Identifying Scientific and Technical “Unicorns”

The “unicorn” is a legendary creature that has been described since antiquity as a beast with a single large, pointed, spiraling horn projecting from its forehead (Wikipedia, 2020). It is imaged as a mythical. Usually, the white animal generally depicted with the body and head of a horse with long flowing mane and tail and a single often spiraled horn in the middle of the forehead. Eileen lee (2013) published a signed article titled “Welcome to the unicorn club: learning from billion-dollar start-ups” on TechCrunch, in which she introduced this “unicorn” into the economy as a company which indicated of rarity and worth. From then on, the title of “unicorn” from the venture capitalist characterization of private start-up companies quickly become popular in Silicon Valley, which has achieved the almost mythical accomplishment of reaching 1 billion dollars or more valuation in ten years (Casanova, Cornelius, & Dutta, 2018). Gradually, venture capitalists have funded today's “unicorn” high-tech companies such as Uber Technologies Inc., Airbnb Inc., Palantir Technologies Inc., etc. in the United States. Several start-up companies emerge (Cbinsights, 2020) and become “unicorn” in China as well, and prominent examples include Xiaomi Inc. and Didi Chuxing. By January 2019, there are 347 “unicorns,” of which 89 are in China and 174 in the United States, with a total valuation of 1.093 trillion dolllars (Wikipedia, 2020). Applying to the smart city mega-developments, the “unicorn” planning (Cugurullo, Datta, & Shaban, 2016; Rebentisch et al., 2020) has been used to convey the potentially massive profits for high-tech companies, the ambition to instantly invigorate local and regional economies and the idealized expectation of overnight success.

As early as 2007, The National Human Genome Research Institute (NHGRI) endorsed a multi-taxon genome-sequencing initiative termed “unicorn” (Ruiz-Trillo et al., 2007), which generated extensive genomic data from animals to summarize the rationale guiding the choice of organisms. In addition, the term “unicorn” assisted the bi ocuration and validation of structure entries (Akune et al., 2016). It were chosen for its mythical connotations into cardiology (Elbarouni et al., 2017), to discuss the anticipated benefits to the broader scientific and technical community. Researchers increasingly gain insights into the medical, health, and life science sectors. Notably, 2018 has brought unprecedented success for the biotech and health-care industry, with 16 companies earning the title of “unicorn”—a valuation mark of the 1 billion dollars (Wharton et al., 2019). Moreover, accompanying the growing number of health-care “unicorns” in 2019, including Babylon Health, Doctolib, and CMR Surgical, today's “unicorns” have focused on defining interventions and services of patients’ health and wellbeing to achieve the rare and almost mythical accomplishment (The Lancet Digital Health, 2019).

Meanwhile, in informetrics, citation analysis (Garfield, 1979) contributed an effective method for estimating the number of important scientific achievements (Bonaccorsi, 2007; González-Betancor & Dorta-González, 2017). Balancing quality and quantity, the idea of “swan” (Zeng et al., 2017) provided an interesting interpretation of key scientific contributions in science, and the “swan groups” (Zhang, Zuccala, & Ye, 2019) contributed a broader metaphor of remarkable academic achievements in science and social science. In this article, we introduce the concept of informetric “unicorn” as a useful metaphor, which is different from the idea of “swan” and “swan group.” While the “swan” has both qualitative contents and quantitative citations, the “unicorn” focuses on the publications that have a very high impact of quantitative citations only, which may be a useful concept for identifying the rare and worthy works in science and technology.

Citation analysis (Garfield, 1979; Meyer, 2000) as a methodology is always one of the most essential emphases in scientometrics, via its citation-based metrics, by analyzing scientific journals (Garfield, 1972; Guerrero-Bote & Moya-Anegón, 2012; Moed et al., 2012; Silva, 2016), scientific and technical papers (Kuan & Cheng, 2014; Narin, 1994), and authors (Grimwade & Garfield, 2002; Hirsch, 2010; Zou & Peterson, 2016), which reveal intrinsic characteristics, distribution rules, and significant influences. Bornmann and Mutz (2015) analyzed modern science's growth rates by the data for the natural sciences, the medical and health sciences. In the authors’ analysis, Wang et al. (2018) tracked the scientific fame of great scientists in physics and revealed the greatest minds had gone but not forgotten. Meanwhile, for patents, Harhoff et al. (1999) found that the higher an invention's economic value estimate was, the more the patent was subsequently cited. Generally, scientific and technical citation research evaluation (Leydesdorff, 2004; Persson, 1986; Tijssen, 2001) provides theoretical guidance and practical experience for national scientific and technical management and policy. However, the citations to papers and patents grow at different pace and pattern. Covering 1996–2000, Glänzel and Meyer (2003) assessed the frequency and characteristics of papers citing patents. The data source for papers was SCI, and for patents was USPTO, and the result showed that only 1.7% of all papers from SCI contain patent references, most of which were from periodicals. Using 4.8 million US patents and 32 million WoS research articles, Ahmadpoor and Jones (2017) found that about 80% of cited scientific publications (i.e. cited at least once by other scientific journals) eventually link forward to a future patent. Patents directly cited only 10% of cited scientific publications.

