Adiposity signals predict vocal effort in Alston's singing mice

Abstract

Advertisement displays often seem extravagant and expensive, and are thought to depend on the body condition of a signaller. Nevertheless, we know little about how signallers adjust effort based on condition, and few studies find a strong relationship between natural variation in condition and display. To examine the relationship between body condition and signal elaboration more fully, we characterized physiological condition and acoustic displays in a wild rodent with elaborate vocalizations, Alston's singing mouse, Scotinomys teguina. We found two major axes of variation in condition—one defined by short-term fluctuations in caloric nutrients, and a second by longer-term variation in adiposity. Among acoustic parameters, song effort was characterized by high rates of display and longer songs. Song effort was highly correlated with measures of adiposity. We found that leptin was a particularly strong predictor of display effort. Leptin is known to influence investment in other costly traits, such as immune function and reproduction. Plasma hormone levels convey somatic state to a variety of tissues, and may govern trait investment across vertebrates. Such measures offer new insights into how animals translate body condition into behavioural and life-history decisions.

1. Introduction

Male sexual displays are among the most conspicuous examples of animal signalling. From the distinctive ‘foot-dancing’ of Plethodon salamanders [1] to the squeaking dives of Anna's hummingbirds [2], advertisement displays seem designed to be noticed. In many species, the exaggerated displays also increase reproductive success [3]. Despite such rewards, not all males display equally. One common explanation suggests that signals impose costs that not all males can afford to pay [4–6]. Some costs are probabilistic, like the risk of injury or energy spent in fight or flight, while others are inescapable, like the nutrients required to properly develop and express the structures and behaviours of a display [5,6]. Signalling costs thus play an important role in sexual signalling. In many models, a link between signal and condition leads to selection on female mate choice [7]. The topic of condition-dependent signalling, however, is much broader than any one hypothesis regarding female mate choice. Regardless of what a sexual signal might convey—genetic quality, health or disease, or simply a male's presence against a noisy background—all males must evaluate how much to invest in a costly display. But how do signallers use condition to modulate their investment in display? This question is central to understanding signalling behaviour, but remains remarkably under-studied.

One limit to our understanding of signalling decisions derives from our incomplete characterization of condition itself. While condition can include a variety of dimensions (e.g. integrity of body or cellular processes [8], epigenetic state [9], resistance to pathogens [10]), one common component is the combination of energy balance and nutritional status that determines the ability of a body to carry out future functions. This is often interpreted as an individual's pool of available energetic resources [11], and in vertebrates is most often measured by estimation of body fat or size, such as residual body mass (RBM) [12]. While these metrics have helped to conceptualize condition, they are relatively coarse proxies for available resources (e.g. in rodents [13]). The indirect nature of these measures may explain why a variety of studies find that condition is a poor predictor of sexual signals [13].

Moving beyond gross measures of weight and length requires consideration of specific physiological parameters of condition. For example, feeding is regulated by a variety of gastrointestinal (GI) hormones released in response to fluctuating levels of circulating nutrients (e.g. proteins, carbohydrates, fats [14,15]). Although transient changes in circulating nutrients or GI hormones may shape display effort, the ability to pay costs on a longer time scale is likely to be signalled by hormones more directly related to storage of energy in fats and carbohydrates. Mammalian regulatory hormones like leptin, adiponectin and insulin are influenced by both sustained food consumption and adipose tissue mass size [14]. Thus, alternative metrics of condition might incorporate measurement of a variety of nutrients and hormones that could capture fluctuation in current nutritive status, or provide information about long-term energy balance. Research on laboratory rodents offers insights into the physiological indicators of condition, but little is known about how physiology affects ecologically relevant behaviour in wild populations.

