Introduction

Synthetic Biology (SynBio) involves designing and constructing biological parts, devices, and systems, as well as the re-design of existing biological systems and organisms for practical purposes. SynBio offers innovative solutions to address global challenges such as renewable energy, climate change, improved medicine discovery, and robust food supply (Guedan et al. 2019; Jatain et al. 2021; Kim et al. 2020; Zhang et al. 2020). Recent progress in SynBio has significantly expanded the strategies, tools, and scales for modifying existing biological systems or creating new systems for various fundamental investigations and applications. In addition, the economic impact of SynBio is also expected to be significant in the coming decades (Jatain et al. 2021; Maina et al. 2022; Wang and Zhang, 2019). However, SynBio is recognized as an emerging technology with dual-use risks, as its increasing maturity and accessibility could lead to various biosafety and biosecurity concerns (Mackelprang et al. 2022; Sun et al. 2022; Wang and Zhang, 2019). Two main behaviors that cause biosafety and biosecurity issues in SynBio research are malicious use (abuse), such as using pathogens for terrorist attacks (Paris, 2022; Pollard, 2003), and misuse that occurs by accident, such as the discharging of synthetic microorganisms along with liquid waste directly into ecosystems (Devos et al. 2021). The potential biosafety and biosecurity risks are becoming more prominent as incidents occur than before (Colussi, 2013; Rager-Zisman and Nevo, 2012).

To address these issues, various countries have implemented regulatory policies. For instance, the UK’s 2012 Roadmap of Synthetic Biology emphasized responsible and safe development, followed by the 2018 Biological Security Strategy, which advocated integrating biosecurity into SynBio regulations (Chen and Zhang, 2020). In 2018, the United States National Academy of Sciences issued Biodefense in the Age of Synthetic Biology, proposing the relevant requirements for the safety review of SynBio research (National Academies of Sciences, 2018). In 2019, the Johns Hopkins Center for Health Security and Tianjin University Center for Biosafety Research and Strategy launched a Track II dialog with the theme of Biosecurity in the Era of SynBio: Meeting the Challenges in China. In this dialog, the potential biosecurity and bioethical risks in SynBio research and application were discussed (Gronvall et al. 2022; Inglesby et al. 2019). Nonetheless, effective legislative regulatory measures have been based on a deeper understanding of SynBio’s dual-use risks.

Due to the difficulty in identifying abusive behavior from a behavioral perspective, the risk discussed in this study is mainly the misuse behavior. In fact, addressing biosafety issues may sometimes mitigate biosecurity risks and vice versa (WHO, 2021). In our model, abuse risk arises mainly during the artificial intelligence (AI) learning process, where unconscious abuse behavior can occur. Recent scientific debates and conspiracy theories about the artificial origin of viruses have increased public awareness of SynBio’s potential (Garry, 2022; Harrison and Sachs 2022a, 2022b). Abuse of SynBio may involve bioterrorism or warfare by state or nonstate players (Li et al. 2021; Wang and Zhang, 2019), the creation of new viruses or pathogens for terrorist purposes (Ho and Duncan, 2005), and crossing the ethical boundaries across the world (Toumey, 2012). However, the risk factors associated with misuse risks of SynBio research are still ambiguous until recent years. SynBio processes are becoming easily available and user-friendly (Gomez-Tatay and Hernandez-Andreu, 2019). Potential risks caused by the uncertainty of advanced technology, such as protocells and synthetic organisms released in the environment, would interfere with biological functions that are still unknown (Gomez-Tatay and Hernandez-Andreu, 2019; Theo Vermeire et al. 2015). The complexity of participating researchers was also individually found to be associated with the misuse of SynBio (National Academies of Sciences, 2018).

One of the major challenges in analyzing potential dual-use risks in SynBio research and development is the lack of suitable quantitative tools (Rycroft et al. 2019; Trump, 2017). Previously, the LEGO® SERIOUS PLAY® tool was used to analyze the risks associated with SynBio research (McLeod et al. 2018). Moreover, existing qualitative analyses primarily focus more on identifying risk factors for SynBio technologies, with little attention paid to the relationships between these risk factors, especially the dynamic changes over time. The rapid advancement of SynBio research urgently calls for more quantitative and sophisticated methodologies to measure the relevant risks and provide practical solutions. Additionally, in laboratories conducting SynBio research, the human subsystem—which is the most effective of the four subsystems for risk analysis (i.e., human, management, equipment, and environment) proposed by Ma et al. (2021)—has often been overlooked (Dickmann et al. 2015; Ma et al. 2021). In fact, scholars have found that users’ motivations and behaviors are highly correlated with risks but difficult to accurately predict (Scholer et al. 2010). Moreover, machine learning, as an effective tool as black boxes to predict biological system behavior (Ghannam and Techtmann, 2021), has been applied in SynBio systematically (Radivojevic et al. 2020). For example, it can study biomolecule interactions of living cells (Akay and Hess, 2019), predict biological sequence activities, and optimize culture conditions (Faulon and Faure, 2021).

