.S. and 19,334 Chinese biotech publications across the four domains — counted over the top contributing institutions per search (2.2) — covering 66 U.S. and 74 Chinese institutions in the bibliometric dataset. On these institution-level counts (Figure 3), the U.S. leads decisively in the two large domains, gene editing (10,897 vs 10,317) and precision medicine (19,717 vs 8,626). The two AI/ML domains are far smaller and close: the U.S. leads ML + biotechnology (217 vs 159), while China narrowly leads AI + biotechnology among top institutions (232 vs 195). These counts are capped at the top contributors per search and are used for the institution ranking in 3.3; they are not the basis for domain-level direction over time. For that, the full uncapped annual counts in 4.1 (Figures 4, 5, and 13) are used, and they show all four domains U.S.-led in the early years, with China crossing U.S. annual output in gene editing, precision medicine, and ML + biotechnology by 2025 while AI + biotechnology remains U.S.-led. The aggregate U.S. advantage in the institution-level counts is concentrated in precision medicine, which alone accounts for 64% of total U.S. biotech publications in the dataset.

Figure 3. National-level Publication Output, U.S. vs China, across four bio-AI/ML domains, 2015–2026.

From 2015 to 2025, China’s combined biotech publication volume grew from 186 to 6,061; the U.S. grew from 1,043 to 4,771. The crossover is uneven by domain: gene editing, precision medicine, and ML + biotechnology cross parity between 2023 and 2025, while AI + biotechnology remains U.S.-led despite a sharp Chinese surge in 2024–2025 (Figure 4). By 2025, China leads the combined annual total by 27% (Figure 5). Year-over-year growth-rate analysis (Figure 6) reveals that China’s lowest annual growth occurred in 2023 at 9%, against the U.S. lowest at -8% in 2022. This suggests that China’s national innovation system absorbed pandemic-period shocks with greater resilience, an indirect signal of the National Innovation System characteristic. The pace and concentration of growth in AI + biotechnology and ML + biotechnology further operationalize the AI/ML-Driven Acceleration characteristic of 1.2: each of these two domains roughly tripled in Chinese output between 2022 and 2025.

Figure 4. Publication-output Trends by Domain, 2015–2025 (four-panel small multiples; raw publication counts).

Figure 5. Combined Biotech Publication Volume Over Time, China vs U.S. China surpasses the U.S. between 2024 and 2025 (6,061 vs 4,771 in 2025).

Figure 6. Biotech Publication Growth Rate, China vs U.S. (year-over-year % change, all four domains combined, 2016–2025). China stays positive in every year, with its weakest year (9%, 2023) still above the U.S. weakest year (-8%, 2022).

At the institution level, Harvard Medical School holds the top rank with 2,977 publications across the four domains, followed by the Chinese Academy of Sciences (CAS) with 2,655. The next eight ranks are entirely U.S. institutions: NIH, Stanford University, Stanford University School of Medicine, Massachusetts General Hospital, the University of California, San Francisco (UCSF), the University of Pennsylvania, Brigham and Women’s Hospital, and the University of Washington. Following those institutions, Fudan University enters the list at rank 15. The top 20 contains 15 U.S. institutions and 5 Chinese institutions (CAS, University of Chinese Academy of Sciences, Fudan, Chinese Academy of Medical Sciences & Peking Union, and Zhejiang) (Figure 7). In the per-domain head-to-head (Figure 8), CAS leads in two of the four domains: gene editing (1,851), and AI + biotechnology (40). However, it trails Harvard Medical School in precision medicine (746 vs 1,901) and MIT in ML + biotech (18 vs 20). This institutional pattern reflects the Platformization of Biology characteristic of 1.2: CAS-style horizontal scale across multiple platform domains.

Figure 7. Top 20 Institutions by Publication Volume across the four domains (blue = U.S.; red = China Tier 1; yellow = China Tier 2). Fifteen of the top 20 are U.S. institutions; five are Chinese.

Figure 8. Top 5 Institutions Per Domain, head-to-head U.S. vs China.