In this study, we examine the relationship between body condition and display effort in Alston's singing mouse, Scotinomys teguina. Alston's singing mice are small (10–16 g), terrestrial, promiscuous rodents, inhabiting montane cloud forests of Central America [16]. Singing mice are both diurnal and insectivorous, unusual behaviours for murid rodents [16], but their most remarkable behaviour is the trilled singing for which they are named. Most mice produce ultrasonic vocalizations [17,18], but the Scotinomys song is far more elaborate (figure 1). The vocalization comprises a series of rapid, high-bandwidth frequency sweeps that span sonic and ultrasonic frequencies (10–43 kHz). These songs can be extraordinarily long (up to 16 s) and seem to be substantially louder than other murid rodents [16,19]. Because lower frequencies are less readily absorbed by vegetation [20], this suite of characteristics suggests a substantial increase in the active space of the song relative to its close relatives. Females show preferences for male songs with higher note rates [21]. Moreover, the decision to sing is influenced by the presence of females, aggressive experience, reproductive status and perceived predation risk [21–23]. Individual differences in song performance, measured by the ability to increase both frequency bandwidth and note rate [24], are weakly correlated with coarse measures of body condition [21]. This natural history suggests that the singing mouse will be a productive model for exploring the decision mechanisms by which males allocate display effort.

Figure 1.

Structure of Scotinomys song. (a) Spectrogram of a typical male song. Lower panels depict 0.2 s of song at 1 second, 3 s and 5 s into a song. Internote interval (INI) is calculated as time between the ending of one note and the beginning of the next. (b) Parameters measured during song analysis. Panels depict a close-up of a note at 3 s (shaded box in (a)). Left, spectrogram detail of the note with max and min frequency labelled. F_max − F_min was used to calculate note bandwidth. Center, how note frequency modulation (FM) was measured. The frequency with the highest energy was measured in each time bin (open circles), and a quadratic curve was fit to the resulting data. Fma is the curvature of a note, FMb its slope, and FMc its starting frequency. Right, waveform detail of the note with peak amplitude (peak amp), peak amplitude within first quarter of a note (qpeak amp), and note duration (note dur) labelled. Each of the parameters described varied systematically as a function of note position within a song, so note position and note parameters were related to one another with a quadratic function capturing note-to-note variation (see Material and methods electronic supplementary material, table S1). (Online version in colour.)

In this study, we examine natural patterns of variation in song and condition in wild Alston's singing mice. We first characterize the dimensions of acoustic variation in the songs of male singing mice, examining the number of songs, latency to sing, length of songs and an exhaustive variety of measures of amplitude and frequency modulation. To explore the dimensions of body condition, we examine a variety of circulating plasma nutrients and metabolic hormones, as well as the more traditional measure of RBM. Finally, we ask how well dimensions of condition predicted dimensions of vocalization. By doing so, we hope to gain a more nuanced understanding of the relationship between advertisement displays, body condition and the roles of physiological cues in the coordination of the two.

2. Material and methods

(a). Study subjects

Singing mice were captured in Sherman live traps in San Gerardo de Dota, Costa Rica. We collected 53 adult male mice and brought each into the Quetzal Education and Research Center (QERC) for processing. We housed each subject male singly in 28 × 28 × 28 cm³ PVC-coated wire-mesh cages and provided them with cat chow and water ad libitum. In a variety of taxa, capture and handling can cause spikes in stress hormones and heart rate that persist for a few hours [25]. Stress reactivity measures are associated with singing behaviour [23], for example, suggesting that capture and handling might influence our findings. To mitigate this, we gave all subject males at least one day of acclimation after day of capture before recording and sacrifice (mode = 1 day; mean ± s.d. = 2.05 ± 1.29 days). Although this may influence observed behavioural and physiological parameters, our emphasis is on differences among males treated equivalently, and this design allowed high-quality acoustic recording that would not have been feasible outside of the laboratory.

We standardized hind foot and tail measurements by computing Z-scores for each male. We then took the mean of each male's tail and hind foot Z-scores to obtain a composite metric for skeletal length. To determine RBM, we regressed mass on skeletal length and calculated the residual error (in grams) for each mouse [26].

(b). Acoustic recording and analysis

(i). Song recording

We moved subject males in their home cages into a 42 × 42 × 39 cm³ acoustic isolation chamber the evening before recording. The chambers were constructed of expanded PVC, lined with anechoic foam (described in detail in [21,27]). At dawn on the morning of recording, the chambers were closed. To minimize perceived predation risk, we used a dim red LED light to illuminate chambers during recording. We first recorded songs made by subject males in the absence of stimuli for 2 h. Following 2 h recording in silence, we recorded evoked responses of males to each of four stimuli: a conspecific song, a heterospecific song (S. xerampelinus, a congener species [27,28]), noise or silence. Each stimulus within a treatment was presented once a minute for 3 min. Stimulus sets were separated by 20 min and provided in random order. Playback lasted 2 h and was video recorded. Thirty-eight of 53 mice sang within the 4 h of recording. All recording was accomplished between 0500 and 1100 h.

