Background of the study

In 2022, the Japanese government designated the year as the “Startup First Year,” initiating efforts for strengthening the domestic startup ecosystem [1]. A particular focus has been placed on “deep-tech startups,” which integrate advanced and innovative technologies into their business models (for example, BioNTech in Germany and Moderna in the US). Such startups often originate from universities and research institutions, a trend that aligns with the Japanese government’s approach to bolstering its ecosystem with a university-centered approach. The government aims for deep-tech startups to grow into unicorns or decacorns.

However, to implement this policy effectively, one must recognize that the creation and growth trajectories of deep-tech startups differ fundamentally from other startups.

Firstly, deep-tech startups often arise from research outputs—such as intellectual property, data, and prototypes—developed by university or institutional researchers. Launching such startups therefore requires not only entrepreneurs but also the active involvement of these researchers. As Lerner [2] emphasized, innovation is driven not merely by institutional reforms but by effective mechanisms that connect academic research, venture capital, and entrepreneurship, highlighting the pivotal role of researchers in transforming scientific knowledge into economic value. Likewise, Visintin and Pittino [3] demonstrated that university-based spin-offs with founding teams combining academic and managerial expertise outperform those composed solely of academics, underscoring that researchers’ active participation is critical for early-stage growth.

Secondly, for deep-tech startups to achieve sustainable growth, they require longer development cycles and substantial funding from the outset. Unlike lean startup methods [4], the path for deep-tech startups is quite different [5]. For instance, Boston Consulting Group [6] reports that, while deep-tech startups represent a small portion of the overall startup ecosystem, investments in them surged from $15 billion in 2016 to over $60 billion in 2020 [7]. Additionally, these startups now constitute around 20% of venture capital investments, up from about 15% a decade ago [8]. According to EUROPEAN STARTUPS [9], deep-tech startups in Europe, particularly in biotechnology, generally experience longer intervals between funding rounds than other startup types.

In summary, deep-tech startups require researchers, substantial funding and extended development periods from their inception to achieve sustainable growth. Consequently, investing domestic resources in their growth involves higher risk compared to other types of startups, calling for careful consideration in investment decisions. Hence, assessing researchers’ outputs and their profiles at the early stage is crucial for investment decisions.

Deep-tech startups are typically born out of academic research, requiring not only advanced scientific knowledge but also mechanisms that enable effective knowledge transfer from universities to industry. Lockett et al. [10] conceptualized such university-based spin-offs within a knowledge-based view, defining them as independent firms created to commercialize publicly funded research and emphasizing that successful commercialization depends on bridging “knowledge gaps” between scientific and entrepreneurial capabilities. Meyer [11] further highlighted that researchers play diverse roles in the commercialization of academic knowledge, distinguishing between “academic entrepreneurs,” who establish firms to pursue the growth of research-based ventures, and “entrepreneurial academics,” who remain primarily research-oriented while engaging in commercialization. He argued that public support mechanisms can shape these behaviors, emphasizing that managerial expertise and external networks are crucial for successful university spin-offs. Marx and Hsu [12] focused on the phenomenon of similar scientific papers being published in close succession, creating a sample set of “Twin” papers. Their analysis revealed that research outcomes proposed in Twin studies tend to ensure scientific progress and competitive advantage, thereby more easily leading to commercialization by deep-tech startups. Additionally, Goji et al. [13,14] analyzed the relationships between researchers in specific scientific fields and deep-tech startups, finding that central researchers in these fields are more likely to participate in deep-tech startups. Thus, the core research outcomes and the researchers behind them are pivotal for deep-tech startups to successfully secure development funds and achieve continuous growth.

