Virus genomes reveal factors that spread and sustained the Ebola epidemic

Origin, ignition and trajectory of the epidemic

Molecular clock dating indicates that the most recent common ancestor of the epidemic existed between December 2013 and February 2014 (mean 2014.06, 95% credible interval, CI: 2013.96, 2014.14) and phylogeographic estimation assigns this ancestor to the Guéckédou Prefecture, Nzérékoré Region, Guinea, with high credibility (Figure 2). In addition, we find that initial EBOV lineages deriving from this common ancestor circulated among Guéckédou Prefecture and its neighbouring prefectures of Macenta and Kissidougou until late February 2014 (Figure 2). These results support the epidemiological evidence that the West African epidemic began in late 2013 in Guéckédou Prefecture³.

The first EBOV introduction from Guinea into another country that resulted in sustained transmission is estimated to have occurred in early April 2014 (Figure 2), when the virus spread to Kailahun District of Sierra Leone^{5, 6}. This lineage was first detected in Kailahun at the end of May 2014, from where it spread across the region (Figures 4 & 3). From Kailahun EBOV spread extremely rapidly in May 2014 into several counties of Liberia (Lofa, Montserrado and Margibi)¹⁵ and Guinea (Conakry, back into Guéckédou)^{9, 11}. The virus continued spreading westwards through Sierra Leone, and by July 2014 EBOV was present in the capital city, Freetown.

By mid-September 2014 Liberia was reporting >500 new EVD cases per week, mostly driven by a large outbreak in Montserrado County, which encompasses the capital city, Monrovia. Sierra Leone reported >700 new cases per week by mid-November, with large outbreaks in Port Loko, Western Urban (Freetown) and Western Rural districts (Freetown suburbs). December 2014 brought the first signs that efforts to control the epidemic in Sierra Leone were effective as EVD incidence began dropping. By March 2015 the epidemic was largely under control in Liberia and eastern Guinea, although sustained transmission continued in the border area of western Guinea and western Sierra Leone. By the following month prevalence had declined such that only a handful of lineages persisted^{10, 12} (Figure 4).

The last EBOV genome resulting from a conventionally-acquired infection was collected and sequenced in October 2015 in Forécariah Prefecture (Guinea)¹⁰. After this, only sporadic cases of EVD were detected: in Montserrado (Liberia) in November 2015, Tonkolili (Sierra Leone) in January and February 2016, and Nzérékoré (Guinea) in March 2016. All these sporadic cases likely resulted from transmission from EVD survivors with established persistent infections^{12, 17, 18}.

Factors associated with EBOV dispersal

To determine the factors that influenced the spread of EBOV among administrative regions at the district (Sierra Leone), prefecture (Guinea) and county (Liberia) levels we used a phylogeographic generalized linear model (GLM)¹⁶. Of the 25 factors assessed (see Table 3 for a full list and description) five were included in the model with categorical support (Table 1). In summary, EBOV tends to disperse between geographically close regions (great circle distance: Bayes factor (BF) support for inclusion BF>50). Half of all virus dispersals occurred between locations <72 km apart and only 5% involved movement over 232 km (Figure 11a). Both origin and destination population sizes are very strongly (BF>50) positively correlated with viral dissemination, with a stronger effect for origin population size. The positive effect of population sizes combined with the inverse effect of geographic distance, implies that the epidemic’s spread followed a classic gravity-model dynamic. Gravity models, widely used in economic and geographic studies and a natural choice for modelling infectious disease transmission^19–21, describe the movement of people between locations as a function of their population sizes and distance apart. Here we use viral genomes to provide empirical evidence that such a process drove viral dissemination during the EVD epidemic.

In addition to geographical distance, we found a significant propensity for virus dispersal to occur within each country, relative to internationally (Nat/Int effect, BF>50), suggesting that country borders acted to curb the geographic spread of EBOV. When international dispersals do take place, they are more intense between administrative regions that are adjacent at an international border (IntBoSh, BF>50).

We tested whether sharing of any of 17 vernacular languages explains virus spread, as common languages might reflect cultural links including between non-contiguous or international regions, but found no evidence that such linguistic links were correlated with EBOV spread. A variety of other possible predictors of EBOV transmission, such as aspects of urbanization (economic output, population density, travel times to large settlements) and climatic effects, were not significantly associated with virus dispersal. However, these factors may have contributed to the size and longevity of transmission chains after introduction to a region (see below).

Finally, to investigate the potential of ‘real-time’ viral genome sequencing, we considered the degree to which the findings could have been obtained at the height of the epidemic, had sequences been available shortly after samples were taken (see Methods for details). For the factors associated with EBOV dispersal the results were extremely comparable with those for the full dataset with the same five factors being strongly supported and having similar effect sizes (Figure 5).