All vocalizations were recorded at 32-bit, 195.3 kHz resolution with ACO Pacific microphones on Tucker-Davis RX6 hardware. We assessed recording quality by visual inspection of spectrograms at time of recording, and low-quality or truncated recordings were excluded from analyses.

(ii). Song analysis

Spectral and amplitude properties of all recorded songs were assessed in Matlab (code available at www.phelpslab.net). Song lengths were bimodally distributed; we operationally define full songs as those at least 4 s long. Only full songs were considered for analysis. Initial analyses did not reveal significant acoustic differences between songs produced in response to stimuli and spontaneous songs, so we pooled all songs for analysis. Song analysis for the singing mice has been previously described [27], but we briefly summarize our measures here.

Because the position and orientation of mice at the time of singing can vary, absolute amplitude measures were not meaningful. Thus we normalized all recordings to have a peak amplitude value of 1 [27]. We measured four amplitude and timing variables for each note in a song : internote interval (figure 1a), note duration, time to peak amplitude and time to quarter-peak amplitude, defined as the peak amplitude within the first 25% of a note (last panel of figure 1b). Each note in a Scotinomys song is a simple descending frequency sweep. To measure the frequency modulation of individual notes, we performed an FFT for each note. We found the frequency with the maximum energy in each time bin of the note, then fitted a quadratic curve to the resulting frequency sweep (middle panel of figure 1b; FMa corresponds to the curvature of the note, FMb its initial slope and FMc the starting frequency). These four amplitude and three frequency parameters result in seven measures of each note. We find that notes change systematically over the course of a song, getting longer and louder as the song progresses. For each of these seven measures of each note, we related note properties to the note position within a song using a quadratic equation (figure 1; electronic supplementary material, table S1). These 21 measures, combined with the number of notes, are sufficient to synthesize a song, and so represent an exhaustive suite of measures. To these we added a number of measures of the entire song. These include the spectral entropy of the song (entropy is a measure of chaos; a greater entropy song is noisier and more broadband), the song length (in seconds), average spectral bandwidth across notes (note bandwidth = F_max − F_min; first panel of figure 1b), and dominant frequency (electronic supplementary material, table S1). Lastly, we recorded the total number of songs sung spontaneously, the number of songs sung in response to playback and the latency to begin singing in response to playback. If a male did not sing during a playback bout, we recorded his latency to sing as 22 min (1320 s), the duration of each playback bout. To reduce the number of tests performed and simplify song analyses, all measures were entered into a principal component analysis (PCA), described in the Statistical analysis section below. Non-singers were excluded from PCA.

(c). Measures of hormones and nutrients from plasma

Subject males were euthanized between 1400 and 1700 h on the day of recording. The decision to sacrifice animals within a 3 h window minimizes the effect of fluctuations due to circadian rhythms [15]. Hormone concentrations can change drastically in response to the stress of handling, but studies have found that blood sampled within 2 min of handling reflects baseline concentrations in a variety of taxa [25,29]. To minimize plasma stress response, we collected trunk blood within 2 min of sacrifice in heparinized capillary tubes. We then centrifuged blood for 5–10 min at 1100g, and plasma samples were extracted and stored at −20°C until analysis. Plasma assays were completed by the Mouse Metabolic Phenotyping Center at the University of Cincinnati, Ohio (www.uc.edu/labs/mmpc.html). We requested measures of plasma levels of cholesterol, phospholipids, non-esterified fatty acids, glucose and triglycerides, as well as the hormones adiponectin, ghrelin, insulin and leptin. As the amount of plasma collected at sacrifice varied, not all subject males had a sufficient volume to complete all assays; as a result, sample sizes vary across plasma assays (table 1). Thirty-eight mice had plasma sufficient for at least one test; 26 had measures for all but insulin and leptin; 16 had all measures. Ghrelin levels were non-detectable (presumably due to the lability of the hormone [15]), and were excluded from all analyses.

