Habituation in Aedes aegypti mosquito larvae is context specific Free

Learning is crucial for any animal, as it allows adaptation of the individual's behaviour to changing environments (Giles and Rankin, 2009). Furthermore, being able to recognise and to differentiate among different kinds of stimuli reduces energy costs and time consumption (Scott, 2005). Habituation is a non-associative type of learning that allows filtering out of irrelevant stimuli. When an individual faces a repetitive stimulus without any consequence, that stimulus becomes irrelevant. Consequently, a gradual decrease in its behavioural response occurs (Thompson and Spencer, 1966; Groves and Thompson, 1970; Giles and Rankin, 2009; Rankin et al., 2009; Rankin, 2009; Klein, 2012).

Habituation has often been considered among the simplest and most elementary forms of learning. It has been documented in a wide diversity of animals, both vertebrates and invertebrates. Among vertebrates, habituation has been studied in horses (e.g. Christensen et al., 2010), rats (e.g. Askew, 1970), birds (e.g. Dong and Clayton, 2009), fishes (e.g. Best et al., 2008), rabbits (e.g. Whitlow, 1975), cats (e.g. Groves and Thompson, 1970), dogs (e.g. Pullen et al., 2012) and humans (e.g. Dycus and Powers, 1997). Among invertebrates, habituation has been studied in the nematode Caenorhabditis elegans (e.g. Rankin, 2000), molluscs (e.g. Pinsker et al., 1970; Carew and Kandel, 1973; Bailey and Chen, 1983; Fischer et al., 2011), crabs (e.g. Brunner and Maldonado, 1988; Lozada et al., 1990; Tomsic et al., 1998, 2009; Hemmi and Tomsic, 2012; Raderschall et al., 2011) and insects (Tachind fly Drino bohemica: Monteith, 1963; honey bee Apis mellifera: Braun and Bicker, 1992; Gerber and Menzel, 2000; fruit fly Drosophila melanogaster: Glanzman, 2011; Soibam et al., 2013; Twick et al., 2014). More recently, habituation-like behaviour has been reported in the myxomycete slime mould Physarum polycephalum (Boisseau et al., 2016).

The adult mosquito Aedes aegypti is the vector of several human diseases such as dengue, yellow fever, Chikungunya, Zika and Mayaro viruses (Marklewitz and Junglen, 2019). Pre-imaginal stages of this holometabolous insect, i.e. from egg to pupa, develop in clean and calm bodies of water. When confronted with a moving visual stimulus above the water surface (e.g. a shadow), mosquito larvae and pupae perform a stereotyped escape response (fast movements with the abdomen allowing the animal to dive; Holmes, 1911; Thomas, 1949; Leftwich, 1954; Mellamby, 1958). The escape response causes significant energy expenditure, affecting survival and fertility (Timmermann and Briegei, 1993; Lucas and Romoser, 2001). Cognitive abilities of mosquitoes in their pre-imaginal stages have been barely studied. Recently, Baglan et al. (2017) reported habituation leaving a mnesic trace lasting for 1 h in mosquito larvae of A. aegypti. Analogous to the work performed by Brunner and Maldonado (1988) in the crab Neohelice granulata (Chasmagnathus granulatus), Baglan et al. (2017) quantified the reduction of the escape response to the shadow projected by an object moving over the surface of the water (named by the authors as a ‘visual danger stimulus’) subsequent to repeated presentations.

According to Chilaka et al. (2012), it is particularly important to study learning and memory in mosquitoes and other insects that transmit human diseases because their behaviour determines their success as disease vectors (McCall and Kelly, 2002; Alonso and Schuck-Paim, 2006; Bouyer et al., 2007; Vinauger et al., 2014, 2016). Therefore, behavioural studies in mosquitoes must take into account all the developmental stages, as larval performance has an impact on their adult success (McCall and Eaton, 2001; Kaur et al., 2003). So far, studies on cognitive ability in pre-imaginal stages remain scarce, as mentioned above.

Habituation is traditionally understood as a non-associative form of learning. However, during habituation, animals can integrate other stimuli constantly present in their environment. Those stimuli are together known as the context, i.e. the surrounding conditions where learning takes place (Myers and Gluck, 1994). Contextual stimuli can be associated with habituation stimuli. When such an association exists, it determines the contextual specificity of habituation. Context dependency in habituation was demonstrated in the nematode C. elegans (Rankin et al., 1990; Bernhard and van der Kooy, 2000; Rankin, 2000) and in the sea-hare Aplysia californica (Pinsker et al., 1970; Carew and Kandel, 1973). In the crab N. granulata, it was shown that when the context was changed, habituation to a shadow passing over the crab was reversed (Hermitte et al., 1999).

