All animal care and experimental protocols complied with local and federal laws and guidelines and were approved by the appropriate governing body in Berlin, Germany, the Landesamt fur Gesundheit und Soziales (LaGeSo G-0224/20).
Experimental breeding and design
The all-female Amazon molly (Poecilia formosa) is a naturally clonal, live-bearing fish species that gives birth to broods of genetically identical offspring. Like all unisexual vertebrates, Amazon mollies are the result of inter-specific hybridization44,45. As such, this ‘frozen hybrid’ has a heterozygous genome from its ancestral P. mexicana mother and P. latipinna father alleviating concerns about reduced genetic variation and the resulting inbreeding depression often associated with artificially selected isogenic animals. Additionally, despite their clonal nature, the Amazon’s genome shows no evidence of increased mutation accumulation, genomic decay or transposable element activity suggesting the genomes of these animals are evolving in similar ways as sexual species46. They reproduce through gynogenesis where the meiotic process is disrupted so that the eggs contain a full maternal genome. The egg must be fused with a sperm from one of their ancestral species to stimulate embryogenesis, but this paternal DNA is not incorporated into the egg. This provides the opportunity to control when reproduction occurs by controlling the females’ access to male sperm donors.
We placed adult females, as potential mothers of experimental fish, in individual (5-gallon) breeding tanks with two Atlantic molly (P. mexicana) males for one week to act as sperm donors. Amazon mollies give birth to broods of generally ~8-30 individuals. A brood is born at once (i.e. all individuals are born within minutes of each other) and birth generally happens early in the day close to dawn. These parental fish were lab-bred and themselves sisters, so of the same age and lineage, and were kept at similar social densities and under standardized environmental conditions throughout their lives to further minimize potential variation in maternal experience. Each breeding tank contained an artificial plant as refuge and was checked frequently each day for the presence of offspring, especially during the morning hours when births are most likely. Newborn mollies were always found in the morning and then singly netted by trained animal caretakers, into individual experimental tanks where their behavior was automatically recorded for the next 70 days (see below). Moving the fish from the maternal tank to the experimental tanks was done in a standardized manner (i.e. individual fish were netted and placed into small dishes of water and then placed in the tracking tanks to limit exposure to the air) by the same caretakers to minimize variation in experience among individual fish. Altogether, eight mothers provided offspring that completed the entire 10-week experiment (Supplementary Table 1).
Experimental tanks (27 x 27 cm), made of white Perspex, consisted of four equally sized compartments, and were evenly lit from below using 6500K-LEDs. Environmental conditions were highly standardized across tanks: all tanks were on the same 11:13 (L:D) light schedule, water depth was maintained at 10 cm depth, temperature was maintained at 25 ± 1 °C by a room air conditioning system, and fish received a standardized amount of powdered flake fish food (TetraMin™) twice daily. Opaque blinds surrounded the tanks to further limit outside disturbances. All experimental tanks were connected to the same filtration system where water could mix in the sump tank, allowing chemical cues to be shared across all experimental fish. Previous work has shown exposure to just chemical cues of conspecifics is sufficient in preventing the developmental of pathological behavior that could be associated with development in complete isolation14. We initially placed a total of 40 newborn individuals into the tracking tanks. At the end of the 10-week experiment, we were able to achieve complete tracking data on 26 individuals; camera malfunctions prevented data collection on four individuals, two individuals jumped into neighboring tanks causing the loss of data of all four individuals as we could not verify their identity; four newborn individuals escaped through holes in the water outlet of the tanks; and four individuals died as newborns. All results in the manuscript are on these 26 animals, though including data from all 40 (e.g. patterns of individual variation on the first day post birth) did not change the results or their interpretation (see Supplementary Table 2).
Behavioral tracking
We developed a custom recording system using Raspberry Pi computers, which are an upcoming low-cost, highly adaptable solution for many applications in the biological sciences25. Specifically, we created a local network of Raspberry Pi 3B + ’s, each connected to a Raspberry Pi camera positioned exactly above an experimental tank, commanded by a lab computer, and connected to the server on the institute network (Supplementary Fig. 1). We programmed the Raspberry Pi’s using pirecorder26 to take timestamped photos every 3 s across the daily light period, each day, for 10 weeks, and store them automatically in dedicated, automatically named folders on the server. Image settings and resolution were thereby optimized to minimize file size while assuring image quality. After the experimental period, we created videos of all the recorded images of each fish of each day. These videos were subsequently tracked with the Biotracker software27, using background subtraction, providing the x, y coordinates of each fish in each frame. We then processed the data, including scaling and converting the coordinates to mm, and, for each frame, computed fish’s swimming speed (cm/s) and distance from the tank walls (cm). We then summarized these variables both on an hourly and daily basis to compute fish’s median swimming speed, inter-quartile range of swimming speeds, activity (proportion of time spent moving >0.5 cm/s), and median border distance. To quantify fish’s body size over time, we randomly selected five photos per week of each compartment, making sure the fish was away from the compartment walls and did not show strong body curvature, and then used ImageJ software to measure total body length (mm) from the tip of the snout to the end of the body. By averaging the measurements of the five images, we acquired one body size measurement per week.