In recent years, quantitative methodologies for risk analysis have gradually become available (McCord et al. 2022; Stanley Kaplan, 1981; van Dijk et al. 2021). Given that many serious accidents stem from the long-term effects of systematic issues, system dynamics has emerged as a valuable quantitative approach for understanding the entire development process of accidents from a systematic and dynamic perspective (Feng et al. 2020; Garbolino et al. 2016). Notable examples include the Chernobyl nuclear disaster (Waldmann and Rabes, 1997), the Fukushima nuclear accident (Steinhauser et al. 2014), and the Bhopal gas leak case (Broughton, 2005). In a recent study, Stringfellow et al. (2022) employed a system dynamics approach to model potential strategies for reducing opioid use disorder and overdose deaths in the United States (Stringfellow et al. 2022). System dynamics is particularly valuable for analyzing risks in complex environments, especially for risks in laboratories of SynBio research. Thus, adopting system dynamics for SynBio risk analysis is a prudent choice.

More importantly, establishing an effective governance system for SynBio necessitates a quantitative assessment of all potential dual-use risk factors and their interactions. To achieve this, our study designed a model-based analysis consisting of three rounds of Delphi interviews to explore the risk from the research process of SynBio research from the perspective of human behavior. Through three rounds of Delphi interviews, we identified and prioritized the risk factors involved, with the “safety atmosphere” emerging as the most critical factor. Furthermore, we developed a quantitative system dynamics model to analyze seven scenarios of biotechnology, information technology, and management subsystems. The results indicated that, higher safety requirements and a better safety atmosphere result in a lower error rate. Higher safety awareness and the reasonable storage of hazardous chemicals can significantly improve infrastructure safety. To our knowledge, this study firstly presents a quantitative and dynamic model for dual-use risk analysis of SynBio research, laying a crucial foundation for building a more sophisticated quantitative framework for detailed risk analysis of dual-use SynBio research in the near future.

Literature review

Risk of SynBio

The research process in SynBio is divided into four phases: Design, Build, Test and Learn. The Design phase primarily involves both manual and AI design methods. Asin-Garcia et al. (2020) previously advocated ‘safety-by-design’ approach, suggesting that current risk assessment expertise may not be sufficient for the long term. Risks associated with manual design include the potential for increased horizontal gene transfer, enhanced transmission capabilities of specific pathogens, and the synthesis of new viruses (Pei et al. 2022). AI design is becoming more prevalent in the Design phase (Rives et al. 2021), such as designing intricate synthetic gene circuits that perform complex sensing and response functions (Lopatkin and Collins, 2020). In the Design phase, researchers can access shared knowledge, such as methods for producing reactive COVID-19 (Thao et al. 2020; Xie et al. 2021). The design and production of viruses like smallpox and SARS-Co-2 have sparked debate, as some argue that the open availability of viral information could lead to increased risks and more severe societal harm if not properly managed (Cortegiani et al. 2021).

In the Build phase, risks include the potential leakage of synthetic parts, elements, or even organisms, which can lead to ecosystem disruptions and the loss of the original characteristics of living creatures (Pei et al. 2022; Trump et al. 2020). Additionally, the growing accessibility of SynBio technology and the significantly reduced costs raise concerns about potential abuse risks. Due to the high correlation between the implementation of the construction process and operational ability, some problems exist in laboratories, including a lack of in-depth biosafety education, evaluation standards for biosafety operations, and loose personnel management. Moreover, there are unknown risks that remain unidentified, such as the knockout of noncoding bits, and these may pose unknown risks once released into the ecosystem (DiCarlo et al. 2015).

The Test phase for synthetic organisms enriched in the culture medium by selection or screening has been accelerated in recent years. Traditional intracellular testing has been replaced by cell-free system testing, and low-speed testing has been replaced by high-throughput testing using microfluidic devices (Lawson et al. 2019; Meng and Ellis, 2020; Silverman et al. 2020). These methods have reduced testing costs, making it easier for malicious users to cross the bottleneck of testing technology and engage in unethical behavior. Sanitary conditions may indirectly affect the safety of SynBio laboratory. A cluttered workspace threatens the safety of both employees and visitors. Many small changes could occur and be neglected, such as the movement of instruments, the placement of new equipment, etc.

In the Learn phase, AI tools can assist in learning existing component design rules and propose new synthesis rules (Rives et al. 2021). Given the widespread use of AI tools, including machine learning and deep learning, potential risks during the Learn phase primarily arise from the universality, judgment, and ethics of the new machine-generated rules, as well as the high uncertainty of the resulting products (Abdulkareem and Petersen, 2021; Liel and Zalmanson, 2020; Martin, 2019). However, the reliability of machine-generated rules has yet to be fully established.

