Coordinated gas release among the physostomous fish sprat (Sprattus sprattus)

Previous experimental studies suggest that the production of sound associated with expelling gas from an open swimbladder may play a role in communication. This would suggest non-random gas release. We used deployed echosounders to study patterns of gas release among a fjord population of sprat (Sprattus sprattus). The echosounder records concurrently revealed individual fish and their release of gas. The gas release primarily occurred at night, partly following recurrent temporal patterns, but also varying between nights. In testing for non-randomness, we formulated a data-driven simulation approach. Non-random gas release scaled with the length of the analyzed time intervals from 1 min to 6 h, and above 30 min the release events in more than 50% of the intervals were significantly connected.

www.nature.com/scientificreports/ heard from a few fish swimming in loose aggregation was accompanied by a brief pause by all individuals 17 . The captive schooling herring showed, however, no behavioral response to playbacks of such sounds during daytime 17 . Solberg and Kaartvedt 9 applied submerged, upward-facing echosounders to study sprat during overwintering in Oslofjorden, Norway. Gas bubbles are excellent acoustic targets and gas release became clearly depicted in the echograms. The sprat carried out rapid and short excursions to the surface, apparently to fill the swim bladder. Each fish surfaced an estimated 3-4 times daily, with ~ 70 daily burst of gas release per fish. Both types of events were nocturnal. Here, we analyze data from the field campaign of Solberg and Kaartvedt 9 in more detail, using gas release as a proxy for sound production. The fjord froze over from January to April and we analyze data from ice-free conditions in early winter. Specifically, we tested the hypothesis of connected gas release and associated sound production among fishes, expecting that gas release would appear in non-random patterns if related to communication. We compare the observed patterns with model results for the outcome in the case of random releases. For this, we formulate a model assessing connectivity among gas bubble releases allowing for a formal test of non-randomness.

Results
Patterns of gas release. The gas release had a strong non-random component with marked daily peaks, confined to the night (Fig. 1A,B). The nocturnal discharge of gas largely followed recurrent patterns, but also changed during the registration period. To illustrate the main patterns within days, we here combine results on hourly discharge from days 1-6 and 10-18, respectively (Fig. 2). During the first period, a minor initial pulse close to and after sunset (14:56-14:45 GMT) was followed by an exponential increase in gas discharge after midnight, peaking 1-2 h prior to dawn (sunrise at 07:07-07:19 GMT during this first period). During the second part, the afternoon peak was larger, and the subsequent pattern more dome shaped, with an apparent slight decline in gas release towards dawn (Figs. 1, 2). However, such averaging between nights conceals some marked variations, including a strong peak just after sunset on 29 Nov, (day 18; ref Fig. 1), this day with particularly low release of gas bubbles at the end of the night.
The test for connectivity among gas releases revealed increasing rejection of the null hypothesis of random release when increasing the time window from 1 min to 6 h (Fig. 3). With the S-CON test developed for this study (see "Methods"), the average rejection increased from 0% at 1-min intervals to 93.8% at 6-h intervals (assuming a connectivity interval of 30 s, Fig. 3). Furthermore, the connectivity was non-random at the 5% significance level in more than 50% of the windows equal to or longer than 30 min.   www.nature.com/scientificreports/ Relationship between gas release and fish abundance. To exclude the possibility that apparent connectivity would be a mere result of fluctuating fish abundance, we tested whether the number of released bubbles is a function of fish biomass, as assessed by acoustic abundance estimates. The general least squares analysis revealed a significant relationship between gas release and fish abundance as measured by the surface integrated backscatter coefficient (F 1,420 = 23.37, p < 0.001). However, fish abundance only explained approximately 12% of the variation in the release events.  Figure 3. The percentage of the time intervals where the observed connectivity of the release events was determined to be non-random. Yellow numbers represent the number of analyzed time intervals with 2 or more release events for each interval length. The connectivity interval is the period within following gas releases were considered to be connected.