On research impact, the picture inverts in part. Mean citations per paper are nearly identical at the country level (China 23.3, n=65; U.S. 25.0, n=61), but the distributions differ. The single highest-impact institution in the dataset is Xiangya Hospital Central South University (75.2), followed by Tongji Medical College of Huazhong (42.2) among Chinese institutions, and by Princeton University (59.8) followed by the Buck Institute for Research on Aging (53.8) and Lawrence Berkeley National Laboratory (50.9) among U.S. institutions (Figure 9). On Nature Index fractional count, China leads both at the mean (414 vs 274) and at the maximum (CAS at 3,167 vs Harvard University at 1,152), reflecting Chinese concentration of high-impact biotech publishing in a small number of dominant institutions (full per-institution Nature Index values are available in the reproducibility log).

Figure 9. Citation-Impact Comparison, Top 15 Institutions Per Country (mean citations per paper).

3.2 Strategic Relevance: Security Designations of Chinese Institutions

The strategic-relevance dimension applies only to the 74 Chinese institutions retrieved through the bibliometric search; U.S. institutions are the reference baseline and carry no U.S. government designations by definition. Of those 74 institutions, 65 (88%) carry at least one designation. Designation prevalence is reported over the full set of 74 bibliometrically identified institutions — the strategically relevant search population as a whole — rather than only over the 65 that carry a designation. The prevalence of the six most common codes is reported in Figure 10. Section 1286 captures espionage, intellectual-property theft, or collaboration with the People’s Liberation Army (PLA). This section is the most prevalent designation, applying to 64 of 74 institutions (86%). The BIS Entity List follows at 84% (62 institutions), and direct PLA operational status (the “Extreme” tier) applies to 49 institutions (66%). Section 1260H, which lists Chinese military firms active in the U.S., applies to 33 institutions (45%); the FCC Covered List to 13 (18%); and the Military End User list to 12 (16%).

Figure 10. Risk-flag Prevalence Across the 74 Chinese Institutions in the Dataset.

The concentration of designations among top-ranked institutions is striking. Every one of the top five China-ranked institutions identified in 3.1 carries multiple flags: CAS holds five designations (Extreme, 1260, 1286, EL, CL); Fudan University holds six (Extreme, 1260, 1286, EL, CL, MEU); the Chinese Academy of Medical Sciences & Peking Union Medical College holds three (Extreme, 1286, EL); Zhejiang University holds five (Extreme, 1260, 1286, EL, CL); and Shanghai Jiao Tong University School of Medicine holds three (Extreme, 1286, EL). The 1286-and-EL pair co-occurs in 62 of 74 institutions (84%), making it the dominant joint signal of strategic concern (Figure 11).

The Extreme/PLA tier co-occurs with the Entity List in 46 of 49 cases (94%); its co-occurrence with Section 1286 (48 of 49, 98%) is expected in part by definition, since Section 1286 itself encodes PLA collaboration. Read together, these overlaps indicate that direct PLA-operational status is almost always accompanied by export-restriction-related designations that are defined independently of PLA status. Section 1260H is also highly correlated with the other top-tier designations: it co-occurs with the Entity List in 31 of 33 cases (94%) and with the Extreme tier in 28 of 33 cases (85%). The Department of Defense’s (DoD) Chinese-military-firms list, the BIS Entity List, and direct PLA-operational status thus identify a substantially overlapping, rather than complementary population, with the same institutions appearing across multiple U.S. government designation regimes. 

Figure 11. Risk-flag Co-occurrence Across the 74 Chinese Institutions, shown as the share (%) of the 74 institutions carrying each pair of designations.

Sectoral inspection of the dataset (sector-level tabulation available in the reproducibility log) reveals the highest-tier designations clustering in two sectors: government laboratories (CAS, Chinese Academy of Agricultural Sciences) and academic research institutions (Tsinghua, Peking, Fudan, Zhejiang, Shanghai Jiao Tong). Commercial biotech companies appear primarily in Tier 3 (Entity List or FCC Covered List) or Tier 4 (state-owned enterprises, PRC-listed public companies).