The fact that context dependency of habituation has been observed in diverse organisms led us to hypothesise that this is a generalised phenomenon. Consequently, we asked whether habituation in the larvae of A. aegypti was context-dependent as well.

In the present work, we studied whether mosquito larvae were able to associate visual contextual stimuli with the stimulus inducing habituation. To that end, mosquito larvae were individually trained in a Petri dish positioned over a set visual pattern (the context), with a shadow stimulus presented above. Once habituation was induced, the context was changed and the shadow stimulus was presented again. Context dependency of habituation would be evinced if the larvae performed the escape response.

Fourth-instar larvae of Aedes aegypti Bora strain were used in this work. Mosquitoes were reared in a climate-controlled room at 25°C with 12 h:12 h light:dark cycle. Larvae were obtained from eggs provided by the Laboratoire d'Insectes Nuisibles (MIVEGEC-IRD, Montpellier, France). Eggs were put in small plastic containers (12×6×6 cm l×w×h) with dechlorinated tap water. Larvae were fed with shrimp food (JBL Novo Prawn, Neuhofen, Germany; www.jbl.de). Under these conditions, larvae reached their fourth instar after 5 days and moulted into pupae 5 days later, i.e. it took around 10 days from hatching to pupal moult. In order to avoid using pharate pupae, we did not use larvae older than 8 days. All the animals were reared and treated according to the rules and regulations on ethics applied in the European Union.

The protocol used in the experiments was adapted from Baglan et al. (2017). The experiments were performed in a different room at 24±1°C. Each larva was individually placed into a Petri dish (3 cm diameter and 1.2 cm height) filled with dechlorinated tap water, and the Petri dish was placed on a table directly under a spot light. A square of cardboard was positioned under each Petri dish to provide the visual context. The cardboard was painted with one of the following three patterns: 100% white, 100% black or striped black and white (each stripe was 1 cm wide; black and white each covered 50% of the surface). The Petri dishes were visually isolated from each other by opaque walls (30 cm in height and width). To avoid any disruption caused by vibrations, each Petri dish was placed on a foam block.

The visual stimulus was presented by a mechanical arm; namely, a wooden stick, 18 cm long, with a 3×3 cm square of cardboard at the end. The cardboard square projected a shadow stimulus to be perceived by the naive larva; this signal might be similar to the visual stimulus provided by an aerial predator in nature. Consequently, the naive larva performed the escape response. We used a servomotor associated with a control board (Arduino Uno; http://www.arduino.cc) to automatically deliver the habituation stimulus. The angular speed of the stimulus was 315.8 deg s⁻¹ at 4.8 V.

The protocol of habituation included three phases: acclimatation, training and test. During acclimatation, each larva was deposited in a Petri dish under the spot light and rested for 30 min. Over the course of the training phase, the mechanical arm projected its shadow on the Petri dish for 1.5 s, with an inter-trial interval of 5 min. Training was carried out for a total duration of 1 h and 15 min. The visual stimulus was displayed 15 times, i.e. 15 trials. In every trial, the response of the larva was recorded, i.e. escape response=1, no response=0. The escape response was evinced by fast movements of the abdomen, allowing the animal to dive (see the video in Baglan et al., 2017). A decrease in the escape response over the course of the training phase would indicate habituation (see Baglan et al., 2017). Once the training phase was completed, the Petri dish was very gently raised to avoid any disturbance that could provoke dishabituation (Baglan et al., 2017). The piece of cardboard beneath it (the context) was removed and either replaced by another one or put back, in experimental or control groups, respectively; larvae were then left to rest for 10 min. The test phase then occurred, consisting of one presentation of the visual stimulus.

Three different types of background (white, black or striped) were used as context (i.e. the stimulus that was present during the whole training phase; Nadel and Willner, 1980), for both training and test. All possible combinations represented nine independent experimental groups; each group included 30 replicates. In total, 270 individuals were included (Fig. 1).

The data analyses were performed using a Bayesian model selection procedure on a series of mathematical models describing the habituation process, starting with the simplest model and increasingly adding complexity until the information criterion used worsened or the full model with all the parameters was reached.