Table 1.

Summary of variation in body condition.

condition variable		n	mean	s.d.	CV	min	max
glucose	mg dl−1	38	121	74.1	61.3	14.7	378.5
triglyceride	mg dl−1	35	223	142.6	63.9	52.2	718.3
cholesterol	mg dl−1	32	220	64.6	29.4	18.5	340.7
phospholipids	mg dl−1	32	347	106.4	30.6	41.9	526.4
NEFA	mg dl−1	30	0.66	0.2	31.8	0.3	1.3
adiponectin	ng ml−1	26	153	52.5	34.2	53.4	324.2
leptin	pg ml−1	16	1620	839.6	51.8	9.0	2967.1
insulin	pg ml−1	16	636	723.5	113.7	69.0	2814.1
mass	g	53	12.5	1.2	9.5	10.4	15.7
hind foot	mm	53	16.4	0.6	3.5	15.5	18.0
tail	mm	53	56.2	2.7	4.8	49.0	62.0
RBM	g	53	0	1.2	n.a.	−2.4	2.8

(d). Statistical analysis

(i). Acoustic variation

We characterized axes of acoustic variation by PCA of 29 variables: three descriptors of propensity to sing, five whole-song measures (e.g. song length, overall song dominant frequency), nine descriptors of frequency modulation and twelve descriptors of amplitude modulation (electronic supplementary material, table S1). To prevent over-representation of any male with more than one recording, we calculated the average value of song variables over all long songs for these males and used these values in our analysis. We used varimax rotation to help interpret results from PCA. Lastly, we estimated repeatability of measures via intraclass correlation coefficients using the ICC R package (version 2.3.0, 17 June 2015). For these data and throughout the paper, all statistical analyses were completed in R version 3.2.5 (R Core Team, 2016).

(ii). Condition variation

We calculated Pearson correlation coefficients for all condition variables and visualized the resulting correlation matrix in a heat map to determine how individual measures covaried. For this matrix, we use the Benjamini & Hochberg [30] procedure to identify significant correlations corrected for multiple comparisons. For all analyses, we report α as the critical value needed to obtain a family-wise FDR = 0.10.

Because variables were highly correlated, we used PCA to reduce dimensionality and examine the underlying structure of variation in condition. To improve the power of our analysis, we excluded leptin and insulin from the PCA; including these measures would have reduced our sample size from n = 26 to n = 16. Our final condition PCA was described by RBM, five metabolic measures of current nutritive status (cholesterol, phospholipids, glucose, triglycerides and non-esterified fatty acids), and one hormonal measure of long-term energy balance (adiponectin; table 1). We used a varimax rotation to clarify interpretations of the PCA.

(iii). Condition-dependence of song

To examine how elements of mouse song covary with dimensions of body condition, we performed a series of linear regressions. To reduce the number of regressions run and prevent model over-fitting, we focused on the first two PC scores from each of our analyses, and on a small number of variables which had a clear a priori relationship to song effort (total singing, song length) or condition (RBM, signals of adiposity). This also allowed us to include subjects (e.g. subjects for which we measured RBM but lacked one or more plasma measures) or variables (insulin, leptin) that were excluded from the PCA due to missing data. We also constructed an additional condition composite score by averaging Z-scores for RBM, insulin, adiponectin and leptin for animals that had at least two of these measures. We compared this composite score to the first two dimensions of song variation defined by PCA. Finally, we ran a bidirectional stepwise regression to determine the relative contributions of RBM, insulin, adiponectin and leptin. Interaction terms that did not significantly improve AIC scores were deleted.

3. Results

(a). Variation in singing behaviour

We caught 53 adult male singing mice, and successfully obtained 163 unabridged and clear recordings from 39 of the 53 mice. We retained 124 full songs from 38 adult males for analysis. Among males that sang, we obtained 3.21 ± 2.03 songs (mean ± s.d.) per male. Songs ranged in length from 4.6 to 9.7 s, mean 6.80 ± 1.04 s. Intraclass correlation coefficients are reported in electronic supplementary material, table S2.

