Introduction

Information processing, even at the earliest sensory stages, can be modulated by several influences. One prominent influence is that of arousal-related behavioral states, which have been shown to change neural activity throughout the brain [1–4]. One classic indicator for arousal is pupil size, a metric that is relatively simple to measure and analyze and that has provided fundamental insights into neuromodulatory and cognitive influences on brain activity [2–8]. However, a multitude of factors converge to affect the pupil size signal, and how they are combined into this single indicator of arousal is not known. Likewise, how the various influences on the pupil signal relate to spontaneous and stimulus driven neural activity is not well understood.

An ideal system to investigate arousal-related modulations of neural activity and sensory processing is the visual thalamus. The thalamic dorsal lateral geniculate nucleus (dLGN) is the primary relay of visual signals from the retina to the visual cortex, and dLGN neurons, like those in other thalamic nuclei, have long been known to display prominent patterns of activity associated with arousal [9–12]. In particular, 2 state-related firing modes have been described: burst firing, which is more prevalent during low-arousal states [9] and behavioral inactivity [12,13], and tonic firing, which is observed during alertness. Burst firing in the thalamus is characterized by a high frequency discharge of action potentials after sustained hyperpolarization, and it relies on the activation of low-threshold, transient (T-type) calcium channels (reviewed in [14]). In contrast, tonic firing occurs when the membrane potential is relatively depolarized, and T-type calcium channels are inactivated [14]. Since thalamorecipient circuits in primary sensory cortices are highly sensitive to the temporal coordination of inputs [15–17], the presence of bursts and tonic firing can have different effects on postsynaptic cortical neurons [18]. This has led to the hypothesis that thalamic nuclei use burst and tonic firing modes to gate or alter the flow of information to and between cortical areas according to the arousal state of the animal [14].

While previous studies have often relied on a single variable to define the state of the animal, several lines of evidence suggest that arousal-related modulations of neuronal activity cannot be adequately characterized by assigning mutually exclusive states to experimental epochs. For example, studies simultaneously measuring locomotion and pupil-linked arousal have revealed distinct effects of each in the visual cortex [6,8]. More generally, a substantial fraction of shared variability in visual cortex activity can be explained by high-dimensional sets of behaviors [19], suggesting that the state of sensory systems at any given moment results from a combination of multiple processes [3]. One potential factor that may distinguish these processes is the timescale over which they extend. For example, single arousal-related neuromodulators can have an impact across several timescales [20–23], and multiple neuromodulatory systems can influence neural activity over broadly different timescales [24].

To move beyond relating neural modulations to mutually exclusive states of arousal, we characterized modulations of spiking activity in the dLGN with respect to pupil size dynamics, taking into account that arousal-related processes occur across a wide range of temporal scales. We discovered that both tonic and burst spiking in the dLGN were coupled to fluctuations in pupil size over timescales ranging from seconds to minutes. Across these timescales, tonic spikes preferred opposite phases of the pupil signal compared to bursts. These multi-scale pupil dynamics captured modulations of dLGN activity beyond those explained by the pupil size per se and could occur in the absence of changes in locomotion state or saccadic eye movements. Furthermore, these modulations were also prevalent during presentation of naturalistic movies, despite the presence of rich stimulus-driven neural activity. Finally, we found that opposing phases of pupil dynamics across all timescales were associated with differences in the encoding of naturalistic movies by the dLGN, indicating that pupil-linked neural activity modulations across various timescales contribute to state-dependent differences in the flow of sensory information to the cortex. Our findings support the notion that arousal-related modulation, rather than being a singular process, likely involves an interplay of changes occurring over diverse timescales.

Results

To assess how dLGN spiking activity is influenced by internal state, we paired extracellular silicon probe recordings with video-based analysis of pupil size. During these recordings, mice were head-fixed, but free to run on an air-cushioned styrofoam ball while viewing a static gray screen or a sparse noise stimulus (Fig 1A). We measured pupil size and locomotion speed as indicators of the internal and behavioral states of the animal. Throughout these recordings under largely isoluminant conditions, pupil size, a marker for internal states such as arousal, was in continuous fluctuation (see example in Fig 1B, black trace and S1A₁ Fig, gray and pink; N = 15 recording sessions in 10 mice). Similarly, active and quiescent behavioral states, as measured by locomotion (Fig 1B, green trace), were observed in all experiments. The proportion of time spent in locomotion varied across recordings, ranging from 0.11 to 0.47 (median = 0.27; S1B₁ Fig). Consistent with previous studies [8,12,25], bouts of locomotion were often accompanied by increases in pupil size (S1B₄ Fig). However, we also noticed that the fluctuations in pupil size were generally similar between active and quiescent behavioral states (S1B₅ Fig). In accordance with previous results [6,7], this suggests that diverse internal states may coexist within a behavioral state.

Download:

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

Fig 1. Extracellular recordings from the dLGN reveal diverse arousal-related modulations of firing rate.

