1. Computational and Systems Biology
  2. Neuroscience
Download icon

Prediction of primary somatosensory neuron activity during active tactile exploration

  1. Dario Campagner
  2. Mathew Hywel Evans
  3. Michael Ross Bale
  4. Andrew Erskine
  5. Rasmus Strange Petersen  Is a corresponding author
  1. The University of Manchester, United Kingdom
  2. University of Sussex, United Kingdom
  3. The Francis Crick Institute, United Kingdom
Short Report
  • Cited 30
  • Views 2,186
  • Annotations
Cite this article as: eLife 2016;5:e10696 doi: 10.7554/eLife.10696

Abstract

Primary sensory neurons form the interface between world and brain. Their function is well-understood during passive stimulation but, under natural behaving conditions, sense organs are under active, motor control. In an attempt to predict primary neuron firing under natural conditions of sensorimotor integration, we recorded from primary mechanosensory neurons of awake, head-fixed mice as they explored a pole with their whiskers, and simultaneously measured both whisker motion and forces with high-speed videography. Using Generalised Linear Models, we found that primary neuron responses were poorly predicted by whisker angle, but well-predicted by rotational forces acting on the whisker: both during touch and free-air whisker motion. These results are in apparent contrast to previous studies of passive stimulation, but could be reconciled by differences in the kinematics-force relationship between active and passive conditions. Thus, simple statistical models can predict rich neural activity elicited by natural, exploratory behaviour involving active movement of sense organs.

https://doi.org/10.7554/eLife.10696.001

eLife digest

The brain receives information from the world through the senses. In particular, cells called sensory neurons can detect signals from the environment and relay the information to the brain. A critical test of how well we understand the role of a given sensory neuron is whether it is possible to predict its activity under natural conditions. Previous research has succeeded in predicting the responses of sensory neurons in animals that were anaesthetised. However, it has been difficult to extend this approach to awake animals.

Mice and other rodents rely on their whiskers to tell them about their surroundings. Campagner et al. set out to predict how the sensory neurons that send information from whiskers (or ‘whisker neurons’) to the brain would respond in awake mice that were actively exploring an object in their environment. The approach involved using high-speed video (1,000 frames per second) to film the whiskers while the mice used them to explore a thin metal pole. At the same time, Campagner et al. recorded the electrical activity of the whisker neurons. The videos were used to calculate the forces acting on the whiskers, and then computational models were used to relate the activity of the neurons to the forces.

This approach allowed Campagner et al. to predict the responses of the whisker neurons, even when the mice were exploring the pole freely and unpredictably, simply from knowledge of the forces that were acting on the whiskers.

Together, these findings move the field of neuroscience forward by showing that sensory signals and neuronal responses can be correlated even in an awake animal. A key challenge for the future will be to further extend the approach to investigate how the signal conveyed by sensory neurons is transformed by neural circuits within the brain.

https://doi.org/10.7554/eLife.10696.002

Introduction

A major challenge of sensory neuroscience is to understand the encoding properties of neurons to the point that their spiking activity can be predicted in the awake animal, during natural behaviour. However, accurate prediction is difficult without experimental control of stimulus parameters and, despite early studies of awake, behaving animals (Hubel, 1959), subsequent work has most often effected experimental control by employing anaesthesia and/or passive stimulation. However, the active character of sensation (Gibson, 1962; Yarbus, 1967), based on motor control of the sense organs, is lost in reduced preparations. Recent methodological advances permit a way forward: in the whisker system, it is now possible to record neuronal activity from an awake mouse, actively exploring the environment with its whiskers, whilst simultaneously measuring the fundamental sensory variables (whisker kinematics and mechanics) likely to influence neuronal activity (O'Connor et al., 2010b).

Our aim here was to predict spikes fired by primary whisker neurons (PWNs) of awake mice engaged in natural, object exploration behaviour. The manner in which primary neurons encode sensory information fundamentally constrains all downstream neural processing (Lettvin et al., 1959). PWNs innervate mechanoreceptors located in the whisker follicles (Zucker and Welker, 1969; Rice et al., 1986). They are both functionally and morphologically diverse; including types responsive to whisker-object contact and/or whisker self-motion (Szwed et al., 2003; Ebara et al., 2002). PWNs project to the cerebral cortex, analogously to other modalities, via trisynaptic pathways through the brainstem and thalamus (Diamond et al., 2008).

Here, we show that PWN responses are well-predicted by rotational force ('moment') acting on the whisker, while whisker angle is a poor predictor. Moment coding accounts for spiking during both whisker-object interaction and whisker motion in air. Moment coding can also account for findings in previous studies of passive stimulation in the anaesthetized animal; indicating that the same biomechanical framework can account for primary somatosensory neuron responses across diverse states. Our results provide a mechanical basis for linking receptor mechanisms to tactile behaviour.

Results

Primary whisker neuron activity during object exploration is predicted by whisker bending moment

We recorded the activity of single PWNs from awake mice (Figure 1A,E, Figure 1—figure supplement 1) as they actively explored a metal pole with their whiskers (N = 20 units). At the same time, we recorded whisker motion and whisker shape using high-speed videography (1000 frames/s; Figure 1D, Video 1). As detailed below, PWNs were diverse, with some responding only to touch, others also to whisker motion. Since each PWN innervates a single whisker follicle, we tracked the ‘principal whisker’ of each recorded unit from frame to frame, and extracted both the angle and curvature of the principal whisker in each video frame (total 1,496,033 frames; Figure 1B–E; Bale et al., 2015). Whiskers are intrinsically curved, and the bending moment on a whisker is proportional to how much this curvature changes due to object contact (Birdwell et al., 2007): we therefore used ‘curvature change’ as a proxy for bending moment (O'Connor et al., 2010a). Whisker-pole contacts caused substantial whisker bending (curvature change), partially correlated with the whisker angle (Figures 1E, 4E) and, consistent with Szwed et al. (2003) and Leiser and Moxon (2007), robust spiking (Figures 1E, 2E).

Figure 1 with 2 supplements see all
Electrophysiological recording from single primary whisker units in awake, head-fixed mice and simultaneous measurement of whisker kinematics/mechanics. 

(A) Schematic of the preparation, showing a tungsten microelectrode array implanted into the trigeminal ganglion of a head-fixed mouse, whilst a metal pole is presented in one of a range of locations (arrows). Before the start of each trial, the pole was moved to a randomly selected, rostro-caudal location. During this time, the whiskers were out of range of the pole. At the start of the trial, the pole was rapidly raised into the whisker field, leading to whisker-pole touch. Whisker movement and whisker-pole interactions were filmed with a high-speed camera. (B, C) Kinematic (whisker angle θ) and mechanical (whisker curvature κ, moment M, axial force Fax and lateral force Flat) variables were measured for the principal whisker in each video frame. When a whisker pushes against an object during protraction (as in panel D, red and cyan frames), curvature increases; when it pushes against an object during retraction (as in panels B and C), it decreases. (D) Individual video frames during free whisking (yellow and green) and whisker-pole touch (red and cyan) with tracker solutions for the target whisker (the principal whisker for the recorded unit, panel E) superimposed (coloured curve segments). (E) Time series of whisker angle, push angle and curvature change, together with simultaneously recorded spikes (black dots) and periods of whisker-pole contact (red bars). Coloured dots indicate times of correspondingly coloured frames in D.

https://doi.org/10.7554/eLife.10696.003
Video 1
Video of an awake mouse, exploring a pole with its whiskers with simultaneous electrophysiological recording of a primary whisker neuron.

At the start of the video, the pole is out of range of the whiskers. The whisker tracker solution for the principal whisker of the recorded unit is overlaid in red. White dots represent spikes; orange trace shows whisker angle (scale bar = 40°); blue trace shows whisker curvature change (scale bar = 0.05 mm-1). Video was captured at 1000 frames/s and is played back at 50 frames/s. Related to Figure 1.

https://doi.org/10.7554/eLife.10696.006
Figure 2 with 3 supplements see all
Primary whisker neurons encode whisker curvature, not whisker angle, during active sensation. 

(A) Schematic of the Generalized Linear Model (GLM). (B) For an example unit, whisker angle (top panel), whisker curvature change (middle panel) and simultaneously recorded spike train (bottom panel, black), together with predicted spike trains for the best-fitting angle GLM (bottom panel, orange) and curvature change GLM (bottom panel, blue). Spike trains discretized using 1-ms bins and smoothed with a 100 ms boxcar filter. Prediction performance (Pearson correlation coefficient, PCC) for this unit was 0.59. Inset shows tuning curves for both GLMs, computed by convolving the relevant sensory time series (angle or curvature change) with the corresponding GLM stimulus filter to produce a time series of filter coefficients, and estimating the spiking probability as a function of filter coefficient (25 bins). (C) Analogous to panel B, for a second example unit. Prediction performance PCC for this unit was 0.74. (D) Prediction performance (PCC between predicted and recorded spike trains) compared for GLMs fitted with three different types of input: curvature change alone; angle alone; both curvature change and angle. Each blue/orange/green dot is the corresponding PCC for one unit: large black dots indicate median; error bars denote inter-quartile range (IQR). To test statistical significance of each unit’s PCC, the GLM fitting procedure was repeated 10 times on spike trains subjected each time to a random time shift: magenta dots show these chance PCCs for the unit indicated by the magenta circle; the mean chance PCC was computed for each unit and the large grey dot shows the median across units. Black circles indicate units whose PCC was significantly different to chance (signed-rank test, Bonferroni-corrected, p<0.0025). To facilitate direct comparison between results for curvature change GLM and angle GLM, these are re-plotted in the inset. (E) Left. Firing rate during touch episodes compared to that during non-touch episodes for each unit, compared to corresponding predicted firing rates from each unit’s curvature change GLM. Right. Medians across units: error bars denote IQR; * denotes differences significant at p<0.05 (signed-rank test).

https://doi.org/10.7554/eLife.10696.007
Figure 3 with 1 supplement see all
Primary whisker neurons encode whisker angular acceleration during free whisking. 

(A) Mean response of an example whisking-sensitive unit to whisking amplitude, computed during non-contact episodes (dark green, shaded area shows SEM) with regression line (black). Inset shows regression line slopes (median and IQR) for whisking sensitive (green) and whisking insensitive (grey) units. * indicates statistically significant rank-sum test (p=0.05). (B) Mean response of two example units as a function of angular acceleration. The dark brown unit is the same as that shown in A. (C) Mean response of two example units as a function of whisking phase. The dark pink unit is the same as that reported in A; the light pink unit is the same as that shown as light brown in B. (D) Excerpt of free whisking (orange) along with activity of an example, whisking-sensitive unit (black) and activity predicted by a GLM driven by angular acceleration (brown). The unit is the same as that shown in A. (E) GLM prediction accuracy (PCC) for all whisking sensitive (brown) and whisking insensitive units (grey). Bars and vertical lines denote median and IQR respectively.

https://doi.org/10.7554/eLife.10696.011
Figure 4 with 2 supplements see all
Whisker angle and whisker curvature change are highly correlated during passive whisker deflection, but decoupled during active touch. 

(A) Whisker angle (top) and whisker curvature change (bottom) time series, due to passive, trapezoidal stimulation of C2 whisker in an anaesthetized mouse, estimated as mean over 10 repetitions. Note that error bars (showing SEM) are present but very small. (B) Corresponding data for low-pass filtered white noise (hereafter abbreviated to ‘white noise’) stimulation of the same whisker. (C) Cross-correlation between curvature change and angle during white noise stimulation, for C2 whisker. (D) Cross-correlation between angle and curvature change at zero lag, for both passive stimulation under anaesthesia and awake, active sensing (median of absolute cross-correlation for each unit; error bar denotes IQR). (E) Joint distribution of whisker angle and whisker curvature change in awake, behaving mice (1 ms sampling). Different colours denote data corresponding to different recorded units. Inset: Analogous plot for passive, white noise whisker deflection in an anaesthetised mouse. Different colours indicate data from different whiskers. (F) Joint distribution of angle and curvature change for an example recording from an awake behaving mouse, with samples registered during touch and non-touch distinguished by colour (1 ms sampling). (G) Touch data of F classified according to pole position (dot colour).

https://doi.org/10.7554/eLife.10696.013

To test between candidate encoding variables, our strategy was to determine how accurately it was possible to predict PWN activity from either the angular position or curvature change of each recorded unit’s principal whisker. To predict spikes from whisker state, we used Generalised Linear Models (GLMs; Figure 2A). GLMs, driven by whisker angle, have previously been shown to provide a simple but accurate description of the response of PWNs to passive stimulation (Bale et al., 2013) and have mathematical properties ideal for robust parameter-fitting (Truccolo et al., 2005; Paninski et al., 2007).

