J Comp Physiol A (2010) 196:315–320 DOI 10.1007/s00359-010-0514-8 SHORT COMMUNICATION The structure and size of sensory bursts encode stimulus information but only size affects behavior Gary Marsat • Gerald S. Pollack Received: 9 December 2009 / Revised: 11 February 2010 / Accepted: 19 February 2010 / Published online: 7 March 2010 Ó Springer-Verlag 2010 Abstract Cricket ultrasound avoidance is a classic model system for neuroethology. Avoidance steering is triggered by high-firing-rate bursts of spikes in the auditory command neuron AN2. Although bursting is common among sensory neurons, and although the detailed structure of bursts may encode information about the stimulus, it is as yet unclear whether this information is decoded. We address this question in two ways: from an information coding point of view, by showing the relationship between stimulus and burst structure; and also from a functional point of view by showing the relationship between burst structure and behavior. We conclude that the burst structure carries detailed temporal information about the stimulus but that this has little impact on the behavioral response, which is affected mainly by burst size. Keywords Sensory coding
Ultrasound avoidance
Bursting
Temporal coding
Cricket Introduction Echolocating bats hunt flying insects using ultrasonic probes, and crickets, like many nocturnally flying insects, respond to ultrasound stimuli with avoidance responses. G. Marsat Department of Cellular and Molecular Medicine, University of Ottawa, Ottawa, ON K1H 8M5, Canada G. S. Pollack (&) Department of Biology, McGill University, Montreal, QC H3A 1B1, Canada e-mail: gerald.pollack@mcgill.ca AN2 is a first-order, ultrasound-tuned, bilaterally paired auditory interneuron that serves as a command neuron for avoidance steering during flight (Nolen and Hoy 1984). Responses of AN2 to long-lasting, dynamic stimuli consist, after a rapid adaptation phase, of both isolated spikes, separated by relatively long and variable interspike intervals, and of bursts, consisting of groups of spikes separated by short intervals (\6.5 ms). Recent work showed that bursts, but not isolated spikes, in AN2 signal the occurrence of salient peaks in ultrasound amplitude and result in behavioral responses (Marsat and Pollack 2006). Although bursts in sensory neurons are known to function as feature detectors (Gabbiani et al. 1996), it is unclear whether they should be considered as unitary, all-or-none events, or whether the fine structure of bursts carries additional information about the features they encode (Krahe and Gabbiani 2004). Recent studies suggest that some stimulus information can be encoded in the burst structure, either in the size of the burst or the pattern of interspike intervals (Kepecs et al. 2002; Kepecs and Lisman 2003; Oswald et al. 2007; Arganda et al. 2007; Eyherabide et al. 2008, 2009). These analyses show that this information is present in the spike train and thus is potentially available to a decoder, but it is unknown whether the information is actually used by post-synaptic networks. We used combinations of neural and behavioral recordings to clarify this issue. We show that information about the size and shape of the amplitude envelope of an auditory stimulus is encoded both in the burst size and interspike-interval structure of the response of AN2. We consider the network between AN2 and the motor output as a whole and show that this decoder uses the information encoded in the burst size but that the behavioral output is not influenced by the interspike-interval structure of the burst. 123 316 J Comp Physiol A (2010) 196:315–320 Materials and methods Experimental procedures are described in detail in Marsat and Pollack (2006); the results presented here are from the same dataset. Briefly, adult Teleogryllus oceanicus were tethered dorsally and placed ventral-side up in a wind stream to elicit flight behavior. Abdominal steering movements associated with ultrasound avoidance were monitored by casting a shadow of the abdomen onto a photocell array, the voltage output of which varied roughly linearly with lateral displacement of the abdomen. Signals were digitized for off-line analysis (16 bits; 10 kHz sampling rate, PCI-6251 AD/DA board, National Instruments, Austin, TX). Behavior recordings of abdomen movements were high-pass filtered with a cutoff frequency of 0.5 Hz and averaged across repetitions of the same stimulus to calculate burst-triggered averages (see below). Twenty to forty minutes after the behavioral assay, the cricket was prepared for in vivo electrophysiology (see Marsat and Pollack 2005). AN2 was recorded extracellularly from the cervical connectives with stainless-steel hook electrodes. Recordings were amplified with a Grass P15 amplifier (Astro-Med, West Warwick, RI) and digitized as described above. Sound stimuli (duration 15 s) had carrier frequency of 30 kHz and Gaussian amplitude envelopes (low-passfiltered at 200 Hz; standard deviation 6 dB). Stimuli were generated digitally at sampling rate of 120 kHz. Mean stimulus level was 85 dB SPL, corresponding to approximately 20–25 dB above threshold. Analysis included the data from crickets in which AN2 frequently fired bursts of more than