(A) Schematic of the experimental setup. (B) Pupil size (black) and locomotion speed (green) overlaid on the firing rate (blue, binned in 2.5 s windows) and spike bursts (red asterisks) from an example dLGN neuron. (C1) Spike counts (min-max normalized) across pupil sizes (min-max normalized) for 140/156 dLGN neurons with significant modulation (top, one-way ANOVA p ≤ 0.05 across 10 pupil size bins) and 16/156 dLGN neurons without significant modulation (bottom). Neurons were sorted by the pupil size with the highest spike count (black dots). The example neuron from (B) is shown in magenta. The dashed white line indicates the 90th percentile of pupil size. (C2) Same as (C1) but for bursts of spikes. Neurons were sorted by the pupil size with the highest burst count. The dashed white line indicates the 10th percentile of pupil size. (D) The pupil size trace in (B) (gray traces) separated into CPDs (colored traces) occurring over different timescales. The components are described by their characteristic timescale (D1, left) and power (% total, D1, right), as illustrated for the trace indicated by the dashed box in (D2). (D2) Top: The characteristic timescale (gray dashed line) is computed as the amplitude-weighted mean of the component’s instantaneous frequency. Middle: The component’s power is defined as fraction of total power density, computed from the squared amplitudes of all CPDs derived from the same pupil signal. Bottom: The phase of the CPD describes if the pupil is undergoing dilation (−π to 0) or constriction (0 to π) at the timescale defined by the component. (D3) Gray dots: The characteristic timescale of each CPD plotted against its power for CPDs extracted from all recordings (N = 14). Colored dots: Components from the example recording in (B, D). The inset box plot (top) shows the range and IQR of the components with the maximum power from each recording. CPD, components of pupil dynamics; dLGN, dorsal lateral geniculate nucleus.

https://doi.org/10.1371/journal.pbio.3002614.g001

We observed that fluctuations in pupil size during both locomotion and quiescence were accompanied by changes in the spiking activity of dLGN neurons. For example, the neuron shown in Fig 1B generally increased its firing rate when the pupil was large. To characterize the relationship between pupil size and firing rates, we binned the spiking activity of individual dLGN neurons in 250 ms windows and examined mean spike counts across pupil sizes. We observed that the relationship between pupil size and dLGN firing rate varied between neurons, and was often non-monotonic (Fig 1C₁), reminiscent of previous observations in cortex [7]. Indeed, among the 89.7% of recorded dLGN neurons showing significant modulation across pupil sizes (140/156 neurons; one-way ANOVA, p ≤ 0.05), the majority (78/140) had their peak firing rate outside the top decile of pupil size. Indeed, peak firing rates were observed across the entire range of pupil sizes (Fig 1C₁). Similarly, bursts of spikes (see Fig 2A) were also linked to pupil size in most dLGN neurons (Fig 1C₂; 93/145), but, unlike overall firing rates, burst rates tended to be highest at the smallest pupil sizes (72/93; only 21/93 had non-monotonic modulation profiles). To assess the degree to which the relationship between pupil size and spiking activity originated from changes in retinal illumination, we repeated these analyses for experiments conducted in darkness. We found that even in darkness, firing rates and bursting were modulated across pupil sizes in the majority of dLGN neurons (firing rate: 89/94 neurons; burst rate: 44/83 neurons), with many neurons showing high firing rates when the pupil was large (S1C₁ Fig) and bursting when the pupil was small (S1C₂ Fig), suggesting that under these stimulus conditions the overall relationship between pupil size and spiking activity originates from nonvisual factors. We therefore conclude that firing rates and bursting depend on pupil size in the majority of dLGN neurons.

Download:

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

Fig 2. Tonic spikes and bursts in the dLGN are coupled to pupil dynamics across multiple timescales.

(A) Bursts (red) were defined as ≥ 2 spikes occurring after > 100 ms of silence and with an inter-spike interval less than 4 ms [30]; all remaining spikes were classified as tonic spikes (blue). For the phase coupling analysis below, all spikes in a burst were treated as a single burst event. (B) Top: example component of pupil dynamics (CPD). Middle: the corresponding phases of the example CPD. Bottom: simultaneously recorded tonic spiking (blue) and bursting (red) from an example neuron. Right: the phase distributions for bursts and tonic spikes from the example neuron-CPD pair (R^ˆ2: coupling strength; colored dots: preferred phase). The statistical significance of coupling to a CPD was determined using a permutation test by shuffling 300 ms bins of spiking activity (see Materials and methods; asterisks: * for p ≤ 0.05, ** for p ≤ 0.01, *** for p ≤ 0.001). (C) The preferred coupling phase of bursts and tonic spikes for all neuron-CPD pairs with significant coupling. Significant tonic spike coupling was observed in 98.1% of neurons (153/156) and burst coupling in 87.3% of neurons (110/126). (D) Left: Burst—tonic spike phase differences for each neuron-CPD pair in (C) with significant coupling for both types of spiking (N = 284 neuron-CPD pairs), sorted by CPD timescale. Right: distribution of burst—tonic spike phase differences (mean = 2.8, V-test for non-uniform distribution with a mean of π: V = 140.7, p < 0.001; grid lines indicate proportion of 0.25). (E) Coupling strength distributions for all significantly coupled neuron-CPD pairs, binned by CPD timescale (horizontal bars: median coupling strength; Kruskal–Wallis one-way ANOVA across timescale bins for tonic spike coupling strengths: H = 66.4, p = 5.8 × 10⁻¹³, N = 681 neuron-CPD pairs; burst coupling strengths: H = 32.1, p = 5.8 × 10⁻⁶, N = 320 neuron-CPD pairs; Wilcoxon rank-sum test for burst versus tonic spike coupling strengths: W = 3.3 × 10³, p = 2.0 × 10⁻³³, N = 284 neuron-CPD pairs). (F) Coupling strength decay across the multiple CPDs to which single neurons were coupled. Coupling strengths were measured after removal of periods of phase coupling between CPDs (S3B Fig; see Materials and methods) and normalized to the highest coupling strength for each unit (mean ± SEM; left: tonic spiking; right: bursting). Only neuron-CPD pairs with significant coupling after removal of CPD phase coupling were included (blue arrow: tonic spiking mean = 3.8 CPDs per neuron; red arrow: bursting mean = 2.1 CPDs per neuron). (G) Distribution of the preferred coupling timescale (CPD with the strongest coupling) of neurons recorded in each recording session for tonic spiking (left) and bursting (right). There is some variability in timescale preference between recording sessions (one-way ANOVA for tonic spiking: F = 3.2, p = 2.5 × 10⁻⁴; bursting: F = 2.6, p = 2.5 × 10⁻³), sessions are sorted by the timescale to which most neurons had their strongest tuning. CPD, components of pupil dynamics; dLGN, dorsal lateral geniculate nucleus.

https://doi.org/10.1371/journal.pbio.3002614.g002

Returning to Fig 1B, beyond the relationship to pupil size per se, we also found instances where firing rate increases were coupled to dilating phases of the pupil dynamics (Fig 1B, marked by 1). These dilation-related firing rate changes could occur while the pupil was relatively constricted (Fig 1B, marked by 2) or dilated (Fig 1B, marked by 3). Indeed, even in darkness, for a given pupil size the firing rate variability was often larger than the mean (median Fano factor = 1.4; W = 4.4 × 10³, p = 6.4 × 10⁻¹⁶, N = 94 neurons; S1C₃ Fig), suggesting the presence of additional modulatory processes not captured by pupil size. We thus sought to develop a framework that could capture this breadth of modulations in dLGN spiking activity by focusing on the multi-scale dynamics of pupil size.

To gain a better understanding of these modulations of dLGN firing rates, we explored the multi-scale dynamics of the pupil signal, aiming to extract information about changes in internal state beyond the size of the pupil per se. Indeed, previous studies dating to the beginning of pupillometry have suggested that pupil dynamics can be a relevant indicator of arousal [26–28]. The relationship between pupil dynamics and internal states has also been explored in mouse visual cortex [6], where associations between neuromodulatory signaling and pupil dynamics were reported [24]. Building on these studies (see also [7]), we observed that pupil size dynamics could be described over a variety of temporal scales: from minutes-long changes to cycles of dilation and constriction lasting 10s of seconds, to quicker changes on the order of seconds. We therefore employed a data-driven approach, called empirical mode decomposition (EMD, see Materials and methods; [29]), to split the pupil signal without prior assumptions into components capturing its underlying dynamics (Fig 1D). Each component of pupil dynamics (henceforth, CPD) is described by its characteristic timescale (Fig 1D₁, left) and its relative power (Fig 1D₁, right). The characteristic timescale describes the average period of the pupil dilation-contraction cycles captured by the component (Fig 1D₂; see Materials and methods). The extracted CPDs spanned several orders of magnitude in their characteristic timescale, capturing dilation-contraction cycles lasting from several minutes (10⁻³ = 0.001 Hz) to just a second (10⁰ = 1 Hz). Components with high power were found across this entire range, underscoring the multi-component nature of pupil size dynamics (Fig 1D₃). Additionally, the broad distribution of the CPD timescale with the highest power from each recording (Fig 1D₃, top) illustrates the diversity of pupil dynamics across recording sessions. Importantly, the set of CPDs extracted from a single pupil recording progress through dilation-contraction cycles largely independently from each other (S1D Fig; see Materials and methods), indicating that CPDs capture distinct aspects of pupil size dynamics.

Having captured pupil dynamics at multiple timescales, we went on to characterize their relationship to arousal-related dLGN activity. We separated the spiking activity from each dLGN neuron into tonic spiking and bursting (Fig 2A; bursts were defined as ≥2 spikes with ≤4 ms ISI preceded by ≥100 ms without spikes [30]). Consistent with previous findings in awake animals, we found that bursts were relatively rare, accounting for only 3% of spikes on average (median burst ratio for the 93% of neurons with bursting; S2A₂ Fig). Despite their small contribution to the total spike count (S2A₁ Fig), bursts provide an extracellular marker for the membrane potential status of thalamic neurons (indicating prolonged hyperpolarization) and play a role in determining rhythmic cortical states ([13]; see also S2B and S2C Fig). After separating burst and tonic spikes, we performed a phase coupling analysis for each neuron and simultaneously recorded CPD, collecting the phase of the CPD at the time of tonic spikes or bursts—considering all spikes in the burst as one single event (Fig 2B and 2C). The preferred phase of each spike type indicates whether they mainly occur during pupil dilation (-π to 0) or contraction (0 to π) as captured by the CPD. Meanwhile, the coupling strength indicates the degree to which spikes and bursts adhere to the preferred phase and was computed with a bias-corrected metric that allows comparison between neurons with different rates of tonic spiking and bursting (see Materials and methods; S2A Fig). The statistical significance of the coupling was assessed using a permutation test that accounts for short-timescale spiking patterns that typically inflate coupling strength metrics (see Materials and methods; S2B Fig).

We found that both tonic spiking and bursts in dLGN neurons were coupled to the extracted components of pupil dynamics (CPDs) across a wide range of timescales (Fig 2C). Significant coupling of to at least 1 component of pupil dynamics was measured in 98.1% of dLGN neurons (153/156) for tonic spiking and 87.3% (110/126) for bursting. Examining the phase coupling for all neurons and simultaneously recorded CPDs (Fig 2C), we found that this coupling was not limited to a single temporal scale, but rather occurred over time scales spanning seconds to several minutes (1 to 0.001 Hz). Across this wide range of time scales, tonic spikes, and bursts consistently preferred opposite phases of pupil dynamics (Fig 2D; mean tonic spike—burst preferred phase difference = 2.8; V-test for non-uniform distribution with a mean of π: V = 140.0, p ≤ 0.001, N = 284 neuron-CPD pairs with significant coupling of both spike types). The coupling strengths for both tonic spiking and bursting differed across temporal scales (Fig 2E; Kruskal–Wallis one-way ANOVA for tonic spiking: H = 68.6, p = 2.0 × 10⁻¹³, N = 681 neuron-CPD pairs; bursting: H = 34.7, p = 1.7 × 10⁻⁶, N = 320 neuron-CPD pairs), although bursting consistently displayed stronger coupling than tonic spiking (Fig 2E; Wilcoxon rank-sum test: W = 3.4 × 10³, p = 7.1 × 10⁻³⁴, N = 284 neuron-CPD pairs), possibly because bursts are more exclusive to a certain membrane potential state of dLGN neurons. The same analyses performed on data collected in darkness also revealed phase coupling across temporal scales, with significant coupling observed in almost all neurons (S3A₁ Fig; tonic spiking: 96.8%, 91/94 neurons; bursting: 64.8%, 35/54 neurons). In darkness, the opposing phase preference between bursting and tonic spiking was largely preserved (S3A₂ Fig; mean tonic spike—burst preferred phase difference = 2.5; V = 25.8, p = 5.1 × 10⁻⁵, N = 88 neuron-CPD pairs with significant coupling of both spike types), and coupling was also consistently stronger for bursting than tonic spiking (S3A₃ Fig; Wilcoxon rank-sum test: W = 58.0, p = 2.7 × 10⁻¹⁵, N = 88 neuron-CPD pairs), together suggesting that coupling to CPDs was likely driven by changes in internal state rather than the changes in retinal illumination caused by pupil size fluctuation.

Next, we asked whether the coupling we observed across multiple temporal scales resulted from different neurons being modulated at different time scales, or if modulation at multiple temporal scales was present within the spiking of single neurons. Many neurons showed significant coupling to more than 1 CPD for both tonic spiking (mean = 4.4 CPDs with significant coupling per neuron) and bursting (mean = 2.5 CPDs per neuron). To ensure that potential phase relationships between the CPDs themselves (S1D Fig) did not underlie the observed coupling to multiple temporal scales, we repeated the phase coupling analysis after removing periods of time in which the components themselves are coupled (S3B Fig; see Materials and methods). Neurons retained their coupling to more than 1 temporal scale (tonic spiking mean = 3.8 CPDs per neuron; bursting mean = 2.1 CPDs per neuron), and coupling strengths across timescales remained unchanged for both tonic spiking (S3B₃ Fig; median = 0.0089, N = 563 neuron-CPD pairs versus original median = 0.0081, N = 682 neuron-CPD pairs; Mann–Whitney U test: U = 1.9 × 10⁵, p = 0.34) and bursting (median = 0.0583, N = 250 neuron-CPD pairs versus original median = 0.0528, N = 320 neuron-CPD pairs; U = 3.9 × 10⁴, p = 0.56). Among these neurons with multi-scale coupling, the gradual decay in coupling strengths from the strongest to the weakest indicated that modulations at the non-preferred timescales were not negligible (Fig 2F). Considering again the full dataset (Fig 2C–2E), the specific timescale to which a neuron was most strongly coupled was stable for individual neurons across subsamples of the data (S3D Fig) but varied between neurons such that strong coupling was observed across the entire range of timescales measured (Fig 2G). Notably, part of this variability could be attributed to the mouse and/or recording session (one-way ANOVA for tonic spiking timescale preferences across recordings: F = 3.2, p = 2.5 × 10⁻⁴; bursting: F = 2.6, p = 2.5 × 10⁻³), specifically, to the frequency of switches between locomotion and quiescence in a given session (Pearson’s R = 0.71, p = 4.3x10^-3, N = 14 recording sessions; S3C₁ Fig). However, in almost all recording sessions, we found the CPD associated with the strongest coupling (preferred timescale) of neurons from the same recording was distributed across more than 1 component (tonic spiking: 15/15 sessions; bursting: 14/15 sessions), indicating that the timescale of strongest modulation was not only influenced by factors common to the recording session, but also by neuron-specific factors. We thus conclude that the spiking activity of individual dLGN neurons can be coupled to pupil dynamics across multiple independent temporal scales, and that there is diversity in the timescale to which neurons are most strongly coupled.

Our phase coupling framework introduced above (Fig 2) not only captures modulations associated with aspects of pupil dynamics, but it is also capable of capturing modulations usually related to pupil size per se. For example, in Fig 1D, it is apparent that peaks and troughs of certain components coincide with large and small pupil sizes. To disentangle the influence of pupil size from CPD phase coupling, we used a subsampling approach to minimize differences in the distribution of pupil size between the phases of each CPD (Fig 3A; see Materials and methods). We then assessed phase coupling considering only the spikes that occurred in these subsampled periods, noting that, especially for slower CPDs, sometimes only a small proportion of the original data could be retained for analysis (median proportion retained (IQR) for timescales <0.1 Hz: 52% (39, 61); timescales >0.1 Hz: 82% (82, 88)). Despite this reduction in available data, we found that coupling to CPDs in most neurons was largely preserved for tonic spiking (significant coupling observed in 93.9% of neurons, 139/148; mean = 3.3 CPDs per neuron) and burst events, albeit in a smaller proportion (63.6% of neurons, 77/121; mean = 1.0 CPD per neuron). For neurons that retained phase coupling, the characteristic opposing phase preferences of bursts and tonic spikes within neurons was also preserved (Fig 3B; mean = −2.9; V-test for non-uniformity and a mean of π: V = 20.4, p = 6.4 × 10⁻⁴, N = 80 neuron-CPD pairs). Overall, coupling strengths decreased for both tonic spikes (Fig 3C; overall median = 0.0081, N = 681 neuron-CPD pairs versus size-matched median = 0.0046, N = 489 neuron-CPD pairs; Mann–Whitney U test: U = 2.0 × 10⁵, p = 4.4 × 10⁻⁷) and bursts (overall median = 0.0528, N = 320 neuron-CPD pairs versus size-matched median = 0.0250, N = 115; Mann–Whitney U test: U = 2.2 × 10⁴, p = 1.0 × 10⁻³). This coupling strength decrease, however, was only present for a middle-range of temporal scales, suggesting that pupil size per se may not contribute to slow (timescale <0.03 Hz) or fast (>0.1 Hz) coupling in dLGN neurons. We then examined how the preferred timescale of individual neurons was affected after controlling for the influence of pupil size per se, finding that many neurons shifted the timescale to which they were most strongly coupled (Fig 3D, faded slices; tonic spikes: 55.9%; bursts: 30.2%). For bursting, such shifting of timescales was particularly prominent among neurons whose activity was monotonically related to pupil size (S3E Fig; Chi-squared test: χ² = 9.5, p = 8.7 × 10⁻³; see also Fig 1C). Strikingly, however, we also found a sizeable proportion of neurons whose strongest coupling increased after controlling for pupil size (Fig 3D, highlighted slices; tonic spikes: 25.5%; bursts: 15.1%), suggesting that for these neurons effects related to pupil size per se were masking other firing rate modulations related to pupil dynamics. We thus conclude that the majority of dLGN neurons undergo multi-scale modulations of bursting and tonic spiking beyond those associated with pupil size per se.

Download:

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

Fig 3. Coupling of dLGN spiking to pupil dynamics persists after controlling for pupil size.

(A) Comparison of the difference in pupil size (min-max normalized) between the phase bins from each CPD sorted by CPD timescale before (gray) and after subsampling (black; see Materials and methods), illustrating the efficacy of the size matching procedure. (B) Distribution of burst—tonic spike phase differences measured after the size-matching procedure (N = 80 neuron-CPD pairs; mean = −2.94, V-test for non-uniform distribution with a mean of π: V = 20.4, p = 6.4 × 10⁻⁴; gray: distribution from Fig 2D; grid lines indicate proportion of 0.25). (C) Coupling strengths (solid lines: median; dashed lines: bootstrapped SE of the median) measured before (faded lines; Fig 2E) and after (bold lines) the size-matching procedure for tonic spikes (blue; overall median = 0.0081, N = 681 neuron-CPD pairs vs. size-matched median = 0.0046, N = 489 neuron-CPD pairs; Mann–Whitney U test: U = 1.96 × 10⁵, p = 4.37 × 10⁻⁷) and bursts (red; overall median = 0.0528, N = 320 neuron-CPD pairs vs. size-matched median = 0.0250, N = 115; Mann–Whitney U test: U = 2.22 × 10⁴, p = 1.04 × 10⁻³). (D) Proportion of neurons with significant phase coupling from Fig 2 that kept the same preferred coupling timescale (bold slices; 38.6% for tonic spiking, 30.2% for bursting), shifted their preferred coupling timescale (faded slices; 55.9% for tonic spiking, 30.2% for bursting), or lost their coupling (gray slices; 5.5% for tonic spiking, 39.6% for bursting). The highlighted slices show the proportion of neurons that had an increase in coupling strength after controlling for pupil size (25.5% for tonic spiking, 15.1% for bursting). CPD, components of pupil dynamics; dLGN, dorsal lateral geniculate nucleus.

https://doi.org/10.1371/journal.pbio.3002614.g003

We next investigated whether coupling between dLGN spiking and pupil size dynamics could be attributed to modulations driven by transitions in behavioral states. Previous studies have shown that changes in behavior, such as the transition from quiescence to locomotion, are accompanied by firing rate changes in dLGN neurons [12,31,32]. Consistent with these findings, we observed a decrease in tonic spiking in a 5 s window surrounding the offset of a locomotion bout (S4A₁ Fig; see also [8]), and a sharp increase in tonic spiking, preceded by a slight decrease in bursting, in a 4 s window surrounding the onset of a locomotion bout (S4A₂ Fig). Periods of locomotion were also associated with an increased pupil size (S1B₄ Fig). Unsurprisingly then, 55.3% of CPDs (52/94) had a small positive correlation with locomotion speed (S4B Fig; permutation test: p ≤ 0.05, see Materials and methods). Given these findings, we asked if the CPD-linked spiking modulations we observed might be driven by these locomotion-correlated components. As a first step, we compared phase coupling across components, and observed that significant coupling for tonic spikes was equally likely regardless of whether the component was correlated to locomotion or not (neuron-CPD pairs with significant coupling: correlated = 73.5% (305/415) versus uncorrelated = 75.9% (341/449); Chi-squared test: χ² = 0.56, p = 0.45). For bursting, locomotion-correlated components were even less likely to have significant coupling (correlated = 38.2% (128/335) versus uncorrelated = 50.3% (186/370); Chi-squared test: χ² = 9.9, p = 1.7 × 10⁻³). Thus, while locomotion can drive changes in spiking activity, and locomotion speed is partially reflected in certain CPDs, these relationships do not appear to be the sole cause of coupling to pupil dynamics.

To explicitly remove the influence of behavioral state changes on the coupling between pupil dynamics and spiking activity, we computed the phase coupling taking into account only the spiking activity that occurred within a given behavioral state. We also excluded spiking activity in the transitional windows between behavioral states (S4A Fig; quiescence to locomotion: −2 to 2 s; locomotion to quiescence: −1 to 4 s). We first focused on periods of quiescence (Fig 4A₁), which constituted on average 72% of the recordings (median proportion of time outside of locomotion bouts; S1B₁ Fig). During these quiescent periods, where variance in locomotion speed was low (S1B₃ Fig), we found that the coupling between spiking activity and pupil dynamics was largely the same as when measured across the entire recording (Fig 4A₂; significant tonic spike coupling in 97.3% (143/147); burst coupling in 77.1% (84/109) of neurons). Neurons retained the anti-phase relationship between tonic spiking and bursting (Fig 4A₃; mean tonic spike—burst phase difference = 2.9, N = 221 neuron-CPD pairs; V-test for non-uniform distribution with a mean of π: V = 111.1, p < 0.001), with coupling to multiple timescales of pupil dynamics (tonic spiking mean = 4.2 CPDs per neuron; bursting mean = 2.3 CPDs per neuron), and similar coupling strengths across all timescales compared to the coupling measured across the whole recording for tonic spiking (Fig 4A₄; quiescence median = 0.0100, N = 591 neuron-CPD pairs versus overall median = 0.0081, N = 681 neuron-CPD pairs; Mann–Whitney U test: U = 1.9 × 10⁵, p = 2.6 × 10⁻²) and bursting (quiescence median = 0.0564, N = 256 versus overall median = 0.0528, N = 320 neuron-CPD pairs; Mann–Whitney U test: U = 3.9 × 10⁴, p = 0.28).

Download:

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

Fig 4. The coupling between dLGN spiking and pupil dynamics is not driven by overt behaviors.

(A) Locomotion speed for an example period of quiescence (gray bar: period of quiescence). Dashed lines: period during which spikes were taken for the subsequent phase coupling analysis; note that the 4 s after the offset of the first locomotion bout and 2 s prior to the next locomotion bout were not included in this time window. (A2) Preferred coupling phase of tonic spikes (blue) and bursts (red) for all neuron-CPD pairs with significant coupling during quiescence. Significant tonic spike coupling was observed in 97.3% of neurons (143/147; mean = 4.2 CPDs per neuron) and burst coupling in 77.1% of neurons (84/109; mean = 2.3 CPDs per neuron). (A3) Distribution of the preferred phase differences between tonic spiking and bursting for neuron-CPD pairs with significant coupling of both spike types (mean = 2.9, N = 221 neuron-CPD pairs; V-test for non-uniform distribution with a mean of π: V = 111.1, p < 0.001; grid lines indicate proportion of 0.25). (A4) Coupling strengths (median ± bootstrapped SE) measured during quiescence (bold lines) or across the whole recording (faded lines) for tonic spikes (blue; quiescence median = 0.0100, N = 591 neuron-CPD pairs vs. overall median = 0.0081, N = 681 neuron-CPD pairs; Mann–Whitney U test: U = 1.9 × 10⁵, p = 2.6 × 10⁻²) and bursts (red; quiescence median = 0.0564, N = 256 vs. overall median = 0.0528, N = 320 neuron-CPD pairs; Mann–Whitney U test: U = 3.9 × 10⁴, p = 0.28). (B1) Locomotion speed for an example bout (green bar: period of locomotion). Dashed lines: period during which spikes were taken for the subsequent phase coupling analysis; note that the first and final 1 s of the bout were not included in this time window. (B2–5) Same as (A2–5) but for phase coupling measured during locomotion bouts. Significant tonic spike coupling was observed in 85.3% of neurons (122/143; mean = 2.5 CPDs per neuron; median coupling strength during locomotion = 0.0254, N = 359 neuron-CPD pairs; Mann–Whitney U test: U = 7.2 × 10⁴, p = 2.6 × 10⁻²⁸) and burst coupling in 77.9% of neurons (60/77; mean = 1.7 CPDs per neuron; median coupling strength during locomotion = 0.0927, N = 133 neuron-CPD pairs; Mann–Whitney U test: U = 1.3 × 10⁴, p = 6.2 × 10⁻¹⁰). Tonic spiking and bursting in neuron-CPD pairs with significant coupling of both spike types tended to occur at opposing phases (mean tonic spike—burst phase difference = 3.1, N = 87 neuron-CPD pairs; V-test for non-uniform distribution with a mean of π: V = 56.7, p < 0.001). (C1) Eye position for an example recording period with saccadic eye movements (dark purple: azimuth; light purple: elevation; gray bars: the time windows without saccades). Dashed lines: period during which spikes were taken for the subsequent phase coupling analyses, note that the −2 to 2 s of activity surrounding each saccade was excluded. (C2–5) Same as (A2–5) but for phase coupling measured during periods with no saccadic eye movements. Significant tonic spike coupling was observed in 97.4% of neurons (151/155); mean = 4.1 CPDs per neuron; median coupling strength without saccades = 0.0090, N = 634 neuron-CPD pairs; Mann–Whitney U test: U = 2.1 × 10⁵, p = 0.11) and burst coupling in 81.7% of neurons (94/115; mean = 2.5 CPDs per neuron; median coupling strength without saccades = 0.0545, N = 284 neuron-CPD pairs; Mann–Whitney U test: U = 4.5 × 10⁴, p = 0.99). Tonic spiking and bursting in neuron-CPD pairs with significant coupling of both spike types tended to occur at opposing phases (mean tonic spike—burst phase difference = 2.9, N = 203 neuron-CPD pairs; V-test for non-uniform distribution with a mean of π: V = 131.5, p < 0.001). CPD, components of pupil dynamics; dLGN, dorsal lateral geniculate nucleus.

https://doi.org/10.1371/journal.pbio.3002614.g004

Having shown that phase coupling across timescales persisted when measured only during periods of quiescence, we next repeated the same analyses focusing on periods with locomotion bouts (Fig 4B₁) and found similar coupling characteristics as for the other states (Fig 4B₂; significant tonic spike coupling in 85.3% (122/143), burst coupling in 77.9% (60/77) of neurons). Tonic spikes and bursts showed preferences for opposing phases (Fig 4B₃; mean tonic spike—burst phase difference = 3.1, N = 87 neuron-CPD pairs; V-test for non-uniform distribution with a mean of π: V = 56.7, p < 0.001). There was a notable increase in coupling strengths for both tonic spikes (Fig 4B₄; locomotion median = 0.0254, N = 359 neuron-CPD pairs; Mann–Whitney U test: U = 7.2 × 10⁴, p = 2.6 × 10⁻²⁸) and bursts (locomotion median = 0.0927, N = 133 neuron-CPD pairs; Mann–Whitney U test: U = 1.3 × 10⁴, p = 6.2 × 10⁻¹⁰). Given the higher locomotion speed variability during locomotion bouts (S1B₃ Fig), we hypothesized that particularly strong dLGN modulation in this state might be linked to these changes in overt behavior. We therefore identified CPDs that were correlated to locomotion speed only within bouts of locomotion (64.9% of CPDs, 61/94; S4C Fig). Consistent with this hypothesis, we found that the locomotion-correlated components were more likely to drive significant phase coupling for both tonic spikes (neuron-CPD pairs with significant coupling: correlated = 77.2% (477/618), uncorrelated = 68.7% (169/246); Chi-squared test: χ² = 6.3, p = 0.01) and bursts (correlated = 49.6% (256/516), uncorrelated = 30.7% (58/189); Chi-squared test: χ² = 19.3, p = 1.1 × 10⁻⁵). Tonic spike coupling strengths were also higher for locomotion-correlated components (correlated median = 0.0099, N = 477 versus uncorrelated median = 0.0054, N = 169; Mann–Whitney U test: U = 4.6 × 10⁴, p = 7.1 × 10⁻³). Thus, while transitions between quiescence and locomotion bouts do not drive dLGN-pupil phase coupling, movement-related modulations may underlie a particularly strong coupling between dLGN spiking and pupil dynamics within periods of locomotion.