This sectoral skew is consistent with the bibliometric finding in 3.1: the highest-publishing Chinese institutions are predominantly academic and government laboratories, which are also the institutions most likely to carry Tier 1–2 PLA-related designations.

The “Seven Sons of National Defense” — Beihang University, Beijing Institute of Technology, Harbin Engineering, Harbin Institute of Technology, Nanjing University of Aeronautics and Astronautics, Nanjing University of Science and Technology, and Northwestern Polytechnical University — all appear in the dataset and are all classified as Tier 1. Every one of the seven carries at least the Extreme, Section 1286, and BIS Entity List designations; five additionally carry Military End User status (Beihang, Harbin Institute of Technology, Nanjing University of Aeronautics and Astronautics (UAA), Nanjing University of Science and Technology (UST), and Northwestern Polytechnical). Three of the seven (Beijing Institute of Technology, Harbin Institute of Technology, Northwestern Polytechnical) appear in the Track B bibliometric search results, indicating direct biotech publication output from PLA-affiliated universities.

Taken together, these findings operationalize the Geopolitical Biotech Rivalry characteristic of 1.2: the leading Chinese biotech-publishing institutions are not merely productive research entities; their strategic alignment is documented by the U.S. government across multiple, mutually reinforcing designation regimes.

3.3 Composite ranking and tier assignment

Applying the composite scoring framework of 2 to the 140 institutions in the dataset yields a four-tier distribution: 3 institutions in Tier 1, 35 in Tier 2, 48 in Tier 3, and 54 in Tier 4. All three Tier-1 positions are held by Chinese institutions — Chinese Academy of Sciences (composite 0.612), Xiangya Hospital Central South University (0.520), and Fudan University (0.500). Within Tier 2, China holds 21 of the 35 positions and the U.S. holds 14. By contrast, in Tier 4 the U.S. holds 33 of 54 positions to China’s 21, indicating that the U.S. cohort is concentrated in the middle and lower tiers while the Chinese cohort populates the upper and middle tiers more densely. The highest-scoring U.S. institutions are Princeton University (0.406), Stanford University (0.404), Dana-Farber Cancer Institute (0.393), Stanford University School of Medicine (0.363), and Harvard University (0.353), all in Tier 2 and all separated from the Tier-1 cluster by approximately 0.1 composite-score units.

The three Tier-1 Chinese institutions reach the top tier through different dimensional drivers, illustrating the value of a multi-dimensional composite over any single bibliometric metric. CAS is activity-driven (A_norm = 1.000, I = 0.119, S = 0.824 — maximum publication volume, moderate per-paper impact, near-maximal strategic relevance). Xiangya is impact-driven (A = 0.036, I = 1.000, S = 0.529 — maximum citation impact despite low aggregate output). Fudan is security-driven (A = 0.311, I = 0.439, S = 1.000 — maximum strategic-relevance score from six co-occurring designations). The Activity-vs-Impact scatter (Figure 12) makes the cluster structure visible: the Tier-1 institutions occupy the extreme upper-right (CAS) and extreme upper-left (Xiangya) corners of the A–I plane, while Tier-2 institutions cluster in a denser middle band. Re-ranking on the research-only composite (Activity and Impact, with the security weight redistributed) leaves the Chinese Academy of Sciences and Xiangya as the top two institutions overall, while Fudan falls to sixth, behind Stanford, Princeton, and Dana-Farber. Two of the three Tier-1 positions therefore hold on research dimensions alone; Fudan’s reflects the strategic-relevance overlay the composite is designed to capture.

Figure 12. Activity vs Impact Scatter, colored by composite-score tier.

The stability analysis (2.4) returned ρ_min = 0.752 and ρ_mean = 0.919, placing the result in the marginal range (0.70 < ρ_min ≤ 0.85). The worst-case scenario pair is S8 (Activity-only) vs S9 (Security-only), reflecting the natural divergence between two extreme single-dimension views. The discrete four-tier classification is, however, fully stable: no institution changes tier across any of the nine scenarios. Because the tier assignments are invariant across all nine weighting scenarios (2.4), the study’s tier-level conclusions do not depend on the specific default weights: the weights affect only the finer ordering within a tier, while the stability analysis establishes that the tier assignments, on which the conclusions rest, are robust to the weighting choice. The external-validity check passes — 7 of 8 pre-specified ground-truth institutions appear in Tier 1 or Tier 2, the single exception being MIT, which falls to Tier 3 because of relatively low publication volume in the four target domains under the AFFILCOUNTRY filter.

