Bruno Kessler Foundation, Trento, Italy

Department of Mathematics, University of Trento, Trento, Italy

Laboratory for the Modeling of Biological and Socio-technical Systems, Northeastern University, Boston 02115, MA, USA

Computational Epidemiology Laboratory, Institute for Scientific Interchange (ISI), Torino, Italy

Institute for Quantitative Social Sciences at Harvard University, Cambridge, MA 02138, USA

Abstract

Background

The recent work on the modified H5N1 has stirred an intense debate on the risk associated with the accidental release from biosafety laboratory of potential pandemic pathogens. Here, we assess the risk that the accidental escape of a novel transmissible influenza strain would not be contained in the local community.

Methods

We develop here a detailed agent-based model that specifically considers laboratory workers and their contacts in microsimulations of the epidemic onset. We consider the following non-pharmaceutical interventions: isolation of the laboratory, laboratory workers’ household quarantine, contact tracing of cases and subsequent household quarantine of identified secondary cases, and school and workplace closure both preventive and reactive.

Results

Model simulations suggest that there is a non-negligible probability (5% to 15%), strongly dependent on reproduction number and probability of developing clinical symptoms, that the escape event is not detected at all. We find that the containment depends on the timely implementation of non-pharmaceutical interventions and contact tracing and it may be effective (>90% probability per event) only for pathogens with moderate transmissibility (reproductive number no larger than R_{0} = 1.5). Containment depends on population density and structure as well, with a probability of giving rise to a global event that is three to five times lower in rural areas.

Conclusions

Results suggest that controllability of escape events is not guaranteed and, given the rapid increase of biosafety laboratories worldwide, this poses a serious threat to human health. Our findings may be relevant to policy makers when designing adequate preparedness plans and may have important implications for determining the location of new biosafety laboratories worldwide.

Background

The risk associated with the accidental laboratory escape of potential pandemic pathogens is under the magnifying lens of research and policy making communities

Here, we perform a quantitative analysis of (accidental) post-release scenarios from a BSL facility, focusing on the likelihood of containment of the accidental release event. Although BSL 4 agents, such as Ebola virus and Marburg virus, are considered the most dangerous to handle because of the often fatal outcome of the disease, they are unlikely to generate global risk because of their inefficient mechanism of person-to-person transmission and other features of the natural history of the induced diseases _{0}. We analyzed different scenarios by assuming transmissibility comparable to that observed in past influenza pandemics, for example, the 2009 H1N1 virus (namely R_{0} or effective transmissibility in the range 1.2 to 1.6 _{0} = 1.8 or higher _{0} of SARS was estimated to be slightly larger than that of influenza, namely in the range 2 to 3 _{0} (in the range of 5 to 10 _{0} (about 1.5

Methods

In order to provide a quantitative assessment of the containment likelihood and the detailed modeling of interventions we used a stochastic microsimulation model structurally similar to the one used elsewhere (see _{0} of a modified influenza strain (typical values for past influenza pandemics are in the range 1.3 to 2 _{0} varying from 1.1 to 2.5, accounting for the possible larger transmissibility of the modified virus with respect to past influenza viruses. The resulting doubling time of simulations without intervention is shown in Figure

**Supporting material.** Model details, additional results.

Click here for file

Doubling time

**Doubling time.** Average doubling time (dots) and 95% CI (vertical lines) as a function of R_{0}. For each value of R_{0} results were obtained by analyzing 100 uncontrolled (no intervention) simulated epidemics.

Once the initial conditions for the outbreak were set the model generated stochastic ensemble estimates of the unfolding of the epidemic. The infection transmission chain can be analyzed at the level of each single individual and all the microscopic details of the progression of the epidemic in the population can be accessed for each stochastic realization of the escape event. The escape events were identically initialized in a BSL facility in the Netherlands (see Figure

Study area

**Study area.** The map shows population density of the Netherlands (colors from yellow to dark brown indicate increasing densities, from 1 to 3,500 inhabitants per km^{2}), the location of the laboratory in a randomly chosen simulation (in Rotterdam, red point), the location of the workers houses (blue points), the location of workplaces and schools attended by household members of laboratory workers (green). Black concentric circles indicate distances of 10 km, 20 km, 30 km from the laboratory. The inset shows the probability of commuting to (at) a certain distance by laboratory workers.