We measured the propensity to sing in three ways: the number of songs produced in two hours in the absence of stimuli (spontaneous rate: n = 53, range = 0–9 songs, mean ± s.d. = 1.87 ± 2.18 songs); the number of songs produced in two hours in response to playback stimuli (response rate: n = 53, range = 0–7 songs, mean ± s.d. = 1.74 ± 1.68 songs); and the average latency to begin singing during playback experiments, in seconds (latency: n = 53, range = 240.2–1320.0 s, mean ± s.d. = 971.5 ± 318.2 s).

The song PCA generated seven components with λ > 1.0 and two with λ > 5.0. We retained the first two components (44.96% of variance). To make interpretation of PCA results more intuitive, we report the additive inverse of PCA results (i.e. we changed the sign of all results [31]). The first component explained 25.96% of overall variance with λ = 7.53. Song length, total songs, the note number, amplitude modulation and the beginning note duration loaded strongly and positively on PC1 while internote interval_a and note duration_b loaded strongly and negatively on PC1 (electronic supplementary material, table S3, figure 2a). Males that scored strongly and positively on PC1 produced more songs, were quicker to begin singing, made longer songs and produced songs with shorter inter-note intervals than males that loaded strongly and negatively on PC1 (figure 2b,c). We interpreted this component as individual differences in ‘song effort’.

Figure 2.

Variation in Scotinomys song. (a) Biplot resulting from PCA showing how song variables loaded in PC space. In red are parameters that loaded strongly on PC1, or ‘song effort’. In blue are parameters that loaded strongly on PC2, or ‘frequency modulation’. (b) Individual singers plotted on principal component space. Songs that loaded strongly positively or strongly negatively on PC1 but close to zero on PC2 were selected as ‘PC1-loading’ exemplar songs (red). (c) Waveforms of PC1 exemplar songs (from individuals in red in (b)). Top panels depict waveform of full song; lower panels depict waveform of 0.2 millisecond samples at 1, 3, 5 and 7 s into the songs. The left panels illustrate PC1-negative exemplar song, and the right panels illustrate PC1-positive exemplar song. Abbreviations for parameters provided in electronic supplementary material, table S1. (Online version in colour.)

The second component explained 19.01% of variance (λ = 5.51). Songs that loaded strongly and positively on PC2 began with sharply curved notes that gradually flattened over the course of the song. PC2-positive songs were also characterized by greater entropy, began at lower dominant frequencies, and ended on higher dominant frequencies than PC2-negative songs. We interpreted this component as differences in ‘frequency modulation’ (figure 2a).

(b). Dimensions of variation in condition

We examined nine distinct measures of condition, supplementing RBM (a traditional measure of energy stores) with information from eight plasma measures (table 1). Reference values for all plasma measures fall within the range of those reported in laboratory mice except adiponectin, a hormone released by mammalian adipose tissue (Mus: 3000–15 000 ng ml⁻¹ [32]; Scotinomys: 50–350 ng ml⁻¹).

PCA for retained condition measures revealed three components with λ > 1 (86.24%; figure 3b; electronic supplementary material, table S3). The first component (47.86%, λ = 3.35) included a variety of plasma nutrients—glucose, triglycerides, non-esterified fatty acids, cholesterol and phospholipids. The second component explained (21.95%, λ = 1.54) was dominated by adiponectin and RBM. The third component (16.43%, λ = 1.15) included glucose, NEFA, RBM, cholesterol and phospholipids, with cholesterol and phospholipids loading opposite others variables. Varimax rotation indicated that glucose, triglycerides and non-esterified fatty acids loaded most strongly on PC1, while cholesterol and phospholipids loaded most strongly on PC3.

Figure 3.

Patterns of covariation among measures of condition. (a) Correlation heatmap of condition measures. The upper diagonal displays pairwise correlations, the lower diagonal the absolute values of correlations. All pairwise correlations except glucose and RBM were significant before FDR correction (p < 0.05); correlations that survived the multiple-test correction are denoted with a filled circle. (b) PCA plot showing loading of condition measures. Points have been slightly offset for ease of reading.