Theory of planned behavior

Based on the Theory of reasoned action (TRA), Ajzen and Fishbein proposed the Theory of planned behavior (TPB), which asserts that behavior is influenced by behavior attitudes, subjective normative factors, and perceived behavioral control (Ajzen, 1991). TPB is commonly used to explain the mechanisms of behavior and predict personal actions by analyzing behavioral attitudes, subjective norms, and behavioral intentions (Ajzen and Kruglanski, 2019; Yuriev et al. 2020; Shin et al. 2018).

TPB is widely applied in the risk assessment process. Many studies have shown that the theory can predict and explain behaviors such as consumers’ use of social media for transactions and attitude towards risk, green hotel visit intention, and green consumer behavior of the young (Verma and Chandra, 2018; Taufique and Vaithianathan, 2018). Bae and Chang (2020) utilized TPB to investigate how risk perception related to COVID-19 influenced behavioral intention towards ‘untact’ tourism. In this paper, we explore the misuse and abuse behaviors of some SynBio technology users, which are highly correlated to human behavior mechanisms and largely depend on their intentions, making TPB become an ideal framework for explanation. In this paper, we applied TPB as a theoretical foundation to design the Delphi interview questionnaires and identify the risk factors contributing to the potential misuse and abuse behavior of SynBio technology users.

Methodology

System dynamics

The laboratory scenarios of SynBio generally involve complex interactions among biological components, chemical reactions, environmental factors, and experimental protocols. System dynamics is a modeling approach particularly suited for studying such intricate systems (Sterman, 2000). It significantly aids in comprehending the intricate dynamics and formulating effective policies (Arranz, 2024; Stringfellow et al. 2022). However, systematic risk-oriented quantitative studies on SynBio risk scenarios have yet to be conducted. The distinct advantages of employing system dynamics for risk analysis in Synbio laboratories are outlined as follows.

First, SynBio laboratory scenarios can involve numerous variables and intricate causal relationships (Pei et al. 2022). System dynamics enables us to represent these complexities through interconnected feedback loops and delays, making it easier to capture the dynamic nature of the system. Second, system dynamics models are designed to illustrate the feedback loops and interdependencies among different components in a system (Bouloiz et al. 2013). For SynBio, these interactions can be crucial for understanding how changes in one part of the system can impact other parts over time (Pei et al. 2022). Third, SynBio experiments often involve processes that evolve over time. System dynamics models explicitly account for time delays, accumulation, and changes in rates, which are essential for understanding the temporal behavior of biological systems (Yang et al. 2018). In summary, utilizing system dynamics to study SynBio laboratory scenarios offers a structured and dynamic framework for understanding, analyzing, and designing complex biological systems. It aids researchers in exploring various scenarios, predicting outcomes, and making informed decisions about system design and interventions.

Vensim is a commonly used software for modeling and simulating system dynamics (Loh et al. 2020). In this study, Vensim has been selected for its superior model quality, robust data integration capabilities, flexible distribution options and advanced algorithms. The simulation process can be divided into three phases: identification, modeling, and evaluation. During the identification phase, it is essential to pinpoint relevant risk variables and develop qualitative causal loop diagrams. Following a systematic review in this study, experts were carefully chosen. Three rounds of Delphi online surveys were conducted to identify the SynBio laboratories’ risk factors. In the modeling phase, the formulation of the model (quantitative stock and flow models) was established based on the hypothesis and model testing, simulation, and model calibration. Thereafter scenario analysis is conducted. In the evaluation process, the risk was evaluated, and policy recommendations were provided. The results of scenario analysis and model evaluation are presented in the “Results” section of this study.

Figure 1 shows the simplified and basic versions of a causal loop, which was determined through expert discussions. In the reinforcing loop, the increased frequency of misuse and abuse of SynBio technology in the laboratory decreases the experimental frequency of technology users due to laboratory safety regulations, resulting in a decline in the average experience level of the laboratory. This will worsen the situation and require addressing the trust, training, and reliable systems (Dickmann et al. 2015). Moreover, according to the laboratory management regulations, team members who are responsible for safety cannot leave in time due to the pressure of experimental arrangements, resulting in a decline in the laboratory leaving rate.

Fig. 1: Overview of the risk analysis framework and causal loop example of the SynBio Laboratory.
figure 1

This figure shows that the risk analysis is based on system dynamics and three rounds of Delphi and consists of three steps, including hazard identification, modeling, and evaluation.

Full size image

In the balancing loop, the increased frequency of misuse and abuse will draw the attention of managers, who will then increase the frequency of safety training to correct misuse and abuse in the laboratory (Keckler et al. 2019). Thus, the safety awareness of the team will increase, preventing further expansion of misuse and abuse in the laboratory. This can be regarded as a means of regulation within the system dynamics.

Hazard identification—Delphi interview

In this study, a structured Delphi method was employed with experts in SynBio, biosafety and management to identify the risk factors in a SynBio laboratory. As a typical technique widely used in effective group-based judgment and decision-making (Belton et al. 2019), Delphi is a useful qualitative research technique for collecting data from respondents (Paul et al. 2021). The experts who participated had published articles in top journals such as Nature and Science. The Delphi method is an iterative series of surveying experts until a consensus is reached.