three spikes in response to these stimuli (n = 14 out of 22 neurons). Our analysis relies on comparing the timing of the behavioral and neural responses to the same stimulus waveform (for details see Marsat and Pollack 2006). The bursting response contains very little noise thus it is highly reproducible from trial to trial (Fig. 1; also see Marsat and Pollack 2006) but the behavioral response is more noisy. To remove the noise from the behavioral response, we average responses to 20–30 repetitions of the stimulus. We deduce the relative timing between AN2 bursts and the averaged behavioral response by referring both of these to the shared stimulus timing. We excluded from our analyses the first second of responses (both behavioral and physiological), during which AN2’s firing rate and behavioral responses adapt in order to satisfy the ergodicity requirement of the analysis (Marmarelis and Marmarelis 1978). While insects respond quickly to the onset of ultrasound, the insect can remain within hearing distance of a hunting bat for several seconds (Simmons 2005); thus the adapted response is behaviorally relevant. Spike trains, expressed as series of ones (for time points at which spikes occur) and zeros, were down- 123 Fig. 1 Exemplars of neural and behavioral responses to RAM stimuli. Stimuli consist of an ultrasonic carrier (30 kHz, not shown) and a random amplitude modulation (RAM; standard deviation: 6 dB) envelope. Neural responses of AN2 to the stimulus excerpt shown on top are displayed as a raster plot where black dots represent bursting responses and grey dots isolated spikes. The behavioral response (bottom) is quantified as the amplitude of the lateral displacement of the abdomen (arbitrary units; see ‘‘Materials and methods’’). We show two segments of recording in the same cricket that exemplify the relationship between stimulus, neural response and behavior. We underlined a prominent peak in the stimulus that gives rise to large neural and behavioral responses sampled to 2 kHz. The points where isolated spikes occurred were set to 0 so as to keep only spikes within bursts (ISI B 6.5 ms, as determined from ISI histograms; see Marsat and Pollack 2006). Burst structure was classified according to the criteria described in the results and figures. The timings of the first spike of the bursts in a given category were used to calculate the corresponding burst-triggered average (BTA) of the stimulus amplitude or of the behavioral response. BTAs were characterized by three measures: maximum (amplitude of the peak), half-width (the width at half maximum) and total power, which is essentially the area under the peak of the BTA (see inset in Fig. 2a). Total power was calculated as: Pe t¼b ðBTAðtÞ  MinÞ; where b is the time of the local minimum preceding the peak of the BTA, e the time of the minimum after the peak and Min is the minimum value of the BTA between b and e, inclusive (see Fig. 2a, inset). We also measured the onset slope of the BTA by fitting a linear function through the steepest portion of the rising phase. Matrices of data were created where each entry contains the value of the measure being analyzed (e.g., the maximum) for the BTA of a single experiment and the corresponding independent variable (e.g., number of spikes in the burst). These datasets contained only 186 entries out of the 266 possible entries (14 experiments 9 19 categories) because very large bursts did not occur in some experiments. For each of these datasets the linear best fit between the variables was calculated along with the R2 value and the probability that the slope is not different from zero. All analyses were performed with Matlab 2008 (Mathworks, Natick, MA) including the curve fitting toolbox. J Comp Physiol A (2010) 196:315–320 317 a d b e c f Fig. 2 Average stimulus amplitude envelope and behavioral response associated with bursts of different structure. Burst were divided in groups based either on their mean interspike interval (a, d), their duration (b, e) or the number of spikes they contain (c, f). The figures display the shapes of the average amplitude modulations that precede the bursts in one of these groups (a–c) or the average behavioral response that follows them (d–f). Zero on the x axes represents the timing of the first spike in the burst and on the y axes, the mean sound amplitude (a–c) or behavioral response (d–f). The inset in a illustrates the calculation of the stimulus power measure used later (Figs. 3, 4), which is essentially calculating the area under the curve (shaded area) Results Visual inspection of the response of AN2 to amplitude modulated ultrasound stimuli suggests that peaks in sound amplitude with different modulation patterns can trigger bursts of different size and structure (Fig. 1). Small bursts of two or three spikes are most common but longer bursts containing many spikes also occur. The features of the stimulus that trigger these large bursts seem to be large or broad peaks in amplitude (e.g., the underlined portion of the stimulus in Fig. 1). Most importantly we observe that peaks in amplitude that lead to large bursts also lead to large steering movements in the behaving animal (Fig. 1). In order to explore these qualitative observations, we