20 July 2000

Nature 406, 302 - 306 (2000) © Macmillan Publishers Ltd.

Transformation from temporal to rate coding in a somatosensory thalamocortical pathway

EHUD AHISSAR, RONEN SOSNIK & SEBASTIAN HAIDARLIU

The anatomical connections from the whiskers to the rodent somatosensory (barrel) cortex form two parallel (lemniscal and paralemniscal) pathways^{1, 2}. It is unclear whether the paralemniscal pathway is directly involved in tactile processing, because paralemniscal neuronal responses show poor spatial resolution, labile latencies and strong dependence on cortical feedback^3-5. Here we show that the paralemniscal system can transform temporally encoded vibrissal information into a rate code. We recorded the representations of the frequency of whisker movement along the two pathways in anaesthetized rats. In response to varying stimulus frequencies, the lemniscal neurons exhibited amplitude modulations and constant latencies. In contrast, paralemniscal neurons in both thalamus and cortex coded the input frequency as changes in latency. Because the onset latencies increased and the offset latencies remained constant, the latency increments were translated into a rate code: increasing onset latencies led to lower spike counts. A thalamocortical loop that includes cortical oscillations and thalamic gating can account for these results. Thus, variable latencies and effective cortical feedback in the paralemniscal system can serve the processing of temporal sensory cues, such as those that encode object location during whisking. In contrast, fixed time locking in the lemniscal system is crucial for reliable spatial processing.

The lemniscal pathway of the rat trigeminal system ascends through the ventral posterior medial nucleus (VPM) of the thalamus to the barrels in layer 4 of the cortex and to layers 5b and 6 (refs 6, 7). The paralemniscal pathway ascends through the medial division of the posterior nucleus (POm) of the thalamus to layers 1 and 5a and to the septa between the barrels in layer 4 (refs 7 , 8). To reveal the processing differences between the two pathways, we recorded from single-units (n = 710) and multi-units (n = 540) of the major stations along these pathways while we stimulated (moved) the whiskers. First we analysed single- and multi-units separately. Analysis of the single-units revealed that response patterns were usually similar for neighbouring neurons. Therefore, spikes of all single- and multi-units that were recorded simultaneously from a single electrode were pooled, and these pools were referred to as 'local populations'.

First, we examined responses to stimuli that mimic natural whisking conditions⁹: 50-ms pulses of air puffs applied to one or two rows of whiskers at 8 Hz. Typical recordings along both pathways are shown in Fig. 1. Brainstem neurons appeared simply to relay primary afferent responses¹⁰: the response to a single pulse was constant along the train, as evident from the raster display and from the overlap between the peristimulus time histograms (PSTHs) computed for the first cycle (thick lines) and for the last three cycles (thin lines). However, as expected^{4, 11}, thalamic neurons did not behave as relays, and their outputs exhibited specific transformations of the brainstem signals ( Fig. 1, middle rows). The transformation at the VPM resembled an adaptation process during which, following the first stimulation cycle, the spike count (per stimulus cycle) decreased as the train continued, until a 'steady state' (a period during which the spike count was steady) was reached. Recordings from the POm revealed a different stabilization process. The latency of the response onset increased during the 8-Hz train until it stabilized at a significantly longer steady-state value (Fig. 1, POm raster display and PSTHs). These latency shifts resulted in decreased spike counts in the POm neurons because offset latencies did not change. In the barrel cortex there was a similar dissociation for the response patterns of the lemniscal and paralemniscal projection targets (Fig. 1, top rows). Whereas the responses of layer 4 barrel neurons were similar to those of VPM neurons, the responses of layer 5a neurons resembled those of POm neurons. These response patterns were so characteristic for single- and multi-units at each station that the recording location could often be predicted based on the physiological responses, even before histological processing.

Next, we tested the dependency of the thalamic transformations on the frequency of whisker movement using three frequencies (2, 5 and 8 Hz), revealing another dissociation between the two pathways. Whereas steady-state latencies of the lemniscal pathway were almost constant for all the tested frequencies, those of the paralemniscal pathway increased with increasing frequencies. The pathway-specific response patterns are demonstrated by the PSTHs that were averaged, for each station, across all neurons recorded from well localized sites (Fig. 2). At a stimulation frequency of 2 Hz (below the natural whisking range), latencies at all stations barely changed between the first stimulus cycles (left insets) and the steady-state periods (centre PSTHs). With stimulation frequencies of 5 and 8 Hz (both of which are within the whisking frequency range¹²), brainstem and lemniscal latencies were also essentially constant (Fig. 2, centre PSTHs and right insets), in agreement with other findings¹⁰. However, steady-state latencies of the paralemniscal pathway increased with increasing stimulus frequency (Fig. 2, centre PSTHs and right insets; see also ref. 3).