(c). Condition dependence of elaborate song

RBM was weakly correlated with multiple measures of singing behaviour, including PC1 of song (‘song effort’: n = 38, R² = 0.09, p = 0.066; figure 4a), latency to sing (n = 53, R² = 0.050, p = 0.11), spontaneous song rate (n = 53, R² = 0.12, p = 0.0094) and evoked songs (n = 53, R² = 0.092, p = 0.027). RBM was not correlated with PC2 of song (‘frequency modulation’: n = 38, R² = 0.0028, p = 0.75).

Figure 4.

Male condition predicts variation in song effort. (a) Residual body mass weakly predicts song effort (R² = 0.09, p = 0.066). (b) ‘Long-term reserves’, PC2 of our condition PCA, predicts song effort. (c) A condition composite score (average Z scores for RBM, adiponectin, leptin and insulin) strongly predicts song effort. Animals with complete datasets are in solid dots; open circles denote animals with fewer than four but at least two measures. (d) Of our plasma hormones, leptin was the strongest predictor of song effort. (Online version in colour.)

Among condition measures, PC1 (‘circulating nutrients’) did not predict either PC1 or PC2 of song (‘song effort’, ‘frequency modulation’), nor any individual measure of propensity to sing (all R² < 0.06, p > 0.25). PC2 of condition, ‘long-term reserves,’ predicted PC1 of song (‘song effort’: n = 19, R² = 0.29, p = 0.016; figure 4b) but not ‘frequency modulation’ (R² < 0.001, p = 0.97).

Because two interesting plasma measures (insulin, leptin) were excluded from our PCA analysis, we asked whether adding these data would improve our ability to predict song effort. We calculated a composite score for body condition based on the average Z-scores of RBM and at least one of our three hormonal measures (adiponectin, insulin and leptin). We found that this composite score was a strong predictor of song effort (n = 27, R² = 0.49, p = 0.00064; figure 4c) but did not predict frequency modulation (R² < 0.001, p = 0.90). Bidirectional stepwise regression produced a model that included RBM, adiponectin and leptin: effort = 2.34 × leptin − 4.09 × adiponectin + 0.95 × RBM (AIC = 26.7). Leptin had the largest coefficient, and removing leptin from the model caused the largest increment in AIC (leptin, AIC = 40.9; RBM, AIC = 33.4; adiponectin, AIC = 32.3). Consistent with this analysis, leptin positively predicted song effort (n = 13, R² = 0.39, p = 0.025; figure 4d) but did not predict frequency modulation (n = 16, R² < 0.001, p = 0.94).

4. Discussion

In this study, we investigated the relationship between body condition and display effort in Alston's singing mouse. We supplemented a traditional field measure of condition, RBM, with information from circulating nutrients and metabolic hormones. We then examined recordings of wild mice and described the dimensions of variation in song structure and display effort. Lastly, we used these data to examine whether dimensions of body condition predict display effort. We found that both song structure and condition varied extensively in the field, and that measures of ‘song effort’ were associated with multiple measures of adiposity. We now discuss our findings in more detail.

Songs varied substantially in all dimensions we examined. PCA revealed two latent variables that explained 45% of between-individual variation. The first component (26%) was explained by the number of songs sung, the latency to sing, the length of songs and a variety of amplitude-modulated parameters related to note durations and inter-note intervals. We interpret this factor as a measure of ‘song effort’. This interpretation is consistent with work in a variety of taxa. For example, it is well documented that song length, call rate, syllable rate and amplitude correlate with motivation and condition in anurans, mammals and birds [33–35].

The second component was explained by variables describing the slope and shape of frequency sweeps, and how they changed over the course of the song. One possible interpretation is that variation in spectral characteristics may signal individual identity. We find, for example, that measures that describe the shape of notes, the dominant frequency of the vocalization and the tonal versus broadband nature of the song (entropy) were the most repeatable song measures (see electronic supplementary material). Indeed, in a variety of vertebrates, spectral characteristics are more stable within individuals and more variable between individuals, allowing for between-individual discrimination (birds [36]; frogs [37]). This pattern also holds for a variety of mammals, including fruit bats [38], foxes [39], coyotes [40], wild dogs [41], weasels [42], ground squirrels [43], deer [44] and humans [45]. Frequency parameters are under the physical constraints imposed by the vocal organs, musculature and other structures necessary for sound production [46,47]. Individual differences in these structures can thus lead to individual variation in spectral components without associated variation in condition. We hypothesize that frequency modulation characteristics of singing mouse song may serve as vocal signatures of identity. This interpretation, however, requires additional experimental tests; for example, habituation/dishabituation tests and other playback experiments could reveal whether individual differences in frequency modulation are being used by audiences of male song.