We first conducted a pre-experiment and invited experts from relevant fields to conduct descriptive polishing of the survey questionnaire, making sure that the specific scenarios could be well understood. The terms and definitions were drafted and presented in the pre-experiment process. Kendall’s coefficient of concordance (W) is a useful statistical tool that measures the agreement among experts (Ongbali et al. 2024). During the formal Delphi process, the degree of expert opinion concentration was expressed by the full score rate. Moreover, the coefficient of variation (CV) was used to assess the consistency of the experts’ opinions and determine the end of the Delphi process. The specific calculation formula for the CV and Kendal’s coefficient of concordance is as follows:

$${{\rm{Coefficient}}\; {\rm{of}}\; {\rm{Variation}}}=\frac{{{\rm {SD}}}}{M}$$
(1)
$$W=\frac{12\mathop{\sum }\nolimits_{j-1}^{k}{R}_{j}^{2}-3{b}^{2}{k(k+1)}^{2}}{{b}^{2}k({k}^{2}-1)}$$
(2)

Formula (1) calculates coefficient of variation, in which SD represents standard deviation, M is the mean of the experts’ opinion. Formula (2) shows the agreement of the experts’ opinion, where W is Kendall’s W value, \(b\) is the number of evaluations, \(k\) is the number of indicators, and \({R}_{j}\) is the total rank assigned to the \({j}{{{\rm {th}}}}\) observation indicator (Rezaei et al. 2021). Kendall’s W ranges from 0 (no agreement) to 1 (complete agreement) (Gómez-Polo et al. 2022). According to English and Kernan (1976), Shah and Kalaion (2009), when the CV is <0.8, the survey rounds can be stopped.

The weighted average of experts’ ratings is obtained by calculating the average of the product of experts’ familiarity with the risk scenario and the risk level assessment of the scenario. The experts’ familiarity with the risk scenario is evaluated by the experts by self-assessment, using 7-point Likert scale. By calculating the weighted average of experts’ importance ratings, the results are shown in Table 1. The consensus exercise of data and causal relations in our model is basically led by the scientific committee that is composed of synthetic biologists, laboratory safety management experts, and biosafety experts.

Table 1 The weighted average of experts’ importance ratings.
Full size table

Modeling

Dynamic hypothesis, model testing, and simulation

The model is a continuous-time differential equations model based on system dynamics. System dynamics is a method to design effective policies through understanding the dynamic complexity and the sources of policy resistance. Consequently, we developed a continuous-time differential equations model based on system dynamics. In the context of a SynBio laboratory the research environment was conceptualized as a complex system, delineated by the management boundaries of the laboratory. The endogenous feedback, which changes over time cyclically, is called feedback loops (Apostolopoulos et al. 2018). The model’s dynamic function is enhanced by both reinforcing and balancing loops. The analysis period is tentatively set as 10 years, which is adequate for conducting policy analyses within laboratory contexts (Stringfellow et al. 2022). The entire risk analysis process is situated within a laboratory setting, recognized as the most critical scenario in SynBio research. Furthermore, quantifying malicious behavior to determine evaluation indicators presents significant challenges, particularly since these behaviors are primarily associated with the AI component of our model.

Beyond the intrinsic risks linked to SynBio technology, human factors constitute a substantial segment of risk factors (Zeng et al. 2022). According to an analysis of high-biocontainment laboratory safety accidents based on data from the Federal Select Agent Program (FSAP) and the National Institutes of Health (HIN), human error is the main cause of potential exposure in BSL-3 laboratories. Through three rounds of Delphi interviews, we attempted to obtain the risk factors that exist in the entire Design-Build-Test-Learn (DBTL) framework of SynBio research. A quantifiable stocks and flow model was built based on discussions with SynBio experts and practitioners in experimental safety, with the aim of clarifying the interrelationships among various risk factors. The resulting flow chart consists of five stocks, namely, “Experiment Frequency”, “Safety Awareness”, “Total Experience of SynBio Laboratory”, “Number of Technical Users in the Laboratory”, and “Misuse Frequency of SynBio Laboratory”. Following Ma et al.’s (2021) study, we performed a detailed analysis of the risks involved in the most important risk subsystem (human subsystem and management subsystem) (Ma et al. 2021). The price of the experiment instrument (and its availability) is exogenous, and we would not try to display the dynamic process endogenously.

Model calibration

The model was calibrated using recent accident reports, the details of which are available through the literature and semi-structured interviews. These reports include the time, process and reason for the accidents that occurred in SynBio laboratories in the last decades. The parameters of the model mainly include (i) literature and expert judgment (Delphi), (ii) accident report data, and (iii) formal estimation. To better fit the real history of the accident in the laboratory, all feedback loop strengths and transition rates are estimated formally (see Delphi part). Efforts were made to achieve the best historical replication to validate the model. Because accident report data are not in a uniform format, we must stimulate the model through expert judgment. We have communicated and carried out revisions with experts for two years to achieve a uniform format.

