Using a clustering algorithm to identify patterns of valve-gaping behaviour in mussels reared under different environmental conditions

Physiological adaptations for inhabiting transitional environments with strongly variable abiotic conditions can sometimes be displayed as behavioural shifts. A striking example might be found in bivalve species that inhabit estuaries characterised by fluctuations in environment. The opening and closing of their valves, so called gaping activity, represents behaviour that is required for two key physiological functions: food intake and respiration. Linking valve-gaping behaviour to environmental drivers can greatly improve our understanding and modelling of bivalve bioenergetics. Nowadays large data sets on gaping activity can be collected with automated sensors, but interpretation is difficult due to the large amount of environmental drivers and the intra-individual variability. This study aims to understand whether an unsupervised machine learning method (k-means clustering) can be used to identify patterns in gaping activity.Two commercially important congener mussels, Mytilus galloprovincialis and Mytilus edulis inhabiting two transitional coastal areas, the Venice Lagoon and the Wadden Sea, were fitted with sensors to monitor valvegaping, while a comprehensive set of environmental parameters was also monitored. Data were analysed by applying three times a k-mean algorithm to the gaping time series. In the 1st analyses, the algorithm was applied to the overall gaping time series, including daily variations. We identified at both sites three clusters that were characterised by different average daily gaping aperture. The algorithm was subsequently reapplied to relate daily means of gaping to environmental conditions, being temperatures, oxygen saturation and chlorophyll levels. This 2nd analyses revealed that mean gaping aperture was mainly linked to food availability. A 3rd follow-up analysis aimed at exploring daily patterns. This third analysis again revealed consistent patterns amongst the two sites, where two clusters emerged that showed different degrees of oscillatory behaviour. There was however no obvious relationship between this fine scale oscillatory behaviours and environmental variables, but in the Venice Lagoon there was a site effect. Overall, we show that clustering algorithms can disentangle behavioural patterns within complex series of big data. The latter offers new opportunities to improve sitespecific bioenergetic bivalve models by rephrasing the clearance and respiration terms based on the mean gaping aperture, provided that further laboratory experimentations are conducted to extrapolate parameters linking aperture with energy inputs and outputs.