For each recorded unit (median 69,672 frames and 550 spikes per unit), we computed the GLM parameters that best predicted the unit’s spike train given the whisker angle time series, using half the data as a training set for parameter-fitting (8 total fitted parameters - 5 for stimulus filter, 2 for history filter, 1 bias; Figure 2—figure supplement 3). We then assessed prediction performance using the other half of the data as a testing set: we provided the GLM with the whisker angle time series as input and calculated the predicted spike train, evoked in response (Materials and methods). We then compared the recorded spike train to the GLM-predicted one (Figure 2B–C) and quantified the similarity between the smoothed spike trains using the Pearson correlation coefficient (PCC). This is a stringent, single-trial measure of model prediction performance (Figure 2—figure supplement 1B). We then repeated this entire procedure for the whisker curvature time series. Although angle GLMs predicted spike trains of a few units moderately well (2/20 units had PCC > 0.5), they performed poorly for the majority (median PCC 0.06, IQR 0.019–0.3; Figure 2B–D, orange). This was unlikely to be because of non-linear tuning to whisker angle, since quadratic GLMs fared only marginally better (median PCC 0.097, IQR 0.042–0.31; p=0.044, signed-rank test, Figure 2—figure supplement 1A). In contrast, we found that, at the population level, the curvature GLMs were substantially more accurate than the angle GLMs (median PCC 0.52, IQR 0.22–0.66; p=0.0044, signed-rank test; Figure 2B–D, blue) with prediction accuracy up to PCC 0.88. Curvature GLMs also predicted spikes during touch episodes significantly more accurately (median PCC 0.57, IQR 0.23–0.72) than did angle GLMs during non-touch episodes (median 0.06, IQR 0.02–0.35; p=0.005, signed-rank test). At the level of individual units, 90% had above chance PCC and we termed these ‘curvature-sensitive’ (Materials and methods). Of the curvature-sensitive units, 61% were sensitive to positive curvature change and 39% to negative curvature change (Materials and methods).

The result that curvature predicted PWN responses better than angle was robust to the number of fitted parameters: a GLM sensitive to instantaneous curvature (4 parameters: 1 stimulus filter parameter, 2 history filter parameters and 1 bias) exhibited very similar prediction accuracy (Figure 2—figure supplement 1C). The result was also robust to time-scale: prediction accuracy based on curvature was significantly greater than that based on angle for smoothing time-scales in the range 1–100 ms (signed-rank test, p<0.05, Bonferroni-corrected).

Although the activity of most units was better predicted by whisker curvature change than by whisker angle, there was significant variability in prediction performance, and there were a few units for which the angle prediction performance was appreciable (Figure 2D). However, we found that this could largely be attributed to redundancy. When a mouse whisks against an object, curvature change and angle fluctuate in concert (Birdwell et al., 2007; Bagdasarian et al., 2013; Pammer et al., 2013; Figures 1E, 4E and Figure 4F–G). When we fitted GLMs using both curvature change and angle as input, these GLMs predicted the spike trains no more accurately (median PCC 0.53, IQR 0.40–0.62; p=0.067, signed-rank test; Figure 2D) than GLMs based on curvature change alone. Moreover, on a unit-by-unit basis, for 65% of units, curvature change GLMs predicted spikes better than angle (signed-rank test, p<0.05, Bonferroni-corrected); only for 5% of units did angle predict spikes better than curvature change. GLMs based on curvature change also predicted spike trains more accurately than GLMs based on ‘push angle’ – the change in angle as the whisker pushes against an object (Figure 1E; median PCC 0.25, IQR 0.04–0.45; p=0.006, signed-rank test). Moreover, prediction accuracy of GLMs fitted with both push angle and curvature change (median PCC 0.52, IQR 0.2–0.69) inputs was no better than that of GLMs fitted with curvature alone (p=0.43, signed-rank test).

In principle, neurons might also be sensitive to the axial force component (parallel to the whisker follicle) and/or lateral force component (orthogonal to axial) associated with whisker-object contact (Figure 1B–C, Figure 1—figure supplement 2; Solomon and Hartmann, 2006; Pammer et al., 2013). We restricted our analysis to bending moment since, under our experimental conditions, axial/lateral force components were near-perfectly correlated with bending moment (Figure 2—figure supplement 2) and bending moment is likely to have a major influence on stresses in the follicle (Pammer et al., 2013).

To further test the curvature-encoding concept, we asked whether curvature GLMs could account for the response of PWNs to whisker-pole touch. To this end, we parsed the video data into episodes of ‘touch’ and ‘non-touch’. Units fired at a higher rate during touch than otherwise (Szwed et al., 2003; Leiser and Moxon, 2007). Without any further parameter-adjustment, the curvature-based GLMs reproduced this effect (Figure 2E): the correlation coefficient between recorded and GLM-predicted firing rate for touch episodes was 0.97. Collectively, the above results indicate that, during active touch, the best predictor of whisker primary afferent firing is not whisker angle but rather the bending moment.

Primary whisker neuronal activity during whisking is predicted by moment

During free whisking - in the absence of whisker-pole contact - whisker curvature, and therefore bending moment, changed little (Figure 1E, Figure 4F); consistent with previous studies (Knutsen et al., 2008; Quist et al., 2014). Yet, 50% of recorded units (‘whisking-sensitive units’) were significantly modulated by whisking amplitude (Figure 3A). Consistent with Szwed et al. (2003), PWNs were diverse: 45% were curvature-sensitive (significant PCC for curvature based GLM) but not whisking-sensitive; 45% were both curvature- and whisking-sensitive; 5% were whisking-sensitive but not curvature-sensitive.

The presence of whisking sensitivity suggests that moment due to whisker bending is not the only force that influences PWN activity. A likely candidate is the moment associated with the rotational acceleration of a whisker: this moment is proportional to the whisker’s angular acceleration (Quist et al., 2014; Materials and methods). Consistent with this possibility, we found that whisking-sensitive units were tuned to angular acceleration (Figure 3B) and that 50% of these were phase-modulated (Figure 3C). Angular acceleration tuning was diverse: some units fired to acceleration in a particular direction (rostral or caudal), whilst others responded to acceleration in both directions (Figure 3B, Figure 3—figure supplement 1). Moreover, for whisking-sensitive units (but not whisking-insensitive ones), quadratic GLMs trained on data from non-touch episodes were able to predict spikes using whisker angle acceleration as input (Figure 3D–E; whisking-sensitive units, median PCC 0.37, IQR 0.18–0.58; whisking-insensitive, median PCC -0.0071, IQR -0.035–0.041; p=0.0017 rank-sum test for whisking-sensitive vs whisking-insensitive units). For 70% of whisking-sensitive units, directional selectivity for acceleration was consistent with that for curvature. These findings indicate that, in the absence of whisker-object contact, responses of PWNs to whisking itself can be accounted for by sensitivity to the moment associated with angular whisker acceleration.

Relation between kinematics and mechanics is different in active vs passive touch and has implications for neural encoding

We found, during active object exploration, that curvature change, but not whisker angle, predicts PWN firing. In apparent contrast, studies using passive whisker stimulation have reported that PWNs encode whisker angle and its temporal derivatives (Zucker and Welker, 1969; Gibson and Welker, 1983; Lichtenstein et al., 1990; Jones et al., 2004; Arabzadeh et al., 2005; Bale and Petersen, 2009; Lottem and Azouz, 2011; Bale et al., 2013). We wondered whether the discrepancy might be due to differences in whisker mechanics between passive and active stimulation conditions. To test this, we analysed the relationship between angle and curvature change during active touch and compared it to that during passive whisker stimulation. During active pole exploration, angle and curvature change were, over all, only loosely related (median correlation coefficient 0.20, IQR 0.079–0.39, Figures 4D–E). Important contributory factors were that the angle-curvature relationship was both different for touch compared to non-touch (Figure 4F) and dependent on object location (Figure 4G). In contrast, during passive stimulation, whisker angle was near perfectly correlated with curvature change (for C2, correlation coefficients 0.96 and 0.94 respectively; similar results for C5; Figures 4C–D, Figure 4E, inset and Figure 4—figure supplement 2); consistent with properties of cantilevered beams (Birdwell et al., 2007). Simulations confirmed that, due to the tight relationship between the variables, a unit tuned purely to curvature change can appear tightly tuned to angle (Figure 4—figure supplement 1). The implication is that apparent sensitivity to whisker angle under passive stimulation conditions can be accounted for by moment-tuning.

Discussion

Prediction of spikes fired by sensory neurons under natural conditions

In the endeavour to understand how neurons encode and process sensory information, there is a basic tension between the desire for tight experimental control and the desire to study animals under natural, unconstrained conditions. Theories of sensory encoding suggest that neural circuits have evolved to operate efficiently under natural conditions (Simoncelli and Olshausen, 2001; Reinagel, 2001). Previous studies have succeeded in predicting/decoding spikes evoked by passive presentation of natural sensory stimuli to anaesthetised/immobilised animals (Lewen et al., 2001; Arabzadeh et al., 2005; Pillow et al., 2008; Mante et al., 2008; Lottem and Azouz, 2011; Bale et al., 2013), but it has been difficult to extend this approach to encompass natural, active movement of the sense organs. Here, we have addressed this general issue, taking advantage of experimental possibilities recently created in the whisker system (O'Connor et al., 2010a), and the ability of computational methods, such as GLMs, to uncover stimulus-response relationships even from data with complex statistical structure (Paninski et al., 2007; Fairhall and Sompolinsky, 2014). Our main finding was that responses of PWNs, recorded as an awake mouse actively explores an object with its whiskers, can be predicted from the forces acting on the whiskers. Given that, for each unit, we were attempting to predict the entire ~70 s time course of activity, the variability of the behaviour of untrained mice (O'Connor et al., 2010a), and the lack of trial-averaging as a noise reduction strategy, it is remarkable that we found model prediction correlation coefficients up to 0.88. A challenge of studying neural coding under unconstrained, awake conditions is that sensory variables tend to correlate. A valuable feature of the GLM training procedure is that it takes such correlations into account. We found that, although whisker angle predicted spikes for a subset of units, this effect was very largely explained by a curvature-coding model, together with the correlation between angle and curvature.

Mechanical framework for tactile coding

Pushing a whisker against an object triggers spiking in many PWNs (Szwed et al., 2003; Szwed et al., 2006; Leiser and Moxon, 2007). Biomechanical modelling by Hartmann and co-workers accounts for this by a framework where the whisker is idealised as an elastic beam, cantilever-mounted in the skin (Birdwell et al., 2007; Quist et al., 2014). When such a beam pushes against an object, the beam bends, causing reaction forces at its base. Our data are in striking agreement with the general suggestion that mechanoreceptor activity is closely related to such reaction forces. Our results show that curvature change associated with contact-induced whisker bending, and acceleration associated with whisker rotation, predict PWN spiking. Our results also provide a mechanical basis for previous findings: our finding of subtypes of curvature-only and curvature-acceleration PWNs is consistent with previous reports of ‘touch’ and ‘whisking-touch’ units (Szwed et al., 2003; 2006). Thus, a common framework accounts for diverse PWN properties.

Our finding that whisker angle predicts PWN spikes poorly indicates that whisker angle can change without modulating mechanotransduction in the follicle. This is consistent with evidence that, during artificial whisking, the follicle-shaft complex moves as a rigid unit (Bagdasarian et al., 2013). In apparent contrast, previous studies using passive stimulation in anaesthetised animals have consistently reported a tight relationship between whisker kinematics and PWN response. In the cantilever whisker model, passively induced changes in whisker angle correlate highly with whisker bending. We confirmed that this applies to real whiskers in vivo and demonstrate that moment-sensitive units can thereby appear angle-tuned. In this way, moment-encoding can account for primary neuron responses not only during active touch but also under passive stimulation. More generally, our results highlight the importance of studying neurons under natural, active sensing conditions.

In this study, we considered PWN encoding under conditions of pole contact, since this is well-suited to reaction force estimation (O'Connor et al., 2010a; Pammer et al., 2013) and involves object-stimulus interactions on a ~100 ms time-scale that is conducive to single-trial analysis. Since whisker bending is ubiquitous in whisking behaviour, it is likely that our finding of curvature sensitivity is a general one. However, prediction performance varied across units, suggesting that other force components may also be encoded. Other experimental conditions – for example, textured surfaces – may involve multiple force components (Quist and Hartmann, 2012; Pammer et al., 2013; Bagdasarian et al., 2013) and/or encoding of information by spike timing on a finer time-scale (Panzeri et al., 2001; Petersen et al., 2001; Arabzadeh et al., 2005; Bale et al., 2015).