For both pathways, the magnitude of the steady-state responses (measured by the PSTH area, that is, spike counts per cycle) decreased with increasing frequencies. Again, in the lemniscal pathway, spike-count decrements were caused by amplitude reduction, whereas in the paralemniscal pathway they resulted primarily from latency shifts. As a result, during steady-state periods of whisker stimulation, whisker frequencies were represented by spike counts (rate coding) in both pathways, and by latencies (temporal coding) in the paralemniscal pathway. These representations were reproduced during each stimulus train, as the responses to the first cycle in each train were identical for all frequencies (Fig. 2, left insets). The averaged latency and spike-count representations of the whisker frequency ( Fig. 2, right insets) depict typical representations of single local populations in each of the stations. These pathway-specific representations occurred for all well localized recording sites, except for 1 out of 14 in POm and 1 out of 13 in layer 5a, in both of which the latency increased between 2 and 5 Hz but not between 5 and 8 Hz. Part of the amplitude adaptation could be attributed to anaesthesia¹⁰. However, anaesthesia could not account for the latency shifts in the paralemniscal system, because latency shifts due to anaesthesia are an order of magnitude smaller^13-15. Furthermore, as these latency shifts developed during each stimulus train, they probably reflect a dynamic process.

Paralemniscal processing could be conducted in parallel to, or in series with^{2, 4, 5}, lemniscal processing. If processing is in series, the onset of POm activity should lag the onset of lemniscal cortical activity. However, as seen in Fig. 3, which shows the average responses to the first cycle of mechanical stimulation applied to the principal whiskers, the average POm response rises before those of layers 4 and 5b. Furthermore, the shortest and modal latencies to the first stimulus cycle in the POm (7 and 8 ms, respectively) were shorter than those in layer 4 (8 and 9 ms) and layer 5b (8 and 9 ms). Thus, the onset of POm activity must be driven directly by the onset of brainstem activity, which is transferred along the paralemniscal pathway in parallel with the activity conveyed along the lemniscal pathway. These results are consistent with the 1–2-ms delay between VPM and POm activation by brainstem electrical stimulations¹⁶. Of course, POm activity following the onset of activation could be affected by mixed lemniscal and paralemniscal cortical feedback. However, the similarity between the transient and steady-state response patterns within each pathway (Figs 1, 2) indicates that the sustained activation may also be pathway-specific, and that it is dominated, at least to a first approximation, by closed thalamocortical loops in each pathway.

A striking feature of the paralemniscal response pattern is that while the onset latency increased, the offset latency remained constant, which caused a reduction in the duration of the response and hence reduced spike counts. The latency increments cannot simply be attributed to VPM–POm interactions. These interactions are inhibitory¹⁷, which predicts a positive correlation between the response strength of the VPM neurons and the latencies of the POm neurons, whereas we observed the opposite (Fig. 2). However, an alternative explanation, based on closed-loop operation of the paralemniscal thalamocortical system, is consistent with our results. This explanation suggests that thalamocortical circuits implement phase-locked loops, a computational algorithm that is optimal for temporal decoding^{18, 19}.

An implementation of a POm–cortex phase-locked loop that is consistent with the findings presented here is based on a simple closed-loop thalamocortical circuit²⁰ that includes cortical inhibition^{20, 21}, cortical oscillatory mechanisms^22-24 and thalamic gating²⁵. Briefly, this circuit establishes a negative feedback loop in which the latency of the cortical oscillations determines the latency and spike count of the thalamic neurons, and the thalamic spike count determines the subsequent latency of the cortical oscillations. Within such a circuit, thalamic and cortical latencies do not depend directly on the brainstem latency, but rather on the latency of the cortical oscillations. Consistent with the model and results presented here, cortical oscillating neurons display increasing latencies with increasing stimulus frequencies²⁴ (see Supplementary Information for a more detailed description of the model).

When properly tuned, such a simple thalamocortical circuit should be able to track changes in the input frequency with a delay of a single cycle, and to represent these changes in the latencies and spike counts of its thalamic and cortical neurons¹⁸ (see Supplementary Information ). To achieve this, thalamic neurons must function in their gating mode. However, following a quiescent period, thalamic neurons shift to a non-gating mode in which their outputs do not appear to depend on the cortical signal²⁵. This can account for the relatively short latency of POm neurons to the first stimulus cycle (Figs 1–3) and the relatively long period required for stabilization (Fig. 1). In the awake animal, thalamic neurons probably shift into a gating mode upon whisking¹³.

