Bee and frog numbers are in decline, and we need to know why. Thus, understanding the dynamics of various species – the study of population biology and ecology – is an important component of BBSRC science, especially where this impacts agriculture. This kind of problem is in fact a classic subclass of problem common in systems biology, where many components may interact, we have little knowledge of the parameters or even the network topology of the system, and where the best we can usually do is to measure system variables. Since it is the parameters of the system that control and determine the time evolution of the (dependent) variables, how can we make progress? The answer is by using inferencing methods (including the methods of data mining and machine learning) that permit one to infer the structure and parameters simply from the measurement of such variables. This is then a data-driven or hypothesis-generating strategy (Kell & Oliver, 2004). The results of the hypothesis-generation step are then hypotheses that can be tested by making the most important inferred parameters independent variables in a subsequent experiment.

Rohr and colleagues (2008a) give a superb recent example of this approach, that might be applied more generally. The number of individuals in many natural species is declining over time, and there may be many reasons (multiple causes) for this. Obvious ones include predation, lack of food supply, toxins, parasites, temperature changes, immunosuppression and so on, not all of which may act directly (e.g. temperature change may have no direct effect upon an individual but may affect the number of its predators).

Rohr and colleagues (2008a) were interested in the decline of the northern leopard frog (Rana pipiens). However, rather than testing a specific hypothesis, they measured 240 separate variables that they considered might have had influence, and found two in particular (the amount of the herbicide atrazine and its metabolite desethylatrazine, and to a lesser extent the levels of inorganic phosphate) that were most correlated with the abundance of larval trematodes (parasitic flatworms) in the frogs. They then determined, using mesocosm experiments, that tanks containing atrazine had fourfold higher levels of snails – an intermediate host for the trematodes – and that the tadpoles that the mesocosms contained were both immunosuppressed relative to controls and contained threefold higher parasite loads. The specificity of these interactions is shown by the fact that there were no effects of atrazine on the survival of the green frog (R. clamitans) in these same mesocosms. (Note too that phosphate exhibited its influence only when atrazine was present.) A separate paper by Rohr and colleagues (2008b) indicated that climate change was probably not involved causally in amphibian decline.

The authors stressed that only with these inferencing methods – they used a variant of path analysis – could they effectively test the numerous implicit hypotheses that might be used to describe the system, and concluded that their studies “illustrate the value of quantifying the relative importance of several possible drivers of disease risk while determining the mechanisms by which they facilitate disease emergence”.

This last is a particularly important point. Imagine three separate effects that could each account for a decline in 10% of a species, but that also interact synergistically. The individual effects would be almost unmeasurably small (i.e. well within the typical fluctuations observed) but if all occurred together might lead to a decline in a population of 50%, a number that would be easily measurable and would be seen as catastrophic. Something similar may be happening with both wild bees (Goulson et al., 2008) and with honeybees.

Honeybees are also in decline throughout the world, and this is important not only for honey production but, of much more economic significance, in the pollination of both wild plants (Biesmeijer et al., 2006) and of various crops. A particularly worrying threat in the USA is known as colony collapse disorder (CCD). In this syndrome, beehives simply lose almost all their workers, who die outside the hive. (This contrasts with the effects of varroa mites that can contribute to bee colony death by immunosuppressing bees, but this death is seen within the affected hives.) There was preliminary evidence that CCD might be due to a transmissible agent, as irradiation of equipment from infected beehives broke the transmission of infection that otherwise occurred. In the absence of any hypothesis (bees are prone to many diseases), Cox-Foster and colleagues (2007) used modern (‘next generation’) high-throughput nucleic acid sequencing methods to determine which organisms were present in hives suffering CCD but not in controls. Sequences of a virus related to the Israeli Acute Paralytic Virus (IAPV) of bees were most strongly implicated, although other organisms, such as the microsporidian fungus Nosema ceranae, were also present in all CCD hives (as well as many without CCD). This knowledge has allowed the dynamics of the various subpopulations of IAPV (Palacios et al., 2008)to be determined, while N. ceranae has been found to be highly pathogenic in the absence of IAPV (see also papers by Higes et al. (2006), by Martín-Hernández et al. (2007), and by Klee et al. (2007)). A particularly attractive result was the fact that the infection underlying the CCD caused by N. ceranae could be controlled using the antifungal fumagillin.

This work raises several issues regarding how best to understand the very damaging decline (one in three hives lost last year) of the British honeybee population (wild bumblebees are also in decline). First, we should recognise that there is unlikely to be a unitary cause – varroa mites, bacteria, viruses, fungi, temperature changes, food shortages, herbicides/pesticides/insecticides and other sources of stress may be well tolerated if the challenges are presented alone, but when they appear together (the ‘perfect storm’) may overwhelm an organism’s defences, and the exact combination in any given instance may well vary considerably. Second, the combinatorial problems of hypothesis-dependent science here are large (twenty different agents that may occur at just two different levels – present or absent 0 give 220 i.e. >1,000,000 possible manipulations to cover this space fully at just the two levels), so hypothesis-generating approaches seem to be preferred. In addition, we will need greater numbers of cases for minimising false discoveries (Broadhurst & Kell, 2006)(the numbers in many of the above studies were quite small). At all events, an experimental approach similar to that of Rohr and colleagues may well prove appropriate for unravelling the problems of UK bee decline, and thereby doing something about them. This is also a topic that lends itself nicely to Web 2.0-style mass collaboration.

Finally, I end by noting that those who are considering starting bee-keeping or the study of bee biology are, of course, known as wannabees.

Related posts (based on tags and chronology):