Discussion
The sprat primarily released gas at night, in accordance with previous experimental studies of Pacific herring 4 . Our records did not cover the lower part of the water column, yet nocturnal gas release was confirmed by simultaneous measurements encompassing the vertical range of the entire sprat population (see "Methods"). Most discharge events were significantly connected at any time scale exceeding 30 min (Fig. 3). We therefore reject the null hypothesis of random gas release and interpret the results as strong support of the alternative hypothesis of non-random release. Thus, this finding from this field study of sprat is in line with conclusions from the experimental studies of pacific herring 4 . Rejection rates for the null hypothesis become less, the smaller the considered time window becomes. Within short time windows, even purely random points will tend to form peaks because of lack of space. Consequently, the statistical tests at this time scale will only be able to separate highly dependent short-sequenced release events from a random process. Nevertheless, even among the 754 5-min intervals which had more than one release event, 17% appeared with dependency between the gas bubble releases. This is 3.6 times more than expected at the 5% level if all the 5-min intervals contained purely random release.
There is predation risk associated with filling the swim bladder at the surface, manifested in that the sprat only surface for refills at night 9 . Continued gas release requires repeated access to surface air each night 4,9 , thus apparently increasing mortality risk. A particular urge to refill the swim bladder appeared when the study site froze over later in the same winter. The frequency of surfacing increased markedly after the fjord became ice covered, with apparent frantic, yet unsuccessful search for air under the ice as the records of gas release decreased sharply 9 . Evidently, gas is a valuable resource. This inference is underlined by the swim bladder of herring being equipped with guanine crystals, likely to hamper leakage of gas 3 .
Sprat is negatively buoyant at depth, manifested by a continuous sink-and-rise swimming 18 . Although fat plays a significant role for the buoyancy of clupeids 7 , the deliberate release of gas probably involves a metabolic cost for the negatively buoyant sprat. Other costs are conceivable, given the multitude of swimbladder functions [1][2][3][4] . Nevertheless, the sprat consistently released gas in spite of the costs, signifying there is a gain in gas release.
The pulsed release patterns suggest that release was either stimulated by common external factors, or that the gas release-considered here as a proxy for sound production of the fishes-stimulated each other. Solberg and Kaartvedt 9 could not identify any external potential trigger for the gas release, but we cannot exclude the possibility of individuals reacting to an external stimulus that remains to be identified. The sprat in the Oslofjord exhibit diverse DVM patterns, including "mid-night sinking" and an intermittent dawn ascent prior to sunrise 18,19 . During the current study, dawn ascent was apparent the first weeks of November. The timing of gas release suggests some association with this event (cf. Fig. 1). However, the sprat also released gas at other times of night. The nocturnal distribution became shallower by the end of November 9 . The sprat then released gas in a nonrandom fashion throughout the night (Fig. 1). While fish abundance expectedly would affect the number of gas release events, it could only explain 12% of the variation in the data so that fluctuating fish abundance in the acoustic beam did not explain the non-random gas releases observed in our data.
Reciprocal stimulation of gas release would be in accordance with the experimental studies by Wilson et al. 4 . A filled swim bladder holds potential for producing sound, which Wilson et al. 4 suggested herring might use for keeping contact in darkness. More than 800 fish species are known to produce sound 20 and mounting evidence suggests that (intraspecific) acoustic communication can affect their success in various ways 21,22 . However, knowledge about the ecology of sound communication is still very limited 23 . The observed non-random connectivity of gas release-i.e. sound production-recorded here concords with the hypothesis of communication purposes. Further research is needed to provide first hand evidence of acoustic communication through gas release.