Recognizing that deep-tech startups, like other startups, fundamentally depend on skilled management teams and a supportive ecosystem to thrive, Japan should additionally focus on identifying researchers with research outcomes that can drive such growth and connecting them with entrepreneurial opportunities. However, Japan’s startup ecosystem remains relatively small on a global scale, with venture capital-backed cases in 2021 totaling 1,915—about one-eighth of the United States’ total and approximately two-fifths of China’s [15]. Therefore, to foster more high-growth deep-tech startups within Japan, it is essential and efficient to identify and support researchers and research outcomes with strong potential for growth. As Mazzucato [16] argued, because the state itself can act as a primary risk-taking agent in innovation, this approach is particularly relevant for Japan.

Purpose

Given Japan’s limited number of startups, establishing deep-tech startups involving researchers with outstanding research achievements is an efficient means to generate more unicorns or decacorns. This requires first identifying “researchers with outstanding achievements that contribute to growth” (hereafter referred to as “involved researchers”). It then requires initiating discussions around their precise definition and the development of an identification model. We therefore begin by analyzing actual growth outcomes of Japanese deep-tech startups and the characteristics of their involved researchers. Since prior works by Hsu and Goji rely on global datasets, directly applying their findings to Japan—where research funding environments and startup economics differ—could lead to imprecise assessments. Those studies primarily use publication metrics to evaluate researchers and define growth simply as “startup founded,” which may not align with Japan’s context.

Therefore, in this study we utilize Japan-specific data to analyze the relationship between deep-tech startup growth and the characteristics of involved researchers, and to develop an identification model. We focus on university-originated startups not because we assume systematic differences from other Japanese deep-tech startups, but because reliable, government-maintained datasets exist only for the university-originated subset, making it the sole nationally standardized source for analysis in Japan.

This study will actively incorporate machine learning techniques into the analysis. The datasets of startups and researchers contain multivariate data with various parameters. Machine learning is a powerful tool for analyzing such multivariate data, and its application has recently become more prominent in startup research. For instance, Zbikowski et al. [17] developed a machine learning model to predict business success using the Crunchbase database, which focuses on U.S.-based startups. Additionally, Cao et al. [18] discussed the use of deep learning (DL) to predict startup success, while Qian et al. [19] developed a model linking company information, region, and venture capital funding to successful outcomes, such as acquisitions and IPOs, utilizing machine learning techniques. These examples demonstrate the increasing use of machine learning in startup analysis, and this study aims to develop an efficient identification model using similar techniques. In particular, we apply an AutoML tool to efficiently construct interpretable tree-based classification models.

Building on this foundation, the present study quantitatively analyzes the relationship between the growth of university-originated startups in Japan and the involved researchers, with the goal of constructing an identification model for research outputs and researcher characteristics that contribute to the growth of deep-tech startups. Specifically, we first collect data on the growth performance of Japanese university-originated startups, along with publication records and KAKENHI (Grant-in-Aid for Scientific Research) information of the associated researchers. Next, machine learning techniques are applied to the collected data to develop and evaluate a classification model that estimates startup growth based on the research outputs of the involved researchers. Finally, based on the model evaluation results, we identify researcher characteristics that significantly contribute to the growth of deep-tech startups. These characteristics are then used to compare and analyze the performance of researchers at universities that rank highly in the number of university-originated startups they produce.

Data cleaning

Fig 1 summarizes the data integration and cleaning process used to construct the final analytical sample.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. Relationships Between Each Dataset.

https://doi.org/10.1371/journal.pone.0346137.g001

The “University-Originated Venture Database” contained 967 listings, which were matched with “STARTUP DB” using corporate numbers as the primary identifier. As a result, 357 startups were successfully linked between the two databases. A total of 610 entities were excluded at this stage. This reduction reflects structural differences in database definitions rather than data loss. Specifically, the University-Originated Venture Database maintained by the Ministry of Economy, Trade and Industry (METI) allows the inclusion of a broad range of organizational forms, such as general incorporated associations, and does not require listed entities to have conducted external fundraising. In contrast, STARTUP DB focuses on venture companies with identifiable corporate registration and financial activity, resulting in a narrower and more investment-oriented subset.

