Grand challenges and emergent modes of convergence science

The paper investigates how convergence science—integrating multiple knowledge and disciplinary domains—has evolved within human brain science (HBS), a frontier area with numerous major funding initiatives since 2013 (e.g., US BRAIN Initiative, EU Human Brain Project, Japan Brain/MINDS, China Brain Project). The authors outline the historical context of convergence and large-team science (e.g., Manhattan Project, Human Genome Project) and the socio-cognitive challenges associated with integrating distant disciplines. They motivate HBS as a live testbed to evaluate operational modes of convergence in a contemporary, global, and competitive research ecosystem. The study formalizes two dimensions of cross-domain integration: cognitive (cross-topic via MeSH subject areas) and social (cross-disciplinary via departmental affiliations mapped to CIP). It poses five research questions: RQ1—how to define and measure convergence; RQ2—status and impact of HBS convergence; RQ3—distribution of convergence across subdomains and which discipline–topic combinations are overrepresented; RQ4—temporal and geographic variation in convergence; RQ5—how propensity and citation impact depend on convergence mode and how flagship funding post-2013 relates to shifts in prevalence and impact. The overarching hypothesis is that multiple operational modes of convergence have emerged, including a prevalent cross-topic without cross-discipline "shortcut" mode that may be incentivized by funding pressures.

Theoretical background frames convergence within the broader interdisciplinarity spectrum (Nissani, 1995; Barry et al., 2008), distinguishing cognitive integration of knowledge and social integration via collaboration across disciplines (Wagner, 2011). Classic convergence is characterized by integrating distant epistemic domains and benefits from diversity and collective intelligence (Page, 2008), multi-disciplinary teams (National Research Council, 2014), and recombinant innovation (Fleming, 2001; Fleming & Sorenson, 2004). The literature indicates innovators more successfully exploit local expertise and hedge uncertainty via cross-disciplinary collaboration for distant exploration (Fleming, 2004). Biomedical convergence is increasingly facilitated by techno-informatics capabilities (Yang et al., 2021), mirroring a triple-helix of supply (biology), demand (health), and technological capabilities (techno-informatics) (Petersen et al., 2016). Prior work documents team science growth and its relation to impact (Wuchty et al., 2007), challenges in cross-disciplinary collaboration (Cummings & Kiesler, 2005, 2008), roles of specialists/generalists (Melero & Palomeras, 2015; Haeussler & Sauermann, 2020), mobility (Petersen, 2018), and science policy considerations (Fealing, 2011; National Research Council, 2005, 2014; Roco, 2013).

Data: Integrated publication- and author-level data from Scopus, PubMed, and Scholar Plot (Majeti, 2020), covering 1945–2018, comprising 655,386 publications linked to 9,121 Scopus Author profiles. Article-level variables include publication year y_p, team size k_p, MeSH keywords W_p, and citations c_p (Scopus, counted through late 2019). Author-level variables include academic age T_ap. Regions are assigned from author affiliation locations. Pre- (2009–2013) and post- (2014–2018) periods are used for comparisons. Topical classification (MeSH to SA): Mapped article MeSH descriptors (branches A,B,C,E,F,G,J,L,N) into 6 Subject Area clusters via O_SA(W_p) → SA_p = [1: Psychiatry & Psychology (F); 2: Anatomy & Organisms (A,B); 3: Phenomena & Processes (G); 4: Health (C,N); 5: Techniques & Equipment (E); 6: Technology & Information Science (J,L)]. N_SA,p counts SA categories per article. Disciplinary classification (CIP): Mapped each author's home department (from Scopus profiles, supplemented by web/Scholar Plot) to NCES CIP codes, aggregated into 9 clusters: [1] Neurosciences, [2] Biology, [3] Psychology, [4] Biotech & Genetics, [5] Medical Specialty, [6] Health Sciences, [7] Pathology & Pharmacology, [8] Engineering & Informatics, [9] Chemistry/Physics/Math. Article-level CIP_p is the count vector over coauthors; N_CIP,p counts distinct CIP categories. Scholars are assumed to have a single principal CIP. Regional classification: Count vector R^p over North America (NA), Europe (EU), Australasia (AA), Rest of World. Mono- vs cross-domain assignment: For feature vectors SA_p and CIP_p, operator O_SA(F_p) or O_CIP(F_p) returns X (cross-domain) if multiple categories are present or M (mono-domain) otherwise; O_SA&CIP identifies articles with both X_SA and X_CIP. Diversity measure: For v_p ∈ {SA_p, CIP_p}, compute normalized upper-diagonal outer-product D_p(v,v) = U(v⊗v)/||U(v⊗v)|| capturing weighted dyadic co-occurrences, with off-diagonal mass f_pp = 1 − Tr(D_p) representing Blau-like cross-domain diversity in [0,1). Annual means f̄_v(t) summarize topical and disciplinary diversity by region and configuration. Configurations: Broad (any cross-domain), Neighboring (short-distance interfaces: SA [1]×[2–4]; CIP [1,3]×[2,4–7]), and Distant (long-distance neuro-psycho-medical techno-informatic interfaces: SA [1–4]×[5,6]; CIP [1,3,5]×[4,8]). Citation normalization: For each publication year, compute z_p = (ln(C_p+1) − μ_t)/σ_t leveraging approximate log-normality (Radicchi et al., 2008), yielding stationary N(0,1) z-scores with σ̄ ≈ 1.24 over 1970–2018. Model A (propensity for X): Logistic regression modeling odds log(P(X)/P(M)) = β_0 + β_y y_p + β_x'X, estimating annual growth in P(X) across configurations and modes (X_SA, X_CIP, X_SA&CIP), controlling for confounds (e.g., team size). An indicator I_2014+ tests post-2013 shifts. Model B (impact premium): Author fixed-effects regression of normalized impact z_p = α_a + γ_X_SA I_X_SA + γ_X_CIP I_X_CIP + γ_X_SA&CIP I_X_SA&CIP + β'X + ε, estimated separately for Broad, Neighboring, Distant configurations. Percent citation premium computed as 100·σ·γ for each X indicator. Difference-in-differences estimates δ_X+ quantify pre/post-2013 changes in citation premiums. Controls include log team size, topical breadth (major MeSH count), career age, among others.