Factors associated with local EBOV proliferation

The analysis above identified predominantly geographical and administrative factors that predict the degree of importation risk, i.e. the likelihood that a viral lineage initiates at least one infection in a new region. However, the epidemiological consequences of each introduction—the size and duration of resulting transmission chains—may be affected by different factors. Thus we investigated which demographic, economic and climatic factors might predict cumulative case counts¹ for each region (Bayesian GLM; see Methods) and found these were associated with factors related to urbanization (Table 2): primarily population sizes (PopSize, BF 29.6) and a significant inverse association with travel times to the nearest settlement with >50,000 inhabitants (tt50K, BF 32.4). These results confirm the common perception that, in contrast to previous EVD outbreaks, widespread transmission within urban regions in West Africa was a major contributing factor to the scale of the epidemic of the Makona variant.

As the epidemic in West Africa progressed there were fears that increased rainfall and humidity might prolong environmental persistence of EBOV particles, increasing the likelihood of transmission²². Although we found no evidence of an association between EBOV dispersal and any aspects of local climate, we find that regions with less seasonal variation in temperature, and more rainfall, tended to have larger EVD outbreaks (TempSS, BF >50 and Precip, BF 4.4 respectively).

Effect of international travel restrictions on EBOV dispersal

Porous borders between Liberia, Sierra Leone and Guinea may have allowed unimpeded EBOV spread during the 2013–2016 epidemic^23–25. Our results indicate that international borders were associated with a decreased rate of transmission events compared to national borders (Figure 6), but that frequent international cross-border transmission events still occurred. These events were concentrated in Guéckédou Prefecture (Guinea), Kailahun District (Sierra Leone) and Lofa Country (Liberia) during the early phases of the epidemic (Figure 7a), and between Forécariah Prefecture (Guinea) and Kambia District (Sierra Leone) in the later stage (Figure 7b). These later EBOV movements significantly hindered efforts to interrupt the final chains of transmission in late 2015, with EBOV from such chains moving back and forth across this border^{10, 12, 26}.

Sierra Leone announced border closures on 11 June 2014, followed by Liberia on 27 July 2014, and Guinea on 9 August 2014, but little information is available about what these border closures actually entailed. Although we show that the relative contribution of international spread to overall viral migration was lower after country borders were closed (mean Nat/Int coefficient increasing from 1.15 to 2.83 between August and September 2014; 80.0% posterior support; (Figure 4b), it is difficult to ascertain whether the border closures themselves were responsible for the apparent reduction in cross-border transmissions, as opposed to concomitant control efforts or public information campaigns. However, even if border closures reduced international traffic, particularly over longer distances and between larger population centres, by the time Sierra Leone and Liberia closed their borders the epidemic had become firmly established in both countries (Figure 4).

Why did the epidemic not spread further?

A few EBOV exportations were documented from Guinea by road transport into Mali and Senegal^{27, 28} and by air from Liberia to Nigeria and USA^{29, 30}. However, apart from these limited exceptions, the West African Ebola virus epidemic did not spread into the neighbouring regions of Côte d’Ivoire, Guinea-Bissau, Mali, and Senegal. By extending our GLM (the supported predictors and their estimated coefficients) to include these regions we were able to address whether they were spared EVD cases through good fortune, or because they were associated with an inherently lower risk of EBOV spread and transmission. We estimated the degree to which these, apparently EVD-free, regions had the potential to be exposed to viral introductions from affected regions (see Methods).

Overall, the contiguous regions in unaffected neighbouring countries were all predicted to have low numbers of EBOV introductions (Figures 8a and 9a) based on the phylogeographic history of the sampled cases. They were not, however, predicted to have particularly low levels of transmission if an outbreak had been seeded (Figures 8b and 9b). Thus, it is likely that some of these regions were at risk of becoming part of the EVD epidemic, but that their geographical distance from areas of active transmission and the attenuating effect of international borders prevented this from occurring. The Kati Cercle in Mali and Tonkpi Region in Côte d’Ivoire are to some extent exceptions to this general result, being more susceptible to EBOV introductions under the gravity model because of their large populations (1 million and 950,000, respectively), (Figure 8a) and predicted to have experienced many cases had EVD become established (Figure 8b).

Metapopulation structure and dynamics of the EVD epidemic

After the initial establishment of transmission in Sierra Leone and Liberia, Guinea experienced repeated reintroductions of viral lineages from disparate transmission chains from both countries (Figure 4). Our analysis reveals that there were at least 21 (95% CI: 16 – 25) re-introductions into Guinea from April 2014 to February 2015. An early epidemic lineage was established around the Guinean capital, Conakry, and persisted for the duration of the epidemic (GN-1 in Figures 2 & 4). However, the continual ‘seeding’ of EBOV into Guinea without a clear peak in transmission suggests that elsewhere the virus may have been failing to maintain transmission. There were also numerous introductions into Sierra Leone over a similar time period (median: 9, 95% CI: 6 – 12) but the resulting transmission chains constituted a tiny proportion of the country’s EVD cases, with the bulk of transmission resulting from one early introduction (Figure 4a).