To examine funding trends over time, these 357 companies were divided into two groups based on their founding years, and statistical significance was tested. With 2015 identified as a key threshold year (p < 0.05), only startups founded from 2015 onward (194 companies) were selected for analysis, as illustrated in Fig 1. This year corresponds to a major institutional and financial shift initiated by the enforcement of the Industrial Competitiveness Enhancement Act (April 1, 2014) and the launch of the Public-Private Innovation Program by the Ministry of Education, Culture, Sports, Science and Technology (MEXT). Under this program, four major national universities—the University of Tokyo, Kyoto University, Osaka University, and Tohoku University—each established university-affiliated venture capital funds with a total investment capacity of approximately 100 billion JPY, beginning in FY 2014. This policy context does not imply that our analytical sample was restricted to these four universities; the startups in the final dataset originated from a broader range of universities. Therefore, the post-2015 subset represents startups founded under this new policy and funding environment, ensuring analytical consistency within a unified institutional context [21].

Among the 194 companies analyzed, a total of 252 involved researchers were identified, indicating that some startups are associated with more than one researcher. For model construction and evaluation, we did not aggregate researcher-level metrics within a startup (e.g., by simple or weighted averages); instead, we created a one-to-one startup–researcher dataset by selecting one researcher per startup via repeated random sampling, as described below.

Target variables

We classified university-originated startups into Positive (growing) and Negative (not growing) groups using total funding amounts from STARTUP DB. To achieve this, we linked the University-Originated Venture Database with STARTUP DB.

Additionally, to refine our identification, we filtered startups by the main technology field of their core product or service as listed in the Venture Database. The thresholds for funding amounts and technology fields are summarized in Table 1.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Example of table.

https://doi.org/10.1371/journal.pone.0346137.t001

We imposed a time cutoff to ensure that only information available by the startup’s founding year was used. For KAKENHI, we collected projects up to 10 years before foundation. For publications, we retrieved two datasets: (i) all papers before the founding year and (ii) papers from the five years immediately preceding foundation to reduce bias from cumulative achievements.

For certain predictor variables in Table 2 that are not directly output from databases, calculations were made post-data retrieval. Specifically, variables related to research delegation, contracted research, collaborative research in the KAKEN data, and co-author information in the publication data were compiled into separate “Collaborative Research Data” and “Co-Author Data” sets. In the “Collaborative Research Data,” peers who collaborated on the same project (hereafter, collaborative researchers) were extracted from KAKEN, and co-authors were identified from Scopus, ensuring no duplication with the primary researcher. The “Co-Author Data” included the frequency of co-authorship per pair, and the “Collaborative Research Data” recorded the project counts and total research budgets, categorized by project role (principal researcher, collaborative researcher, or otherwise). Total counts of co-authors and collaborative researchers were then calculated for each primary researcher.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 2. Predictor variables.

https://doi.org/10.1371/journal.pone.0346137.t002

Predictor variables

Predictor variables consisted of researchers’ publication metrics (e.g., citation counts, JIF scores) from Scopus and their KAKENHI funding information (e.g., total budget, number of projects). Table 2 lists all selected variables.

Construction and evaluation method of the machine learning model

We used the Tree-Based Pipeline Optimization Tool (TPOT) to build a classification model relating the target variables (Table 1) to the predictor variables (Table 2). TPOT is an open-source AutoML tool that automates pipeline development [22]. We selected TPOT because it is open source—facilitating easy setup of our research environment—and because it integrates seamlessly with Python’s scikit-learn library. While other AutoML tools may outperform TPOT in some respects, TPOT offers a good balance between cost-effectiveness and system requirements.

TPOT’s flexibility in choosing interpretable tree-based algorithms is a key advantage. Although deep learning models can achieve high accuracy, they often obscure the relationships between target and predictor variables. TPOT can incorporate deep learning methods, but this study prioritized interpretability and thus limited algorithm selection to the tree-based methods listed in Table 3.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 3. Example of table.

https://doi.org/10.1371/journal.pone.0346137.t003

To construct a consistent one-to-one dataset for modeling, one researcher was randomly selected per startup. This random sampling was repeated five times, and the average model performance across these iterations was used to evaluate the model’s stability. In each iteration, the sampled dataset contains at most one researcher per startup; therefore, no within-startup averaging of researcher metrics was performed. This approach ensured that model evaluation did not depend on any particular random draw of researchers from startups with multiple participants.