We tested the performance of the POm–cortex circuits in single-cycle temporal decoding by applying frequency-modulated (FM) stimuli as follows. To shift the thalamic neurons into a gating mode and to allow 'locking in' of the thalamocortical circuits¹⁸, the whisker stimulation frequency was first kept constant at 5 Hz for 2 s. Then the stimulus frequency was modulated for an additional 6 s (40% modulation depth, 0.4 Hz, initial phase 90°). The entire process was repeated 36 times for each simultaneous recording, and the responses of each local population were averaged across repetitions.

In the brainstem, the response to each stimulus cycle was the same, regardless of the instantaneous frequency (the reciprocal of the current period) of the stimulus (Fig. 4a). In contrast, once 'locked in' to the stimulus, paralemniscal thalamic and cortical neurons directly represented the input frequency, by both latency and spike-count, with almost no delay (see pooled layer 5a representation in Fig. 4c). The spike-count and latency representations were usually linearly related, reflecting the transformation of latency to spike count observed in this system. Linear regression estimates of spike counts from latencies could explain most of the spike count variability for layer 5a (r² = 0.66  0.2, n = 6, P  < 0.0001) and POm (r ² = 0.51  0.12, n  = 9, P < 0.0001) local populations. As expected from the steady-state data, whisker frequency was represented by spike count alone in the lemniscal pathway (see pooled barrel representation in Fig. 4b). The lemniscal spike-count representation was not sensitive to the fast frequency transition (at t = 2 s) and exhibited a smaller dynamic range than the paralemniscal one (modulation depths of the spike-count representations were 12.5  15% (n = 13) for VPM and layer 4 barrel neurons, and 25.5  14.5% (n = 15) for POm and layer 5a neurons (P < 0.05, two-tailed t-test)). Thus, the temporal information of the stimulus was not reliably re-coded by lemniscal firing rates, even though the temporal information was faithfully replicated by lemniscal firing times.

During exploration and active touch, rats move their whiskers back and forth at frequencies between 4 and 11 Hz^{9, 12, 13, 19}. These whisking movements yield spatial and temporal encoding of the external spatial information. For example, object location (relative to the face and initial position of whiskers) is encoded by the identity of the activated whiskers (spatial encoding) and by the temporal interval between protraction onset and perturbation by the object (temporal encoding)²⁶. Our results indicate that these two components may be processed separately and in parallel, in the lemniscal and paralemniscal pathways, respectively. The loss of time locking to stimulus onset in the paralemniscal system is a reflection of the temporal-to-rate code translation process, in which latency shifts produce a rate-coded representation of the whisker frequency. The fixed time locking preserved along the lemniscal system probably reflects a lack of temporal processing within the whisking frequency range. On the other hand, the fixed time locking is required for reliable spatial processing during whisking, because, owing to whisker movements, temporal uncertainties in afferent firings result in uncertainties about whisker location. Thus, counterintuitively, fixed time locking is crucial for spatial processing, whereas variable latencies serve temporal processing. The response durations of the paralemniscal neurons are well tuned for temporal processing in the frequency range of whisking, as these response durations match the duration of a single protraction phase ( 50 ms) and thus allow the decoding of all possible temporal intervals during protraction. Note that decoding of temporal information at higher frequencies, such as those that encode the texture of objects scanned by the whiskers⁹, probably cannot be achieved by the slow paralemniscal system, but can be achieved by lemniscal thalamocortical circuits in which high frequencies can be coded (and translated to rate) by fine modulations of the response latency.

Our data indicate that the paralemniscal pathway of the rat somatosensory system is optimally tuned for temporal processing of vibrissal information around the whisking frequency range ( 8 Hz), and that temporally encoded information can be translated to a rate code by thalamocortical loops. These findings point towards a parallel processing scheme in which spatially encoded (and perhaps high-frequency, temporally encoded) information is decoded by the lemniscal system, whereas low-frequency, temporally encoded information is decoded by the paralemniscal system. Sensory pathways dedicated to temporal processing have been reported in the electric fish and in the auditory brainstem of birds and mammals²⁷. Here we provide evidence for a separate sensory pathway for temporal processing that involves the thalamus and cortex.