It is axiomatic that mechanoreceptors are sensors of internal forces acting in the tissue within which they are embedded (Abraira and Ginty, 2013) and therefore valuable to be able to measure mechanical forces in the awake, behaving animal. In general, including the important case of primate hand-use, the complex biomechanics of skin makes force-estimation difficult (Phillips and Johnson, 1981). In contrast, for whiskers, the quasi-static relationship is relatively simple: the bending moment on a whisker is proportional to its curvature. This has the important implication that reaction forces can be directly estimated from videography in vivo (Birdwell et al., 2007; O'Connor et al., 2010a; Pammer et al., 2013). Our results are the first direct demonstration that such reaction forces drive primary sensory neuron responses – likely involving Piezo2 ion channels (Woo et al., 2014; Poole et al., 2015Whiteley et al., 2015) – and provide insight into how sensitivity to touch and self-motion arises in the somatosensory pathway (Szwed et al., 2003; Yu et al., 2006; Curtis and Kleinfeld, 2009Khatri et al., 2009; O'Connor et al., 2010b; Huber et al., 2012; Petreanu et al., 2012; Peron et al., 2015).

Moment-based computations in tactile behaviour

Extraction of bending moment is a useful first step for many tactile computations. Large transients in bending moment signal object-touch events, and the magnitude of bending is inversely proportional to the radial distance of contact along the whisker (Solomon and Hartmann, 2006). As illustrated by our results on the statistics of active touch, if integrated with cues for whisker self-motion, whisker bending can be a cue to the 3D location of an object (Szwed et al., 2003; 2006; Birdwell et al., 2007; Bagdasarian et al., 2013; Pammer et al., 2013). Bending moment can permit wall following (Sofroniew et al., 2014) and, if integrated across whiskers, can in principle be used both to infer object shape (Solomon and Hartmann, 2006) and to map the spatial structure of the environment (Fox et al., 2012; Pearson et al., 2013).

Summary and conclusion

We have shown that the responses of primary whisker neurons can be predicted, during natural behaviour that includes active motor control of the sense organ, from forces acting on the whiskers. These results provide a bridge linking receptor mechanisms to behaviour.

Materials and methods

All experimental protocols were approved by both United Kingdom Home Office national authorities and institutional ethical review.

Surgical procedure

Request a detailed protocol

Mice (C57; N=10; 6 weeks at time of implant) were anesthetized with isoflurane (2% by volume in O2), mounted in a stereotaxic apparatus (Narishige, London, UK) and body temperature maintained at 37°C using a homeothermic heating system. The skull was exposed and a titanium head-bar (19.1 × 3.2 × 1.3 mm; O'Connor et al., 2010a) was first attached to the skull ~1 mm posterior to lambda (Vetbond, St. Paul, MN), and then fixed in place with dental acrylic (Lang Dental, Wheeling, IL). A craniotomy was made (+0.5 to -1.5 mm posterior to bregma, 0-3 mm lateral) and sealed with silicone elastomer. Buprenorphine (0.1 mg/kg) was injected subcutaneously for postoperative analgesia and the mouse left to recover for at least 5 days.

Behavioural apparatus

Request a detailed protocol

Mice were studied in a pole exploration apparatus adapted from O'Connor et al., 2010a , but were not trained on any task. A mouse was placed inside a perspex tube (inner diameter 32 mm), from which its head emerged at one end, and immobilised by fixing the head-bar to a custom mount holder. The whiskers were free of the tube at all times. The stimulus object was a 1.59 mm diameter metal pole, located ~3.5 mm lateral to the mouse’s snout. To allow control of its anterior/posterior location, the pole was mounted on a frictionless linear slide (NDN 2-50.40, Schneeberger, Roggwil, Germany) and coupled to a linear stepper motor (NA08B30, Zaber, Vancouver, Canada). To allow vertical movement of the pole into and out of range of the whiskers, the pole/actuator assembly was mounted on a pneumatic linear slide (SLS-10-30-P-A, Festo, Northampton, UK), powered by compressed air. The airflow was controlled by a relay (Weidmüller, Richmond, VA). In this way, the pole moved rapidly (~0.15 s) into and out of range of the whiskers. The apparatus was controlled from Matlab via a real-time processor (RX8, TDT, Alachua, FL).

Electrophysiology

Request a detailed protocol

We recorded the activity of PWNs from awake mice in the following way. To permit reliable whisker tracking (see below), before each recording session, A, B and E whisker rows were trimmed to the level of the fur, under brief isoflurane anaesthesia. The trigeminal ganglion was targeted as previously described (Bale et al., 2015). The silicone seal was removed and a 3/4 shank tungsten microelectrode array (FHC, Bowdoin, ME, recording electrodes 8 MΩ at 1 kHz, reference 1 MΩ; tip spacing ~500 μm) was lowered through the brain (angle 4° to vertical in the coronal plane) using a micromanipulator (PatchStar, Scientifica, Uckfield, UK) under isoflurane anaesthesia. Extracellular potentials were pre-amplified, digitised (24.4 kHz), filtered (band pass 300–3000 Hz) and acquired continuously to hard disk (RZ5, TDT). The trigeminal ganglion was encountered 6–7 mm vertically below the pial surface and whisker-response units identified by manual deflection of the whiskers with a small probe. Once a well-isolated unit was found, the whisker that it innervated (the ‘principal whisker’, PW) was identified by manual stimulation. To define the PW, we deflected not only untrimmed whiskers, but also the stubs of the trimmed whiskers. Any unit whose PW was a trimmed whisker was ignored. At this point, anaesthesia was discontinued. Once the mouse was awake, we recorded neuronal activity during repeated presentations of the pole (‘trials’). Before the start of each trial, the pole was in the down position, out of reach of the whiskers. The pole was first moved anterior-posteriorly to a position chosen randomly out of a set of 11 possible positions, spanning a range ± 6 mm with respect to the resting position of the base of the PW. A trial was initiated by activating the pneumatic slide relay, thus moving the pole up into the whisker field, where it remained for 3 s before being lowered. At the end of a recording session, the microelectrode array was withdrawn, the craniotomy sealed with silicone elastomer, and the mouse returned to its home cage.

High-speed videography

Request a detailed protocol

Using the method of O'Connor et al. (2010a) to image whisker movement/shape, whiskers ipsilateral to the recorded ganglion were illuminated from below using a high-power infrared LED array (940 nm; LED 940-66-60, Roithner, Vienna, Austria) via a diffuser and condensing lens. The whiskers were imaged through a telecentric lens (55-349, Edmunds Optics, Barrington, NJ) mounted on a high speed camera (LTR2, Mikrotron, Unterschleissheim, Germany; 1000 frames/s, 0.4 ms exposure time). The field of view of the whiskers was 350×350 pixels, with pixel width 0.057 mm.

Response to touch and non-touch events

Request a detailed protocol

Mouse whisking behaviour during the awake recording was segmented into ‘touch', and 'non-touch’ episodes. Touches between the PW of each unit and the pole were detected manually in each frame of the high-speed video. A frame was scored as touch if no background pixels were visible between the pole silhouette and the whisker. Any frame not scored as a touch was scored as non-touch. Touch and non-touch firing rates for a given unit were computed by averaging activity over all corresponding episodes.

Whisker tracking

Request a detailed protocol

Since the trigeminal ganglion lacks topography, it is difficult to target units that innervate a specific whisker, and therefore desirable for a whisker tracker to be robust to the presence of multiple rows of whiskers. However, since neurons in the ganglion innervate individual whiskers, it is sufficient to track only one whisker (the PW) for each recorded neuron. To extract kinematic/mechanical whisker information, we therefore developed a whisker tracker (‘WhiskerMan’; Bale et al., 2015) whose design criteria, different to those of other trackers (Perkon et al., 2011; Clack et al., 2012), were to: (1) be robust to whisker cross-over events; (2) track a single, target whisker; (3) track the proximal segment of the whisker shaft. The shape of the target whisker segment was described by a quadratic Bezier curve r(t,s) (a good approximation away from the zone of whisker-object contact; Quist and Hartmann, 2012; Pammer et al., 2013): r(t,s) = [x(t,s), y(t,s)], where x, y are horizontal/vertical coordinates of the image, s = [0,..,1] parameterises (x,y) location along the curve and t is time. We fitted such a Bezier curve to the target whisker in each image frame using a local, gradient-based search. The initial conditions for the search were determined by extrapolating the solution curves from the previous two frames, assuming locally constant, angular velocity. The combination of the low-parameter whisker description and the targeted, local search made the algorithm robust to whisker cross-over events. The ‘base’ of the target whisker was defined as the intersection between the extrapolated Bezier curve and the snout contour (estimated as described in Bale et al., 2015). The solution curve in each frame was visually checked and the curves manually adjusted to correct occasional errors.

Estimation of kinematic/force parameters

Request a detailed protocol

The whisker angle (θ) in each frame was measured as the angle between the tangent to the whisker curve at the base and the anterior-posterior axis (Figure 1C). Whisker curvature (κ) was measured at the base as κ=x'y''x''y'(x'2+y'2)3/2, where x', y' and x'', y'' are the first and second derivatives of the functions x(s) and y(s) with respect to s (Figure 1C). Since reaction force at the whisker base reflects changes in whisker curvature, rather than the intrinsic (unforced) curvature (Birdwell et al., 2007), we computed ‘curvature change’ Δκ = κ - κint, where κint, the intrinsic curvature, was estimated as the average of κ in the first 100 ms of the trial (before pole contact; O'Connor et al., 2010a). During free whisking, whisker angle oscillated with the characteristic whisking rhythm, but curvature changed little. The small changes in whisker curvature during free whisking were consistent with torsional effects (Knutsen et al., 2008). We estimated the number of whisking cycles from the duration of touch/non-touch episodes and the whisking frequency: median 419 whisking cycles per unit during touch periods; 415 during non-touch periods.

Under conditions of whisking against a smooth surface, such as in the present study, the quasi-static framework of Birdwell et al. (2007) applies. Δκ, measured, at the base of a whisker, in the horizontal plane, is proportional to the component of bending moment in that plane. We used Δκ as a proxy for bending moment. Bending moment (M), Axial (Fax) and lateral force (Flat) at the whisker base were calculated, during periods of whisker-pole contact, using the method of Pammer et al. (2013), using published data on areal moment of inertia of mouse whiskers (Pammer et al., 2013), along with whisker-pole contact location (see Figure 1—figure supplement 2 for details). Pole location, in the horizontal plane, in each frame, was identified as the peak of a 2D convolution between the video image and a circular pole template. To localise whisker-pole contact, the whisker tracker was used to fit the distal segment of the whisker close to the pole, seeded by extrapolation from the whisker tracking solution for the proximal whisker segment, described above. Whisker-pole contact location was defined as the point where this distal curve segment was closest to the detected pole centre. Pole and contact locations were verified by visual inspection.

As expressed by Newton’s second law of rotational motion, the moment – or torque – of a rigid body, rotating in a plane, is proportional to the body’s angular acceleration. During free whisking, a whisker behaves approximately as a rigid body and, for the whiskers considered in this study, their motion is predominantly in the horizontal plane (Bermejo et al., 2002; Knutsen et al., 2008). Thus, to assess whether such a moment is encoded by PWNs, we measured angular whisker acceleration during free whisking as a proxy. Acceleration was calculated from the whisker angle time series after smoothing with a Savitzky-Golay filter (polynomial order 5; frame size 31 ms).

Push angle – the change in angle as a whisker pushes against an object - was measured during touch epochs. For each touch episode, we determined the value of the angle in the frame before touch onset and subtracted this from the whisker angles during the touch.

Passive whisker deflection

Request a detailed protocol

To determine how whiskers move/bend in response to passive deflection under anaesthesia, a mouse was anesthetized (isoflurane 2%) and placed in the head-fixation apparatus. Individual whiskers (C2 and C5 trimmed to 5 mm) were mechanically deflected using a piezoelectric actuator as previously described (Bale et al., 2013; 2015). All other whiskers were trimmed to the level of the fur. Each whisker, in turn, was inserted into a snugly fitting plastic tube attached to the actuator, such that the whisker entered the tube 2 mm from the face. Two stimuli were generated via a real-time processor (TDT, RX8): (1) a 10 Hz trapezoidal wave (duration 3 s, amplitude 8°); (2) Gaussian white noise (duration 3 s, smoothed by convolution with a decaying exponential: time constant 10 ms; amplitude SD 2.1°). During the stimulation, the whiskers were imaged as detailed above (1000 frames/s, 0.2 ms exposure time).

