Limits...
Prediction of primary somatosensory neuron activity during active tactile exploration.

Campagner D, Evans MH, Bale MR, Erskine A, Petersen RS - Elife (2016)

Bottom Line: Primary sensory neurons form the interface between world and brain.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.Thus, simple statistical models can predict rich neural activity elicited by natural, exploratory behaviour involving active movement of sense organs.

View Article: PubMed Central - PubMed

Affiliation: Faculty of Life Sciences, The University of Manchester, Manchester, United Kingdom.

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.

No MeSH data available.


Related in: MedlinePlus

Effect on GLM performance of quadratic input terms, simulated repeated trials and minimal stimulus filters. (A) Angle GLM prediction performance is robust to addition of quadratic stimulus-dependence. Prediction accuracy (PCC) for standard angle GLM (same data as Figure 2D of main text) in comparison to quadratic GLM (see Materials and methods). Black dots denote medians, error bars IQR. (B) Single-trial GLM prediction accuracy is limited by neuronal response variability. Prediction accuracy (PCC) for simulated neurons. Each simulated neuron is the best-fitting GLM, based on instantaneous curvature change, for its corresponding recorded unit (see Materials and methods). Prediction accuracy is quantified both using the single-trial approach of the main text and using a repeated-trial method only possible by virtue of using a simulation. Black dots denote medians, error bars IQR. (C) Prediction accuracy of curvature-based GLMs is accounted for by tuning to instantaneous curvature change. A GLM performs a temporal filtering operation on its sensory stimulus input and the sensory feature(s) which it encodes is determined by this ‘stimulus filter’. The stimulus filters can, in principle, be complex, but we found that the ability of a GLM to predict spikes (lower left) from curvature change was fully explained by the simple case where the action of the stimulus filter is simply to multiply the sensory input by a gain factor (median 0.55, IQR 0.26–0.66; p=0.35 signed-rank test). Recorded spike train (upper left) and curvature-predicted spike trains (lower left) both for a ‘curvature history’ GLM with a length 5 stimulus filter, as used in Figure 2D of main text, and for an ‘instantaneous curvature’ GLM with a length 1 stimulus filter. Data for unit 2 of main text Figure 2C. Prediction accuracy of the curvature history GLM compared to that of the instantaneous curvature GLM for every recorded unit (right). (D) Tuning curves for curvature change (blue) and angle (orange) of unit 1 and unit 2 in Figure 2.DOI:http://dx.doi.org/10.7554/eLife.10696.008
© Copyright Policy
Related In: Results  -  Collection

License
getmorefigures.php?uid=PMC4764568&req=5

fig2s1: Effect on GLM performance of quadratic input terms, simulated repeated trials and minimal stimulus filters. (A) Angle GLM prediction performance is robust to addition of quadratic stimulus-dependence. Prediction accuracy (PCC) for standard angle GLM (same data as Figure 2D of main text) in comparison to quadratic GLM (see Materials and methods). Black dots denote medians, error bars IQR. (B) Single-trial GLM prediction accuracy is limited by neuronal response variability. Prediction accuracy (PCC) for simulated neurons. Each simulated neuron is the best-fitting GLM, based on instantaneous curvature change, for its corresponding recorded unit (see Materials and methods). Prediction accuracy is quantified both using the single-trial approach of the main text and using a repeated-trial method only possible by virtue of using a simulation. Black dots denote medians, error bars IQR. (C) Prediction accuracy of curvature-based GLMs is accounted for by tuning to instantaneous curvature change. A GLM performs a temporal filtering operation on its sensory stimulus input and the sensory feature(s) which it encodes is determined by this ‘stimulus filter’. The stimulus filters can, in principle, be complex, but we found that the ability of a GLM to predict spikes (lower left) from curvature change was fully explained by the simple case where the action of the stimulus filter is simply to multiply the sensory input by a gain factor (median 0.55, IQR 0.26–0.66; p=0.35 signed-rank test). Recorded spike train (upper left) and curvature-predicted spike trains (lower left) both for a ‘curvature history’ GLM with a length 5 stimulus filter, as used in Figure 2D of main text, and for an ‘instantaneous curvature’ GLM with a length 1 stimulus filter. Data for unit 2 of main text Figure 2C. Prediction accuracy of the curvature history GLM compared to that of the instantaneous curvature GLM for every recorded unit (right). (D) Tuning curves for curvature change (blue) and angle (orange) of unit 1 and unit 2 in Figure 2.DOI:http://dx.doi.org/10.7554/eLife.10696.008

Mentions: 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).