At the dataset level, the cross-country comparison can be summarized across the empirical metrics already reported above. The U.S. retains advantages in total publication volume (31,026 vs 19,334) and mean publications per institution (470 vs 261). China leads in dataset institution count (74 vs 66) and Nature Index mean fractional count (414 vs 274), with median citations per paper at near-parity. The composite ranking thus combines two genuinely comparative narratives that no single metric captures alone: the U.S. retains aggregate volume leadership while China holds concentration of high-impact research output and dataset-level institution count, layered with the strategic-relevance designations characterized in 3.2.

4.1 Implications of the 2024–2025 Crossover

The bibliometric record between 2015 and 2025 shows a recurring pattern across three of the four domains studied: a steady Chinese rise that crosses U.S. annual publication output between 2023 and 2025. By 2025, China leads in annual output in gene editing (ratio 1.36), precision medicine (1.23), and ML + biotechnology (1.16), while AI + biotechnology remains U.S.-led (0.91) but with a sharply narrowing margin (Figure 13). Two features of these trajectories carry implications for the next three to five years. First, the crossover is not domain-isolated: it occurs in three of four target domains nearly simultaneously, which is more consistent with an underlying capability shift than with a single-domain anomaly. Second, the growth-rate trajectory through the COVID-19 disruption, China sustaining 9% in its weakest year against the U.S. -8% in its weakest year, indicates that the Chinese trajectory absorbed a major external shock without inversion. AI + biotechnology is the sole domain still U.S.-led in 2025, at a China-to-U.S. ratio of approximately 0.91 (Figure 13) with a steepening slope through 2024–2025; if the recent rate of convergence holds, it could cross parity within the next one to two years, though projection beyond the observed window should be read with caution because the trajectory is accelerating rather than linear.

Figure 13. China-to-U.S. Publication-Volume ratio Over Time, Four Domains, 2015–2025. Parity line at ratio = 1.0; three of four domains cross parity between 2024 and 2025. This figure visualizes the trajectory metric central to the manuscript’s thesis.

4.2 Innovation-System Context

These bibliometric patterns are most coherently read within the broader innovation-system context developed in the team’s parallel research efforts, which identified six accelerators and seven decelerators of bio-AI/ML convergent innovation (Bostick, 2026). Four of those factors map most directly to the trajectories in 3 and 4.1. First, sustained R&D investment growth: China expanded R&D spending by 9% per year between 2018 and 2023, against U.S. growth of 4.7% per year (Long, 2025a), broadly consistent with the publication-output trajectory through the same window. Second, STEM workforce scale: in 2022, China produced approximately 50,970 STEM PhDs against approximately 33,820 in the U.S. — a gap of roughly 17,000 graduates per year, up from approximately 500 in 2007 (Long, 2025b) — expanding the labor base for bibliometric output. Third, unified national strategy: China’s sequenced Five-Year Plans, Made in China 2025, and the 15th FYP create a coherent multi-decadal incentive structure, while U.S. biotech policy spans multiple agencies without a single national lead (NSCEB, 2025). Fourth, genomic-data infrastructure: the Beijing Genomics Institute (BGI) group’s globally distributed sequencing capacity provides Chinese researchers with data-volume advantages at the input side of the bio-AI/ML pipeline (Wang et al., 2020; Wang et al., 2025). These four factors, jointly observed in the historical record, support reading the 3 trajectories as structurally driven rather than cyclically driven.

Seven limitations of the methodology in 2 are acknowledged here.

First, Scopus coverage is not symmetric across countries. Scopus provides directly comparable cross-country counts but may undercount Chinese publications that appear in domestic journals not indexed by the platform, particularly China National Knowledge Infrastructure-only outlets. The bilateral comparison should therefore be read as Scopus-indexed output, not as a complete census of Chinese biotechnology research.