Electrophysiological data analysis

Spike sorting

Request a detailed protocol

Single units (N=20) were isolated from the extracellular recordings as previously described, by thresholding and clustering in the space of 3–5 principal components using a mixture model (Bale and Petersen, 2009). A putative unit was only accepted if (1) its inter-spike interval histogram exhibited a clear absolute refractory period and (2) its waveform shape was consistent between the anaesthetised and awake phases of the recording.

Responses to whisking without touch

Request a detailed protocol

To test whether a unit responded to whisking itself, we extracted non-touch episodes as detailed above and computed time series of whisking amplitude and phase by band-pass filtering the whisker angle time series (6–30 Hz) and computing the Hilbert transform (Kleinfeld and Deschênes, 2011). Amplitudes were discretised (30 equi-populated bins) and the spiking data used to compute amplitude tuning functions. Phases for bins where the amplitude exceeded a given threshold were discretised (8 equi-populated bins) and used to construct phase tuning functions. To determine whether a unit was significantly amplitude-tuned, we fitted a regression line to its amplitude tuning curve and tested whether the slope was statistically significantly different to 0 (p=0.0025, Bonferroni-corrected). To determine whether a unit was significantly phase-tuned, we computed the maximum value of its phase tuning curve and compared this to the distribution of maxima of chance tuning functions. Chance tuning functions were obtained by randomly shifting the recorded spike sequences by 3000–8000 ms and recomputing tuning functions (500 times). A unit was considered phase-tuned if its tuning function maximum (computed using amplitude threshold of 2°) exceeded the 95th percentile of the shuffled distribution.

Acceleration tuning curves were quantified, for each unit, as follows. First, an acceleration tuning curve was estimated (as above). Units typically responded to both positive and negative accelerations, but with unequal weighting between them. To capture this, we fitted the following regression model to the tuning curve:

ri = μ0 + μ1ai + μ2iai

Here, for each bin i of the tuning curve, ri was the firing rate and ai was the acceleration; μ0, μ1 and μ2 were regression coefficients; the term i (i=1 if ai<0, i=0 otherwise) allowed for asymmetric responses to negative and positive acceleration. Based on its best-fitting regression coefficients (p=0.05), units were classified as: having ‘preference for negative acceleration’, if μ2 was significantly >0; having ‘preference for positive acceleration’, if μ2 was significantly <0; as having ‘no preferred direction’ if both μ1 was significantly >0, and μ2 was not significantly different from 0; and as ‘not acceleration sensitive’ if neither μ1 nor μ2 were significantly different from 0.

Generalised Linear Model (GLM)

Request a detailed protocol

To investigate how well PWNs encode a given sensory variable (e.g., whisker angle, curvature), we fitted single unit activity to a GLM (Nelder and Wedderburn, 1972; Truccolo et al., 2005; Paninski et al., 2007), using methods similar to Bale et al., 2013. For each unit, a ‘stimulus’ time series (x) (whisker angle or whisker curvature change) and a simultaneously recorded spike time series (n) were discretized into 1 ms bins: xt and nt denote respectively the stimulus value and spike count (0 or 1) in bin t.

GLMs express how the expected spike count of a unit depends both on the recent stimulus history and on the unit’s recent spiking history. The standard functional form of the model we used was:

(1) yt = fkTxt + hTnt + b

Here nt, the output in bin t, was a Bernoulli (spike or no-spike) random variable. The probability of a spike in bin t, yt, depended on three terms: (1) the dot product between the stimulus history vector xt= xt-Lk+1,,xt  and a ‘stimulus filter’ k (length Lk = 5); (2) the dot product between the spike history vector n = (nt-Lh+1,…,nt) and a ‘spike history filter’ ht (length Lh= 2); (3) a constant bias b, which sets the spontaneous firing rate. f() was the logistic function f(z)=(1+ez)1. The preferred direction of the GLM is determined by the sign of the stimulus filter. Positive (negative) k coefficients tend to make positive (negative) stimuli trigger spikes. Since we found that GLM performance was just as good with Lk = 1 as Lk = 5 (Figure 2—figure supplement 1C), we used results from the Lk = 1 case to define selectivity to curvature change direction: positive kimplies selectivity for positive curvature change; negative k selectivity for negative curvature change. When a whisker pushed against an object during protraction, curvature increased; when it pushed against an object during retraction, it decreased.

To consider whether units might encode multiple sensory variables (e.g., both whisker angle and whisker curvature change), we used a GLM with multiple stimulus history terms, one for each sensory variable:

yt=fk1Txt;1+k2Txt;2+htTn*+b

Here the indices 1, 2 label the sensory variables.

Training and testing of the GLM were done using a cross-validation procedure. For each unit, half of the trials were assigned randomly to a training set and half to a testing set. The training set was used to fit the parameters (k, h and b), while the testing set was used to quantify the similarity between the spike train of the recorded unit and that predicted by the GLM. GLM fitting was achieved by finding the parameter values (k, h and b), which minimized a cost function consisting of the sum of the negative log-likelihood and a regularizing term αk2. For all units, model prediction performance on the test set was robust to variation of α over several orders of magnitude: α was therefore set to a standard value of 0.01. To quantify the performance of the model, the sensory time series of the testing set was used as input to the best-fitting GLM to generate a ‘predicted’ spike train in response. Both real and predicted spike trains were then smoothed by convolution with a 100 ms box-car filter and the similarity between them quantified by the Pearson correlation coefficient (PCC). For each unit, the entire training/testing procedure was repeated for 10 random choices of training/testing set and the final prediction accuracy defined as the median of the 10 resulting PCC values. Data from these 10 samples were also used to test whether an individual unit exhibited statistically significant prediction performance for different sensory features. To test whether the results were robust to the smoothing time-scale, the above procedure was repeated for a range of box-car smoothing filters (1, 5, 10, 20, 50, 70 ms). To test whether a given ‘actual’ PCC was statistically significant, we tested the null hypothesis that it could be explained by random firing at the same time-averaged rate as that of the recorded unit. To this end, the recorded spike sequences were randomly shifted by 3000–8000 ms and the training/testing procedure above applied to this surrogate data. This was repeated 10 times and the resulting chance PCCs compared to the actual PCC using a signed-rank test, p=0.0025 (Bonferroni-corrected). This analysis was used to classify units as being ‘curvature-sensitive’.

Quadratic GLM

Request a detailed protocol

To test whether the units might exhibit nonlinear dependence on the stimulus parameters, we adapted the GLM defined above (Equation 1) to include quadratic stimulus variables (Rajan et al., 2013). This was important to assess whisker angular acceleration during free whisking, since a subset of units exhibited U-shaped acceleration tuning functions (Figure 3B). Given a stimulus time series xt, the quadratic stimulus history vector was [xt-Lk+1,…,xt,x2t-Lk+1,…,x2t]. Fitting methods were otherwise identical to those detailed above.

Effect of angle-curvature correlations on apparent neuronal stimulus encoding in the passive stimulation protocol

Request a detailed protocol

If, in a given recording, sensory variable X correlates with sensory variable Y, a neuron responsive purely to X will tend to appear tuned to Y. To investigate whether such an effect might produce apparent sensitivity to whisker angle in the passive stimulation paradigm, we simulated the response of curvature-tuned neurons to the whisker curvature change time series measured during passive white noise stimulation. To minimise free parameters, constrained GLMs (4 free parameters) were used, sensitive either to instantaneous curvature (k=[γ]) or to its first order derivative (k=γ[-1 1]), where γ was a signed, gain parameter. Parameters (h, b, γ) were adjusted to produce two spike trains (one for training, the other for testing) with a realistic white noise induced firing rate (~50 spikes/s; Bale et al., 2013). We then attempted to predict the simulated, curvature-evoked (training) spike train by fitting GLMs (length 5 stimulus filter, 8 free parameters) using as input either angle or curvature change. Cross-validated model accuracy was computed as the PCC between the predicted spike train and the testing spike train (both smoothed by convolution with a 5 ms box-car).

Effect of single-trial approach on GLM prediction performance

Request a detailed protocol

The objective of encoding models, such as GLMs, is to obtain an accurate description of the mapping between a stimulus and the neuronal spike trains it evokes. Since the random component of a neuron’s response is inherently unpredictable, the best any model can do is to predict the probability of the spike train. To enable this, encoding models have generally (with few exceptions; Park et al., 2014) been applied to a ‘repeated-trials’ paradigm, where a stimulus sequence (e.g., frozen white noise) is repeated on multiple ‘trials’ (Arabzadeh et al., 2005; Lottem and Azouz, 2011; Bale et al., 2013; Petersen et al., 2008; Pillow et al., 2008). Model accuracy can then be quantified, largely free of contamination from random response variability, by comparing (using PCC or otherwise) the trial-averaged response of the model to the trial-averaged response of the neuron.

In contrast, in the present study of awake, actively whisking mice, the precise stimulus (time series of whisker angle/curvature) was inevitably different on every pole presentation: there were no precisely repeated trials to average over. Our standard model performance metric (PCC) was computed by comparing the response on a single long, concatenated ‘trial’ with the corresponding GLM predicted response. Such a PCC is downwards biased by random response variability.

To gauge the approximate magnitude of this downward bias, we used a simulation approach. By simulating the response of model neurons, we could deliver identical, repeated trials and thereby compare model prediction performance by a metric based on trial-averaging with that based on the single-trial approach. To this end, for each recorded unit, we used the best-fitting curvature change GLM to generate 100 trials of spike trains evoked by the curvature time series measured for that unit. Data from the first of these trials was used to fit the parameters of a minimal ‘refitted GLM’ (stimulus filter length 1, spike history filter length 2; bias; total 4 free parameters), and the single-trial performance quantified, using the approach of the main text (Figure 2—figure supplement 1B, left). Next, we used the refitted GLM to generate 100 repeated trials of spike trains evoked by the curvature time series. Repeated-trials performance was then quantified as the PCC between PSTHs obtained by trial-averaging (Figure 2—figure supplement 1B, right).

References

  1. 1
  2. 2
  3. 3
  4. 4
  5. 5
  6. 6
  7. 7
  8. 8
  9. 9
  10. 10
  11. 11
  12. 12
  13. 13
  14. 14
    Tactile SLAM with a biomimetic whiskered robot
    1. C Fox
    2. M Evans
    3. M Pearson
    4. T Prescott
    (2012)
    IEEE International Conference on in Robotics and Automation (ICRA) pp. 4925–4930.
  15. 15
  16. 16
  17. 17
  18. 18
  19. 19
  20. 20
  21. 21
  22. 22
  23. 23
  24. 24
  25. 25
  26. 26
  27. 27
  28. 28
  29. 29
    Generalized linear models
    1. JA Nelder
    2. RWM Wedderburn
    (1972)
    Journal of the Royal Statistical Society 135:370–384.
    https://doi.org/10.2307/2344614
  30. 30
  31. 31
  32. 32
  33. 33
  34. 34
  35. 35
  36. 36
    Simultaneous localization and mapping on a multi-degree of freedom biomimetic whiskered robot
    1. MJ Pearson
    2. C Fox
    3. JC Sullivan
    4. TJ Prescott
    5. T Pipe
    6. B Mitchinson
    (2013)
    In IEEE International Conference On Robotics and Automation (ICRA) pp. 586–592.
  37. 37
  38. 38
  39. 39
  40. 40
  41. 41
  42. 42
    Tactile spatial resolution. III. a continuum mechanics model of skin predicting mechanoreceptor responses to bars, edges, and gratings
    1. JR Phillips
    2. KO Johnson
    (1981)
    Journal of Neurophysiology 46:1204–1225.
  43. 43
  44. 44
  45. 45
  46. 46
  47. 47
  48. 48
  49. 49
  50. 50
  51. 51
  52. 52
  53. 53
  54. 54
  55. 55
  56. 56
  57. 57
  58. 58
  59. 59
  60. 60

Decision letter

  1. David Kleinfeld
    Reviewing Editor; University of California, San Diego, United States

In the interests of transparency, eLife includes the editorial decision letter and accompanying author responses. A lightly edited version of the letter sent to the authors after peer review is shown, indicating the most substantive concerns; minor comments are not usually included.