Methods
Study area. The study was carried out in Bunnefjorden, Norway. The fjord froze over from January to April and we here analyze data from ice-free conditions in early winter (12 Nov to 2 Dec 2009). Bunnefjorden is a 150 m deep inner branch of the Oslofjord, and a 57 m deep sill at the entrance restricts water exchange with the outer part of the fjord. Klevjer and Kaartvedt 24 provide a map of the study area. The fjord branch normally becomes hypoxic in the lower part of the water column. During the current study, oxygen contents were 2-3 ml l −1 between 15 and 60 m, while waters below 70-80 m were severely hypoxic and devoid of fishes 9 .
Studies of overwintering sprat have been undertaken in Bunnefjorden during several winters, and the biology of sprat as well as the identity of the main acoustic targets in the fjord are well established 9,18,24 . In the winter of the current study, catches from 33 trawl samples were dominated by sprat; with ~ 40 times higher catches than the next most abundant species, herring (Clupea harengus) 9 .
Study design. Solberg and Kaartvedt 9 and Solberg et al. 18 provide details on methods, and we here only give a summary of the acoustic setup. In short, upward-looking Simrad EK 60 echosounders kept in pressure-proof casings were deployed at the bottom (150 m) and in buoys (80 and 30 m) for enhanced resolution in shallower part of the water column. Cables for electricity and transfer of data to a PC on shore enabled continuous operation of the systems. We here use the data from the shallowest echosounder (200 kHz) that provided superior resolution in near-surface water, though did not cover the full depth range of the population distribution. Echograms from the deeper located echosounders covering the whole (inhabited) water column and showing the full diel population behavior are given in Solberg and Kaartvedt 9 and Solberg et al. 18 .
Records of gas release. Released gas appeared as ascending lines in the echogram (Fig. 4). We quantified the release as explained by Solberg and Kaartvedt 9 . We only included ascending traces connected to the acoustic record of a fish, but without enumerating the release per individual fish. Since the same fish may release several www.nature.com/scientificreports/ bursts of bubbles within a short time interval, we here pooled any sequences of gas release within a 10-s period as one event. This procedure will also exclude cases with several different individuals releasing bubbles in the course of this short time interval, yet we chose this conservative approach not to generate an artificial high connection of gas releases between the fishes.
Analyses of data. The frequency of gas releases varied with time, both within a day and between the weeks.
Such patterns compare to service systems like call centres and hospital emergency rooms 25 that can be modelled as a Poisson process 26,27 . We therefore started our analysis with the statistical procedure suggested by Brown et al. 28 in their influential analysis of the call dynamics in a banking call centre. The first step is to subdivide the day into time intervals, which are short enough to consider event rates as approximately constant. Here we chose to investigate alternative periods of respectively 1, 5, and 30 min, as well as 1, 2, 4, and 6 h. At the longest interval, the peaks in the gas release intensity are expected to be the result of a non-constant Poisson parameter, and therefore more likely to induce rejection of the null-hypothesis of a homogenous random process. In contrast, we expect to find higher concordance with a random process for the short intervals of 5 min. In assessing connectivity among gas bubble releases, we formulate a new model allowing for a formal test of non-randomness (summarized in Fig. 5). We name this approach the simulated connectivity test (S-CON test), which we implemented in R 29 , with the code being available in the Supplementary appendix. If there is a common physiological reason or some form of communication among sprat, a burst of gas release is likely followed by subsequent releases. Thus, it is reasonable to assume that the total number of releases within a short time interval like 30 s would be effective in detecting dependencies between the releases. We therefore define the concept of connectivity as follows: Let the gas be released at times T 1 , T 2 , …, T n and define the connectivity at each event as the numbers of records within the following 30 s. The average connectivity in any considered window of the investigated timeperiod (for example a window of 1 h) is defined as the average connectivity of all cases of connectivity within the considered window (see also Fig. 5).
In order to test the null hypothesis of no dependency between gas bubble releases, we compare the measured average connectivity in the data set with the simulation of 1000 random placements of the total number of observations in a given time window. For example, if we consider a window of 30 min with 15 release events having an average connectivity of 2.1, we performed 1000 random placements of 15 points between 1 and 1800. In this way, we get 1000 simulated values of the average connectivity, from which we pick out the critical 95th percentile, following the common significance level of 0.05 in biology. If the observed average connectivity is larger than this critical value, we reject the null-hypothesis and conclude that the releases of gas bubbles are dependent random variables. Thus, if our example obtains a critical value of 1.7, the null-hypothesis of random arrival times of bubbles is rejected (because the observed value of 2.1 is larger than the critical value of 1.7).
Since a dependency between the fish will induce a higher concentration of release events than produced by random releases, we expect the average connectivity to be quite sensitive to the alternative hypothesis of dependent arrival times. Also, note that the concept of connectivity has a combinatory nature, so we need only require that the considered window contains at least two releases of gas bubbles. In contrast, alternative approaches using Kolmogorov-Smirnov tests 28 are based on the cumulative distribution function and therefore require at least five observed bubble releases.
To test the dependency of the results on the chosen time interval, we also ran the analysis using connectivity intervals of 25 and 35 s, which revealed some variability to the estimates of non-random bubble release (Fig. 3) but did not influence the general pattern. We also tested whether the interval within which we consider subsequent bubbles to be part of one single release event influences our results. The more we consider sequential   Average connectivity Distribution of simulated connectivities 95th percentile comparison to the 5% critical value Figure 5. Illustration of the steps related to the simulated connectivity test (S-CON test). Bubbles occurring within 10 s are pooled into single release events. We then determine the connectivity of each release eventaka the number of release events within the following 30 s time window. From these, we calculate the average connectivity for a specified period (1, 5, 30 min, 1-, 2-, 4-and 6-h intervals). In a final step, we compare the observed average connectivity to the critical value which is defined as the 95th percentile of 1000 simulations of random placements of the same number of releases. If the observed connectivity is larger than the critical value, we reject the null hypothesis of random gas releases for the specified time interval.