A series of Hidden Markovian models (Eddy, 2004), based on the power law of practice (Snoddy, 1926), was used to explain the dynamics of learning. Our explanations were based on the experimental data. According to this approach, learning does not occur at a constant speed but is slowed down over the course of the habituation trials. In the beginning, learning occurs faster; however, further improvements are more difficult. Power law models propose that learning decreases linearly with the logarithm of the number of practice trials taken, thus producing an exponential decay, in which the response tends asymptotically towards a minimum value (generally zero). In this case, the escape response is expected to decrease as the larva is subjected to successive tests until it approaches a minimum baseline value (see Results for details).

The model we applied is well known and it has been widely tested in human learning (see Fitts and Posner, 1967; Anderson, 1982; Card et al., 2017). More recently, Evans et al. (2018) developed Bayesian versions of this model.

Once the tests were finished, the habituation could decrease or not, according to the response d during the test. The full development of the model is shown in the Appendix (see Table A1).

The main objective of the analyses was to define whether or not the model parameters differed among treatments (test settings, and their combinations), and to what degree. Once we defined which parameters were specific to the treatments and which parameters were general to the learning model, the values of the parameters were calculated.

The model describing learning is non-linear. Hence, the parameters could interact with each other in many possible ways. As a consequence, we used a process of selection of models through information indices to obtain an optimal model instead of simple statistical tests of the null hypothesis. These parameters would be treatment specific only in the event that a given parameter varied between treatments, and would be common to all in cases where the parameter did not vary between treatments.

Consequently, the data were analysed using a stepwise model selection procedure under a full Bayesian approach. We proposed a series of explanatory models that were generated from the null model, becoming more complex by adding parameters to each step, thus increasing complexity. The development of these models continued until the model complexity exceeded the increase in complexity caused in the model by adding a new parameter. The deviance information criterion (DIC; Gelman et al., 2004) was used as a decision rule for accepting/rejecting the proposed models. When a newly more complex model was rejected, because the DIC increased, the procedure stopped, and we kept the last selected model as the explanatory model.

The distribution of parameters of the models was calculated using the Markov Chain Monte Carlo (MCMC) algorithm with the pymc module for stochastic modelling in Python version 2.3.8 (Patil et al., 2010).

Non-informative uniform a priori distributions between 0 and 1 were used for all the parameters, except the autocorrelation coefficient, which was between −1 and 1. A total of 200,000 iterations of the MCMC were used, of which the first 100,000 were discarded as a burn-in, and the remaining were used to estimate the a posteriori distributions.

Using a Bayesian approach provides a more straightforward way to fit and select among a great number of models, unlike frequentist statistical methods, and the same approach has already been used by ecologists in animal behaviour studies (Ellison, 2004; McNamara et al., 2006; Pietrantuono et al., 2015, 2017).

Additionally, the generalised coefficient of determination (GCD) for binary data, according to Cox and Snell (1989) and Magee (1990), was calculated as a measure of the fitness of each of the proposed models to the data of the experiments.

A total of 31 models (plus null model) were developed. Through the model selection process, we obtained a model that, according to the information criteria used (minimum DIC), was optimal in terms of explanatory power and complexity (model no. 18, Table A1). The resulting model contained six parameters (h, d_c, R_aw, R_ab, R_as and ϕ), detailed in Table A1, and its explanatory power in terms of GCD was 0.9611. This means that this model is capable of explaining 96% of the variance of the study data. As explained above, all parameters that differed among test contexts, testing or their interactions, were included in the model as separate parameters (where different contexts are indicated by subscript j), while the parameters that did not differ were replaced by a single common parameter for all treatments.

The parameters calculated for this model (see Table 1) show that, on the one hand, the habituation rate (h) was the same for all contexts, whilst, on the other hand, the asymptotic response (R_a) depended on the context, being similar between the white and striped background, but higher with the black background. As explained in Materials and Methods, the asymptotic response had an inverse relationship with habituation, so the white and striped backgrounds favoured habituation in the long term. As can be seen in Fig. 2, the final responses with the white and striped backgrounds were similar, and the response was greatest with the black background, with similarly sloped curves in all cases.

Finally, autocorrelation was strong, as the autoregressive coefficient (ϕ) was estimated to be 0.3 (see Table 1) and could alter the results in unpredictable ways if not taken into account.