Thank you for submitting your work entitled "Prediction of primary somatosensory neuron activity during active tactile exploration" for peer review at eLife. Your submission has been favorably evaluated by Eve Marder (Senior editor), a Reviewing editor, and three reviewers.

The Reviewing editor and the reviewers agree that this is a potentially very important work that defines the features of whisking and vibrissa contact that cause a trigeminal ganglion neuron to spike. Further, while this latter topic has been explored extensively in anesthetized animals, there is limited work in the awake case. Thus a study of parameters that lead to spiking in the awake case would be a welcome contribution to the vibrissa field. Yet there are major weaknesses in the current submission that must be addressed, with additional analysis and potentially new experiments, before we can proceed further.

Major issues:

The choice of the angular variable was absolute angle – which is likely to be a poor choice. The analysis should be redone in terms of deflection angle. The change in position of the mystacial pad also needs to be taken into account in the analysis, as noted by multiple reviewers. Related to this is a need to re-evaluate claims, as noted by reviewer 3 "[…]only 50% of the data shows a drop in performance when angle is used instead of curvature. The other 50% seems to perform as well for angle as it does for curvature. The conclusion that curvature is therefore the single most important parameter seems not to be supported by the data." This requires reanalysis, yet also may require some additional data to increase the spike count.

The application of the GLM was done with relatively few spikes, about 550 or 275 each for training and testing sets for 8 feature parameters and an unstated number of spike history parameters, which gives the editor pause. It would be thus imperative for the author to show the feature vector and the history term for at least a few units, as opposed to just predictions. In fact, the feature vector is a main point of such an analysis. Further, while the recorded data show clear cycle-by-cycle whisks (Figure 1), and "whisking" cells in the trigeminal ganglion faithfully respond on a cycle-by-cycle basis, these fast changes do not appear in the predicted spike rates of Figure 2. The reason for this omission, as pointed out by Reviewer #1, needs to be explained.

Further on the topic of analysis, Reviewer #2 notes "The authors use the poor performance of angle GLMs during active pole exploration as evidence that curvature changes are what drive PWNs. But it is known that touch dominates PWN spiking responses, so including touch periods when assessing angle GLM decoding will trivially result in very low GLM angle performance. More interesting would be comparing the performance of angle GLMs during non-touch, free whisking periods with performance of curvature GLMs during touch periods." The critical issue is to determine if there is a big difference between passive and active states or if, within strong statistics, there is not a big difference. This requires reanalysis, yet also may require some additional data to increase the spike count.

Please read through the attached thorough and very thoughtful reviews by three of your colleagues and please address all of the issues raised in a cover letter to accompany the resubmitted manuscript.

Reviewer #1:

This is an elegant work, addressing a crucial question – what do sensory neurons code during active exploration and touch – in a professional manner. The paradigm is simple and clear, and the paper is well written (for the most part), expressing clear thinking and straight-forward reasoning. This impressive work can potentially advance the understanding of sensory coding significantly. Yet, in its current form there is a danger that the paper will instead increase the confusion in the field – this is due to several major flaws that need to be carefully addressed.

1) The choice of the angular variable for analysis. The authors analyze the angle of the whisker relative to the head – let's call this the "absolute angle" here. There are 2 problems with it. One, portions of the pad rotate with significant angles during active whisking such that the absolute angle of a whisker changes but this has no effect on the shaft-follicle mechanical interactions (the entire complex moves together). This can be seen in the supplementary video of the paper, when examining the pad. Thus, the angle should be measured relative to the pad surrounding the whisker and not to the fixed head. Second, the relevant angular variable in the sensory coding game is most likely the change in angle upon contact. Both the "push angle" (Quist and Hartman, 2012; Bagdasarian et al., 2013) and the "Angle absorption" (Bagdasarian et al., 2013) carry meaningful information. This analysis of relative angular changes upon touch will also make the angle-curvature comparison more symmetric (currently the change in curvature upon touch is compared with the absolute angle).

2) After reading through the Results section it turns out that this study actually (re) revealed two types of cells – those termed by Szwed et al. (2003) as "Whisking" cells and "Touch" cells. As in Szwed et al. (2003), the former respond to whisking in air and are sensitive to the phase of whisking and the latter respond to touch and are sensitive to curvature changes. This fact should be described at the outset (Abstract, Introduction, Discussion) and compared with the relevant previous reports.

What seems to be missing here are two complementary analyses – the sensitivity of "Whisking" cells to touch and of "Touch" cells to whisking in air. Thus, the fractions of pure whisking and touch cells, and that of a combined "whisking-Touch" type (Szwed, 2003) is not clear. True, the cell count is not high (I believe it is total of 20, although this was hard to dig – please state it at the outset) but the cost of this should not be in flattening all types to one common denominator.

Importantly, the point in the paper where the reader realizes that half of the cells are "Whisking" cells is a confusing point, reflecting back on the initial analysis. For clarity, the separation between cell types should be clarified at the beginning.

3) The Abstract statement "[…]we found that primary neuron responses were poorly predicted by kinematics but well-predicted by rotational forces acting on the whisker[…]" is not supported by the data. In fact, the insisting on a single mechanical variable does not make much sense, is not convincing and, as said, is not consistent with the data presented in the paper. I strongly recommend re-considering it. First, the paper shows that half of the cells (the whisking cells) are actually sensitive to a kinematic variable, acceleration. Indeed, it is associated with force but aren't all kinematic changes associated with forces? Also, selecting whisker acceleration instead of other angular variables, such as phase and velocity, and even angle itself, seems to be arbitrary. As for the Touch cells, indeed the curvature is correlated with various angular variables, but the parameters of these correlations depend on the interactions with external objects (see Bagdasarian et al., 2013), interactions that are not investigated here. In fact, Bagdasarian et al. showed that relying on a single mechanical variable must lead to ambiguity about external features.

4) The paper deals only with slow dynamics of coding – in time scales of seconds and resolution > 100 ms. Analysis at higher temporal resolutions (as was impressively done by the Petersen lab previously) is probably not possible in the current challenging setup of TG recording in awake animals. Yet, perceptual processing depends crucially on within-cycle millisecond time scales. This should be emphasized at the outset and discussed in relation to candidate sensory variables and relevant external features. It seems that while this slow time course may be relevant to features such as object radial distance (in the case of touch) and intensity of whisking (combination of whisking amplitude and frequency, which determine average acceleration throughout the cycle – see Figure 3D), but not object azimuthal position, texture or shape and not phase within the whisking cycle. Also, the choice of 100 ms should be justified, and the dependency of the results on this choice should be described.

5) Figure 1B & C show a whisker that pushes against the object during retraction. The video and Figure 1D shows the "standard" contact, during protraction. The authors should make it clear whether their analysis was based on both directions. If so, this comment becomes a major one – the authors must include the direction as one of the analyzed variables and describe the dependency of the various findings on it. Also, curvatures are very strong in this study (Figure 2B, movie). Please refer to it and compare to free-head conditions in which often the minimal impingement principle (Prescott et al., 2013) applies. Please discuss the implications on the predominance of curvature coding in this study.

Reviewer #2:

Key findings:

1) PWNs are relatively insensitive to absolute whisker angle but highly sensitive to curvature change.

2) The degree to which PWNs are tuned to curvature change predicts their response to inertial force during free whisking.

These results are well supported by the data, and the data is valuable, nicely collected and presented. However, the results don't change the general understanding of PWN coding and thus are not novel. The paper focuses on overturning a straw-man characterization of the literature, that PWNs are tuned to absolute whisker angle, not deflection forces.

It is unfair to characterize the current results as "at odds with passive stimulation studies (Gibson 1983, Lichtenstein 1990[…])". The classic studies refer to PWN tuning to angle of deflection not absolute angle. These particular studies had no ability to assess PWN tuning in the absence of deflection. In Bale (2013), again the positional tuning was in the context of positional deflection not free whisking angle. Indeed, Leiser (2007) showed that firing rates are 10x higher in PWNs during contact than during awake free whisking. The logical interpretation of this and many other cited studies of PWN coding is that deflection-induced forces (often quantified as deflection angle) are the primary driver of PWN spiking, not whisker position absent deflection.

The authors use the poor performance of angle GLMs during active pole exploration as evidence that curvature changes are what drive PWNs. But it is known that touch dominates PWN spiking responses, so including touch periods when assessing angle GLM decoding will trivially result in very low GLM angle performance. More interesting would be comparing the performance of angle GLMs during non-touch, free whisking periods with performance of curvature GLMs during touch periods.

In the study, active touch occurs at multiple pole positions, while passive deflections have only one starting position. Thus the comparison of curvature and angle coupling between active and passive conditions (Figure 4) is apples to oranges. For example, if the mouse must position his whisker 10 degrees more protracted to contact the pole in one position vs. another during active sensing, the correlation between angle and curvature will be degraded when averaged across pole positions. Including non-touch periods in the analysis further degrades the correlation. Thus the poor cross-correlation for the awake condition in Figure 4D is trivial.

The more interesting and fair comparison is the extent to which active control of whisker position impacts the relationship between curvature and push angle. Push angle is defined as the angle through which the whisker is rotated into the object (see Quist and Hartmann, 2012 or Hires, 2013 for details). Active control could alter the rigidity of the follicle, impacting follicle stresses and thus spiking activity of PWNs. This should be detectable via comparing the difference in push angle/curvature coupling (i.e. the slope of touch trajectories in 4E, assuming curvature was measured at the same radial distance) between active and passive states.

Additional comments:

The data in the paper are interesting and do have potential to address some open questions that would increase the importance and novelty of the work. Some possible ideas that reanalysis could address, (in order of increasing interest):

1) Do PWNs that are tuned to acceleration direction show the same directional selectivity to deflection direction?

2) Do force components (Faxial, bending moment) differentially drive PWNs?

3) Do PWN responses to passive vs. active touch exhibit different sensitivity to deflection angle or whisker curvature change?

Detailed justification:

1) This would be a simple expansion of the analysis of Figure 3 to show correlation of directional tuning between touch and whisking across the population of whisking sensitive neurons. This would make the Figure 3 result more compelling.

2) Using Faxial and Bending Moment as independent predictors in a GLM could determine if PWNs specialize for these components during active touch. Longitudinal deflection of whiskers causes robust responses in PWNs (Zucker and Welker 1969, Stuttgen 2008). Axial and lateral/moment ratios are used for radial object localization (Solomon 2011, Pammer 2013). This could bridge those physiology and behavior results.

3) Quantifying a non-trivial difference between passive and active touch, particularly if reflected in spiking activity would make the paper much more interesting. In cortex differences have been seen (e.g. airpuff of whiskers when awake elicits much weaker dendritic responses in Figure 1—figure supplement 1 than active touch, despite the much greater deflections air puffs evoke Xu 2012 Nature). Are these differences inherited from PWNs due to different mechanical coupling or sensitivity between these states?

Reviewer #3:

The manuscript of Campagner et al. investigates the whisker parameters (angle and curvature) that allow reliable prediction of spiking of primary whisker neurons upon passive or active touch. The manuscript is potentially interesting, although I have some concern about experimental setup and the validity of comparisons between passive and active conditions. Additionally, even though curvature reliably predicts spiking in awake rats for a subset of the data, the range of quantified reliability is large and not discussed.

1) The major conclusion (curvature much better predicts spiking than angle) is based predominantly on Figure 2C. The full range of reliability measures for curvature is 0.1 – 0.9. The authors put a lot of emphasis on the fraction of high values (max 0.88), but completely ignore the lower measures. Vice versa, the high values for angle GLMs are only briefly mentioned and emphasis put on poor predicting values. It seems very relevant to discuss the entire range for both conditions. Additionally, only 50% of the data shows a drop in performance when angle is used instead of curvature. The other 50% seems to perform as well for angle as it does for curvature. The conclusion that curvature is therefore the single most important parameter seems not to be supported by the data. Since the authors also describe W-sensitive neurons (subsection "Primary whisker neuronal activity during whisking is predicted by moment”), it seems more optimal to present the data in W-sensitive, curvature-sensitive and angle-sensitive fractions of the population data (how many neurons were recorded from in n=10 animals?).

2) Angle changes as a function of curvature as presented in Figure 4—figure supplement 1. This is very informative for the interpretation of Figure 4E and I would suggest moving Figure 4—figure supplement 1 into the main manuscript. Since angle changes dramatically during touch for individual pole positions (up to 20 degrees change in whisker angle for a fixed pole position), it can be concluded that angle is not independent from curvature and this probably underlies the range of reliability measures in curvature GLM and angle GLM. The authors should better discuss how the angle-curvature inter-dependence influences their model.