classified each recorded burst based either on its duration (time between first and last spike), the number of spikes, or the mean interval between spikes. These groups of bursts were then used to calculate averages of the stimulus that preceded them and of the behavioral response that followed. These burst-triggered averages characterize the relationship between the amplitude and temporal features of the stimulus or behavioral response (i.e., their shapes) and bursting. All three burst characteristics vary systematically with the preceding stimulus peak. Interspike interval is shorter (i.e., instantaneous firing rate is higher) the larger and narrower the amplitude peak that triggers the burst (Fig. 2a; Table 1). Similarly, longer bursts or bursts with more spikes, are correlated with larger and broader peaks in the stimulus (Fig 2b, c; Table 1). Note that different bursts can be triggered by stimuli with only subtle, but statistically significant (Table 1), differences in shape. This is, for example, the case of bursts with short to medium durations (Fig. 2b) which are correlated to slight changes in the descending phase of the BTA 4–6 ms before burst onset. We can also observe a clear relationship between two aspects of the burst and the subsequent behavioral response (Fig. 2d–f): behavioral responses increase with increasing burst duration and increasing spike count (Table 1). To explore more quantitatively the relationship between the structure of bursts and of the stimulus, we plot three characteristics of the burst-triggered averages as a function of both number of spikes and burst duration (Fig. 3). The ratio of these reflects the average firing rate within the burst. We sorted the responses into 19 groups, as indicated by the 19 rectangles in each panel of Fig. 3. For each group, we plot the half-width, the maximum, and the total power of the preceding amplitude peak (Fig. 3a–c). We also display the result of a similar analysis, taking the maximum of the spike triggered average of the behavior as a quantification of the strength of the steering response (Fig. 3d). We find only weak correlations between the stimulus maximum or half-width and the burst duration or number of spikes (Table 1). If we consider the overall power contained in the peak of the burst triggered average, a clear correlation is observed with both burst duration and number of spikes. Furthermore, the strength of the behavioral response is also correlated with these two aspects of the bursts. This analysis does not allow separating which aspect of burst structure is best correlated with different aspect of the stimulus. The main issue comes from the fact that the duration of the burst is correlated with the number of spikes it contains. Due to this correlation the shaded squares representing each group of bursts are located on a diagonal. However, the breadth of this diagonal shows that for a given number of spikes, or a given duration, the burst can vary due to a difference in average ISI. Therefore, we re-evaluated the correlations as a function of two new 123 318 J Comp Physiol A (2010) 196:315–320 Table 1 Regression analysis of correlations between burst structure, stimulus structure and behavior Number of spikes Burst duration (ms) 2 Burst size Burst density p R2 0.02 <1025 0.27 [0.1 0.01 <1025 0.29 0.39 <1025 0.39 [0.1 0.007 0.25 <1025 0.31 [0.1 0.005 p R 0.1 [0.1 [0.1 0.008 Stimulus power <1025 Behavior magnitude <1025 2 p R 0.0003 \0.04 \0.01 0.05 0.3 <1025 0.3 <1025 p R Stimulus maximum \10-5 Stimulus half-width 2 The 4 columns represent the categorizing parameter of the bursts (see ‘‘Results’’ for details) and the 4 rows the burst-triggered average (BTA) parameters quantified. The goodness of fit of the linear regression between the two variable (quantified by the R2 value) and the probability that the slope of the linear fit is not different from zero (i.e., no correlation) are given. Bold characters highlight cells with stronger correlations (R2 [ 0.2) a c b d Fig. 3 Correlations between burst structure, stimulus structure and behavior. Each panel quantifies one aspect of the burst-triggered average (BTA): its half-width (a) maximum (b) or total power (c) for the stimulus BTA and overall magnitude (quantified as the peak’s maximum) for the behavioral-response BTA (d). 19 burst categories were defined based on the number of spikes they contain (x axes) and their duration (y axes). The definitions of the 19 different categories are shown at the edge of the color-coded squares. The darkness of the square is proportional to the value of the measure examined in each panel. Note that the top-most squares include all bursts of more than 17 ms and the right-most squares all bursts of 8 spikes and more. The arrows in the top-left corners of the panels illustrate two other dimensions along which burst structure varies: the burst density (see left panels) is the burst duration divided by the number of spikes and it increases from the top-left quadrant to the bottom-right whereas the burst size (burst duration 9 number of spikes) increases from the bottom-left to the top-right quadrant independent variables: the burst size, defined as the product of the number of spikes in the burst and the