Results

Risk factor identification in DBTL cycle and evaluation

In the Delphi process, the alternative risk factors provided to experts were selected based on the literature (Anderson et al. 2012; Asin-Garcia et al. 2020; Collins, 2014; Keiper and Atanassova, 2020; Krenz et al. 2016; Lambalk, 2019; Le Feuvre and Scrutton, 2018; Meng and Ellis, 2020). To further ensure that there is no missing risk factors, we analyzed the DBTL processes of SynBio research. The subjective and objective risk factors are shown in Fig. 2.

Fig. 2: Source of risk factors.
figure 2

Generic risk structure influencing the risk of SynBio laboratory from DBTL to TPB in system dynamic model.

Full size image

Based on a three-round Delphi survey and semi-structured interviews with 15 experts, we calculated the weights of risk factors based on the familiarity of experts and obtained the importance ranking of risk factors. Kendall’s W is a test for assessing the agreement among experts’ ratings (Schmidt, 1997). The coefficient of variation and Kendall’s concordance were used to ensure the reliability of the results (Kim et al. 2021; Rossmann et al. 2018). Kendall’s W of the consultation was statistically significant (P < 0.01), which shows that the consultation results are scientific and reliable (see Table 2). Moreover, we conducted the statistical test of the coefficient of variation (CV). According to English and Kernan (1976), Shah and Kalaion (2009), as long as the CV is <0.8, there is no need to conduct an additional survey rounds. In our study, the CV of the third-round Delphi range from 0.058 to 0.11. Detailed explanation of risk factors is shown in Table 3.

Table 2 Demographics of participants, concentration, and coordination in each Delphi round.
Full size table
Table 3 Explanation of dual-use risk factors analyzed.
Full size table

Causal loop example of the Synbio research process

In the balance loop, the increased misuse and abuse frequency of synthetic biotechnology in the laboratory will attract the attention of the safety response team. The managers in charge of the security of the laboratory would increase the frequency of safety training to correct the misuse of the laboratory. Thus, the safety awareness of the team will increase, avoiding the further expansion of the misuse and abuse frequency of the laboratory. It can be regarded as a means of regulation in system dynamics.

Model and projections

We assessed the effects of seven risk factors, which have been picked out according to the importance ranking, on the error rate, safety awareness and infrastructure safety of the SynBio laboratory over ten years. Error rate of the SynBio laboratory is the overall number of accidents that occur in the laboratory each year, which is considered as a key indicator of measuring laboratory safety in our article. These risk factors belong mainly to three categories: (i) biotechnology, (ii) information technology, and (iii) management risk factors. The definition and source of the risk factors discussed are presented in Table 3. We adopted different effect bearings for different risk factors, given the unequal impact of each loop in the system. Moreover, we verified the strategy and examined 9 packages of risk factors to identify synergies. The package is a combination of strategies that can be taken to improve the safety of SynBio laboratory, such as lowering the error rate of the laboratory. The simulation lasted ten years, thus fully reflecting the dynamics of the laboratory system.

The seven scenarios we discussed can be divided into three subsystems: the equipment subsystem, the machine learning subsystem, and the management subsystem. Figure 3 illustrates the feedback loops of the system. There are nine risk factors in the equipment subsystem (red): the higher the level of laboratory safety management, the less severe the consequence of errors will be (#15). Better ventilation system safety, storage of hazardous chemicals, safety layout rationality, and wiring rationality would increase the infrastructure safety of the laboratory (#31, #34–36). More experiment frequency would cause more instrument failure due to normal deterioration (#43), thus decreasing the laboratory capacity and infrastructure safety (#32, #38).

Fig. 3: Causal loop of the SynBio Lab.
figure 3

The loop numbers are referred to and discussed in the text.

Full size image

A higher level of machine learning technology can increase laboratory capacity due to its outstanding computing and learning ability (green, #42). There are risk factors and loops related to laboratory management (Fig. 3, blue). Considering the empirical decay rate and instrument failure, the error rate is likely to increase. The misuse and abuse of the technology by users mainly strengthens the relationship between the total experience of the SynBio laboratory and the error rate (#10–16). The objective risk factors (#21–25) reinforce the influence of the number of SynBio technology users and experiment frequency on the total experience and error rate.

Meanwhile, relevant physical tests have been conducted on the system. For example, actual quantities such as several SynBio technology users in the lab, the experiment frequency, and the level of safety awareness have been verified to be non-negative. At the same time, the dimensions were checked using the inspection tools integrated in Vensim with <Check Units>. Further testing involved randomly altering variables to monitor the model’s behavior, ensuring that it behaves realistically under any condition.

In general, the simulation indicates that the experience of the laboratory is over ten years. Through the buildup of experiment records, there is an indication that the frequency of misuse and abuse in the laboratory initially increases, then decreases, and finally fluctuates within a tolerable range of error rate. A better laboratory safety management process ought to be a dynamic feedback balance process. As shown in Fig. 4, the total experience of the laboratory shows an upward trend, with a decrease in the middle stage of the laboratory’s initial design, and this may be caused by the loss of personnel and the fact that the accumulation rate of experience is less than the attenuation rate. The error rate of the SynBio laboratory fluctuates within an allowable range, indicating a state of dynamic balance. The range of fluctuation for the laboratory’s error rate gradually increases over time. Similarly, safety awareness and infrastructure safety undertake a state of dynamic balance, with the safety awareness of SynBio exhibiting more frequent fluctuations.