We evaluated model performance using 5-fold cross-validation: in each fold, the model was trained on 80% of the data and evaluated on the held-out 20%, and we report the mean of the evaluation metrics across folds. We report F1 score and ROC-AUC as the primary complementary metrics, while Accuracy is provided to support interpretability. To mitigate class imbalance, we downsampled the larger group to match the smaller group’s size via random sampling. This adjustment was applied in Cases 1, 2, and 4, as the number of Negative (non-growing) startups exceeded that of Positive (growing) ones, as indicated in the “Number of Verification Data/ Remaining Data” row in Table 1. Feature importance scores were computed to assess each predictor variable’s influence.

Model evaluation results

Following the criteria in Cases 1–4 outlined in Table 1, four classification models were constructed and evaluated for accuracy using TPOT. The model performance metrics for Cases 1–4 are presented in Table 4, feature importance scores are illustrated in Fig 2.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 4. Model performance metrics for Cases 1 to 4.

https://doi.org/10.1371/journal.pone.0346137.t004

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. Feature Importance of Models Built in Cases 1 to 4.

https://doi.org/10.1371/journal.pone.0346137.g002

As shown in Table 4, the model’s performance improved as the distinction between the compared groups became clearer. Accuracy and F1 score increased when the funding threshold was raised from Case 1 to Case 2, indicating that a larger gap in funding scale enhanced the model’s ability to discriminate between growing and non-growing startups. Furthermore, when the dataset was restricted to specific technology-field subsets (Cases 3 and 4; biomedical vs. Non-Biomedical), both accuracy and AUC were higher than in the non-restricted cases, suggesting that analyzing more homogeneous technology domains contributed to improved classification performance. However, this binary split should be interpreted as a pragmatic, exploratory adjustment under limited sample size and should not be regarded as a statistically sufficient control for broader sectoral heterogeneity; future work with larger samples should incorporate more granular technology-sector controls.

Regarding feature importance, Fig 2 shows that a few predictor variables consistently recorded high scores across all cases. This indicates that a limited number of predictor variables predominantly influence the model’s classification decisions.

Creation of breakdown trees from the machine learning model

Based on the feature importance results shown in Fig 2, both commonalities and differences were observed in the predictor variables emphasized by the model across each case. It should be noted that the feature importance scores presented here represent statistical associations identified by the model and do not imply direct causal relationships between researcher characteristics and startup growth. Therefore, we discuss these variables as factors that most influence the classifier’s decisions (feature importance), not as causal determinants of startup growth.

In all cases, variables derived from KAKENHI data, particularly Funding Amount, Maximum Funding Amount, Average Funding Amount, and Number of Research Projects, consistently exhibited high importance scores. These findings suggest that the scale and continuity of research funding are fundamental factors for distinguishing growing startups.

In addition to these funding-related variables, publication-related metrics—especially quality-oriented indicators such as Max JIF (5 years) and Avg JIF (5 years), alongside citation counts and h-index—also contributed to the models, with the prominence of these quality-oriented indicators becoming more apparent in the field-stratified analyses (Cases 3 and 4).