In all three countries, repeated seeding of administrative regions seems to have been a large factor in the longevity of the EVD epidemic (Figure 10). As such, regional case numbers were generally the result of multiple overlapping introduction events followed by within-region spread and occasional onward transmission to other regions. This suggests a metapopulation model in which the epidemic’s persistence was driven by introduction into novel contact networks rather than by mass-action susceptible-infectious-removed (SIR) dynamics^{31, 32}. We found that, on average, EBOV migrates between administrative regions at a rate of 0.85 events per lineage per year (95% CI: 0.72, 0.97). Assuming a serial interval of 15.3 days³³, this rate translates to a 3.6% chance (95% CI: 3.0%, 4.1%) that over the course of a single infection, the transmission chain moved between regions. Given the key role that virus dispersal played in sustaining the epidemic, the detection and isolation of these relatively low proportion of mobile cases may have a disproportionate effect on the control of an EVD epidemic.

From our spatial phylogenetic model we conclude that many regions experienced numerous independent EBOV introductions (Figure 11b). However, these introductions gave rise to clusters of cases that were generally small (a mean cluster size of 4.3 and only 5% larger than 17 in our sample; Figure 11c) and of limited duration (a mean persistence time of 41.3 days with only 5% greater than 181 days; Figure 11d). Here, we define a ‘cluster’ as a group of sequenced cases in a region that derive from a single introduction event and define ‘persistence’ as the time between the introduction event and the last sampled case in the cluster. These definitions are conservative regarding sampling intensity as we expect additional samples would have split clusters apart rather than join them. Furthermore, introductions that were not detected will be disproportionately smaller, and so the cluster size estimate will be biased upwards. Segregating these observations by country (Figure 12a) shows that districts of Sierra Leone had more introductions and Guinea generally had smaller clusters but persistence was similar between the three countries. Considering only introductions that occurred before October 2014 to those that occurred after, the number of introductions per location was comparable whereas those that occurred early generally resulted in larger and more persistent clusters (Figure 12b).

Thus, with 5.8% sampling, we arrive at a conservative estimate of approximately 75 regional cases per introduction event. Although larger population centres, in particular capital cities, generally experienced more introductions (Figure 13a), the cluster sizes are less strongly associated with population size (Figure 13b), further highlighting the role of virus movement into urban areas as major factor for the high case loads in large population centres. Frequent cluster extinction, despite a small fraction of individuals being infected, suggests that individual outbreaks were constrained by the degree of connectedness among contact networks. Thus, it appears that the West African EVD epidemic was sustained by frequent seeding that resulted in numerous small local clusters of cases, some of which went on to seed further local clusters.

Viral genomics as a tool for outbreak response

The 2013–2016 EVD epidemic in West Africa has unfortunately become a costly lesson in addressing an infectious disease outbreak in the absence of preparedness of both the exposed population and the international community. Our work demonstrates the value of pathogen genome sequencing in a public healthcare emergency and the value of timely pre-publication data sharing to identify the origins of imported disease case clusters, to track pathogen transmission as the epidemic progresses, and to follow up on individual cases as the epidemic subsides.

It is inevitable that as sequencing costs decrease, accuracy increases, and sequencing instruments become more portable, real-time viral surveillance and molecular epidemiology will be routinely deployed on the front lines of infectious disease outbreaks^{10, 12, 14, 34–36}. Although we have shown here that the broad pattern of EBOV spatial movement was discernible from virus genomes derived from samples collected only up until October 2014, there was a notable hiatus in sequencing at this time³⁵ and the genomes in the present data set from that time were sequenced retrospectively from archived material. The West African EVD epidemic has demonstrated that a steady sequencing pace^34–36, local sequencing capacity^{10, 12, 14} and rapid dissemination of data⁷ are key ingredients in generating actionable sequence data from an infectious disease outbreak. However, as viral genome sequencing is scaled up and approaches the time-scale of viral evolution, the analysis techniques will increasingly represent the bottleneck for timely communication of information for outbreak response.

The analysis of the comprehensive EBOV genome set collected during the 2013–2016 EVD epidemic, including the findings presented here and in other studies^{7, 9, 11–15, 37, 38} provides a framework for predicting the behaviour of future disease outbreaks caused by EBOV, other filoviruses, and perhaps other human pathogens. However, many open questions remain about the biology of EBOV. As sustained human-to-human transmission waned, West Africa experienced several instances of recrudescent transmission, often in regions that had not seen cases for many months as a result of persistent sub-clinical infections^{17, 18, 39}. Although, in hindsight, such sequelae were not entirely unexpected⁴⁰, the magnitude of the 2013–2016 epidemic has put the region at ongoing risk of sporadic EVD re-emergence. Similarly, the nature of the reservoir of EBOV, and its geographic distribution, remain as fundamental gaps in our knowledge. Resolving these questions is critical to predicting the risk of zoonotic transmission and hence of future EVD outbreaks.

Methods

Supplementary Material