3) Passive stimulation is achieved by trimming the whiskers to 5 mm (methods). Under these conditions, it is (in my experience) impossible to induce meaningful curvature changes. The authors should better explain the experimental conditions if their experimental design allows accurate curvature measurement with a whisker trimmed to 5 mm and capillary 2 mm on whisker (Figure 4).

https://doi.org/10.7554/eLife.10696.017

Author response

The Reviewing editor and the reviewers agree that this is a potentially very important work that defines the features of whisking and vibrissa contact that cause a trigeminal ganglion neuron to spike. Further, while this latter topic has been explored extensively in anesthetized animals, there is limited work in the awake case. Thus a study of parameters that lead to spiking in the awake case would be a welcome contribution to the vibrissa field. Yet there are major weaknesses in the current submission that must be addressed, with additional analysis and potentially new experiments, before we can proceed further.

Major issues:

The choice of the angular variable was absolute angle – which is likely to be a poor choice. The analysis should be redone in terms of deflection angle. The change in position of the mystacial pad also needs to be taken into account in the analysis, as noted by multiple reviewers.

As detailed in response to Reviewer 1 (point 1) and Reviewer 2, we have re-analysed the video data to extract push angle and performed new GLM analysis to consider how push angle is encoded. However, we also argue (Reviewer 1, point 1) that including analysis of absolute angle in the manuscript is useful, since it offers novel, direct evidence in favour of the view developed on behavioural grounds (Bagdasarian et al. 2013) that absolute angle only weakly influences mechanoreceptors in the follicle.

Related to this is a need to re-evaluate claims, as noted by reviewer 3 "[…]only 50% of the data shows a drop in performance when angle is used instead of curvature. The other 50% seems to perform as well for angle as it does for curvature. The conclusion that curvature is therefore the single most important parameter seems not to be supported by the data." This requires reanalysis, yet also may require some additional data to increase the spike count.

As detailed in response to Reviewer 3 (point 1), we have performed new statistical tests on the angle vs. curvature prediction performance, for each unit, on an individual basis. We found that the cases where angle-based GLMs performed slightly better than curvature-based GLMs could be attributed, in all but a single case, to chance. This analysis confirms our original report that curvature predicts spikes more accurately than angle.

The application of the GLM was done with relatively few spikes, about 550 or 275 each for training and testing sets for 8 feature parameters and an unstated number of spike history parameters, which gives the editor pause. It would be thus imperative for the author to show the feature vector and the history term for at least a few units, as opposed to just predictions. In fact, the feature vector is a main point of such an analysis.

As suggested, we now illustrate stimulus filter and spike history filters for example units in Figure 2—figure supplement 3. We agree with the editor’s concern and were at pains, in this study, to avoid over-fitting. To this end, we used regularization and cross-validation. We also kept the number of free parameters to a minimum: in most cases, 8 was the total number of parameters – and the Results have been revised to clarify this: “(8 total fitted parameters) and: 5 for stimulus filter, 2 for history filter, 1 bias”). In fact, when we cut down the number of free parameters even further (to 4), we found statistically indistinguishable spike prediction results (p=0.35, signed-rank test). This argues that the spike prediction performance we report is not artificially inflated by over-fitting. Thus, we are confident that we have not over-estimated the predictive power of the models.

Further, while the recorded data show clear cycle-by-cycle whisks (Figure 1), and "whisking" cells in the trigeminal ganglion faithfully respond on a cycle-by-cycle basis, these fast changes do not appear in the predicted spike rates of Figure 2. The reason for this omission, as pointed out by Reviewer #1, needs to be explained.

In this study, we primarily considered coding at a time-scale corresponding approximately to one whisking cycle (~100 ms); we did not consider faster time-scales. As detailed in response to Reviewer 1 (point 4): (1) focus on the ~100 ms time-scale is justified due to the relatively slow dynamics that characterize whisking against a smooth pole; (2) new analysis, where we vary the time-scale shows that our main result that curvature predicts spikes better than angle is robust across time-scales; (3) in our experiment, where we do not have the benefit of repeated trials over which to average, it is extremely difficult to test the capacity of a model to predict spikes at millisecond precision.

Further on the topic of analysis, Reviewer #2 notes "The authors use the poor performance of angle GLMs during active pole exploration as evidence that curvature changes are what drive PWNs. But it is known that touch dominates PWN spiking responses, so including touch periods when assessing angle GLM decoding will trivially result in very low GLM angle performance. More interesting would be comparing the performance of angle GLMs during non-touch, free whisking periods with performance of curvature GLMs during touch periods." The critical issue is to determine if there is a big difference between passive and active states or if, within strong statistics, there is not a big difference. This requires reanalysis, yet also may require some additional data to increase the spike count.

We have carried out the analysis suggested by Reviewer 2. As detailed below, the results support our original conclusion: performance of angle-based GLMs remains low, even when restricted to data from non-touch episodes. We agree that it would be interesting to compare responses in active vs. passive states but please note that it was not the aim of the current study to compare the responses to passive and active pole-contact and we do not have data that speaks to this point. To address the point would require an entirely new study. We have revised the introductory text to clarify the aims of our study.

Please read through the attached thorough and very thoughtful reviews by three of your colleagues and please address all of the issues raised in a cover letter to accompany the resubmitted manuscript.

Reviewer #1:

This is an elegant work, addressing a crucial question – what do sensory neurons code during active exploration and touch – in a professional manner. The paradigm is simple and clear, and the paper is well written (for the most part), expressing clear thinking and straight-forward reasoning. This impressive work can potentially advance the understanding of sensory coding significantly. Yet, in its current form there is a danger that the paper will instead increase the confusion in the field – this is due to several major flaws that need to be carefully addressed.

We thank the reviewer for stating that “this impressive work can potentially advance the understanding of sensory coding significantly”. We address his/her concerns below.

1) The choice of the angular variable for analysis. The authors analyze the angle of the whisker relative to the head – let's call this the "absolute angle" here. There are 2 problems with it. One, portions of the pad rotate with significant angles during active whisking such that the absolute angle of a whisker changes but this has no effect on the shaft-follicle mechanical interactions (the entire complex moves together). This can be seen in the supplementary video of the paper, when examining the pad. Thus, the angle should be measured relative to the pad surrounding the whisker and not to the fixed head. Second, the relevant angular variable in the sensory coding game is most likely the change in angle upon contact. Both the "push angle" (Quist and Hartman, 2012; Bagdasarian et al., 2013) and the "Angle absorption" (Bagdasarian et al., 2013) carry meaningful information. This analysis of relative angular changes upon touch will also make the angle-curvature comparison more symmetric (currently the change in curvature upon touch is compared with the absolute angle).

In response, we have conducted new analysis to consider encoding of push angle and have revised the manuscript to include it (detailed below). Our rationale for considering absolute angle in this study is that it is important for the animal to know location of its whiskers relative to the head, and hence it is interesting to assess how well absolute angle (and quantities such as amplitude, phase and set point derived from it) might be encoded. Including analysis of this quantity connects our study to a substantial previous literature (Szwed et al. 2003; Leiser and Moxon 2007; Khatri et al. 2009; Petreanu et al. 2012; Huber et al. 2012; Hires et al. 2015; Moore et al. 2015; Peron et al. 2015). Our result that Primary Whisker Neurons (PWNs) encode absolute angle poorly is fully consistent with the reviewer’s argument that absolute angle has relatively little effect on follicle mechanoreceptors. Yet, to the best of our knowledge, this point has never previously been demonstrated electrophysiologically in the awake animal. We take the point that this connection was not discussed in the original submission and have revised the Discussion to do so:

Discussion: “Our finding that whisker angle predicts PWN spikes poorly indicates that whisker angle can change without modulating mechanotransduction in the follicle. This is consistent with evidence that, during artificial whisking, the follicle-shaft complex moves as a rigid unit (Bagdasarian et al. 2013).”

To address the reviewer’s suggestion to consider push angle, we have re-analysed our video data to extract push angle during touch episodes. For each recorded unit, we trained a GLM to predict PWN spikes from push angle (using data from touch episodes). In the revised text, Figure 1 now includes push angle; and the Results have been revised to report that:

Results: “GLMs based on curvature change also predicted spike trains more accurately than GLMs based on “push angle” – the change in angle as the whisker pushes against an object (Figure 1E; median PCC 0.25 IQR 0.04-0.45; p=0.006, signed-rank test). Moreover, prediction accuracy of GLMs fitted with both push angle and curvature change (median PCC 0.52, IQR 0.2-0.69) inputs was no better than that of GLMs fitted with curvature alone (p = 0.43, signed-rank test)”.

2) After reading through the Results section it turns out that this study actually (re) revealed two types of cells – those termed by Szwed et al. (2003) as "Whisking" cells and "Touch" cells. As in Szwed et al. (2003), the former respond to whisking in air and are sensitive to the phase of whisking and the latter respond to touch and are sensitive to curvature changes. This fact should be described at the outset (Abstract, Introduction, Discussion) and compared with the relevant previous reports.

What seems to be missing here are two complementary analyses – the sensitivity of "Whisking" cells to touch and of "Touch" cells to whisking in air. Thus, the fractions of pure whisking and touch cells, and that of a combined "whisking-Touch" type (Szwed, 2003) is not clear. True, the cell count is not high (I believe it is total of 20, although this was hard to dig – please state it at the outset) but the cost of this should not be in flattening all types to one common denominator.

In order to address this concern, we have, first, revised the text to introduce the previously demonstrated existence of subtypes of PWN cell:

Introduction: “They are both functionally and morphologically diverse; including types responsive to whisker-object contact and/or whisker self-motion (Szwed et al. 2003; Ebara et al. 2002).”

Second, we have performed new analysis to determine whether there are corresponding subtypes in our data. “Whisking sensitive” units were defined in the submitted version of the manuscript (Figure 3A). In the revised manuscript, we now define “curvature-sensitive” units as those for which the curvature-based prediction of spikes is statistically significant (Figure 2E left, black circles; Peron et al. 2015). We found that:

Results: “At the level of individual units, 90% had above chance PCC and we termed these ‘curvature-sensitive’”.

Results: “Consistent with Szwed et al. (2003), PWNs were diverse: 45% were curvature-sensitive (significant PCC for curvature based GLM); 45% were both whisking and curvature sensitive and 5% were whisking sensitive but not curvature-sensitive.”

The Discussion section has been revised to consider these findings:

Discussion: “Our results also provide a mechanical basis for previous findings: our finding of subtypes of curvature-only and curvature-acceleration PWNs is consistent with previous reports of ‘touch’ and ‘whisking-touch’ units (Szwed et al. 2003; 2006).”

Importantly, the point in the paper where the reader realizes that half of the cells are "Whisking" cells is a confusing point, reflecting back on the initial analysis. For clarity, the separation between cell types should be clarified at the beginning.

This has been done:

Results: “As detailed below, PWNs were diverse, with some responding only to touch, others also to whisker motion”.

Concerning cell count, this is indeed 20 and this is stated in the first sentence of Results. “We recorded the activity of single PWNs from awake mice (Figure 1A, E, Figure 1—figure supplement 1) as they actively explored a metal pole with their whiskers (N = 20 units).” For clarity, we now also report the cell count in the Materials and methods section (subsection “Electrophysiological data analysis”, first paragraph).

3) The Abstract statement "[…]we found that primary neuron responses were poorly predicted by kinematics but well-predicted by rotational forces acting on the whisker[…]" is not supported by the data. In fact, the insisting on a single mechanical variable does not make much sense, is not convincing and, as said, is not consistent with the data presented in the paper. I strongly recommend re-considering it. First, the paper shows that half of the cells (the whisking cells) are actually sensitive to a kinematic variable, acceleration. Indeed, it is associated with force but aren't all kinematic changes associated with forces? Also, selecting whisker acceleration instead of other angular variables, such as phase and velocity, and even angle itself, seems to be arbitrary. As for the Touch cells, indeed the curvature is correlated with various angular variables, but the parameters of these correlations depend on the interactions with external objects (see Bagdasarian et al., 2013), interactions that are not investigated here. In fact, Bagdasarian et al. showed that relying on a single mechanical variable must lead to ambiguity about external features.