We assumed the warning to be issued at the time T_{w} corresponding to the first identification of one of the initial cases. Two key parameters determine the efficacy of subsequent interventions: the first one is probability (P_{c}) of identifying initial infections, which is related to the virus specific probability of developing clinical symptoms and the probability of individuals to be actually concerned and report their health status. The second one is the time (T_{i}) required to link the initial infections to an accidental release of the modified influenza strain in the laboratory (and not, for instance, to other circulating seasonal influenza viruses) and to activate the containment interventions.

Once the PPP escape event has been detected we considered the following set of containment interventions: (i) isolation of the laboratory, (ii) laboratory workers’ household quarantine, (iii) contact tracing of cases and subsequent household quarantine of identified secondary cases, (iv) school and workplace closure both preventive, on a spatial basis, at the very beginning of the epidemic, and reactive during the entire epidemic.

For contact tracing, we assumed that once one case is detected, infected close contacts (that is household, school and workplace contacts) of the case are detected with probability P_{c} and can transmit the infection for a certain time (T_{t}) before isolation and household quarantine. Cases generated through random contacts in the general population are detected with lower probability (P_{g}). We also assume that undetected cases may self-report their health status with a certain probability (P_{r}). Parameters characterizing interventions along with reference values and explored ranges are described in Table

**Variable**

**Description**

**Reference (range)**

P_{c}

Infected close contacts detection probability

0.6 (0.4 to 1)

P_{g}

Infected random contacts detection probability

P_{c} × 0.5 (0.1 to 1)

P_{r}

Infected random contacts self-reporting probability

P_{g} × 0.8 (0.5 to 1)

T_{i}

Delay from initial warning to intervention

3 (0 to 30) days

T_{t}

Delay from case detection to household quarantine

1 (0 to 4) days

T_{p}

Duration of schools and workplaces closure

21 (0, 7, 14, 21, 28) days

D_{p}

Radius for schools and workplaces closure

30 (0, 5, 10, 20, 30, 50, 50>) km

F_{s}

Fraction of closed schools

0.9 (0 to 0.9)

F_{w}

Fraction of closed workplaces

0 (0 to 0.5)

Contact tracing

**Contact tracing. (A)** Probabilities of detecting first and second generation cases (the latter conditioned to the detection of first generation cases) triggered by a traced index case. **(B)** Example of network of cases triggered by the initial infected laboratory worker (undetected in this example; the initial warning is triggered by a secondary case in the laboratory), and probability of case detection at time of intervention (T_{w} + T_{i}).

Results and discussion

Below we discuss the likelihood that the escape of PPP virus will spread into the local population and the ensuing outbreak will be contained by non-pharmaceuticals interventions that are likely the only ones to be available in the early stage of the outbreak.

Proportion of escape events that will trigger an outbreak

In order to set a baseline for our investigation it is worth stressing that there is a certain probability that the epidemic goes extinct without any intervention. In general, it is very difficult to estimate this probability, as it depends, beyond other factors, on seeding location (for example, urban vs rural) and contact network of the initial case. In our simulations, all these factors did not vary much as we simulated the initial epidemic seeding to occur always in a BSL facility in a populated area, thus drastically reducing the uncertainty of estimates. The probability of observing an epidemic outbreak in the absence of any interventions (no intervention scenario) is shown in Figure _{0} = 1.1 to values larger than 80% if R_{0} >2.