Fig 3 presents examples of violin plots for predictor variables, showing that Positive startups tended to exhibit higher values than Negative ones, particularly in variables related to research funding and publication quality. This visualization indicates that researchers associated with Positive startups generally possess both stronger funding records and higher-quality publication profiles.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3. Examples of violin plots for predictor variables.

https://doi.org/10.1371/journal.pone.0346137.g003

These variables also showed strong mutual correlations, as indicated by the Spearman correlation coefficients, and their detailed relationships are illustrated in Fig 4 (Correlation Matrix). Because Funding Amount, Maximum Funding Amount, and Average Funding Amount are closely related measures derived from the same underlying KAKENHI funding records, the strong correlations among these variables are expected and primarily indicate redundancy/overlap among related funding indicators (i.e., potential multicollinearity), rather than any “quality” implication. Publication-related metrics also showed relatively high internal correlations; in particular, citation- and journal-based indicators (e.g., Max/Avg JIF (5 years), JCI-related variables, and h-index) are conceptually related measures of research excellence, so stronger correlations among them are expected and should be interpreted as overlap and potential multicollinearity rather than as evidence of “quality.” These indicators were less strongly correlated with quantity-oriented measures such as the number of papers or number of co-authors.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 4. Correlation Matrix.

https://doi.org/10.1371/journal.pone.0346137.g004

Focusing on these results, breakdown trees were created for each case, as shown in Fig 5, to analyze the internal structure of the model. To enhance interpretability, all cases were examined with particular attention to Funding Amount and Max JIF (5 years), which consistently appeared among the top predictor variables. For the breakdown-tree visualization only, when multiple researchers were associated with a single startup, we selected a single representative researcher per startup to facilitate interpretation of the tree paths. Specifically, we chose the researcher with larger values for the Positive class and the researcher with smaller values for the Negative class to create a clear one-to-one mapping for illustrative purposes. This heuristic was used solely to improve interpretability of the breakdown trees and was not used for model training or performance evaluation, which relied on repeated random sampling as described above.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 5. Breakdown trees in Cases 1 to 4.

https://doi.org/10.1371/journal.pone.0346137.g005

In Case 1, the initial split was on Funding Amount: of 49 cases with ≥10 million JPY, 45 (≈92%) were Positive. The next split used Max JIF (5 years): among the remaining, 41 of 45 with MAX JIF ≥ 10 were Positive. The rest were classified by more complex combinations.

In Case 2, Funding Amount was again the primary discriminator: researchers associated with startups raising more than 15 million JPY were classified as Positive in ≈ 89% of cases. The next most influential variable was Max JIF (5 years) ≥ 13, identifying a further subset of Positive entries.

In Case 3, corresponding to the biomedical domain, the initial split was based on Max JIF (5 years) ≥ 20, which correctly classified the majority of Positive entries (≈ 82%). Among the remaining data, Funding Amount > 20 million JPY further identified additional Positives, indicating that both publication quality and research funding jointly contributed to classification in this field.

In Case 4, representing the Non-Biomedical domain, the first split used Funding Amount > 37.5 million JPY, followed by Max JIF (5 years) > 20. Researchers satisfying both conditions were classified as Positive in about 80% of cases, highlighting the combined effect of substantial research funding and high-impact publications.

In summary, across all cases, the model performed stepwise classification primarily based on quantitative variables related to Funding Amount and Max JIF (5 years), which were consistently defined as the first or second split in every case. Within the model framework, these two variables were central to predicting startup growth (i.e., to the model’s classification decisions). Such threshold-based separations also enhanced the interpretability and transparency of the decision process.

Logistic regression analysis for interpretation

To support interpretation of the machine-learning results, we additionally conducted logistic regression analyses using the binary growth labels defined for each case (Table 1). This complementary analysis was performed to assess the statistical significance of key predictor variables identified by the ML models.

Model specification and evaluation procedure

For each analysis case (defined below), the target variable was the binary startup growth label as defined in Table 1 (1 = positive/growing; 0 = negative/non-growing). The set of predictor variables followed the definitions described in Section 3.3. To keep this complementary analysis consistent with the ML interpretation, we screened candidate predictor variables based on the feature-importance ranking (Fig 2) and focused on a parsimonious subset of up to three predictor variables. To mitigate multicollinearity, we further removed highly correlated variables by referring to the correlation structure shown in Fig 4; therefore, the final number of predictor variables could differ across cases, especially in field-stratified subsets.