While citation analysis penetrates different disciplines, such as economics and management (Laengle et al., 2017; Merigo & Yang, 2017), sociology (Moed, Luwei, & Nederhof, 2002; White, Boell, & Yu, 2009; White, 2015), and biomedicine (Bornmann & Daniel, 2006; Comins & Leydesdorff, 2017; Garfield, 1991), highly-cited analysis shows its unique advantages to generate interesting and meaningful exploration in different discipline areas, where we mention that Bornmann and Leydesdorff (2015) did research in the domain of top-cited papers to investigate the development of the BRICS countries and scientific excellence.

On the one hand, researchers analyzed the content of highly-cited papers in the field of biomedicine. Davis and Cunningham (1990) suggested creative thought in neurosurgical research by citations analysis of journals. Bornmann and Marx (2014) gave information about a researcher's productivity and the impact of their publications based on 10% percentiles of citations, who worked in the natural and life sciences. Moral-Munoz et al. (2018) and Perez-Cabezas et al. (2018) identified and conceptualized microbiology and rheumatology by highly cited papers. Ye et al. (2013) provided new insight into the relationship between Nobel awards and landmark papers in physiology or medicine. In terms of highly cited papers and Nobel Prize-winners’ discoveries are reasonably similar, Rodriguez-Navarro (2016) suggested that the United States’ research success was almost three times that of Europe, which had also published in Nature and Science.

On the other hand, researchers focused on the tendency of highly-cited papers in the field of biomedicine. Garfield (1979) looked at 37 “core” primary journals from 1968–1977 to track the trends in biochemical literature. Based on information extracted from the Science Citation Index database, he found that the biochemical literature was still growing faster than the scientific literature. There is a significant variation of citations for papers in different research fields (Bornmann & Daniel, 2008), e.g. biology papers, on average, receive a larger number of citations than mathematics papers. Ponomarev et al. (2014) studied Medline annual data sets for each of 1995–2004 and found that research fields related to medicine and biology had a stable citation threshold. Boyack et al. (2018) analyzed in-text citations of more than five million articles from PubMed Central Open Access Subset and Elsevier journals. They found that the reference distributions for biomedical and health sciences were more highly cited than other types of papers. However, citation analysis in medical patents is relatively simple. Huang, Zolnoori, and Balls-Berry (2019) analyzed more than 5 million US patent documents between 1995 and 2017, which provided a deep understanding of the focuses and trends of technological innovations in biomedicine.

In summary, serving as a functional linkage between ongoing scientific efforts with prior endeavors (de Solla Price, 1965; Garfiled, Malin, & Small, 1978; Radicchi Fortuno, & Castellano, 2008), citations (especially highly-cited papers) quantify the scholar impact of research, assess the utility of scientific and technical achievements, originate outside of traditional medicine (e.g. medical principles and clinical treatments) and become three enabling biomedical innovation forces.

Therefore, we pay attention to highly-cited scientific papers and technical patents, to explore a new pattern and way to reveal highly-impact achievements.

Essential Science Indicators (ESI; Essential Science Indicators, 2020) is a publication-and-citation-based research analytic tool provided by Clarivate Analytics, which delivers the in-depth coverage you need to effectively evaluate the impact of countries, ranks significant trends and top performers, analyze and benchmark research institutes (Csajbók et al., 2007; Fu et al., 2011; Harzing, 2015). Identifying for WoS-indexed item, the period for ESI counts is ten years. Based on a 10-year rolling file, highly-cited papers (Citation Thresholds, 2020; Highly Cited Papers, 2020) have a clear overview Family form ESI in the Clarivate website, which reflects the top 1% of papers by field and publication year. With consideration of comparative 10-year data for both papers and patents and the definition of “unicorn” in ten years, we chose a 10-year time window in the study.

Although, citations have been inflating over time (Persson, Glänzel, & Danell, 2004), and later papers on average are cited more than earlier ones. Comparing the scientific and technological impact of research, the “unicorn” will have an informetric feature as rarity and highly citations just in the first ten years after its publication, as in Figure 1. Empirically, most scientific discoveries happened before technical invention. Therefore, we set T_p1 as the publishing time of scientific paper and T_p2 as the technical patent's publishing time. As patent applications should do after paper publication, the T_p2 should be the patent's publishing time.