Fig. 4: Results of the system.
figure 4

The error rate (a), safety awareness (b), and infrastructure safety (c) of the laboratory are shown above. Seven scenarios discussed (hygiene, safety atmosphere, storage of hazardous chemicals, average experience of newcomers, advanced level of machine learning technology, tolerable error rate, and safety requirement) are shown in a different color.

Full size image

The results of our study provide a quantitative analysis of the error rate, safety awareness, and infrastructure safety. The error rate is a measure of the possibility of errors occurring in the laboratory, while safety awareness represents the level of safety consciousness among technology users, with a fluctuation range of 0–100. Infrastructure safety has no defined limit, as it may exceed the threshold under certain conditions. The quantifiable stock and flow model of the SynBio laboratory was constructed, and we used the figures and equations that were elicited through multiple discussions with the experts. After the model validation, we carried out the simulation of the model and the detailed effect of risk factors and packages on error rate, safety awareness and infrastructure safety are shown in Tables 4 and 5, respectively.

Table 4 Effect of risk factors on average of error rate, safety awareness and infrastructure safety.
Full size table
Table 5 Effect of packages on error rate, safety awareness and infrastructure safety.
Full size table

Effects of the equipment subsystem

In our results, sanitary condition, as a part of daily biotechnology management, has little impact on the error rate and infrastructure safety of laboratories. However, laboratory hygiene has a slight impact on the safety awareness of the laboratory. Hazardous chemicals are considered a major concern for laboratory safety (Krenz et al. 2016), and their storage plays a significant role in improving infrastructure safety, as shown in Fig. 4c.

Effects of machine learning subsystem

The construction of SynBio involves four phases: Design (D), Build (B), Test (T), and Learn (L) (i.e., DBTL cycle), which can usually be divided into dry experiments (D and L) and wet experiments (B and T) (HamediRad et al. 2019; Vavricka et al. 2020). The Learn and Design phases incorporate machine learning (ML) and computer-aided design (CAD). We found that the SynBio laboratory’s advanced level of machine learning can affect the infrastructure safety of the laboratory. Higher machine learning capability may lead to lower infrastructure safety (Fig. 4c). AI accelerates research by reducing the number of traditional experiments (such as cloning, sequencing, etc.), while increasing the diversity of experiments conducted in the laboratory (natural product genome and metabolome mining, directed mutations, etc.) (Mullowney et al. 2023; Long et al., 2022). Compared with safety awareness and error rate, the advanced level of machine learning has a more significant impact on infrastructure safety.

Effects of management subsystem

In contrast to biotechnology and machine learning subsystems, there are more management means to intervene in the safety of SynBio. As shown in Fig. 4, we conducted a comprehensive analysis of safety within the SynBio laboratory, focusing on key factors such as the safety atmosphere, average experience of newcomers, tolerable error frequency, and safety requirement of the management subsystem. The mobility of laboratory staff differs from that of enterprise employees, exhibiting greater complexity. Typically, the laboratory staff comprise undergraduates or post-graduates, resulting in a relatively fixed experience level concerning experimental tasks. Nevertheless, the nature of the laboratory work tends to be fragmented, often resulting in limited horizontal communication among team members. Novice technology users may be inclined to pursue experiments out of curiosity (Shariff and Norazahar, 2012), with their inexperience being a crucial factor contributing to human errors. However, the results obtained after modeling and analyzing the system dynamics show that the average experience of newcomers has a negligible effect on the error rate, safety awareness and infrastructure safety.

Our findings reveal that the safety atmosphere has a significant impact on the SynBio laboratory, particularly in stabilizing and reducing the error rate, as well as stabilizing the safety awareness of the laboratory (as shown in Fig. 4a, b). The necessity of promoting a biosafety culture has been established, especially regarding the comprehension of biosafety practices (Munson et al. 2018). The stronger the safety atmosphere, the lower the fluctuation of the error rate of the SynBio laboratory and the lower the average error rate. It can be seen that the safety atmosphere has a stabilizing effect on the overall laboratory error rate. Similarly, a stronger safety atmosphere will also stabilize the safety awareness of the SynBio laboratory, leading to an improvement in overall safety. Additionally, a safe atmosphere can help stabilize infrastructure safety (Fig. 4c). The more stringent the safety requirement of the laboratory, the more stable the laboratory safety in general. The safety requirement has a similar but even more positive impact on the safety awareness and error rate of the laboratory than the safety atmosphere.

Conversely, elevated tolerable error frequencies introduce volatility into the SynBio laboratory’s operational framework. Analyzing tolerable error frequency enables an examination of the laboratory’s risk attitudes. A risk-prone attitude may lead to more deviations from the average error rate and safety awareness.