Because the dataset construction involved random selection steps (e.g., when multiple researcher profiles could be linked to a startup), we adopted a repeated sampling strategy to assess robustness. Specifically, we repeated the sampling-and-estimation process R = 100 times. Within each repetition, we performed 5-fold cross-validation to evaluate predictive performance and to reduce sensitivity to a single train/test split. The primary performance metric was the area under the receiver operating characteristic curve (AUC). We report the distribution of AUC across repetitions using the median and interquartile range (IQR), as well as the 2.5th and 97.5th percentiles.

Within each cross-validation fold, continuous predictor variables were standardized (z-score) using statistics estimated from the training data and then applied to the corresponding test data to avoid information leakage. In addition, to assess statistical significance for the selected predictor variables, we fitted a logistic regression model to the dataset used within each repetition (after the sampling step) and report odds ratios (ORs) with 95% confidence intervals and p-values. We summarize the OR estimates across repetitions using the median, IQR, and the 2.5th–97.5th percentile range.

Results of complementary logistic regression analyses

Following the case definitions in Table 1, we conducted complementary logistic regression analyses using the same binary target variable as in the ML classification. Model performance was evaluated under repeated sampling (R = 100) with 5-fold cross-validation, and the ROC–AUC distribution was summarized using the median, interquartile range (IQR), and the 2.5th–97.5th percentile range.

Overall, logistic regression achieved moderate discrimination in the all-field analyses (Cases 1–2), while performance decreased in the Biomedical-stratified analysis (Case 3) and remained moderate in the Non-Biomedical-stratified analysis (Case 4) (Table 5). This indicates that part of the separability captured by the ML models can be reproduced by an interpretable linear log-odds model, but that predictive capacity differs across cases and modeling approaches. In particular, the ML models showed higher discrimination in the field-stratified settings (Table 4), suggesting that non-linear relationships and/or interactions captured by the ML approach contribute to improved classification performance beyond a linear specification.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 5. Logistic regression performance (AUC) across repeated sampling (R = 100).

https://doi.org/10.1371/journal.pone.0346137.t005

To facilitate interpretability and to align this analysis with the ML interpretation, we focused on a reduced set of predictor variables that were emphasized by the ML feature-importance results (Fig 2), while considering the correlation structure among researcher indicators (Fig 4) when finalizing the predictor variable set. We report odds ratios (ORs) with 95% confidence intervals and summarize their distributions across repetitions (Table 6). Across cases, predictor variables reflecting research excellence (e.g., publication-quality metrics such as JIF-related indicators, citation-based indicators, and h-index) tended to show positive associations with the probability of being classified as Positive, consistent with the ML models’ emphasis on a limited number of high-importance variables.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 6. Logistic regression results (odds ratios) across repeated sampling (R = 100).

https://doi.org/10.1371/journal.pone.0346137.t006

However, statistical significance was not uniformly stable across repetitions for all predictor variables. This is expected given the limited effective sample sizes (especially in stratified cases) and the multicollinearity among indicators that capture similar conceptual dimensions of researcher excellence (Fig 4). Accordingly, we interpret the logistic regression results as complementary evidence of association—providing effect-size summaries and a robustness check on directional consistency—rather than as estimates of causal effects.

Definition of researcher characteristics associated with startup growth

The breakdown tree and feature-importance analyses (Cases 1–4) indicated that two major groups of variables characterize researchers associated with startup growth: (1) research funding–related factors, particularly Funding Amount, Maximum Funding Amount, and Average Funding Amount derived from KAKENHI data; and (2) publication-quality indicators, such as Max/Avg JIF (5 years), JCI-related variables, and h-index.

These variables consistently ranked highest in the model and appeared as the primary splitting factors in the breakdown trees. Based on these findings, we discuss the researcher characteristics that contribute to startup growth, focusing on Funding Amount and publication-quality metrics.