duration of the burst, and the burst density defined as the inverse of the average interval between spikes of a burst. The arrows in the top-left corners of the plots (Fig. 3) roughly show the directions along which these variables change. An interesting situation occurs when considering burst density as a variable: the half-width of the burst-triggered average is negatively correlated with burst density, but the maximum is positively correlated (Figs. 3a, b, 4a; note that in Fig. 4 we plot the data as a function of the mean burst ISI because of the clear meaning of this measure, but that mean ISI is inversely proportional to burst density). No strong correlations were found between stimulus half-width or power and burst size (Fig. 4a). A broad but not very intense peak in stimulus amplitude would give rise to a low density burst, whereas a narrow but intense peak will give rise to a dense burst. This means that for a burst of a given size the burst density can allow to distinguish between broad and narrow stimulus peaks of similar overall power. However, 123 J Comp Physiol A (2010) 196:315–320 319 Fig. 4 Correlations between burst size or density and the stimulus or the behavior. a Correlations between three characteristics of the stimulus BTA and the density (blue) or size (red) of the associated bursts. Note that burst density is expressed as mean ISI, i.e., (burst density)-1. Each dot represents the data for a single category of burst (i.e., density and size) in each experiment; the lines show the best linear fit. Correlation statistics are listed in Table 1. b Behavioral response correlates with burst size but not burst density. Each dot represents the maximum of the BTA of the behavioral response calculated for a single category of burst (density or size) in each experiment; see Table 1 for statistics. Points shown in color are results for four representative cells the overall power contained in the peak is not correlated with burst density but only with burst size (Fig 4a). Most importantly, the strength of the behavioral response is also correlated with burst size but not density (Fig. 4b; Table 1). We showed previously (Marsat and Pollack 2006) that the strength of the behavioral response was influenced by the average burst rate of a neuron. The correlations we show here are not determined by average burst rate because for a given neuron, thus a given burst rate, the correlations described are clearly present (see colored points in Fig 4b). We performed similar analyses of the behavioral response while using different aspects of the BTA of steering movements, onset slope or half-width, to quantify the magnitude of the behavioral response but also failed to find a correlation between these measures and burst density (data not shown; R2 \ 0.008 in all cases). Discussion Our analysis reveals that the structure of the burst can carry information about the shape of the stimulus that it encodes. In particular we observed a correlation between the size of the burst and the intensity of the stimulus: broader and/or higher peaks in ultrasound amplitude elicit longer bursts containing more spikes. This finding corroborates results in other systems. Similar to AN2, burst size in grasshopper auditory receptors signals the overall intensity of the amplitude peak (i.e., height and/or breadth; Eyherabide et al. 2008, 2009). In the visual system of cats, burst size is correlated with stimulus orientation (cortex; DeBusk et al. 1997; Martinez-Conde et al. 2002) or the slope of stimulus upstrokes (Kepecs et al. 2002; Kepecs and Lisman 2003). The number of spikes in bursts of tactile neurons of the leech vary according to the velocity of the stimulus (Arganda et al. 2007). The above-mentioned studies stress the importance of the size of the burst (i.e., number of spikes and/or burst duration) as the information-carrying variable within a burst. An alternate, but not exclusive, approach focuses on the relevance of the interval between spikes of a burst for information coding. Previous studies on neural coding noted that the intervals between successive spikes (i.e., relative timing), rather than their absolute timing, carry more information about the stimulus (de Ruyter Van Steveninck and Bialek 1988). Furthermore, short interspike intervals such as occur in bursts, carry information more efficiently than long ones (Reich et al. 2000). Interval 123 320 codes within bursts are therefore a potentially important aspect of information coding in these neurons. Such a code has been found in the sensory neurons of the electric fish which produce bursts most often limited to two spikes. Oswald et al. (2007) showed that the interspike interval of these doublets reliably encodes the amplitude of the stimulus peak. Our results indicate that the interspike intervals in AN2 bursts are correlated with stimulus features. For a burst of a given number of spikes, the average interval in the burst could provide the information to discriminate between a tall, narrow stimulus peak and a shallow, broad peak. The ultimate goal of our study was not only to determine whether the two neural codes detailed above can be observed in AN2, but also whether the information they encode is actually used by the