• Cross-domain prevalence and impact (RQ2): Cross-domain articles (both SA and CIP) have grown over time; two-category mixtures dominate. Articles with above-average citations exhibit higher cross-domain frequencies than below-average ones, while mono-domain articles show the opposite inequality.
• Diversity trends (RQ4): SA diversity shows a steady upward trend across all regions and configurations. CIP diversity growth is weaker and in some regions declines recently, especially for Broad and Distant in NA and AA; EU shows steady increase in Distant CIP diversity, aligning with HBP’s computational framing.
• Overrepresented discipline–topic interfaces (RQ3): The structure–function nexus (SA 2↔3) is notably overrepresented among articles that are simultaneously cross-discipline and cross-topic (X_CIP ∧ X_SA), via teams combining CIP 1, 2, 4, and 9. Cross-disciplinary teams integrate more topics: on average, mono-disciplinary teams (N_CIP=1) span 2.2 SA, while teams with N_CIP≥3 span 19% more SA overall; within Distant X_SA&CIP configurations, teams with N_CIP≥3 span roughly 32% more SA than mono-disciplinary teams.
• Propensity for convergence modes (Model A, RQ5): Annual growth in P(X_SA) is ~3% per year, outpacing growth in P(X_CIP). Growth in P(X_SA&CIP) is higher for Distant than Neighboring, reflecting necessity of cross-disciplinary expertise over longer epistemic distances. After 2013, likelihoods for X_CIP and X_SA&CIP decline significantly (on the order of −30%).
• Citation premiums (Model B, RQ5): Broad configuration: X_CIP yields ~8.6% citation premium (γ≈0.07, p<0.001); X_SA ~5.9% (γ≈0.05, p<0.001); combining both X_SA&CIP yields ~16% premium (γ≈0.13, p<0.001), suggesting additive benefits. Neighboring: X_SA premium larger (~11%; γ≈0.088), while X_CIP and X_SA&CIP similar to Broad. Distant: X_SA&CIP premium smaller (~5.2%; γ≈0.04, p<0.001); X_SA alone is negative relative to M, indicating that the cross-topic shortcut is counterproductive at long distances; hierarchy favors cross-disciplinary collaboration for distant integration.
• Post-2013 shifts in impact: Difference-in-differences shows reduced citation premiums after 2013 for Broad and Neighboring (e.g., Broad X_SA: pre ~19% vs post ~8%). In Distant, combining both modes retains an advantage over single-mode convergence (DiD positive for X_SA&CIP vs X_SA or X_CIP).
• Qualitative exemplars of Distant convergence: MRI methodological advances; genomics/biotechnology applications; neurally controlled robotics; AI/big data in imaging illustrate typical X_SA&CIP products spanning wide topical coverage and mixed technical/non-technical disciplines.

Findings demonstrate multiple operational modes of convergence with distinct dynamics and payoffs. Cross-topic integration has surged, often without parallel cross-disciplinary collaboration—a shortcut mode that is increasingly prevalent and appears incentivized by competitive funding contexts post-2013. While cross-topic blending can succeed at neighboring epistemic distances, it underperforms for distant integrations where cross-disciplinary team composition is essential, as evidenced by positive premiums for X_CIP and especially for X_SA&CIP. The densification of collaboration networks indicates maturing interfaces conducive to convergence, yet team composition—not merely team size—drives effective integration. The observed decline in both propensity and citation premiums for certain modes after 2013 suggests flagship initiatives may unintentionally steer research toward expedient but suboptimal integration patterns. Policy mechanisms that explicitly value and require cross-disciplinary team configurations could realign incentives, reduce the shortcut tendency, and better realize convergence’s promise for complex, multi-level problems like the brain’s structure–function mapping.

The study introduces a generalizable framework quantifying convergence along topical (MeSH-based) and disciplinary (CIP-based) axes, distinguishing classic convergence (cross-topic plus cross-discipline) from mono-topic/cross-discipline and cross-topic/mono-discipline (shortcut) modes. Applying this to a comprehensive HBS dataset reveals that shortcut convergence is increasingly crowding out classic convergence, despite the latter’s higher impact. Cross-disciplinary collaboration is crucial for integrating across distant epistemic domains, while cross-topic-only strategies perform poorly in those cases. Evidence indicates that post-2013 flagship brain initiatives correlate with increased shortcut patterns and reduced impact premiums for certain convergence modes. The authors provide open data and code, encouraging replication across other fields to assess generality. Future research should refine measurement of multi-domain expertise within individuals, examine causal mechanisms linking funding designs to team assembly behaviors, and test targeted policy interventions that incentivize true cross-disciplinary integration.

The classification assigns each scholar a single principal disciplinary affiliation (CIP), not capturing individuals with significant multi-domain expertise. Departmental mappings rely on Scopus profiles and supplemental web sources, which may introduce annotation errors. Topic classification depends on MeSH descriptors; while expert-assigned and structured, MeSH coverage and orthogonality may vary across subfields and time. Analyses focus on human brain science and may not generalize universally without replication in other domains. Citation-based impact, even normalized, captures only one dimension of research value. Observational models (including DiD around 2013) identify associations but cannot fully establish causality of funding initiatives on convergence behaviors or impact.