Among the KAKENHI-related variables, Funding Amount had the highest feature importance. This variable refers to the cumulative amount of KAKENHI received as a principal investigator, and it was found that cumulative funding of over 10 million yen is a strong positive characteristic. In contrast, other funding-related variables, such as the number of co-investigators or the amount of commissioned research funding, had minimal impact on the classification model. These results suggest that experience in leading research projects—as evidenced by principal investigator achievements—is particularly emphasized.

As for publication-related variables, those representing the quality of research outputs—such as Max/Avg JIF (5 years), JCI-related indicators, and h-index—were found to be important across all case analyses. In particular, these quality-oriented variables contributed notably to classification performance in Cases 3 and 4, where the data were stratified by technology field. This tendency likely reflects the generally higher impact levels of journals in biomedical research compared with Non-Biomedical areas. In Cases 1 and 2, where no field distinction was made, publication-quality variables showed weaker discriminative value. This may reflect cross-field heterogeneity in citation practices, suggesting that field-normalized indicators (e.g., field-normalized citation rates) could be more appropriate when modelling publication-related metrics in mixed-field samples. However, when the analysis was stratified by field—as in Cases 3 and 4—these indicators became more discriminative and effective as metrics. Based on these insights, JIF can be regarded as a representative indicator for defining approximate thresholds: ≥ 20 for the biomedical field and ≥ 10 for Non-Biomedical fields. The threshold for the biomedical field is derived from the results of Case 3, while the threshold of JIF ≥ 10 for Non-Biomedical fields reflects the original data scope of Case 4, which only included papers published in journals above that level.

In summary, key characteristics of researchers associated with the growth of university-originated startups include securing a cumulative Funding Amount of at least 10 million yen as a principal investigator, and/or being the author of papers with JIF ≥ 20 in the biomedical field or JIF ≥ 10 in the Non-Biomedical field. These characteristics suggest that researchers with highly regarded academic outputs are more likely to contribute to startup growth.

Identification of researcher performance at universities that rank highly in the number of university-originated startups they produce in Japan

Based on these characteristics, we examined the research performance of top six universities identified by the Ministry of Economy, Trade and Industry’s “Survey on the Status of University-Based Ventures” [23]: The University of Tokyo, Keio University, Kyoto University, Osaka University, University of Tsukuba, and Tohoku University. This university-level assessment is intended as a supplementary application of the identified characteristics and is conceptually separate from the policy-context description in Section 3, where four major national universities were mentioned as an illustrative example of the Public-Private Innovation Program. Accordingly, the mention of those four universities does not imply that the startup-level analytical sample was restricted to any specific subset of universities.

Fig 6 (a) shows the distribution of cumulative KAKENHI grants (threshold: 10 million JPY) for our model dataset. Overall, about 25% of researchers exceeded 10 million JPY. In the biomedical and healthcare fields, roughly 30% passed the threshold. Among researchers from the top six universities, this proportion increased to around 40%.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 6. KAKENHI grant amounts distribution.

https://doi.org/10.1371/journal.pone.0346137.g006

Next, we extracted data from “KAKEN” for researchers at the top six universities who received Section I (Biomedical & Healthcare) awards in FY 2024. The results, shown in Fig 6 (b), indicate that approximately 70% of these researchers had received grant amounts exceeding the threshold, revealing a large gap compared to the model dataset.

Fig 7 (a) shows the distribution of JIF values (thresholds: 10 and 20) by technological domain for the dataset used in model construction. Overall, papers with JIF < 10 constituted the majority, particularly in AI & IoT, Robotics, and Electronics. In contrast, fields such as Biomedical and Environment & Energy had more papers with JIF ≥ 10. As a representative example from a Non-Biomedical domain, we extracted JIF values for researchers at the top six universities whose 2024 KAKENHI awards fell under Materials categories (Medium Category: 18–19 and 26–30). The lower panel of Fig 7 (b) indicates that, in this Materials subset, JIF ≥ 10 again accounted for the majority—demonstrating a distribution distinct from the model dataset.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 7. JIF Distribution.

https://doi.org/10.1371/journal.pone.0346137.g007