Methods
Animal procedures and electrophysiology Experimental procedures were similar to those used previously^{24, 28}. Briefly, 43 adult male Wistar albino rats (300  25 g) obtained from the Animal Breeding Unit of The Weizmann Institute of Science were anaesthetized (urethane, 1.5 g kg^-1, injected intraperitoneally) and mounted in a modified stereotaxic device that allows free access to the somatosensory cortex and vibrissae. We monitored the depth of anaesthesia by assessing corneal and limb-withdrawal reflexes, and maintained it at stage III/3–4 (as determined by the dominant frequencies in the power spectra of local field potentials in the cortex¹⁴) by injections of urethane (10% of initial dose). Atropine methyl nitrate (0.3 mg kg ^-1, intramuscular) was administered before general anaesthesia to prevent respiratory complications. Body temperature was maintained at  37 °C.

The skull was exposed, and openings were made to allow electrode penetrations to different neuronal stations in the barrel cortex, thalamus and/or brainstem. Four to eight microelectrodes were driven into one or two stations simultaneously using one or two microdrive systems²⁸. We isolated single units by spike templates and multiple units by amplitudes using spike sorters (MSD-2; Alpha-Omega). We recorded 96 single-units and 62 multi-units in the brainstem, 333 and 234, respectively, in the thalamus, and 281 and 244, respectively, in the cortex. The number of neurons recorded by a single electrode (a local population) was estimated to be 5  1.3 (mean  s.d.) neurons. The care and treatment of rats were in accordance with the animal welfare guidelines of The Weizmann Institute of Science.

Whisker stimulation Whisker stimulations consisted of pulses of compressed air, generated by a pneumatic pressure pump (Medical Systems) and delivered through 3 m of stiff, thick-walled tubing (6.5 mm outer diameter, 3.5 mm inner diameter). A micropipettor tip was attached to the end of the tubing to reduce its opening to a diameter of 0.7 mm and was positioned 10–20 mm caudally to the most caudal whiskers of the stimulated rows. We usually stimulated one or two rows of whiskers (including at least the four most caudal whiskers and associated straddlers). Constant-frequency stimuli were applied in blocks of 12( 2) or 24 trains of 3 or 4 s each, with inter-train intervals of 2 or 1 s, respectively. FM stimuli were applied in blocks of 36 trains of 8 s each, with inter-train intervals of 2 s. The stimuli were delivered as constant pulses (see bottom trace in Fig. 1): during the first 50 ms of each cycle the whiskers were protracted (rise time, 30 ms; maximum pressure level at pump output, 0.7 kg cm ^-2) and then allowed to retract (fall time, 10 ms). Pulse profiles were equal for all the frequencies used. Air-puff delays (from pump command to whisker movement) were calibrated using a microphone. In all experiments, latencies evoked by air puffs were compared with those evoked by a mechanical stimulator²⁸, which was attached to a single whisker 5–10 mm from the skin (amplitude, 560 µm; rise and fall times, 5 ms). When mismatches (all were <3 ms) were detected, the mechanical latencies were assumed.

Histology We analysed the brains of the rats histologically to localize all recording sites. At the end of each recording session, we induced electrolytic lesions by passing currents (3–5 µA, 2  2 s, unipolar) through the tips of the electrodes. The brains were sectioned (coronally for cortical and thalamic recordings and parasagittally for brainstem recordings) and stained for cytochrome oxidase activity²⁸. In these preparations, the different cortical layers, barrel borders, thalamic and brainstem nuclei, lesions and electrode tracks were clearly visible (Fig. 1).

Data analysis Only data from well localized recording sites (see Fig. 1) are described. We computed PSTHs using 1-ms bins and smoothed them by convolution with a triangle of area 1 and a base of  8 ms. Neuronal latencies were estimated from the smoothed PSTHs as the time in which the firing rate crossed the half-peak value (after subtracting background activity). Other latency estimations, such as latency to peak or to 0.1 peak value, yielded basically the same results. Spike counts were the average sum (per stimulus cycle) of spikes during the first 100 ms after stimulus onset.

Supplementary information is available on Nature's World-Wide Web site (http://www.nature.com) or as paper copy from the London editorial office of Nature.

References

Acknowledgements. We thank M. Ahissar, and S. Barash for suggestions; A. Arieli, F. F. Ebner, K. O. Johnson, R. Malach and D. J. Simons for comments and discussions; M. E. Diamond for assistance during our initial thalamic recordings; and B. Schick for reviewing the manuscript. This work was supported by the US–Israel Binational Science Foundation and by the MINERVA Foundation, Germany. S.H. was supported by The Centre for Absorption of Scientists, Ministry of Absorption, Israel.

Nature © Macmillan Publishers Ltd 2000 Registered No. 785998 England.