Energy homeostasis is a complex physiological process regulated by a variety of biochemical signals. Some of these signals act primarily in the short term while others act as long-term regulators—for example, circulating caloric nutrients determine transient states of satiety while certain regulatory hormones are released in proportion to adipose tissue mass and have more gradual effects [14]. Our analyses revealed that the statistical structure of variation in body condition reflected these roles. Variation in condition separated into a ‘circulating nutrients' component, characterized by non-caloric (cholesterol and phospholipids) and caloric (glucose, NEFA and triglycerides) nutrients, and a ‘long-term reserves’ component, comprising RBM and the hormone adiponectin. Patterns of correlation also grouped cholesterol and phospholipids with leptin and insulin. Cholesterol and phospholipids are critical components of cell membranes, but are not used to store energy [48]. We speculate that the correlation of leptin, insulin, cholesterol and phospholipids reflects individual adiposity and the associated ability to invest in somatic growth.

To our surprise, adiponectin did not covary particularly strongly with leptin and insulin. Adiponectin is vital to blood glucose metabolism and fatty acid oxidation, and is found in very high concentrations in mammalian plasma [15]. We find surprisingly low concentrations of adiponectin in singing mice. This may suggest differences in adiponectin sensitivity related to Scotinomys's unique insectivorous diet or its cool high-elevation habitat [16]. It is also possible that Scotinomys may harbour modifications in the adiponectin gene, ADIPOQ, leading to differences in interaction with its receptor, or to changes in the recognition of adiponectin by antibodies used in ELISA multiplex kits. While interest in adiponectin is growing, studies remain confined to laboratory rodents and humans, and its role in ecological variation in diet or energy balance remains poorly understood.

We found RBM, a traditional measure of body condition common in field studies, predicted ‘song effort’ in male singing mice, though only weakly (R² = 0.09). This is consistent with a previous study in singing mice, in which Pasch et al. [21] found a modest but significant relationship between RBM and performance scores of wild-caught singing mice (R² = 0.05). RBM does not accurately predict energy reserves, and a variety of studies have critiqued its use as a measure of condition [13,26]. In studies of bird song, RBM has proved to be a similarly limited predictor [49]. Overall the data suggest that RBM does covary with adiposity, but that the relationship is too indirect to be a powerful predictor of display effort.

By contrast to RBM, hormonal signals of energy balance offer direct insights into the body's assessment of its condition. In addition to its role as an energy storage site, adipose tissue acts as an endocrine organ, secreting regulatory hormones like leptin and adiponectin. Adipose tissue also affects the production of hormones from other organs, such as the secretion of insulin from the pancreas [14,15]. The primary role of these hormones is to signal long-term energy state to peripheral tissues and the central nervous system. We found that inclusion of adiposity signals in our condition metric substantially improved our ability to predict male singing behaviour. ‘Long-term reserves’, a latent variable reflecting adiponectin and RBM, strongly predicted song effort (R² = 0.29). A composite index of condition based on insulin, leptin, adiponectin and RBM greatly improved our ability to predict song effort (R² = 0.49). These data reveal a much stronger relationship between energy balance and display effort than has been previously reported.

One limitation of our approach stems from the need to record songs with sensitive equipment in a controlled setting. For example, ad libitum feeding might reduce individual variation in nutrient concentrations, potentially hindering our ability to detect associations with song. This seems not to have been a major limitation, however, because circulating nutrients defined the most variable dimension of condition. Similarly, the short duration of laboratory housing seems unlikely to have altered adiposity. Thus, our recording conditions seem likely to provide insights into naturally occurring variation in condition and song in the wild.