Reference scenario

**Reference scenario. (A)** Probability of outbreak for different values of R_{0} by assuming no intervention scenario (uncontrolled epidemics) and reference scenario. **(B)** Probability of undetected epidemics for different values of R_{0} by assuming reference scenario (in red) and reference scenario with different values of P_{c}. **(C)** Upper panel: overall number of cases in contained outbreaks by assuming reference scenario and R_{0} = 1.5 (not considering autoextinct epidemics). Middle panel: as upper panel but for the number of traced cases. Lower panel: as upper panel but for the number of isolated individuals (including the laboratory’s contact network). A total of 1,000 simulations were undertaken for each parameter set to produce the results shown.

Proportion of undetected escape events

Notably, model simulations suggest that there is a non-negligible probability that the escape event is not detected at all. This may happen when no initial cases are detected among laboratory workers and laboratory workers’ household members, but secondary cases are generated through random contacts in the general population. In this case it is reasonable to assume that it is very difficult to ascertain the accidental release of a PPP from the BSL facility and to put in place timely control measures. As shown in Figure _{0} and it is strongly influenced by the probability of detecting cases. If R_{0} >1.5, it may be as high as 5% when P_{c} = 60% and 15% when P_{c} = 40%. In general, the probability of case detection affects the outcome of intervention options. As we note, to a large extent the detection probability depends on the rate of asymptomatic cases and non-detectable transmissions. In the case of accidental release, the situation is even worse because the probability of detecting cases affects the probability of the timely implementation of the control and containment interventions. As shown in Additional file

Controllability of the escape event

By assuming reference values for the parameters regulating the containment plan, the probability of observing an epidemic outbreak is drastically reduced for all values of R_{0}. In particular, containment is likely to succeed for values of R_{0} below 1.5 (probability of outbreak less than 10%, see Figure _{0} the probability of outbreak is largely due to the probability of not detecting the outbreak itself; when the accidental release of the PPP agent is detected in a timely manner, outbreaks are contained with probability close to 100%. The resources required to contain epidemic outbreaks with reference intervention may vary considerably. As shown in Figure _{0} = 1.5, corresponding to the isolation of about 500 individuals). Even more demanding, especially from the social point of view, is the closure of 90% of schools for 21 days in a radius of 30 km around location of initial cases, as assumed by the reference SSO set. The number of cases observed can be easily related to the fatality associated to the outbreak if the case fatality rate (CFR) of the specific PPP agent is known. Unfortunately, the CFR is often not obviously correlated with the transmissibility of the pathogen. In addition, it is extremely difficult to obtain reliable estimates of the CFR during the early stage of an outbreak. A sensitivity analysis of the fatality of the virus can however be performed by applying plausible CFR to the number of cases observed with our approach.

The timeline of simulated epidemics with R_{0} = 1.5 is shown in Figure

Epidemic timing

**Epidemic timing. (A)** Average number of daily cases as observed in autoextinct simulated epidemics (red points) with R_{0} = 1.5. Vertical lines represent minimum and maximum daily incidence. **(B)** As in **(A)** but for the average cumulative number of cases. **(C,D)** As **(A)** and **(B)** but for contained epidemics by assuming reference interventions. **(E,F)** As **(A)** and **(B)** but for uncontained epidemics by assuming reference interventions. **(G,H)** As **(A)** and **(B)** but for undetected epidemics. A total of 1,000 simulations were undertaken to produce the results shown.

Sensitivity analysis of containment policies

Results are very sensitive to most of the parameters describing intervention options. By restricting our analysis to parameters regulating contact tracing (thus excluding self-reporting of cases and preventive closure of schools and workplaces) we found that the probability of detecting infections among close contacts of cases and time from initial warning to interventions are the two most critical variables (see Additional file _{0} containment is very likely to succeed when P_{c} is larger than 60% (for R_{0} = 1.2) or 80% (for R_{0} = 1.5) even when the delay from initial warning to interventions (T_{i}) is much larger than the one assumed by reference simulations (up to 30 or 10 days for R_{0} = 1.2 and 1.5 respectively). For larger values of R_{0}, containment is feasible only when P_{c} is larger than 60% and T_{i} is no larger than 3 to 5 days. Figure _{0} up to 1.6 to 1.7 could be reasonably expected to be contained.