Effect of combinations of risk factors

As illustrated in Table 6, Package 1 focuses on the equipment subsystem, while Package 3 examines the management subsystem. Package 4 evaluates both the biotechnology and machine learning subsystems, whereas Package 2 presents an example of the pair risk factors from the management subsystem. Packages 5–8 match different risk factors for the management subsystem.

Table 6 Package detail.
Full size table

The results of the package analysis are depicted in Fig. 5. With regard to the laboratory’s error rate, two primary types of effects have been identified. Packages 2, 5, and 8 demonstrate a similar capacity to reduce the vibration amplitude of the error rate. Packages 3, 6, 7, and 9 stabilize the fluctuation of the error rate significantly and reduce its duration, indicating superior performance in mitigating the laboratory’s error rate.

Fig. 5: Results of package of the system.
figure 5

The error rate (a), safety awareness (b), and infrastructure safety (c) of the laboratory are shown above. Nine packages, listed below in Table 2, are shown in different color (packages with little impact are not shown in the figure).

Full size image

Furthermore, Packages 3, 6, 7, and 9 also exhibit enhanced effectiveness in fostering safety awareness within the context of daily laboratory management, thereby contributing to a more stable safety awareness. Notably, there is a peak of infrastructure safety in the sixth year. Among the various packages, Package 9 has a pronounced impact on the infrastructure safety of the system, indicating that the safety atmosphere, storage of hazardous chemicals, and safety requirements would sharply impact infrastructure safety. These analyses show that Package 9 is the best measure to lower the error rate and improve the safety awareness and infrastructure of the laboratory.

The results of the nine packages are listed here, showing that there is no synergistic impact. We conducted an additional pairwise analysis, and no synergies were found. Among all combinations of risk factors, no combination performs much better than Package 9, achieving a maximum increase in infrastructure safety by 80.47%.

Cumulative effects of risk factors and packages

The cumulative effects of risk factors and packages over ten years are shown in Fig. 6. Figure 6a–c shows the pairwise combination results of percentage change of error rate, safety awareness and infrastructure safety. The background of the pink side indicates better performance compared to the baseline scenario (point [0,0]) (see Fig. 6a–c).

Fig. 6: Cumulative effects of strategies.
figure 6

This figure shows the effects of cumulative percentage reduction of error rate (a), safety awareness (b) and infrastructure safety (c). The results for 9 packages of three categories: equipment subsystem (red), machine learning subsystem (green), and management subsystem (blue). The 3D comparison of the package analysis is displayed in the figure (d).

Full size image

All 7 risk factors have rather small effects on cumulative safety awareness and hence indicate no more than a 4% increase. The two greatest effects (Fig. 6b) on safety awareness are safety requirements (3.94%) and safety atmosphere (3.73%). Meanwhile, safety requirements and safety atmosphere reduce the error rate by 9.57% and 8.90%, respectively, whereas every other risk factor reduces the cumulative error rate by no more than 1%.

Packages 3 and 6 (Fig. 6d) have similar cumulative performance. Package 6 included the same risk factors as package 3, except that the tolerable error frequency was removed. Removing the tolerable error frequency means little changes in error rate (−0.39%), safety awareness (0.10%), and infrastructure safety (−0.48%). These two packages are also additive.

Package 9 (Figs. 6d and 5) was the most effective package identified cumulatively. It had a maximum enhancement of 4.77% in safety awareness and 81.34% in infrastructure safety. Meanwhile, Package 9 has the greatest effect on reducing the error rate by −18.69%. However, there are no synergistic impacts, which means these effects are additive.

We also did an additional pairwise analysis, finding no synergies. Among combinations of all 7 risk factors, Package 9 has the best performance.

Discussion

The global laboratory leadership program (GLLP) highlights the imperative of empowering national laboratory systems and enhancing health security, particularly in terms of biosafety (Malik et al. 2022). Swiney et al. (2018) advocated for improving the quality of public debate on SynBio research to better evaluate biotechnology advancements (Swiney et al. 2018). The risk of SynBio research has been discussed in qualitative approaches, such as systematic reviews (Delhove et al. 2020; Elgabry et al. 2020), Lego Serious Play (McLeod et al. 2018), and empirical studies (Swiney, 2020). However, these discussions often neglect the dynamic relationship between risk factors and fail to consider safety systems within SynBio laboratory. To address this gap, we developed a quantitative dynamics system based on the Delphi online survey to analyze the risks and influencing factors in the SynBio research laboratory. By constructing system dynamics models, we explored the relevant risk factors within the three specific laboratory subsystems.

Our findings demonstrate that effective management practices significantly impact misuse and abuse behaviors, aligning with previous research by Chen and Zhang (2020), which emphasizes the importance of daily safety training that cannot be compromised or simplified (Chen and Zhang, 2020). Repeated refresher training has been shown to be an effective strategy for ensuring compliance and promoting a culture of biosafety (Malik et al. 2022). Additionally, Michelotti et al.’s discussion on research performance in high-containment laboratory settings underscores the significance of a highly trained workforce with extensive experience (Michelotti et al. 2018).