Of the four signals of adiposity measured in this study, leptin had the strongest individual relationship with song effort. A growing body of experimental evidence from humans and laboratory rodents demonstrates that leptin administration increases caloric expenditure and physical activity [32]. While studies of leptin are often limited to feeding behaviour and physical activity in laboratory settings (e.g. wheel-running), these studies suggest that circulating leptin may govern decisions about energy expenditure into any active behaviour and its ensuing tradeoffs. Here, we find that greater leptin concentrations predicted mice that sang more songs, sang longer songs, and were quicker to begin singing. Males may use leptin to regulate song effort because songs are expensive to produce (e.g. in birds, see [50]; in contrast see [51]). Alternatively, songs may not be costly but may increase the risk of energetically expensive outcomes, such as fight, flight or injury [52].

Our results suggest that well-known neuroendocrine regulators of energy balance may shape male display decisions, and offer new directions for the study of body, brain and behaviour in a variety of taxa. Measurement of these hormones reveals a signaller's assessment of his own condition. Adipokines have been most studied in mammals, but similar hormones are known in all major groups of vertebrates and invertebrates [53–55]. Birds and rodents, for example, have highly similar leptin structures, and the basic mechanisms of leptin's actions on energy balance are thought to have predated the origins of tetrapods [56]. Similarly, multiple hormones can be assayed from as little as 5 µl of plasma [35]. We envision that measurements of leptin and related hormones will be both meaningful and tractable measures of condition in a wide range of study systems. Lastly, these measures also suggest approaches to the neurobiological architecture of display decisions, an area that is poorly understood.

The relationship between condition and display is widely regarded as a major factor in the evolution of animal signals, particularly those used in mate choice and agonistic contests [5,6]. One popular account is that receivers use displays as ‘honest’ indicators of condition [4,6]. Our data offer insights into this perspective by showing that hormonal measures of energy balance are strongly predictive of display. Our findings resemble results by Emlen et al. [57], who find that developmental insulin-like hormones regulate investment in the horns of male dung beetles. These authors interpret the relationship between hormone levels and ornament size as a mechanism that enforces honest signals. In our view, however, this relationship reveals only that signallers judge the display as costly. There are many reasons females may prefer costly displays—to reduce search costs, avoid contagion or choose good genes [6,7]. Finding that males adjust display effort based on energy balance reveals that males are making trade-offs, but does not allow us to discriminate among the reasons costly signals have evolved. Nevertheless, we hope that by refining measures of condition and display, our work will enable more rigorous and sensitive tests of these hypotheses.

In summary, we have shown that an elaborate advertisement display (the songs of singing mice) varies in at least two major dimensions—an ‘effort’ dimension that is related to stable physiological measures of energy balance, and a more idiosyncratic dimension that contained information about identity but not about condition. Among the many physiological measures of condition, leptin emerged as the most predictive of song effort, a pattern consistent with its signalling of fat stores [14]. Because leptin is a widely used hormone [56,58], it is likely to be useful measures of condition in a wide variety of species, and offers a logical focus for the study of the neural mechanisms governing display decisions. By expanding both the mechanistic depth of field studies and the taxonomic scope of physiological work, researchers may gain fundamental new insights into the mechanisms that underlie condition-dependent behaviours, and the processes by which they evolve. Such work will offer new, more integrative perspectives on a timeless topic in animal behaviour.

Ethics

All animal procedures were approved by the University of Texas IACUC (AUP-2013-00178) and the Costa Rican Ministerio del Ambiente y Energia (SINAC-SE-CUS-PI-R-160-2015).

Data accessibility

The data reported in the paper can be found in the electronic supplementary material.

Authors' contributions

T.T.B. and S.M.P. conceived and designed the study. R.R.W. collected field data; T.T.B. collected field data, performed data analysis and drafted the manuscript; S.M.P. participated in data analysis and drafted the manuscript. All authors gave final approval for publication.

Competing interests

We declare we have no competing interests.

Funding

The research was supported in part by grants from American Society of Mammalogists, Texas Ecolab and the EEB Graduate Program at the University of Texas at Austin (T.T.B.) and NSF IOS 0845455 (S.M.P).