Sensitivity analysis: contact tracing

**Sensitivity analysis: contact tracing. (A)** Probability (×100) of outbreak for different values of R_{0} by assuming reference scenario and by varying T_{i} and P_{c}. **(B)** Probability of outbreak for different values of R_{0} by assuming no intervention scenario, reference scenario, and reference scenarios with different delays in the isolation of traced cases. **(C)** Probability of outbreak for different values of R_{0} by assuming no intervention scenario, reference scenario, and reference scenarios with different probabilities of identifying cases in the general community. A total of 1,000 simulations were undertaken for each parameter set to produce the results shown.

Effectiveness of preventive school and workplace closure

Figure _{0} is larger than 1.4. Figure _{0} is 1.2 (as the overall impact of the strategy is not very relevant), while for values of R_{0} = 1.5 or larger, model simulations show that, as expected, the longer the duration and the greater the distance are the lower the probability of outbreak is: duration of 21 days and distance of 30 km represent a good compromise between feasibility and impact. A distance of 30 km for spatially targeted interventions is remarkably larger than that considered in

Sensitivity analysis: school and workplace closure

**Sensitivity analysis: school and workplace closure. (A)** Probability (×100) of outbreak for different values of R_{0} by assuming reference scenario with additional workplaces closure (F_{w} = 0.5) and by varying D_{p} and T_{p}. **(B)** Probability of outbreak for different values of R_{0} by assuming no intervention scenario, reference scenario, and reference scenarios with different policies regulating school and workplaces closure. A total of 1,000 simulations were undertaken for each parameter set to produce the results shown.

Geographical context analysis

The probability of containing an epidemic outbreak may also depend on the BSL laboratory location and the sociodemographic structure of the population. This is shown in Figure _{0} <1.5. Differences reduce drastically for larger values of R_{0}. Without considering control measures, the probability of observing an epidemic outbreak after virus escape is quite similar to that in the Dutch scenario: slight differences can be observed for low values of R_{0}. Such large differences may be due to dissimilarities in sociodemographic characteristics of French and Dutch populations because, despite a general similarity, some marked country-specific features such as age structure and average household size exist. However, although quantitatively different, the general patterns obtained by varying P_{c} and T_{i} are the same observed in the Dutch case. Detailed results for Paris are discussed in Additional file _{0} <1.5. These differences are ascribable to differences in population density and sociodemographic structure, as discussed in

Geographical variability

**Geographical variability. (A)** Ratio between probability of outbreak in different urban areas and probability of outbreak in Rotterdam (The Netherlands) for different values of R_{0} by assuming reference scenario. **(B)** Ratio between probability of outbreak in urban and rural areas in different countries for different values of R_{0} by assuming reference scenario. Urban areas as in **(A)**; rural areas are low population density areas in Wales (UK, 80 km north of Cardiff), Uppland (SE, 100 km north of Uppsala), Sardinia island (IT, 50 km east of Sassari), Andalusia - Castile la Mancha (ES, 50 km northeast of Cordoba), Centre-Burgundy (France, 80 km southeast of Orleans). Note that the reported value of R_{0} refers to that of simulations carried out for Rotterdam (The Netherlands); comparative results for other countries are obtained by assuming the same transmission rates in the different social contexts (that is the same probability of infection transmission given a contact in a certain setting) as in Rotterdam. A total of 1,000 simulations were undertaken for each parameter set to produce the results shown.