The results of our experiments evinced that the insects included in this study had a power law-type learning curve, in which the response decreased to a level that was higher than zero but lower than the initial response level. Over the course of the 15 training trials, the response to the visual stimulus decreased in all cases (Fig. 2). Therefore, habituation was observed independently of the visual background (context).

Some parameters of the learning curves were shared across all the contexts, while others were context specific (Table 1). All the backgrounds shared the same habituation rate (h), but differed in their asymptotic responses (R_a); as a consequence, the speed of habituation was similar regardless of the background (visual context).

The parameter d (response increase), which was only present (i.e. different from zero) in the case of a change of background, indicates that in the test phase, the insects retained the same R_a of escape/response level as at the end of the training phase when using the same background, as can be seen in Fig. 2A,E,I. In the case of a change in the background between training and test, the response increased. This increase in response may be attributable to the change in background (Fig. 2B–D,F–H).

n all the visual backgrounds applied in the present work, A. aegypti larvae could be habituated to a visual stimulus. In all the cases, the decrease in response due to habituation did not depend on the visual background (context), which was evinced by the fact that the slopes among all the learning curves did not differ. However, the habituation level was lowest when larvae were trained with the black visual context. In all cases, changing the visual context between the training and the test phases induced an increase in the response in the test, thus proving that habituation was context dependent.

As mentioned above, larvae could be habituated in all contexts. At the same time, we observed that, over the course of the training phase, a decrease in escape response occurred at the same rate in the first five trials, i.e. the slopes of the acquisition curves did not differ. In subsequent trials, behaviour differed according to the visual context. When larvae were trained with the black context, the slope decreased from the 6th trial onwards, and the response reached its asymptote at a 10% greater level than in the cases in which larvae were trained with either a striped or a white context. Therefore, habituation performance depended on the visual context. In particular, the larvae exhibited their best learning performance when trained with lighter visual contexts. In those cases, the contrast between the shadow of the moving cardboard square was greater when projected on the white than on the black surface. The lower habituation level observed with the black context could be due to the contrast between the shadow and the background being lower than in the other two contexts. Quantification of the effects of contrast on habituation level should be carried out in future experiments. For example, larvae could be trained on different solid greyscale or striped backgrounds varying the width of the stripes or the white/black area ratio.

In our study, mosquito larvae that had habituated to the visual stimulus on a white background no longer responded to the visual stimulus when tested on a white background, whilst they responded to the same stimulus when tested on a black background. Conversely, the same stimulus presented on a different visual background, i.e. a different context, was perceived as a novel stimulus inducing larvae to escape. Context specificity led us recognise that there must have been some kind of associative component of habituation between the context and the visual stimulus, as has been observed in previous work in crabs (Tomsic et al., 1998; Hermitte et al., 1999; Pereyra et al., 2000).

Thanks to the use of Markovian models combined with Bayesian statistics, it was possible to make an accurate description of the mosquito learning curve. Likewise, it was possible to formulate parameters with great biological significance, improving what has been achieved so far using generic statistical models such as those that are most commonly used to analyse habituation experiments. In this work, we calculated the rate, and the final habituation value, the increase in response produced at the end of the training procedure, and its dependency on the context, all corrected by autocorrelation, using an autoregressive (AR) term. Therefore, we suggest the use of this methodological approach to analyse this type of experiment in the future, which significantly increases the amount of information in a study, while maintaining the same design and experimental effort, using the same laboratory equipment.

Our results contribute to the so far scarcely explored learning ability of the mosquito larvae of A. aegypti, in particular, the ability to associate a constantly present visual stimulus (the context) with a moving object. Are mosquito larvae capable of associative learning? The question remains open.

Comprehensive understanding of the pre-imaginal behaviour of mosquitoes – in particular, the behavioural plasticity – will certainly allow us to better understand their adaptability to different environments. This is an important piece of the jigsaw puzzle of integrative pest control and environmental conservation.

We are very thankful to Prof. Claudio Lazzari, Dr Valeria Fernández-Arhex and Martin Dessart for their collaboration and useful comments. We especially thank María Emilia Amadio and David de Luca for helping us with the experiments and for being so lovely! We thank two anonymous reviewers for their comments on the resubmitted version of the manuscript. The authors would also like to acknowledge Argentine public policies for education, science and technology carried out between 2003 and 2015, which supported the development of this research work.