Second, research impact is operationalized as mean citations per paper rather than as the field-weighted citation impact (FWCI). FWCI requires SciVal API entitlement, which was not available under the project’s API key. Mean citations is therefore not normalized by research area, document type, or publication year, and an institution publishing in a hot field will score higher independent of relative impact within that field.

Third, risk-flag name matching is performed by institution-name lookup against Data Abyss; 93% of Chinese institutions in the dataset matched successfully. The remaining 7% may carry designations that this analysis does not capture.

Fourth, the designation list is a point-in-time snapshot. Risk designations and rankings change over time, and the results in 3 should be read as reflecting the search date (March 2026), not as standing classifications.

Fifth, the disruption index (CD index), originally proposed as a third bibliometric dimension, could not be retrieved at scale from available sources. Mean citations per paper substitutes as the impact proxy.

Sixth, Scopus query result limits cap retrievable results at 5,000 per query. For domain-level searches that exceeded this ceiling, the AF-ID facet was used to identify the top 30 contributing institutions rather than paginating the full result set.

Seventh, the bibliometric searches apply English-language keyword sets uniformly to both countries. Because domain terminology — including in translated abstracts — may appear at different rates across national literatures, domain-level counts may be affected by keyword-language rather than research-volume differences. The uniform query template mitigates but does not eliminate this, and per-domain counts should be read with this caveat.

This study assessed bio-AI/ML convergent capabilities in the U.S. and China across 140 institutions using a four-stage lifecycle framework: source selection, retrieval, three-dimensional composite scoring, and stability-based validation. This was applied to Scopus publication data and Data Abyss security designations covering January 2015 through February 2026. Three findings are carried forward.

First, the empirical baseline established here documents a 2024–2025 crossover: China surpassed the U.S. in combined biotechnology publication volume by 27% in 2025 (6,061 vs 4,771), with the per-domain crossover occurring nearly simultaneously across gene editing, AI + biotechnology, and ML + biotechnology, and with precision medicine on an accelerating convergence trajectory whose extrapolation beyond 2025 is sensitive to the unknown slope of continued growth.

Second, the strategic-relevance dimension reveals a concentration that no purely bibliometric study would surface: 86% of Chinese institutions in the dataset carry Section 1286 designations, 84% appear on the BIS Entity List, and all three Tier-1 positions in the composite ranking are held by Chinese institutions.

Third, the methodological contribution, a reproducible composite-scoring framework with Spearman weight-perturbation stability analysis and an external-validity check, is reusable for periodic re-assessment of bio-AI/ML convergent capability and is fully documented for replication with the input dataset’s SHA-256 hash recorded alongside the analysis code; the implementation is available from the authors upon request.

The national level trajectory analysis presented in this study presents an opportunity for a more granular assessment at the institutional level. Evaluating institutions using the same approach in 4.1 would complement the findings in Figure 13 by characterizing capability development over time through a more nuanced and comprehensive analysis. Future work will extend this framework to institutional level trajectories, enabling systematic examination of capability evolution across institutions over time.

For the test-and-evaluation and DoD data-science community, the lifecycle data-collection approach offers a defensible template for capability-trajectory monitoring across other dual-use domains. The same four-stage structure — source selection with explicit comparability rationale, retrieval with reproducibility logging, multi-dimensional composite scoring with stability assessment, and external validation against ground truth — extends naturally to any domain where publication output and security-designation data co-exist and where periodic re-assessment is required. Future work will operationalize this re-assessment cadence and extend the framework to additional convergent technology domains.

The authors thank Dr. Brian Weiss (outgoing mentor), Dr. Chenyi Ling (mentor), Trevor Casey, Mr. Richard Glaven, Dr. Christopher Boswell, Ms. Lauren Diaz, Dr. Marcus Boyd, Dr. Erin Fitzgerald, Dr. Kelly Avery, and the Institute for Defense Analyses for their support. This work was conducted under the Research for Intelligence & Security Challenges (RISC) program at the Applied Research Laboratory for Intelligence and Security (ARLIS), University of Maryland.

