- Darren Green's home page

Lessons from ants and traffic jams:
Complexity in aquaculture biosecurity

Institute of Aquaculture, University of Stirling

BIOSECURITY applies at all levels in aquaculture, from pre-venting the spread of disease at small scale, such as inter-pond on a farm, up to imports and exports at nationwide scale or above. This broad range of spatial scales poses a challenge when designing efficient, effective biosecurity management strategies.

Understanding epidemics in animal populations is not just a matter of scaling up from understanding a single animal: Populations can behave as complex systems demonstrating emergent behaviour not easily observed in its isolated component parts. This is nicely demonstrated by an ant nest. Studying individual ants alone, we would not be able to predict the existence of the complex structure that is an ant nest. Likewise, observing how individual drivers behave doesn't necessarily predict well where and when traffic jams are likely to occur.

Studying or experimentally modi-fying large-scale systems such as traffic networks, or in our case, entire farming industries is almost always either logistically impossible, or prohibitively expensive; therefore epidemiologists rely on models to explore the behaviour of populations in silico (i.e. by computer or mathematical analysis) in a way that would never be possible in vivo.

Farm-to-farm contact

Live (or dead!) animal movements are an important route for disease spread and introduction, as demon-strated by Gyrodactylus salaris in Norway, viral haemorrhagic septi-caemia (VHS) in Yorkshire (UK) in 2006, or the infamous outbreak of foot-and-mouth disease in the UK in 2001. A recent project brought together researchers at the Institute of Aquaculture at Stirling University, and Marine Scotland, Aberdeen, to examine the network of live fish movements in Scotland. The project aimed to investigate the susceptibility of the salmonid industry to disease spread, and find ways of reducing the likelihood or impact of such epidemics [1]. Similar projects have been applied to other industries such as cattle, pigs, and sheep movements, and in a variety of countries.

We used a contact network approach. This is a powerful epidemiological tool for investigating the potential for disease spread and control. BOX 1 demonstrates the key principles of the approach and explains the concept of nodes (the objects of study, be they farms, people, fish cages, river basins, and so on) and the contacts between them. Even if the underlying nodes appear the same in isolation, different network structures can drive epidemiological processes very dif-ferently. To epidemiologists, there are several aspects of contact structure that are of interest, taking as an example fish farm sites as our nodes.

Box 1: Network representation of an industry
network diagram

These two groups of fish demonstrate two networks, represented by nodes (fishes) and edges (arrows). Nodes represent epidemiological units: these could be individual fish, ponds, farms, or river catchments. Edges represent potentially infectious contact: this could be via live fish movements between farms, movement of fish or infectious agent through water courses, or spread via other carriers such as parasites, birds, or people. In this example, contact is undirected (contact A to B implies also contact B to A) but it is often undirected (e.g. movements naturally occur in a single direction, and water courses will preferentially carry infectious agents downstream).

Despite each fish having the same number of contacts in both networks (orange and green), the way the contacts join fish together differs between them. This means that the spread of infection is markedly different. It could be faster and more extensive for the green network, whereas for the orange network, infection at one fish can infect at a maximum only half the network.

network diagram

The third network above-here with one-way arrows representing directed contact-demonstrates the potential for targeted surveillance. Contact-based spread will be difficult to control within the magenta group, and these fish pose a risk of possible onward spread to the green group. Focusing surveillance on the red contact or red fish readily partitions the network into two subgroups, confining any epidemic. The challenge is to identify the "red" contacts in a complex network such as that pictured in BOX 3.

Sites at risk of being a source of infection. These are likely (but not always-it all depends on the structure of the whole network) to have comparatively many outbound contacts. In the context of sexually transmitted diseases, such individuals (as opposed to sites) could be termed "superspreaders".

Sites at risk of being infected. These may or may not be the same as those at risk of being a source of infection. They are likely to have more inbound potentially infectious contacts, e.g. because they receive fish from elsewhere or lie downstream of other farms in the same water course.

Potential size of an epidemic. Through modelling, we can estimate how, where and how fast a potential introduction of disease is likely to spread, and also when is likely to be a period of high-risk time. Nevertheless, epidemic spread is a random process, and like any random process we can only predict trends, not conclusively say where and when an epidemic will occur.

Contacts at risk of spreading infection. We can think not only about which sites are at risk, but which contacts between pairs of sites pose the highest risk. Reducing the numbers of high-risk contacts, or close surveillance of them, will help contain an epidemic.

Potential for targeted surveillance. For regulators in the European Union, there is a requirement to impose risk-based surveillance for diseases (EU Council Directive 2006/88/EC). By identifying high-risk sites and contacts, regulators can concentrate surveillance effort where it is most likely to be effective.

Potential for targeted control. In reacting to a disease problem, control should ideally be as non-disruptive and cost effective as possible, proportionate to the likely risk of failing to control it. Again, identification of high-risk sites and contacts allows for concentration of control strategies where they are most likely to be effective.