We agree with the reviewer that the quoted sentence of Abstract needed reconsideration – it was misleading. The reviewer’s comment made us realize that part of the rationale for our analysis was poorly explained. Our aim was to assess the influence of two types of rotational force (moment) on PWN activity: first, the moment associated with whisking bending; second, the moment associated with whisker motion. We did this by means of proxies. We used curvature as a proxy for bending moment, since bending moment is closely related to the curvature of a whisker at its base (Birdwell et al. 2007). We used angular acceleration as a proxy for the rotational force acting on the follicle during whisker self-motion, since, through the rotational form of Newton’s second law, these two quantities are proportional, under our experimental conditions (Quist et al. 2014). For this reason, we argue that our choice to use acceleration for our analysis is not arbitrary, but well-motivated by the physics. As the reviewer points out, angular acceleration is a kinematic variable but, in the sense that it is a proxy for moment, it can also be considered a “mechanical” variable. However, we now realize that it was confusing to make a dichotomy between kinematic and mechanical. We have rewritten the text as follows.

Abstract: “Using Generalised Linear Models, we found that primary neuron responses were poorly predicted by whisker angle, but well-predicted by rotational forces acting on the whisker”.

Materials and methods: “As expressed by Newton’s second law of rotational motion, the moment – or torque –of a rigid body, rotating in a plane, is proportional to the body’s angular acceleration. […] Thus, to assess whether such a moment is encoded by PWNs, we measured angular whisker acceleration during free whisking as a proxy.“.

The reviewer’s other concern here is that “insisting on a single mechanical variable does not make much sense”. We agree with the reviewer that mechanical variables other than moment may also be encoded by the system. To address this, we also quantified both axial and lateral contact forces during episodes of whisker-pole contact. We found these variables to be strongly associated with bending moment (Figure 2—figure supplement 2) and not, therefore, under our experimental conditions, to convey independent information. It will, in future work, be interesting to explore different experimental conditions, where different mechanical variables may decouple. We have revised the Discussion to consider this issue.

Discussion: “In this study, we considered PWN encoding under conditions of pole contact, since this is well-suited to reaction force estimation […] by spike timing on a finer time-scale (Panzeri et al. 2001; Petersen et al. 2001; Arabzadeh et al. 2005; Bale et al. 2015).”

4) The paper deals only with slow dynamics of coding – in time scales of seconds and resolution > 100 ms. Analysis at higher temporal resolutions (as was impressively done by the Petersen lab previously) is probably not possible in the current challenging setup of TG recording in awake animals. Yet, perceptual processing depends crucially on within-cycle millisecond time scales. This should be emphasized at the outset and discussed in relation to candidate sensory variables and relevant external features. It seems that while this slow time course may be relevant to features such as object radial distance (in the case of touch) and intensity of whisking (combination of whisking amplitude and frequency, which determine average acceleration throughout the cycle – see Figure 3D), but not object azimuthal position, texture or shape and not phase within the whisking cycle. Also, the choice of 100 ms should be justified, and the dependency of the results on this choice should be described.

We agree with the reviewer that coding time-scales are task-dependent and range widely. As the reviewer notes, the natural time-scale of pole exploration is slow (of order 100 ms). Thus, it is appropriate to use a corresponding time-scale for the analysis of neuronal responses in the current task. However, we take the reviewer’s point that finer time-scales of coding are likely to be important in other behavioural contexts and, as suggested, we have revised the text to discuss this:

Discussion: “In this study, we considered PWN encoding under conditions of pole contact, since this is well-suited to reaction force estimation […] by spike timing on a finer time-scale (Panzeri et al. 2001; Petersen et al. 2001; Arabzadeh et al. 2005; Bale et al. 2015).”

Furthermore, as suggested, we tested the dependency of the results on the time-scale of analysis. We did this by repeating the GLM analysis using a variety of lengths of smoothing filter (1-100ms). In all cases, curvature-based GLMs were better predictors of PWN activity than angle-based ones (signed-rank test, p<0.05, after Bonferroni correction):

Results: “The result was also robust to time-scale: prediction accuracy based on curvature was significantly greater than that based on angle for smoothing time-scales in the range 1-100ms (signed-rank test, p<0.05, Bonferroni-corrected).”

Materials and methods: “To test whether the results were robust to the smoothing time-scale, the above procedure was repeated for a range of box-car smoothing filters (1, 5, 10, 20, 50, 70 ms).”

The reviewer also makes an important, technical point. Almost all previous studies of ms-precise spike-timing (e.g., Optican and Richmond, 1987; de Ruyter et al. 1997; Reinagel and Reid, 2000; Panzeri et al. 2001; Johansson and Birzneiks, 2004; Arabzadeh et al. 2006; Montemurro et al. 2007; Gollisch and Meister, 2008; Lottem and Azouz, 2011; Bale et al. 2015) have been carried out in anaesthetized animals and have relied fundamentally on the ability to repeat the precise same mechanical stimulus many times (“trials”). In this way, one can make very precise measurements of the variability in spike timing (jitter) from trial to trial: noise is attenuated by trial-averaging. As the reviewer correctly implies, it is difficult to analyze the dynamics of coding on the ms time-scale in the awake, behaving animal. In the awake, actively whisking rodent, there are no precisely repeated trials and trial-averaging is not an available option. As we show in Figure 2—figure supplement 1B, the lack of repeated trials was a limiting factor for model prediction accuracy in our study. Our strategy was to take advantage of the natural, ~100 ms time-scale of whisker-object interactions in our pole exploration paradigm, to perform averaging in the temporal domain.

5) Figure 1B & C show a whisker that pushes against the object during retraction. The video and Figure 1D shows the "standard" contact, during protraction. The authors should make it clear whether their analysis was based on both directions. If so, this comment becomes a major one – the authors must include the direction as one of the analyzed variables and describe the dependency of the various findings on it.

The reviewer is correct that our analysis is based on both directions. Our GLM framework encompasses directional selectivity as follows. In a curvature GLM, the preferred direction is determined by the sign of the stimulus filter. Positive stimulus filter coefficients tend to make positive curvature change trigger spikes; negative coefficients make negative curvature trigger spikes. The directional selectivity of a GLM is particularly simple when the stimulus filter is a single number. This is relevant to our study, since we found that such simple GLMs predicted spikes as well as GLMs with a length 5 stimulus filter (Figure 2—figure supplement 1C). When a whisker pushes against an object during protraction, curvature increases; when it pushes against an object during retraction, it decreases. Materials and methods have been revised:

Materials and methods: “Since we found that GLM performance was just as good with Lk= 1 as Lk = 5 (Figure 2—figure supplement 1C), we used […] when it pushes against an object during retraction, it decreases.”

To address the reviewer’s point, we conducted further analysis, using the above approach, to determine the directional selectivity of every curvature-sensitive unit:

Results: “At the level of individual units, 90% had above chance PCC and we termed these ‘curvature-sensitive’ (Materials and methods). Of the curvature-sensitive units, 61% were sensitive to positive curvature change and 39% to negative curvature change (Materials and methods).”

Also, curvatures are very strong in this study (Figure 2B, movie). Please refer to it and compare to free-head conditions in which often the minimal impingement principle (Prescott et al., 2013) applies. Please discuss the implications on the predominance of curvature coding in this study.

The maximum curvatures in our study (~0.2 mm-1) are consistent with previous studies of object localisation and tactile maze navigation in head-fixed mice where curvature has been measured (O’Connor et al. 2010; Sofroniew et al. 2014). Unfortunately, to the best of our knowledge, there are no published studies of free-head mice that report whisker curvature, which makes it difficult to compare our data to free-head conditions. Nevertheless, if we understand correctly, the reviewer’s concern is with the sensitivity of PWNs to smaller curvatures. We can address this point, since our dataset contains a wide range of curvature change values, by constructing tuning curves. We have added example tuning curves to Figure 2—figure supplement 1D which show that neuronal firing rate was modulated not only by strong curvature change (~0.2 mm-1), but also weaker ones (~0.01 mm-1), associated with subtler pole contact. Also, the new curvature graph in the movie shows spikes evoked by small curvature change.

Reviewer #2:

Key findings:

1) PWNs are relatively insensitive to absolute whisker angle but highly sensitive to curvature change.

2) The degree to which PWNs are tuned to curvature change predicts their response to inertial force during free whisking.

These results are well supported by the data, and the data is valuable, nicely collected and presented. However, the results don't change the general understanding of PWN coding and thus are not novel.

We thank the reviewer for saying that “the data is valuable” and that our main findings are “well supported by the data”, but we respectfully disagree with the statement that our results are not novel. We agree that our results do not falsify the Hartmann model of PWN coding but we argue that, given the necessary simplifications in any theoretical model, and the daunting complexity of the awake, behaving situation, it is remarkable to find clear support for the theory. As we discuss in the manuscript (subsection “Mechanical framework for tactile coding”, first paragraph), Hartmann and colleagues (Solomon and Hartmann 2006; Birdwell et al. 2007) proposed that the evidence that PWNs respond to touch (Szwed et al. 2003; 2006) could, in principle, be explained by the hypothesis that mechanoreceptors in the follicle are sensitive to whisker bending moment. However, before our work, no study had ever directly tested the theory by simultaneously measuring both PWN activity and whisker reaction forces. This is significant, since the theory makes simplifying approximations and assumptions (for example, concerning boundary conditions) that it is unsafe to assume will necessarily hold in the awake, behaving animal. Since the whisker system is an important model in neuroscience, we argue that this was a major gap in the literature and that our study makes a substantial contribution to putting our general understanding of somatosensory coding on a more solid, mechanical basis.

The paper focuses on overturning a straw-man characterization of the literature, that PWNs are tuned to absolute whisker angle, not deflection forces.

It is unfair to characterize the current results as "at odds with passive stimulation studies (Gibson 1983, Lichtenstein 1990[…])". The classic studies refer to PWN tuning to angle of deflection not absolute angle. These particular studies had no ability to assess PWN tuning in the absence of deflection. In Bale (2013), again the positional tuning was in the context of positional deflection not free whisking angle. Indeed, Leiser (2007) showed that firing rates are 10x higher in PWNs during contact than during awake free whisking. The logical interpretation of this and many other cited studies of PWN coding is that deflection-induced forces (often quantified as deflection angle) are the primary driver of PWN spiking, not whisker position absent deflection.

We agree with the reviewer that passive whisker stimulation can be understood from a force-encoding point of view and we accept that the quoted sentence from the original submission (“at odds with[…]”) lacked nuance. We have revised the text to remove the sentence (Introduction, last paragraph).

Our contention is that there is an apparent contrast (not, we agree, a contradiction) between passive stimulation studies and our data, and that our study sheds light on why this is so. With passive stimulation, whisker angle (more precisely, change in whisker angle with respect to a whisker’s resting angle) correlates with firing rate (Zucker and Welker 1969; Gibson and Welker 1983; Lichtenstein et al. 1990; Jones et al. 2004; Arabzadeh et al. 2005; Bale and Petersen 2009; Lottem and Azouz 2011; Bale et al. 2013) whereas, in our data, whisker angle did not predict PWN firing rate. Thus, apparently the same variable exerts different effects under the two conditions. The reason that this is not actually a contradiction (consistent with the reviewer’s remark) is, as we show, that passive stimulation does not just change whisker angle but also bends the whisker (Figure 4 and Figure 4—figure supplement 2). This characteristic has, however, not been widely appreciated: passive stimulation is usually carried out on shortened whiskers, which are so stiff that the bending is hard to detect unless, as here, measured with high-resolution imaging.

The authors use the poor performance of angle GLMs during active pole exploration as evidence that curvature changes are what drive PWNs. But it is known that touch dominates PWN spiking responses, so including touch periods when assessing angle GLM decoding will trivially result in very low GLM angle performance. More interesting would be comparing the performance of angle GLMs during non-touch, free whisking periods with performance of curvature GLMs during touch periods.

Our primary aim was to seek input variables that could predict spikes during pole exploration as a whole – and we therefore argue that it is important to test the GLM’s performance on long episodes encompassing both contact and non-contact. However, the comparison suggested by the reviewer adds insight and we thank him/her for the suggestion. We performed new analysis and found that:

Results: “Curvature GLMs also predicted spikes during touch episodes significantly more accurately (median PCC 0.57, IQR 0.23-0.72) than did angle GLMs during non-touch episodes (median 0.06, IQR 0.02-0.35; p=0.005, signed-rank test)”.

In the study, active touch occurs at multiple pole positions, while passive deflections have only one starting position. Thus the comparison of curvature and angle coupling between active and passive conditions (Figure 4) is apples to oranges. For example, if the mouse must position his whisker 10 degrees more protracted to contact the pole in one position vs. another during active sensing, the correlation between angle and curvature will be degraded when averaged across pole positions. Including non-touch periods in the analysis further degrades the correlation. Thus the poor cross-correlation for the awake condition in Figure 4D is trivial.