Although there is no direct correlation between laboratory sanitary conditions and the error rate of the laboratory, most safety experts have concluded that inadequate daily laboratory management is the primary cause of laboratory accidents. Moreover, sanitary conditions are included in numerous chemical health and safety manuals.

In addition, different attitudes toward risk highlight the necessity for specific management protocols. A risk-prone attitude can facilitate technological progress (Loh et al. 2020). Therefore, laboratory managers should occasionally adjust their risk attitudes to promote better development of the facility. As indicated by our findings and previous research (Honore and Ganco, 2023), the experience of newcomers could play varying roles depending on the situation in the laboratory. Thus, necessitating dynamic management of their integration into the lab environment.

Our research introduces a novel approach to mitigate the risks associated with SynBio, especially in managing biosafety and environmental hazards. The proposed framework aims to enhance laboratory operations and management practices. A positive safety atmosphere has been demonstrated to facilitate risk communication, thereby improving risk perception (Zhang et al. 2022). Moreover, its influence on safety knowledge dissemination and the propagation of unsafe behaviors is also highlighted (You et al. 2019). Our findings indicated that a more systematic and macro-level analysis reveals a significant impact of the safety atmosphere on both error rate and safety awareness within the laboratory, while effective storage of hazardous chemicals outstandingly enhances infrastructure safety.

Moreover, our study showed that employing advanced machine-learning techniques may inadvertently compromise laboratory infrastructure safety. Information technology is increasingly integrated into daily laboratory functions (Aronson et al. 2016; Burger et al. 2020). However, there is still a lack of studies focusing on the potential risks it entails. An explainable neural network designed to predict potentially hazardous chemical reactions has been developed (Kim et al. 2021), illustrating that machine learning is becoming an integral tool for laboratory experimentation. As shown in Fig. 7, misuse of SynBio appears predominantly linked to black box elements within DBTL frameworks, whereas abuse of SynBio seems more closely associated with white box components therein. Analysis of nine packages revealed that Package 4—comprising hygiene protocols, proper storage of hazardous chemicals, and advanced machine learning—significantly bolsters infrastructure safety. The results suggested that enhancements in hazardous chemical storage and overall lab hygiene can effectively counterbalance the adverse effects posed by machine learning on instrument safety.

Fig. 7: Black box risk in SynBio research.
figure 7

This figure shows the relationship between the risk of misuse and abuse faced in the research process of synthetic biology and the black/white box.

Full size image

Package 9 demonstrated a significant impact on safety awareness, error rate, and infrastructure safety. This package encompasses the safety atmosphere, storage of hazardous chemicals, and compliance with safety requirements, primarily representing the management subsystem. Recently, two international guidelines, namely “Responsible life science research for global health security” proposed by the World Health Organization (WHO) and “Tianjin Biosecurity Guidelines for Codes of Conduct for Biological Scientists”, were issued to promote exemplary, safe, secure and responsible life sciences research through an integrated approach. These guidelines have established a global framework for the responsible application of life sciences. Thus, it is anticipated that greater emphasis will be placed on human management in SynBio research moving forward (Mackby and Katakam, 2023).

Our study is situated within laboratory contexts. However, SynBio technology extends beyond SynBio laboratories. For example, the Institute of Do-it-yourself Biology (DIYbio) is a movement of citizen scientists (Seyfried et al. 2014). Notably, documented DIYbio activities are absent in China to date. In addition, it is worth noting that applying system dynamics to the risk analysis in SynBio research presents certain limitations. While system dynamics evaluates risks within SynBio laboratories in specific terms, these risks inherently reflect uncertainty (Istiak and Serletis, 2020). The multidimensionality and dynamic nature of risks—as well as their inherent fuzziness (Pan et al. 2020)—render risk modeling within SynBio laboratories both complex and time-consuming.

Moreover, the diverse behaviors exhibited by individuals within laboratory settings are challenging to characterize accurately due to intricate causal relationships involving experimental personnel and equipment management. Additionally, there may exist biases stemming from expert knowledge selection relevant to this study, all experts consulted were based in China which introduces potential national bias into our findings. The system dynamics model established herein necessitates only ranking risk factors’s importance. Consequently, we did not facilitate separate discussions among experts from diverse backgrounds during our Delphi study process. As a result, the level of consensus among viewpoints did not achieve a particularly high standard.

However, this research employed the coefficient of variation (CV) metrics to evaluate whether expert surveys met adoption criteria. Future investigations should consider incorporating insights from experts across different fields, while categorizing their perspectives accordingly could enhance understanding further. To improve risk analysis methodologies within SynBio laboratories going forward, it would be prudent to gather additional information from the DIYbio researchers as well as those operating internationally. Furthermore, integrating the Bayesian network with system dynamics may help mitigate systematic deviations present in current modeling approaches.