Bostick, Kathleen. April 2026. Accelerators and Decelerators of Bio-AI/ML Convergent Innovation. Draft, Version 3.0.

Bostick, Kathleen and Casey, Trevor. 2025. Convergent Threats Through the Intersection of Biotechnology and AI/ML. Presented at ARLIS RISC Symposium, August 5–August 6, College Park, Marlyand, United States.

Eraslan, G., Avsec, Ž., Gagneur, J., and Theis, F. J. 2019. Deep Learning: New Computational Modelling Techniques for Genomics. Nature Reviews Genetics 20(7): 389–403.

Kim, H. K., Min, S., Song, M., Jung, S., Choi, J. W., Kim, Y., Lee, S., Yoon, S., and Kim, H. H. 2018. Deep Learning Improves Prediction of CRISPR–Cpf1 Guide RNA Activity. Nature Biotechnology 36(3): 239–241.

Long, T. 2025a. China Is Catching Up in R&D—and May Have Already Pulled Ahead. Washington, DC: Information Technology and Innovation Foundation (ITIF). https://itif.org/publications/2025/04/09/china-catching-up-rd-may-have-already-pulled-ahead/ (accessed January 5, 2026).

Long, T. 2025b. America’s Innovation Future Is at Risk Without STEM Growth. Washington, DC: Information Technology and Innovation Foundation (ITIF). https://itif.org/publications/2025/09/10/americas-innovation-future-at-risk-without-stem-growth/ (accessed January 16, 2026).

Pauwels E. and Vidyarthiet A., 2017. Who Will Own The Secrets In Our Genes? A U.S. – China Race in Artificial Intelligence and Genomics. Washington D.C.: Woodrow Wilson International Center for Scholars, p.1- 5.
https://www.nist.gov/system/files/documents/2018/10/19/who_will_own_the_secrets_in_our_genes_woodrow_wilson_center.pdf (accessed December 15, 2025)

UK Research and Innovation (UKRI). 2023. Exscientia: A Clinical Pipeline for AI-Designed Drug Candidates. Swindon, UK: UK Research and Innovation. https://www.ukri.org/who-we-are/how-we-are-doing/research-outcomes-and-impact/bbsrc/exscientia-a-clinical-pipeline-for-ai-designed-drug-candidates/ (accessed November 7, 2025).

U.S. Senate 2025. Final Report: Charting the Future of Biotechnology. Washington, D.C.: National Security Commission on Emerging Biotechnology (NSCEB).
https://www.biotech.senate.gov/wp-content/uploads/2025/10/NSCEB-%E2%80%93-Full-Report-%E2%80%93-Sep-30-.25.pdf (accessed June 24, 2025)

Wang, K., Shen, X., & Williams, R. 2020. Sequencing BGI: the Evolution of Expertise and Research Organisation in the World’s Leading Gene Sequencing Facility. New Genetics and Society, 40(3), 305–330.

Wang M., Duan S., Li X., Yang J., Yuan H., Liu C, He G., 2025. Genome-driven Chinese precision medicine: Biobank-scale Genomic Research as a New Paradigm. The Innovation Life 3:100131.

Weerasinghe R.N., Jayawardane A.K.W., 2024. Huang Q., Critical inquiry on National Innovation System: Does NIS fit with developing countries?, Sustainable Technology and Entrepreneurship 3 (1): 100052.

You, L., Tong, R., Li, M., Liu, Y., Xue, J., & Lu, Y. 2019. Advancements and Obstacles of CRISPR-Cas9 Technology in Translational Research. Molecular therapy. Methods & clinical development, 13, 359–370.

Appendix A. Literature-review inclusion criteria and bibliometric search syntax

This appendix documents the criteria and keywords used in the two data streams of the project: the supporting literature synthesis (Appendix A.1) and the bibliometric Track B Scopus searches (Appendix A.2). The two streams are distinct and were applied to different sources.