Impact of additional intervention

We found that results are not very sensitive to the probability of self-reporting (P_{r}) and to the initial set of interventions on the initial network of contacts comprising laboratory workers and laboratory workers’ household members. The reference scenario assumes the closure of the laboratory and the quarantine of the households of laboratory workers. We explored the possibility of extending these interventions and to preventively close all workplaces and schools attended by relatives of laboratory workers. We found that closing the laboratory is the only intervention leading to a certain reduction of the outbreak probability. Additional interventions are of little impact. We report on these findings in Additional file

Conclusions

Our results suggest that containment is likely to succeed by employing social distancing measures only if R_{0} is no larger than 1.5. Containment could be feasible even for larger values of R_{0} in cases of very timely intervention both in recognizing the accidental release and during contact tracing and high probability of detecting secondary cases in the same household, school or workplace as a newly identified case. Overall, these results suggest that success in containing an accidentally released potentially pandemic influenza virus by employing social distancing measures only is uncertain: containment probability for a virus with transmissibility comparable to many of the estimates for the 2009 H1N1 virus (namely R_{0} or effective transmissibility in the range 1.2 to 1.6 _{0} = 1.8 or higher

Our simulations do not account for the possible use of pharmaceutical interventions. While the availability of an effective vaccine is highly questionable in case of accidental release of genetically manipulated influenza viruses from BSL facilities, the use of antivirals at the very beginning of the epidemic is an option that could be considered. If used for treatment of cases and prophylaxis of close contacts (for example, household and school contacts) only, however, the benefit should not be very different from that obtained by assuming household quarantine and reactive school closure, as this paper does. Moreover, it requires a timely administration (within 2 days from symptoms onset

The preventive immunization of laboratory workers (see for instance the Special Immunization Program in the US

In summary, our results suggest that public health authorities should be prepared to put in place a set of social distancing interventions, for example, contact tracing and closure of schools and workplaces on a geographical basis. Moreover, as it is nearly impossible to get accurate estimates of R_{0} (as well as case fatality rate) for a new virus at the very beginning of the outbreak, in order to maximally reduce the risk of a global pandemic the possibility of timely targeting a large fraction of the population with antivirals (as a prophylactic measure on a geographical basis) or establishing quarantine areas should not be set aside, even though this calls for the development of detailed intervention plans and requires public health agencies to put in place containment efforts hardly achievable in most places in the world. Where the pandemic pathogens are concerned, short generation time and asymptomaticity are among the most critical factors that make accidental release of influenza viruses difficult to contain.

Qualitatively, the results do not vary much by considering different seeding locations. However, containment probabilities are affected by several factors, including population density and sociodemographic structure. These findings may have an important impact on policies: our results strongly suggest the location of new BSL facilities worldwide should be carefully chosen, for instance with priority given to rural areas and, when this is not feasible, by taking into account density and structure of the population in urban areas. This may make the difference, especially for pathogens with low to moderate transmissibility. Of course, these decisions should also be based on other factors not considered in this study, for example, population vulnerability to infectious agents, risk factors, structure of the health system, possibility of putting in place a rapid response program. Simulated scenarios emerging from detailed models such as the one presented here may inform quantitatively the process of identifying locations that minimize risk. Finally, it is worth remarking that the presented approach can be generally extended to other pathogens that can be classified as dual use research of concern if we have the appropriate information on the pathogens, mechanism of transmission and natural history of the disease.

Competing interests

The authors declare they have no competing interests.

Authors’ contributions

SM and AV conceived of the study. SM, MA and LF performed the experiments. SM, MA, LF and AV analyzed results and wrote the manuscript. All authors read and approved the final manuscript.

Acknowledgements

We acknowledge support from the DTRA-1-0910039 and NSF CMMI-1125095 awards to AV, the Italian Ministry of Education, University and Research grant PRIN 2009 2009RNH97Z 001 to LF. The views and conclusions contained in this document are those of the authors and should not be interpreted as representing the official policies, either expressed or implied, of the Defense Threat Reduction Agency or the US Government.

Pre-publication history

The pre-publication history for this paper can be accessed here: