Experimental Subjects and Paradigm

Two rhesus monkeys (Macaca mulatta) were subjects in this study. All animal welfare and experimental procedures were conducted according to the U.S. Public Health Service Policy on Humane Care and Use of Laboratory Animals (Public Law 99–158) and the Guide for the Care and Use of Laboratory Animals (National Academy Press, 1996) and were approved by the Arizona State University Institutional Animal Care and Use Committee. Great care was taken to minimize any pain or discomfort during any medical procedures, using proper anesthesia and analgesia under veterinary care when necessary. All housing, feeding, and environmental social enrichment conformed to institutional standards which are AAALAC International accredited.

The experimental apparatus and paradigm have been described in detail elsewhere [8]. Briefly, two Rhesus monkeys (X and B) were trained to reach to targets in a semi-immersive 3D virtual reality environment and to maintain their arm position at these targets either with or without vision of the endpoint of the arm, which was provided as a virtual sphere (‘arm cursor’). The virtual environment was displayed on a 3D monitor (Dimension Technologies Inc.) and was projected onto a mirror embedded in a metal shield that blocked the real arm from view (Fig 1). Arm position was tracked using an active LED based system (Phoenix Technologies Inc.) and eye position was monitored using a remote optical tracking system (Applied Science Laboratories Inc.).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. Experimental apparatus and paradigm (Shi et al. 2013).

A. Schematic of virtual reality setup. B. Sequence of events on a single trial. Grey rectangle indicates the 800 ms analysis epoch that was the focus of this study. During this time period, animals were required to fixate at the center of the display (S) while maintaining their arm position at the peripherally located targets (T).

https://doi.org/10.1371/journal.pone.0166786.g001

At the beginning of each trial, an animal first acquired a spherical starting position that was presented at the center of the vertically oriented workspace. After maintaining this position for 500 ms, a spherical green target was presented at one of eight locations arranged in a square (monkey B) or rectangle (monkey X) and up to 6.4–7.1 cm from the center of the workspace. Onset of the target instructed the animals where to reach and also served as the ‘go’ signal. Once the target was acquired it was extinguished after 300–400 msec and a yellow sphere was presented at the center of the workspace, cueing the fixation position. Once the fixation position was acquired a holding period commenced during which time animals continued to fixate the center of the display while maintaining their arm position at the target for 800–1200 msec. During the holding period, animals viewed their arm on half of the trials (bimodal condition) while on the remaining trials vision was prevented by blanking the arm cursor (unimodal condition). In summary, on a given trial animals executed a center-out, reaction-time reach and saccade to one of eight peripheral targets, made a saccade back to a central fixation point, then maintained their arm position at the target for a variable period in the presence of proprioceptive feedback or both visual and proprioceptive feedback. Five (5) trials were performed for each target in each sensory condition, which were pseudorandomly selected.

Neurophysiological Procedures and Analysis

All experimental procedures were conducted according to the "Principles of laboratory animal care" (NIH publication no. 86–23, revised 1985) and were approved by the Arizona State University Institutional Animal Care and Use Committee. Extracellular recordings (N = 343; 219 from X and 124 from B) were made using varnish-coated tungsten microelectrodes in the superficial cortex of the SPL (area 5), as judged by recording depth and similarity to previous recordings made in this area [18, 33]. Spikes were isolated from the amplified and filtered (600–6000 Hz) signal via a time-amplitude window discriminator (Plexon Inc.). Spike times were sampled at 2.5 kHz.

Analysis of Neurophysiological Data

Data analyses focused on an 800 ms long section of the holding period beginning 400 ms after target acquisition. Firing rates during this period were compared statistically between conditions using the Mann-Whitney U test using a significance level of α = 0.05. Peristimulus time histograms (PSTHs) of the trial averaged rate were constructed using data from each neuron’s preferred and non-preferred locations during the holding period (based on mean firing rates) and were smoothed with a Gaussian kernel (σ = 50 ms).

We previously quantified the effects of multisensory integration on spiking variability using the Fano factor and coefficient of variation of the interspike intervals [8, 34]. Here, we used the power spectra and autocorrelograms of the recorded spike trains to provide further insights. Power spectra were computed from the spike times on each trial using multitaper spectral methods ([35, 36]; http://chronux.org/). Spectra were computed with a 6.25Hz resolution and nine Slepian data tapers. Single trial spectra were normalized by the mean firing rate during the 800ms window before averaging across trials and cells. The variance of the normalized single cell and population spectra were estimated using a Jackknife procedure that involved leaving out a data taper and creating a distribution of spectral estimates from which the variance was determined. We also computed the variance distribution of individual cells using the autocorrelation function, the time domain counterpart to the spectrum, to look for regularity (oscillations) in the spike trains. Autocorrelograms were computed for each trial by subtracting the shuffle predictor (i.e. the average cross-correlogram of the spike trains from one cell during different trials) from a histogram of the intervals between spikes. The resulting autocorrelogram for all five trials was then normalized by the standard deviation (SD) of the shuffle predictor at each time lag [19, 37] putting them in units of shuffle predictor SDs.

Power spectra and autocorrelograms were used to classify the neural spike trains recorded during the 800 ms holding period using previously published criteria [19, 38]. To classify the spectra from individual cells, we first defined an index A as the average height of the central autocorrelogram bins corresponding to time lags < 5 ms [19]. Spike trains were classified as “bursty” if they exhibited a peak in their power spectrum that exceeded the expected Poisson spectrum by 1 SD in the 5–60 Hz frequency band and had a value of A >1 (1 SD higher than the shuffle predictor). We classified spike trains as “refractory” if they exhibited a spectral trough that deviated from the expected Poisson power by 1 SD and had a value of A < -1. Spike trains were classified as “Poisson” if their power spectrum showed no deviations from the expected Poisson power and -1 < A < 1. Spike trains were classified as “oscillatory” if their power spectra exceeded the expected Poisson power by 1 SD in the 13-30Hz range and exhibited a value of A <1. This sorting paradigm, which is based on previous work [19, 38], utilizes complementary but mathematically equivalent time and frequency domain functions to examine temporal structure in spike trains. Time domain correlation functions are known to suffer from estimation bias and variance, and are particularly susceptible to violations of the non-stationarity assumption [39]. Frequency domain analyses, particularly multitaper spectral methods, greatly reduces the problems associated with estimation bias and variance, and is able to detect subtle changes in temporal structure better than their time-domain counterpart [36]. Owing to these advantages and the ability to construct accurate spectral confidence intervals to assess significant structural features, we also sorted the cells using only the spectral criteria.

Regarding the oscillatory classification, we chose to focus on the 13–30 Hz range for two reasons. First, previous work from the PPC showed strong oscillations in this range during performance of various tasks [18, 19, 40]. In addition, the 13–30 Hz band corresponds to the ‘beta’ band of EEG, LFP, and MEG signals, which has been implicated in the maintenance of sensorimotor state [41], including limb position [28], one of the independent variables in the current study. Although the term “beta oscillations” is typically used to refer to oscillations observed in EEG and other continuous time signals, for convenience we will use this term to refer to rhythmic spiking in the same frequency band. Oscillatory population spectra were computed for both for the preferred and non-preferred locations in each condition. Power in the beta-band was compared between conditions using the Mann-Whitney U test and a significance level of α = 0.05.

Behavior

In the experimental sessions described here, animals performed a reach and hold task while receiving either unimodal (proprioceptive) or bimodal (visual-proprioceptive) sensory feedback [8]. Despite consistent difference in firing rates between the two conditions (described below; see also Fig 2), average arm endpoint positions and variability in endpoint positions did not differ between unimodal and bimodal conditions. We have also previously reported no differences in average eye position or variability in eye position during this task [8]. No difference in arm endpoint position was an expected consequence of our experimental design. In this initial exploration of multisensory interactions we wanted to ensure that any differences in firing rates that were observed did not reflect the fact that animals held their limbs at slightly different positions in the two tasks and were instead due to the different visual conditions. As a result, animals were trained to maintain their arm position within a very tight behavioral window in both tasks. However, this behavioral constraint also resulted in endpoint variability being roughly equivalent between the two conditions. It’s currently unclear how to interpret difference in firing rate during multisensory interactions without a concomitant change in behavior. It should be noted however that other studies exploring the neural correlates of multisensory integration have also not observed a concomitant change in perceptual or action-related variables mainly because such variables were not monitored. For example, work by Graziano and colleagues have previously shown differences in firing rates with and without arm vision in a task where the animal was not required to move or make a perceptual judgement [11]. Since it is impossible to monitor all possible behavioral variables in a single experiment, the most logical explanation for the observation of a neural correlate without a corresponding behavioral correlate is a perceptual or behavioral change that is not observed simply because it is not monitored. Given the known sensitivity of SPL neurons to differences in arm configuration [30, 31], the most likely explanation for our results is that animals used either a slightly different arm configuration while maintaining their position with and without visual feedback, or they used roughly the same configuration but exhibited differing degrees of variability in configuration between conditions. Additional experiments will be necessary to distinguish between these and other potential explanations for the changes in neural activity reported here and elsewhere.

Effects of Multisensory Interactions on Firing Rates

As described previously by Shi et al (2013), firing rates were suppressed under bimodal conditions, consistent with a subadditive interaction of visual and proprioceptive inputs. Early work by Stein and colleagues in the SC described additive, subadditive and superadditive responses to multisensory interactions depending on factors such as spatial and temporal congruency. More recent experimental work however suggests that superadditive responses are less common than initially believed, being observed generally for weak responses to unimodal stimuli [42, 43]. This finding is more in line with computational modeling studies employing a probabilistic population coding framework which predicts that cortical areas should exhibit largely additive or subadditive responses [13, 14]. Interestingly, recent studies of visual-vestibular interactions under conditions designed to test for optimal integration of these inputs showed a predominance of subadditivity in the dorsal medial superior temporal area [15]. Similarly, studies of other cortical regions employing different experimental paradigms [12, 44–46] have reported substantial degrees of multisensory suppression (and therefore subadditivity). Although the significance of this relatively ubiquitous suppression is still not entirely clear, it is noteworthy that some studies have provided evidence that suppression of spiking and reduced spiking variability resulting from multisensory integration is associated with greater encoded stimulus information and greater population decoding accuracy [8, 25].

Spike Train Dynamics

A relatively large number of neurons in this study could not be classified as Poisson, bursty, refractory, or oscillatory and were therefore classified as ‘other’. This is due in part to the fact that the employed spike train classification scheme relies on meeting two separate criteria, one based on the power spectrum and the other on the autocorrelation function. Many cells exhibited power spectra that were consistent with bursty, refractory or oscillatory spike trains but did not meet the corresponding autocorrelation criteria or vice-versa. Of the cells with internally consistent spectra and autocorrelations, the vast majority were classified as oscillatory in the beta-band in one or both conditions. This preponderance of oscillatory spiking in the beta-band is consistent with other studies of the PPC using different tasks and behavioral epochs [18, 19]. It should be noted that virtually all of the oscillatory spike trains were accompanied by a spectral trough at low frequencies. Although somewhat nonintuitive, this suppression is consistent with a refractory period, i.e. a decreased probability of spiking shortly after the occurrence of a spike [16]. A substantial number of cells in our population also showed evidence of refractoriness without accompanying oscillations, also consistent with previous investigations [16, 18, 19, 38].

Although some neurons demonstrated power spectra that appeared Poisson-like, none were classified as Poisson when the corresponding autocorrelograms were considered. Although Poisson-like spiking has been reported in several subdivisions of the PPC, the percentage of cells classified as such appears to be task-, epoch-, and area-dependent [17, 19]. The relative dearth of Poisson-like spike trains reported here has important implications for computational models of multisensory integration. That is, recent studies have demonstrated that Bayes-optimal multisensory integration can be implemented as a simple linear summation of cortical activity assuming that neural variability in the cortex is approximately Poisson-like [13]. The present findings however reinforce previous studies that this assumption does not hold for all cortical regions. Although the model of Ma et al. is at least somewhat insensitive to violations of Poisson-like variability, it is not clear how well the model can accommodate neural variability which is highly regular, as observed here. As a result, it may be necessary for future modeling efforts to account for both Poisson-like and non-Poisson neural variability.

Temporal Coding of Multisensory Information for State Estimation

Although the basic premise of temporal coding, i.e. that information is encoded not only in spike rates but also in the precise timing of action potentials and/or ongoing neuronal (LFP) oscillations, has been postulated to play a role in phenomena such as feature binding, its relevance to multisensory interactions has been addressed only relatively recently [22]. Some of the strongest evidence in support of this idea comes from analyses of neural data obtained from non-human primates engaged in tasks involving somato-auditory and audio-visual interactions. For example, interaction of somatosensory and auditory inputs has been shown to result in changes in the relative phase of ongoing oscillations within primary auditory cortex [26]. Similarly, audio-visual interactions are associated with enhanced LFP and spike-field gamma coherence between auditory cortex and the superior temporal sulcus [20, 27]. Regarding changes in spiking, multisensory interactions have been shown to alter spike timing in the vestibulocerebellum of the mouse [23], in the dorsal cochlea nucleus of the guinea pig [24], and in the cortex of non-human primates engaged in tasks involving audio-visual and visual-proprioceptive interactions [8, 25]. The current findings extend those of Shi et al (2013) by showing that changes in spike timing can manifest differently in different subpopulations within a given cortical area, with potentially important implications for transcortical representations of perceptual and motor-related phenomena.

The present findings are consistent with the idea that the SPL is involved in estimating the state of the arm, as suggested by clinical and functional imaging studies in humans [29, 47–49] as well as neurophysiological studies in monkeys [8, 11, 30, 31, 33, 50, 51]. Previous work in monkeys, has also shown that static maintenance of arm position involves a beta-synchronized cortical network that includes the somatosensory cortex, inferior parietal lobule, and primary motor cortex [28]. The strong beta oscillations observed in the present study suggest that the SPL is also a part of this network and support the idea that beta oscillations serve to functionally link ensembles of neurons in cortical and subcortical regions that are believed to be involved in the maintenance of current motor state [21, 22, 41, 52].

Contrary to our stated hypothesis we did not find evidence that beta oscillations were consistently enhanced under bimodal conditions. Instead roughly equal numbers of neurons fired rhythmically in the beta band under only one set of sensory conditions (unimodal or bimodal) or both conditions and in the latter case, beta band power was the same magnitude in the two conditions. The fact that the beta oscillations of some cells were modulated by the presence/absence of visual input at all suggests that these oscillations are associated not only with sustained muscle activation and somatosensory feedback, as has previously been suggested [53], but also with ongoing visual feedback used for state estimation. However, the lack of a consistent bias in the population raises the question as to the functional significance of these findings. One possible interpretation is that different subsets of neurons exhibiting beta oscillations during proprioceptive, visuo-proprioceptive, and both conditions, reflects the differential recruitment of partially overlapping neural ensembles devoted to these respective modes of sensorimotor control during the maintenance of arm position [54]. The idea that control mechanisms for the arm may differ in the presence and absence of vision is supported by numerous psychophysical studies in humans [55–60]. In further support of this idea, functional imaging studies in humans indicate that the recruited nodes of a cortical network for limb position vary depending on available sensory information. For example, in the absence of vision, interaction of tactile and proprioceptive inputs has been shown to involve the right PPC, whereas in the presence of vision, interaction recruits a network of areas including the parietal and premotor cortices [29]. Validating this hypothesis more directly would require recording from single neurons in multiple nodes of the state estimation network simultaneously and looking for evidence of functional coupling between subsets of neurons only under a prescribed set of sensory conditions. Such manipulations are intended to be the focus of future planned series of studies.