A.1 Literature-review inclusion criteria and keywords

The supporting literature synthesis covered 146 articles drawn from PubMed, Scopus, IEEE Xplore, and Wanfang Data. Sources were assessed in three sequential screening phases: preliminary, title-and-abstract, and full-text, in which an explicit inclusion criterion was applied.

A.1.1 Preliminary Screening. A source advanced to the title and abstract screening phase only if it satisfied all of the following:

• English-language, full-text material from 2015 to present;
• Peer-reviewed source describing explicit AI/ML approaches, precision medicine, or structural innovation barriers;
• Source from one of the four databases: Scopus, IEEE Xplore, Wanfang Data, or Data Abyss.

A.1.2 Title and Abstract Screening. A source advanced to the full-text screening phase only if the title or abstract addressed at least one of the following:

• The modularization of one or more biological systems (e.g., CRISPR-Cas9, base or prime editors, zinc-finger nucleases, transcription activator-like effector nucleases [TALENs]);
• The application of AI/ML to precision medicine as a whole;
• Structural innovation barriers, including accelerators such as R&D spending or cross-sector partnerships, and decelerators such as regulatory hurdles or lack of scalable products.

A.1.3 Full-text Screening. A source was selected for inclusion in the literature synthesis if the full text satisfied one or more of the following:

• Discusses the influence of emergent computational or digital technology over a biological platform;
• Includes relevant geopolitical or strategic angle(s) — innovation, future direction, or security implication of a specific biotechnology;
• Discusses the application of AI/ML to precision medicine;
• Demonstrates a correlation between specific structural barriers (institutional, infrastructural, or otherwise) and diminished realization of novel biotech capabilities.

Sources that did not pass the full-text screening were retained, where relevant, as evidence in the limitations or future directions discussion.

A.1.4 Exclusion Criteria. A source was excluded from the literature synthesis if any of the following applied:

• Not available in English;
• Published before 2015;
• Not available in both abstract and full-text form;
• Did not address both AI/ML and precision medicine;
• Did not apply AI/ML in the context of precision medicine;
• Had already been screened in the project’s Summer 2025 literature review (excluded here to avoid double-counting sources across the two review phases; such sources remained available as background where relevant).

A.1.5 Keywords. Keyword sets were constructed per database. Boolean operators AND and OR were applied as appropriate to balance breadth and specificity.

For Scopus and IEEE Xplore, four keyword categories were used in combination:

• AI/ML:AI-assisted, computational, deep learning, digital, neural network;
• Gene editing:base editors, CRISPR, dual-use gene editing, genetic engineering, prime editors, TALENs, diagnostics, biomarkers;
• Biodefense:biological threat, biosecurity, biosurveillance;
• Innovation:bioeconomy, bottlenecks, clinical translation, innovation policy, regulatory hurdles or barriers or oversight.

For Wanfang Data, the search string Artificial intelligence OR Machine learning AND precision medicine was used, with the precision medicine term substituted in turn by each of the following: biomanufacturing, biosensors, diagnostics, gene editing, CRISPR, gene editing + CRISPR, genomic expression, nanoparticles, clinical translation, clinical trial, policy / regulation / oversight, regulatory hurdles or barriers or oversight.

A.2 Bibliometric search syntax (Scopus Track B, B1–B8)

The Track B bibliometric searches that produced the publication-count inputs to 2.3 followed a single query template applied across the four domain keyword sets and two AFFILCOUNTRY values, yielding eight searches (Table 2). The template is:

TITLE-ABS-KEY(“”)

AND AFFILCOUNTRY(“”)

AND PUBYEAR > 2014 AND PUBYEAR < 2027

Table 2. Eight Searches for Bibliometric Search Syntax. are

Each search was executed against the Scopus API with the date filter 2015/1/1–2026/2/28. Per-institution publication counts were extracted via the AF-ID facet, capped at the top 30 institutions per search to remain within the Scopus 5,000-result ceiling (see 5). Full query parameters, response metadata, and SHA-256 hashes of the retrieved result sets are maintained in a reproducibility log alongside the analysis code, available from the authors upon request.