Prediction of primary somatosensory neuron activity during active tactile exploration.

Campagner D, Evans MH, Bale MR, Erskine A, Petersen RS - Elife (2016)

Effect on GLM performance of quadratic input terms, simulated repeated trials and minimal stimulus filters. (A) Angle GLM prediction performance is robust to addition of quadratic stimulus-dependence. Prediction accuracy (PCC) for standard angle GLM (same data as Figure 2D of main text) in comparison to quadratic GLM (see Materials and methods). Black dots denote medians, error bars IQR. (B) Single-trial GLM prediction accuracy is limited by neuronal response variability. Prediction accuracy (PCC) for simulated neurons. Each simulated neuron is the best-fitting GLM, based on instantaneous curvature change, for its corresponding recorded unit (see Materials and methods). Prediction accuracy is quantified both using the single-trial approach of the main text and using a repeated-trial method only possible by virtue of using a simulation. Black dots denote medians, error bars IQR. (C) Prediction accuracy of curvature-based GLMs is accounted for by tuning to instantaneous curvature change. A GLM performs a temporal filtering operation on its sensory stimulus input and the sensory feature(s) which it encodes is determined by this ‘stimulus filter’. The stimulus filters can, in principle, be complex, but we found that the ability of a GLM to predict spikes (lower left) from curvature change was fully explained by the simple case where the action of the stimulus filter is simply to multiply the sensory input by a gain factor (median 0.55, IQR 0.26–0.66; p=0.35 signed-rank test). Recorded spike train (upper left) and curvature-predicted spike trains (lower left) both for a ‘curvature history’ GLM with a length 5 stimulus filter, as used in Figure 2D of main text, and for an ‘instantaneous curvature’ GLM with a length 1 stimulus filter. Data for unit 2 of main text Figure 2C. Prediction accuracy of the curvature history GLM compared to that of the instantaneous curvature GLM for every recorded unit (right). (D) Tuning curves for curvature change (blue) and angle (orange) of unit 1 and unit 2 in Figure 2.DOI:http://dx.doi.org/10.7554/eLife.10696.008
© Copyright Policy
Related In: Results  -  Collection

License
Show All Figures
getmorefigures.php?uid=PMC4764568&req=5

fig2s1: Effect on GLM performance of quadratic input terms, simulated repeated trials and minimal stimulus filters. (A) Angle GLM prediction performance is robust to addition of quadratic stimulus-dependence. Prediction accuracy (PCC) for standard angle GLM (same data as Figure 2D of main text) in comparison to quadratic GLM (see Materials and methods). Black dots denote medians, error bars IQR. (B) Single-trial GLM prediction accuracy is limited by neuronal response variability. Prediction accuracy (PCC) for simulated neurons. Each simulated neuron is the best-fitting GLM, based on instantaneous curvature change, for its corresponding recorded unit (see Materials and methods). Prediction accuracy is quantified both using the single-trial approach of the main text and using a repeated-trial method only possible by virtue of using a simulation. Black dots denote medians, error bars IQR. (C) Prediction accuracy of curvature-based GLMs is accounted for by tuning to instantaneous curvature change. A GLM performs a temporal filtering operation on its sensory stimulus input and the sensory feature(s) which it encodes is determined by this ‘stimulus filter’. The stimulus filters can, in principle, be complex, but we found that the ability of a GLM to predict spikes (lower left) from curvature change was fully explained by the simple case where the action of the stimulus filter is simply to multiply the sensory input by a gain factor (median 0.55, IQR 0.26–0.66; p=0.35 signed-rank test). Recorded spike train (upper left) and curvature-predicted spike trains (lower left) both for a ‘curvature history’ GLM with a length 5 stimulus filter, as used in Figure 2D of main text, and for an ‘instantaneous curvature’ GLM with a length 1 stimulus filter. Data for unit 2 of main text Figure 2C. Prediction accuracy of the curvature history GLM compared to that of the instantaneous curvature GLM for every recorded unit (right). (D) Tuning curves for curvature change (blue) and angle (orange) of unit 1 and unit 2 in Figure 2.DOI:http://dx.doi.org/10.7554/eLife.10696.008
Mentions: 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).

Bottom Line: Primary sensory neurons form the interface between world and brain.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.Thus, simple statistical models can predict rich neural activity elicited by natural, exploratory behaviour involving active movement of sense organs.

View Article: PubMed Central - PubMed

Affiliation: Faculty of Life Sciences, The University of Manchester, Manchester, United Kingdom.

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.

No MeSH data available.


Related in: MedlinePlus