Figure 1

According to the model, the total citations (C_T) of the first 10 years after paper or patent published can be calculated as (1) CT=∫TpTp+10Cp(t)dt, {C_T} = \int_{{T_p}}^{{T_p} + 10} {{C_p}(t)dt} , where T_p is T_p1 or T_p2, and C_p denotes the citation curve of a scientific paper (P_r) or technical patent (P_n) in the publishing year. The Eq. (1) extended discrete citation counting to continuum variable for analysis.

According to our design, C_T should be significantly large, with much more than average level, so that we had to set it as very high citations. As papers received more citations than patents, we also had to differentiate paper and patent. As a result, selecting from the most recent ten years of data, we also proposed an approach as less than 1% of highly-cited papers (top 1% of papers) to quantify how much rarity and worth papers can be considered as “unicorn,” which is to see where science and technology are going and who's leading the way. With consideration of field-independent computation, we applied absolute values as C_T ≥ 5,000 citations for scientific papers and C_T ≥ 500 citations for technical patents in ten years empirically, replacing relative 1% highly-cited ones. Then we would like to introduce the following definitions.

Definition 1. Scientific “unicorn”: a scientific “unicorn” is a publication that received C_T ≥ 5,000 citations in ten years after publication, with an increasing citation curve in the first two years as the start-up emerging.

Definition 2. Technical “unicorn”: a technical “unicorn” is a patent that received C_T ≥ 500 citations in ten years after it published, with an increasing citation curve in the first two years as the start-up emerging.

The definitions are field-independent. When we use these definitions for practical computation, about 50% “unicorns” fall in biomedicine (see Results).

Since we meet an increasing citation curve, it is reasonable to discuss the models of citation curves. Price (1963) established the exponential growth model of scientific and technical publications. However, we knew that rare or important documents incr eased linearly. American science historian and intelligence scientist Rescher (1978) also proposed a hierarchical sliding in dex as the mathematical model (Zhang, Vogeley, & Chen, 2011) by describing the rule of lengthening in scientific and technical publications with introducing a valued index λ in the book of Scientific Progress. When λ=1, it means the entire literature; λ=3/4, it represents meaningful literature; λ=1/2, it means high important Meaningful literature; λ=1/4, it represents very high important literature. (2) CP(t)=(aebt)λ,b>0,λ∈[0,1]. {C_P}\left( t \right) = {\left( {a{e^{bt}}} \right)^\lambda },b > 0,\lambda \in \left[ {0,1} \right]. When λ=0, it means the first class important literatures, and the law of exponential growth is broken. Eq. (2) is defined as linear relation as C_p(t) = lnC₀ + bt. C₀ is the number of publications at the start of statistics. Therefore, we payed attention to the linear model only. (3) Cp(t)=a+bt,b>0, {C_p}\left( t \right) = a + bt,b > 0, where a is the amount at the start point, and C_p(t) is the citations in the t year after publishing, b is the growth coefficient, and t is the time.

In this paper, we processed data with fixed effects to reduce the effect of different publications and then established a linear regression model using programs by Python and R. The ordinary least squares (OLS; Bertoli-Barsotti & Tommaso, 2019) was used. For comparing the error between the linear model and the actual acquisition data, root means squared error (RMSE) was calculated according to the following formula (4) RMSE=∑i=1N(Predictdi−Empiricali)2N, RMSE = \sqrt {{{\sum\nolimits_{i = 1}^N {{{\left( {Predict{d_i} - Empirica{l_i}} \right)}^2}} } \over N}} , in which N is the total number of data.

For measuring the ratio of “unicorn” in a field every year, we introduce the unicorn-ratio (Ur) as an index as follows (5) Ur=PTP×100%, Ur = {{{P_T}} \over P} \times 100\%, where P is the total publications and P_T is the “unicorn” papers in the field. It is expected that Ur is a very low ratio.

The absolute number and the relative percentage of both scientific and technical “unicorns” show rarity, as shown in Table 1.

The results show the proportion of “Biomedical unicorns” from Table 2 in the number of “Absolute unicorn number” from Table 1. There is an apparent disciplinary bias that the ratio of biomedical “unicorns” is respectively 57.58% (95/165) in WoS and 47.32% (106/224) in DII, which means that disciplinary distribution of “unicorns” is asymmetric and biomedical “unicorns” occupied almost 50%.

Table 2 shows the biomedical scientific and technical “unicorns,” where we see that the annual distribution is not average, it changes over time yearly.