There are several ways in which contact networks can be modelled. These vary in complexity between static network models to detailed simulation models of individual movements. Which is appropriate depends in part upon the disease of interest as shown in BOX 2, as well as the particular type and scale of system (e.g. within a cage of fish versus fish farms in a country).

Box 2: Contrasting models for epidemics

Dynamic simulation models: Individuals (sites, etc.) and their pairwise interactions (e.g. movements between sites) are "replayed" in the order they occurred.
When are they appropriate?
Diseases with a fast timescale compared with the rate of inter-site contact.
Where "who-contacts-who" between sites changes quickly.
When time and computer resources are sufficient to allow time for complex modelling.

Static network: Present a snapshot of possible contacts over a defined time window.
When are they appropriate?
For slow spreading diseases where general patterns of contact are important, not specific contacts in short time intervals.
For networks that change slowly over time.
When simple solutions are needed for complex problems.

Compartmental models: Consider numbers of individuals in classes such as "infected" or "susceptible".
When are they appropriate?
For populations that are well mixed. E.g. the population of salmon in a single cage. They are less appropriate at large spatial scales.
As part of a larger metapopulation model, where populations are made up of nested sub-populations. E.g. for a single cage of a larger farm, of a set of farms in a catchment.

Scottish Salmonid industry

The structure of the Scottish live fish movement network (salmon and trout) is shown in BOX 3. In our paper [1], we considered the potential for targeted surveillance by determining to what extent this network structure could be disassembled by removing a small number of contacts between fish farming sites.

Box 3: Live salmonid fish movements in Scotland
network diagram

This network represents-in a non-geographical way-epidemiological "proximity" through live fish movements for over 500 salmonid-producing sites in Scotland, using 2002-4 movement data. Coloured dots represent Atlantic salmon producing sites (orange), rainbow trout (green), brown trout (violet), with mixed sites in other colours. Clusters of sites can be seen, connected through shared trade, either because they are involved in the same sector, or belong to the same company.

We developed a series of "rules" that could be used in the face of an epidemic or proactively to identify the contacts most at risk. These rules varied from simple-such as "contacts linking farms with lots of contacts"-through to ones based on math-ematical analysis of this and other networks.

Our results indicated that the more simple algorithms performed poorly, failing to identify those contacts which would most benefit biosecurity. This demonstrates the advantage of our network modelling approach. Nevertheless, the more successful rules could identify contacts whose removal from the network (reflecting perfect biosecurity, as it were) could reduce the potential maximum size of an epidemic by approximately half by focussing on as few as four contacts. This assumed that an uncontrolled epidemic could spread throughout the network (probably not an un-reasonable assumption in the case of a disease such as FMD in large livestock).

This approach is, however, only useful for designing disease control measures in advance if the network structure remains consistent through time. By this, we don't mean that the properties of the network (such as how many contacts there are) do not change, but rather that "who is connected to who" remains pre-dictable. In [2], we investigated this for not only the fish movement network of Scotland, but also for movements of British livestock. We found a moderate amount of consistency in both networks through time. But how consistent must networks remain over time for network models to be useful? Here, there is no obvious baseline for comparison, but this is where further modelling can come inů

Nevertheless, movements of sal-monid fish in Scotland are certainly seasonal, and this may lead to increased likelihood of the spread of disease (via movements) at those times of the year where trade is increased, all other things being equal [3]. For example, movements from freshwater to saltwater farms were markedly seasonal in the 2002-4 data, peaking in the spring and autumn. Simulated epidemics were smaller in size on this seasonal network compared to simulations we performed where the seasonality was ignored [3].


How to manage and legislate best biosecurity practice in a large and complex industry is a difficult question, requiring modelling and risk analysis. Network epidemic models are a useful tool here and allow us to predict how epidemics spread, how to monitor them, and how to control them.

Suggested reading

The three papers listed here have come about as part of the project described in this article.

[1] Green, D.M., Werkman, M. & Munro, L.A. (2012). The potential for targeted surveillance of live fish movements in Scotland. Journal of Fish Diseases 35: 29-37.

[2] Green, D.M., Werkman, M., Munro, L.A., Kao, R.R., Kiss, I.Z., & Danon, L. (2011). Tools to study trends in community structure: application to fish and livestock trading networks. Preventive Veterinary Medicine 99: 225-228.

[3] Werkman, M., Green, D.M., Munro, L.A., Murray, A.G. & Turnbull, J.F. (2011). Seasonality and heterogeneity of live fish movements in Scottish fish farms. Diseases of Aquatic Organisms 96: 69-82


With thanks to the other researchers involved in the project described in this article, particularly Marleen Werkman of the University of Warwick and Lorna Munro of Marine Scotland.