post-synaptic network. Indeed, a neural code is only relevant to sensory processing if it is decoded post-synaptically. This question has been addressed mostly with modeling studies showing that burst size/duration (Kepecs and Lisman 2004) or burst interspike intervals (Izhikevich et al. 2003) can be decoded by synapses exhibiting mixtures of depression and facilitation. The question, however, has not been addressed experimentally. The results we present here provide insights into this problem by considering the network between AN2 and the motor output as a whole. By correlating the behavioral response and bursts of different sizes (i.e., number of spikes) and densities (i.e., average ISI) we observe that behavior varies only as a function of burst size, not burst density. In other words, features of the stimulus that affect burst size, also affect behavior whereas features affecting burst density but not burst size do not appear to affect behavior. Our results suggest that in this system, a code based on burst size contains more behaviorally relevant information whereas an interval code, although potentially present, has less obvious relevance. We do not know why the information present in the burst density seems unused. One possibility is that any influence of burst density on behavior is too subtle for our measures to detect. Another possibility is that the apparent encoding of burst density is merely a byproduct of burst generation in AN2, and that a mechanism is not present post-synaptically to decode it. However, our results have behavioral implications. We hypothesize that the overall intensity of a bat echolocation pulse as perceived by the cricket can carry information about its distance. The information carried by the pulse shape (for example breadth) might be harder to interpret as it varies with time across bat species and might be harder to relate to the imminence of an attack. This hypothesis implies that, because pulse intensity is an indicator of the closeness of the predator, it is encoded in the burst structure and affects behavior: stronger pulses lead to more spikes and stronger 123 View publication stats J Comp Physiol A (2010) 196:315–320 steering movements. On the other hand, if pulse shape is not as clearly related to the magnitude of the threat, then selection might not have favored mechanisms to decode this irrelevant information. Acknowledgments This work was supported by a Canadian Institutes for Health Research operating Grant (G.S.P.) and fellowship (G.M.) and a Natural Sciences and Engineering Research Council of Canada Discovery Grant (G.S.P.). References Arganda S, Guantes R, De Polavieja GG (2007) Sodium pumps adapt spike bursting to stimulus statistics. Nat Neurosci 10:1467–1473 de Ruyter van Steveninck RD, Bialek W (1988) Real-time performance of a movement-sensitive neuron in the blowfly visualsystem—coding and information-transfer in short spike sequences. Proc R Soc B 234:379–414 DeBusk BC, DeBruyn EJ, Snider RK, Kabara JF, Bonds AB (1997) Stimulus-dependent modulation of spike burst length in cat striate cortical cells. J Neurophysiol 78:199–213 Eyherabide HG, Rokem A, Herz AVM, Samengo I (2008) Burst firing is a neural code in an insect auditory system. Front Comput Neurosci 2:3 Eyherabide HG, Rokem A, Herz AVM, Samengo I (2009) Bursts generate a non-reducible spike-pattern code. Front Comput Neurosci 3:8–14 Gabbiani F, Metzner W, Wessel R, Koch C (1996) From stimulus encoding to feature extraction in weakly electric fish. Nature 384:564–567 Izhikevich EM, Desai NS, Walcott EC, Hoppensteadt FC (2003) Bursts as a unit of neural information: selective communication via resonance. Trends Neurosci 26:161–167 Kepecs A, Lisman J (2003) Information encoding and computation with spikes and bursts. Netw Comput Neural Syst 14:103–118 Kepecs A, Lisman J (2004) How to read a burst duration code. Neurocomputing 58–60:1–6 Kepecs A, Wang XJ, Lisman J (2002) Bursting neurons signal input slope. J Neurosci 22:9053–9062 Krahe R, Gabbiani F (2004) Burst firing in sensory systems. Nat Rev Neurosci 5:13–23 Marmarelis PZ, Marmarelis VZ (1978) Analysis of physiological systems: the white-noise approach. Plenum Press, New York Marsat G, Pollack GS (2005) Effect of the temporal pattern of contralateral inhibition on sound localization cues. J Neurosci 25:6137–6144 Marsat G, Pollack GS (2006) A behavioral role for feature detection by sensory bursts. J Neurosci 26:10542–10547 Martinez-Conde S, Macknik SL, Hubel DH (2002) The function of bursts of spikes during visual fixation in the awake primate lateral geniculate nucleus and primary visual cortex. Proc Natl Acad Sci USA 99:13920–13925 Nolen TG, Hoy RR (1984) Initiation of behavior by single neurons— the role of behavioral context. Science 226:992–994 Oswald AM, Doiron B, Maler L (2007) Interval coding I. Burst interspike intervals as indicators of stimulus intensity. J Neurophysiol 97:2731–2743 Reich DS, Mechler F, Purpura KP, Victor JD (2000) Interspike intervals, receptive fields, and information encoding in primary visual cortex. J Neurosci 20:1964–1974 Simmons JA (2005) Big brown bats and June beetles: multiple pursuit strategies in a seasonal acoustic predator–prey system. Acoust Res Lett Online 6:238–242