Except for the discipline of biomedicine, science and technology “unicorns” in other fields resemble rare. Here we concluded the “unicorns” distribution of other top 5 disciplines in Table 3, where we see that the values are lower than the biomedical “unicorns.”

Individually, most “unicorns” are important leading papers or patents in science or technology, in which we selected representative one case per year for ten years, listed in Tables 4 and 5, respectively.

In Table 4, there are some breakthrough scientific achievements which belong to biomedical “unicorns.” For example, “Initial sequencing and analysis of the human genome” (Lander et al., 2001) is a famous initial paper on the human genome, Takahashi and Yamanaka (2006) reported their discovery of induced pluripotent stem (iPS) cells, and successfully applied to human cells (Takahashi et al., 2007), which led to Shinya Yamanaka's winning the Nobel Prize for Physiology or Medicine in 2012.

In Table 5, most technical biomedical “unicorns” belong to surgical instruments and their appendages, where 91 technical patents are directly or indirectly related to ETHICON ENDO-SURGERY, INC., which indicates that its sewing products are the market leader in the world. Meanwhile, Shelton was a famous inventor in the company. He invented a surgical stapling instrument for the laparoscopic and endoscopic clinical procedure (Hemmelgarn, Setser, & Shelton, 2004), endo-cutter for use during fastening of buttress pads to tissue (Shelton et al., 2006), and a stepper motor with operational modes for surgical cutting and fastening instrument in 2010.

After 2010, the rapid development of computer and network technologies had widely affected the biomedical field. Both “unicorns” in scientific papers and technical patents become much more “technical.” For example, Molecular Evolutionary Genetics Analysis (MEGA) (Tamura et al., 2011) has used statistical methods (Maximum Likelihood, Evolutionary Distance, Maximum Parsimony, Bayesian, and so on) to analyze sequence alignment from 2004 to 2011. The ImageJ (Schneider, Rasband, & Eliceiri, 2012) website got about 7,000 visitors a day, and there were about 1,900 subscribers to the ImageJ mailing list. PHENIX can save significant time and effort, which has provided a comprehensive Python-based system for macromolecular crystallographic structure solution, emphasizing on automation of all procedures instead of traditional performing by hand. Finally, the Pearson correlation coefficient is tested for patent families and citations by SPSS 23, which find r=0.140 and p=0.164 in the confidence interval of 95%. So, the relationship between patent families and citations is almost irrelevant from 2001 to 2010.

According to linear model Eq. (3), we substitute the real yearly values into the regression model for fitting, the linear equation supports more powerful in scientific (red) and technical (blue) “unicorns,” with results are showed in Figure 2.

Figure 2

In t he regression analysis, the statistical sign is significant at p<0.01. All the parameters shown in Table 6.

When we take the data back to the linear mathematical model, we calculate all in the 95% confidence interval, as shown in Table 7. The first ten records of the prediction results are selected and kept in Table 6, and the RMSE of the scientific “unicorn” is 0.2127, while the RMSE of technical “unicorn” is 0.0936.

For any b > 0, according to Eq. (3), we can estimate theoretically (6) CT=∫110(a+bt)dt=at+12bt2|110. {C_T} = \int_1^{10} {\left( {a + bt} \right)dt = at + {1 \over 2}b{t^2}\left| {_1^{10}} \right..} The quadratic curve indicates conic growth, which means that the total citation curve of “unicorns” will be quickly increasing.

Also, we mention the limitations of this research. As this is a purely quantitative study, we do not know the real quality of “unicorns.” Similarly, we do not know whether the company holding “unicorn” patents will necessarily become a “unicorn” company. Comparing with coupled patents (Kuan, Chen, & Huang, 2019), we hope to learn more via patent analysis in the future.

By considering informetric quantity only, we suggest the model for finding scientific unicorn (C_T ≥ 5,000 in 10 years) and technical “unicorns” (C_T ≥ 500 in 10 years), which may be a useful concept for identifying rare and very high impact works in science and technology, particularly in biomedicine.

During 2000–2012, there are 165 scientific “unicorns” in 14,301,875 WoS papers, with ratio 0.0012%, and there are 224 technical “unicorns” in 13,728,950 DII patents, with rate 0.0016%, in which the rate of biomedical “unicorns” are respectively 57.58% in WoS and 47.32% in DII. The rare “unicorns” increased following linear model, the fitting data show 95% confidence with the RMSE of scientific “unicorn” is 0.2127 in WoS while the RMSE of technical “unicorn” is 0.0923 in DII.

Finally, it would be interesting and significant to explore “potential unicorns” on C_T near 5,000 for papers and CT near 500 for patents, which could also belong to remarkable discoveries in scientific and technical fields. The proportion of reduced C_T is less than 10%. We remain “potential unicorns” for future studies.