We agree that the weak cross-correlation between angle and curvature change for the awake condition are explicable in the terms stated by the reviewer and shown by us in figure 4—figure supplement 1 of the original submission. However, we respectfully disagree that this makes these results “trivial” and note that other reviewers found these results insightful and requested that they be moved into the main text (new Figure 4F and G).

The more interesting and fair comparison is the extent to which active control of whisker position impacts the relationship between curvature and push angle. Push angle is defined as the angle through which the whisker is rotated into the object (see Quist and Hartmann, 2012 or Hires, 2013 for details). Active control could alter the rigidity of the follicle, impacting follicle stresses and thus spiking activity of PWNs. This should be detectable via comparing the difference in push angle/curvature coupling (i.e. the slope of touch trajectories in 4E, assuming curvature was measured at the same radial distance) between active and passive states.

We thank the reviewer for the suggestion to consider push angle. In response, we have re-analysed all our video data to extract push angle during touch episodes. Using these data, we have performed new GLM analysis to consider the relation between push angle and PWN spiking in the awake state. For each recorded unit, we trained GLM to predict PWN spikes from push angle. We have revised Figure 1 to show push angle and Results:

Results: “GLMs based on curvature change also predicted spike trains more accurately than GLMs based on “push angle” […] inputs was no better than that of GLMs fitted with curvature alone (p = 0.43, signed-rank test).”

Additional comments:

The data in the paper are interesting and do have potential to address some open questions that would increase the importance and novelty of the work. Some possible ideas that reanalysis could address, (in order of increasing interest):

1) Do PWNs that are tuned to acceleration direction show the same directional selectivity to deflection direction?

We have performed new analyses to address this point (see Materials and methods, subsection “Responses to whisking without touch”, last paragraph and subsection “Generalised Linear Model (GLM)”, first paragraph). For the whisking sensitive units, reported in Figure 3A, we found that:

Results: “For 70% of whisking-sensitive units, directional selectivity for acceleration was consistent with that for curvature.”

2) Do force components (Faxial, bending moment) differentially drive PWNs?

We agree that this is an interesting possibility. However, under our experimental conditions, we found that axial force is strongly (nonlinearly) associated with bending moment (Figure 2—figure supplement 2). The relationships so strong (under our conditions) that we were concerned that it is difficult to separate out contributions to PWN firing, and that it is misleading to make any strong claim about lack of differential drive. For the reviewer’s information, the performance (PCC) of axial force quadratic GLMs was median 0.35 (IQR 0.08-0.52), significantly lower than the PCC for curvature change (signed-rank test, p = 0.01). To resolve this question would require an entirely new study, investigating other stimuli.

3) Do PWN responses to passive vs. active touch exhibit different sensitivity to deflection angle or whisker curvature change?

We agree that this is an interesting question: however, it is beyond the scope of the present investigation and requires a completely new study.

Reviewer #3:

The manuscript of Campagner et al. investigates the whisker parameters (angle and curvature) that allow reliable prediction of spiking of primary whisker neurons upon passive or active touch. The manuscript is potentially interesting, although I have some concern about experimental setup and the validity of comparisons between passive and active conditions. Additionally, even though curvature reliably predicts spiking in awake rats for a subset of the data, the range of quantified reliability is large and not discussed.

1) The major conclusion (curvature much better predicts spiking than angle) is based predominantly on Figure 2C. The full range of reliability measures for curvature is 0.1 – 0.9. The authors put a lot of emphasis on the fraction of high values (max 0.88), but completely ignore the lower measures. Vice versa, the high values for angle GLMs are only briefly mentioned and emphasis put on poor predicting values. It seems very relevant to discuss the entire range for both conditions.

We agree with the reviewer’s concern to clearly report variability in the data and, to this end, the relevant figures (Figure 2D, E and Figure 3E) show not only averages but also all individual data points for every unit. However, we accept that we did not, in the original text, emphasize the variability. In response, we have revised the text to do so. We have also carried out new analysis detailed below.

Results: “Although the activity of most units was better predicted by whisker curvature change than by whisker angle, there was significant variability in prediction performance, andthere were a few units for which the angle prediction performance was appreciable (Figure 2D).”

Discussion:”In this study, we considered PWN encoding under conditions […] performance varied across units, suggesting that other force components may also be encoded.”

Additionally, only 50% of the data shows a drop in performance when angle is used instead of curvature. The other 50% seems to perform as well for angle as it does for curvature. The conclusion that curvature is therefore the single most important parameter seems not to be supported by the data.

We take the reviewer’s point that, for some units, the angle performance (PCC) is higher than the curvature PCC and thank him/her for the opportunity to drill deeper into this point. In principle, these might be genuine differences; alternatively, they might be statistical fluctuations in the measurements due to chance. To address this, we performed a new, cell-by-cell analysis, where we used a resampling technique (detailed, subsection “Generalised Linear Model (GLM)”, last paragraph) to statistically compare angle vs. curvature PCC for each unit, on an individual basis. The new results support the original conclusion that curvature is encoded better than angle:

Results: “Moreover, on a unit-by-unit basis, for 65% of units, curvature change GLMs predicted spikes better than angle (signed-rank test, p<0.05, Bonferroni corrected); only for 5% of units did angle predict spikes better than curvature change.”

Since the authors also describe W-sensitive neurons (subsection "Primary whisker neuronal activity during whisking is predicted by moment”), it seems more optimal to present the data in W-sensitive, curvature-sensitive and angle-sensitive fractions of the population data.

We have performed a new analysis to classify the recorded units as curvature and/or whisking sensitive (see also response to Reviewer #1, point 2) and report the results in the revised text:

Results: “Consistent with Szwed et al. (2003), PWNs were diverse: 45% were curvature-sensitive (significant PCC for curvature based GLM); 45% were both whisking and curvature sensitive and 5% were whisking sensitive but not curvature-sensitive.”

(How many neurons were recorded from in n=10 animals?).

N=20 units were recorded and this is stated in the first sentence of Results. “We recorded the activity of single PWNs from awake mice (Figure 1A, E, Figure 1—figure supplement 1) as they actively explored a metal pole with their whiskers (N = 20 units).” For clarity, we now also report the N number in Materials and methods (subsection “Electrophysiological data analysis”, first paragraph).

2) Angle changes as a function of curvature as presented in Figure 4—figure supplement 1. This is very informative for the interpretation of Figure 4E and I would suggest moving Figure 4—figure supplement 1 into the main manuscript.

This has been done. These plots are now in the main text (Figure 4F and G of the revised manuscript).

Since angle changes dramatically during touch for individual pole positions (up to 20 degrees change in whisker angle for a fixed pole position), it can be concluded that angle is not independent from curvature and this probably underlies the range of reliability measures in curvature GLM and angle GLM. The authors should better discuss how the angle-curvature inter-dependence influences their model.

We take the point that angle-curvature correlation was not mentioned in the Discussion section and have revised the text to do so:

Discussion: “A challenge of studying neural coding under unconstrained, awake conditions is that sensory variables tend to correlate. A useful feature of the GLM training procedure is that it takes such correlations into account. We found that, although whisker angle predicted spikes for a subset of units, this effect was very largely explained by a curvature-coding model, together with the correlation between angle and curvature.”

3) Passive stimulation is achieved by trimming the whiskers to 5 mm (methods). Under these conditions, it is (in my experience) impossible to induce meaningful curvature changes. The authors should better explain the experimental conditions if their experimental design allows accurate curvature measurement with a whisker trimmed to 5 mm and capillary 2 mm on whisker (Figure 4).

In response, we include (Figure 4—figure supplement 2) video stills and raw whisker tracking data below, which we hope convinces the reviewer that measurement of curvature changes under passive stimulation is possible. The curvature changes are, of course, small and there are two main reasons why we had sufficient sensitivity to reliably measure them: (1) we study mouse not rat; mouse whiskers being thinner and less stiff, a given stimulus produces more whisker bending than it would in rat; (2) our whisker tracker achieves accurate solutions by fitting quadratic curves to the base region of the whisker (thus keeping the number of free parameters small, whilst being justified by thin beam physics; Birdwell et al. 2007; Quist and Hartman 2012; Bale et al. 2015) and takes advantage of temporal contiguity to constrain the fit by the solution in the previous frame. Since short whiskers are stiff and therefore have high moment of inertia, even small curvature changes can correspond to substantial reaction forces.

Author response image 1 (top four panels) shows four video frames taken during trapezoidal, passive whisker stimulation. Whisker tracker solutions are overlaid. Curvature change (lower left) and corresponding tracker solutions (lower right) are shown for a 45 ms episode, with coloured dots marking the times of the four example frames, and shading from blue to aqua indicating time. This whisker has negative intrinsic curvature. As the actuator applies force to the whisker, the whisker straightens up and the curvature increases.

https://doi.org/10.7554/eLife.10696.018

Article and author information

Author details

  1. Dario Campagner

    Faculty of Life Sciences, The University of Manchester, Manchester, United Kingdom
    Contribution
    DC, Designed the study, Performed the experiments, Analyzed the data, Developed the experimental methods, Wrote the manuscript.
    Competing interests
    The authors declare that no competing interests exist.
    ORCID icon "This ORCID iD identifies the author of this article:" 0000-0001-9016-4575
  2. Mathew Hywel Evans

    Faculty of Life Sciences, The University of Manchester, Manchester, United Kingdom
    Contribution
    MHE, Analyzed the data, Developed the experimental methods
    Competing interests
    The authors declare that no competing interests exist.
  3. Michael Ross Bale

    1. Faculty of Life Sciences, The University of Manchester, Manchester, United Kingdom
    2. School of Life Sciences, University of Sussex, Brighton, United Kingdom
    Contribution
    MRB, Developed the experimental methods.
    Competing interests
    The authors declare that no competing interests exist.
  4. Andrew Erskine

    1. Faculty of Life Sciences, The University of Manchester, Manchester, United Kingdom
    2. Mill Hill Laboratory, The Francis Crick Institute, London, United Kingdom
    Contribution
    AE, Performed the experiments, Developed the experimental methods.
    Competing interests
    The authors declare that no competing interests exist.
    ORCID icon "This ORCID iD identifies the author of this article:" 0000-0003-4392-1873
  5. Rasmus Strange Petersen

    Faculty of Life Sciences, The University of Manchester, Manchester, United Kingdom
    Contribution
    RSP, Designed the study, Analyzed the data, Developed the experimental methods, Wrote the manuscript.
    For correspondence
    R.Petersen@manchester.ac.uk
    Competing interests
    The authors declare that no competing interests exist.

Funding

Biotechnology and Biological Sciences Research Council (BB/L007282/1)

  • Rasmus Strange Petersen

Wellcome Trust (097820/Z/11/B)

  • Rasmus Strange Petersen

Medical Research Council (MR/L01064X7/1)

  • Rasmus Strange Petersen

The funders had no role in study design, data collection and interpretation, or the decision to submit the work for publication.

Acknowledgements

We thank S Fox, M Humphries, M Loft, R Lucas, M Montemurro and M Maravall for comments on the manuscript/discussion; K Svoboda for sharing behavioural methods; G Caspani, K Chlebikova, B Nathanson and R Twaites for assistance with whisker tracking.

Ethics

Animal experimentation: All experimental protocols were approved by both United Kingdom Home Office national authorities and institutional ethical review. Project licence: 40/3420.

Reviewing Editor

  1. David Kleinfeld, University of California, San Diego, United States

Publication history

  1. Received: August 10, 2015
  2. Accepted: January 6, 2016
  3. Version of Record published: February 15, 2016 (version 1)

Copyright

© 2016, Campagner et al.

This article is distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use and redistribution provided that the original author and source are credited.

Metrics

  • 2,186
    Page views
  • 391
    Downloads
  • 30
    Citations

Article citation count generated by polling the highest count across the following sources: Scopus, Crossref, PubMed Central.

Download links

A two-part list of links to download the article, or parts of the article, in various formats.

Downloads (link to download the article as PDF)

Download citations (links to download the citations from this article in formats compatible with various reference manager tools)

Open citations (links to open the citations from this article in various online reference manager services)

Further reading

    1. Biochemistry and Chemical Biology
    2. Computational and Systems Biology
    Michael Luke Carlson et al.
    Tools and Resources Updated
    1. Computational and Systems Biology
    2. Physics of Living Systems
    Andreas Hofmann et al.
    Research Article