Automated identification of social interaction criteria in Drosophila melanogaster

Abstract

The study of social behaviour within groups has relied on fixed definitions of an 'interaction'. Criteria used in these definitions often involve a subjectively defined cut-off value for proximity, orientation and time (e.g. courtship, aggression and social interaction networks) and the same numerical values for these criteria are applied to all of the treatment groups within an experiment. One universal definition of an interaction could misidentify interactions within groups that differ in life histories, study treatments and/or genetic mutations. Here, we present an automated method for determining the values of interaction criteria using a pre-defined rule set rather than pre-defined values. We use this approach and show changing social behaviours in different manipulations of Drosophila melanogaster. We also show that chemosensory cues are an important modality of social spacing and interaction. This method will allow a more robust analysis of the properties of interacting groups, while helping us understand how specific groups regulate their social interaction space.

Keywords: Drosophila; automated; behaviour; social.

Figure 1.

Establishing Canton-S males' interaction space based on repeated spatial–temporal behaviour. This figure presents an illustration of the method using the entire dataset of Canton-S males (n = 43). (a) The social distance is identified per group based on the large over-representation of close fly–fly distances in ‘real’ compared to ‘null’ data (see the electronic supplementary material, figure S1). The red dotted line indicates the social distance cut-off (see electronic supplementary material). (b–c) The angle and distance between each fly and all other flies' centre of mass is established for (b) all trials and (c) 500 ‘null’ trials (see electronic supplementary material). Numbers around heatmaps indicate angle. (d) The normalized frequency from the ‘null’ dataset is subtracted from the normalized frequency of the ‘real’ dataset, which reveals spatial positions seen more often in our assay than one would expect from non-social organisms. The angle and distance that captures the majority of this over-representation is established and plotted in red (see the electronic supplementary material). (e) Using the angle and distance criteria from (d), we count the number of interactions that last at least a specified time duration. The normalized histogram of ‘null’ data subtracted from the histogram of ‘real’ data is plotted, and the first positive time bin (red dotted line) indicates the time duration for which the putative interactions occur more often in the ‘real’ data.

Figure 2.

PCA of the interaction criteria. Our bootstrapped estimates were used to perform a PCA to determine whether these treatments generated interaction values that grouped in a multivariate sense. (a) Score for all treatments and coefficients of the variables. (b–f) Highlight specific comparisons within (a). (b) Overlap of strains and sexes. (c) Clustering of Canton-S and Canton-S in the dark. (d) Separation of scores for Canton-S, iav¹/Canton-S, and iav¹. (e) Separation of Canton-S, Orco²/Canton-S and Orco². (f) Overlap between poxn^ΔM22-B5SuperA-158 and poxn^ΔXBs6 (when the latter generates valid criteria).