Error checking
We collected up to 924,000 photos on each individual throughout the experimental period resulting in a total of over 24 million data points collected on our experimental animals (N = 26 individuals). To ensure that our tracking software accurately captured the behavior of our fish, we checked for potential tracking errors in two ways. First, we estimated overall error rates. To do this, we selected at random a starting frame from within a day; then we manually checked each of the subsequent 200 frames and identified whether an error was made (fish was not properly located by BioTracker) or not (fish was properly located) by visual inspection of the videos. We estimated the error rate as the number of errors divided by the total number of checked frames. The overall median error rate over the entire observation period was estimated to be 7%. Error rates increased earlier in the observation period when the fish were smaller (Supplementary Note I). As such, as a second step, we manually went through and corrected all frames for the very first day of tracking (i.e. day 1 post-birth) for all fish (~13,200 frames per individual) as this is a critical time period for one of our research questions. This ensured that the resulting behavioral data were completely accurate for this day. This manual correction allowed us the additional opportunity to compare how well our automatically tracked (i.e. not manually corrected) data performed compared to the manually corrected data. We found that the automatically tracked data re-created near identical estimates of among- and within-individual variance components and most importantly the among-individual correlation between the automatically tracked and manually corrected data was over 0.98 for our behavioral variables (Supplementary Note I). This strongly suggests that any errors introduced by our automated tracking software have minimal influence of our behavioral variables at best and do not affect our interpretation of the results.
Statistical analyses
We used linear mixed, or hierarchical, models to partition the behavioral variation across different times periods into its among- and within-individual components. Throughout we focused our analysis on the 26 individuals for which we had complete data for the entire 10-week observation period to ensure comparable variation over time and across models.
Our first question of interest was to test when individual differences in behavior first appeared over the course of the experiment. We started by investigating behavior on the first day post birth (Fig. 1A, Supplementary Table 2) and then planned to proceed in a day-by-day fashion until significant repeatability in behavior was apparent (Supplementary Table 3). We used hourly median swimming speed (11 observations for each of 26 individuals) as our response variable and included ‘hour’ and ‘total length (TL)’ as fixed effects and ‘individual’ was included as our random effect of interest. Including TL as a covariate allowed us to test whether behavior was related to an offspring’s body size on its first day of life. We set the first hour of the day as 0 and mean-centered TL as this would allow the among- (and within-) individual variance components to be estimated at these values (i.e. the earliest possible moment from when we could record behavior in the fish). We estimated the adjusted repeatability of median swimming speed as the variance attributable to individual identity over the total variance not explained by the fixed effects. We additionally estimated both marginal and conditional R-squared values which estimate the variance explained by the fixed effects only and the variance explained by the fixed and random effects combined, respectively. As our individual experimental fish came from different mothers, we first explored a number of different variance structures including random intercepts and slopes for both individual ID and maternal ID. This allowed us to test whether maternal identity explained variation in individual behavior. However, the most supported model included random intercepts and slopes for individual ID and not for mother ID, indicating that our methods to reduce variation among mothers were successful (Table 1). We used median swimming speed as our behavioral variable of interest throughout the main manuscript, as this behavior was tightly correlated with most of our other behavioral variables (Supplementary Fig. 2); though results using the other behavioral variables yielded the same interpretation (i.e. that significant individuality in (any) behavior was present on the very first day post-birth; Supplementary Table 2).
Our second research question was to investigate how individual behavioral variance changed over the course of the entire observation period (70 days). Again, we first explored several different variance structures to test the importance of maternal identity and/or individual identity on behavioral variation. We found support for the inclusion of random slopes at the individual level, but not maternal level (Table 1). This indicates that levels of among- (and within-) individual variation may differ throughout the observation period. To investigate patterns of change in the variance components, we ran a series of models where we centered the observation covariate on different days. Individual intercepts are estimated when all covariates are set to zero, so this allowed us to ‘slice’ the data to estimate the among- and within-individual variance at different time points over the ten weeks. We ran 11 models as we chose to center the data every 7 days (first model was centered on observation 1; 11th model was centered on observation 70). The predicted individual intercepts (best linear unbiased predictors) and estimated variance components from each model are plotted in Fig. 3.
We also closely investigated any potential influence of body size and/or growth rate differences on behavioral expression and individual behavioral variation in this entire 10-week data set. First, we estimated the repeatability of both weekly total length and weekly growth rates to determine if individuals consistently differed in these traits. Then, we ran a series of models with median weekly swimming speed as the response variable and included either weekly total length, weekly growth rate, and/or overall growth rate (estimated over the entire 10 weeks), as our fixed effects of interest. Each model also included the random effects of individual intercepts and slopes. Finally, because body size varies both among individuals (some individuals are on average larger than others) and within individuals (as they grow), we also performed within-individual centering of total length. In this fifth model, we included each individual’s average total length and their weekly deviation from their average length as the two fixed effects of interest. Individual identity and slopes were included as random effects. For all models, we estimated the variance explained by the fixed effects (marginal R2) and the fixed and random effects together (conditional R2). These results are reported in Table 2.
For our third and final research question, we tested whether early-life behavior predicted later-life behavior. To test this, we estimated the among-individual correlation (including ‘individual ID’ as our random effect) in behavior using multivariate mixed models where the daily median swimming speeds in each week were the response variables (7 observations per week per individual; 10 weeks total; Fig. 4A). Then to investigate how the strength of these correlations may change over development, we used a linear model to test whether the correlation strength was predicted by the interaction between the first week included in the correlation and distance to the next week in the correlation (1, 2, 3, 4 or 5 weeks away in time; Fig. 4B).
All models were performed using Markov Chain Monte Carlo estimation with the MCMCglmm package38 in R v3.6.139. We set our models to run 510,000 iterations with a 10,000 burn-in and thinning every 200 iterations. To ensure proper model mixing and convergence, we initially ran 5 independent chains and inspected posterior trace plots of parameter estimates (Supplementary Note II). In a preliminary analysis we tested three different prior settings (Supplementary Note II); results did not change with prior settings so we chose parameter-expanded priors for all models reported here as these are generally considered to be more robust. An R Markdown file with all the results presented here is included in Supplementary Note II.
Reporting summary
Further information on research design is available in the Nature Research Reporting Summary linked to this article.
Source: Ecology - nature.com