Apart from locomotion, other behaviors, such as eye movements, may also be associated with arousal [33] and can induce changes in spiking activity. Indeed, upon visual inspection of eye position data, we noticed that saccades tended to occur during the dilation phase of large, slow pupil fluctuations, but also appeared to be related to smaller, faster fluctuations (S5A Fig), resembling the relationship between tonic spiking and pupil dynamics. To explore this further, we performed the same phase coupling analysis for saccades as we did for spiking and found that saccades had an almost identical phase coupling profile to that of tonic spikes: saccades predominantly occurred during the dilating phases of multiple temporal scales of pupil dynamics (S5B Fig; significant coupling of saccades to at least 1 CPD in 15/15 recording sessions; mean = 3.7 CPDs per session). To further investigate this relationship, we next asked if saccades could drive changes in spiking activity. Consistent with previous findings in primates (reviewed in [34]), we observed that bursting and tonic spiking in dLGN neurons was modulated in a short timescale window surrounding saccades. In a 4-s window surrounding saccades, tonic spiking increased, and bursting decreased, with the exception of a brief increase at the time of the saccade (S5C₁ Fig; N = 118/121 neurons had significant peri-saccadic modulation). With regards to tonic spiking, neurons displayed diverse peri-saccadic activity patterns, among which we identified at least 2 distinct response types (S5C₂ Fig). Despite this diversity in profiles, the overall effect of this modulation was a consistent increase in tonic spiking in 89.0% of saccade-responsive neurons (S5C₃ Fig; modulation strength median = 0.20; Wilcoxon rank-sum test: W = 5.4 × 10², p = 1.6 × 10⁻¹⁵). In contrast, bursting activity tended to decrease (modulation strength median = −0.06; Wilcoxon rank-sum test: W = 1.4 × 10³, p = 6.2 × 10⁻³). We conclude that saccades occur during dilating phases of pupil dynamics and have a marked impact on dLGN spiking activity.

We therefore hypothesized that the changes in spiking activity during the peri-saccadic period might contribute to the observed coupling between tonic spiking and pupil dynamics. We reasoned that, if the peri-saccadic modulation was driving the coupling between tonic spikes and pupil dynamics, then coupling strengths and saccadic modulation strengths should be correlated. Indeed, we found that saccadic modulation strengths could correlate with coupling strengths, but only for fast timescales of pupil dynamics, and only for tonic spiking (S5D Fig). To eliminate the effect of saccade-driven changes in spiking activity, similar to our approach for locomotion, we excluded spiking activity in a window from −2 s to 2 s surrounding saccades (Fig 4C₁) and repeated the phase coupling analysis. We found that the main characteristics of the coupling between spiking activity and pupil dynamics were preserved across temporal scales, even without peri-saccadic activity. Significant coupling for tonic spiking was observed in 97.4% of neurons (151/155; mean = 4.1 CPDs per neuron) and 81.7% of neurons for bursting (94/115; mean = 2.5 CPDs per neuron; Fig 4C₂). Tonic spiking and bursting retained their anti-phase relationship within neurons (mean tonic spike = burst phase difference = 2.9; V-test for non-uniform distribution with a mean of π: V = 144.0, p = 0.0; Fig 4C₃), and coupling strengths were similar (median tonic spike coupling strength without saccades = 0.0090, N = 634 neuron-CPD pairs; Mann–Whitney U test: U = 2.1 × 10⁵, p = 0.11; median burst coupling strength without saccades = 0.0545, N = 284 neuron-CPD pairs; Mann–Whitney U test: U = 4.5 × 10⁴, p = 0.99; Fig 4C₄). We therefore conclude that, although behaviors like locomotion and eye movements are reflected in some components of pupil size dynamics, and can induce changes in spiking, the coupling between spiking activity and pupil dynamics is not dominated by modulations related to these overt behaviors.

So far, we reported a multi-scale coupling of dLGN activity to pupil dynamics in the absence of a patterned visual stimulus, but how stable is this coupling in the presence of a rich visual stimulus? At least 2 factors could potentially disrupt this coupling. Firstly, naturalistic stimuli have been shown to elicit repeated patterns of both tonic and burst firing in the dLGN [32,35], which may dominate internally driven activity fluctuations. Secondly, luminance changes or other salient features of the stimulus could induce pupil size changes that might interfere with those driven by internal state fluctuations. To investigate dLGN spike coupling to pupil dynamics during stimulus viewing, we presented 5-s long naturalistic movie clips to mice while recording pupil size and dLGN activity (Fig 5; data from [32]). We observed that many neurons in the dLGN responded to the movies with repeated patterns of activity, as exemplified by the neuron shown in Fig 5A₁. However, the average response to the stimulus (Fig 5A₁, bottom) often appeared weaker than the trial-to-trial variability in the time-averaged mean firing rates (Fig 5A₁, right). We quantified this difference using the “response variability ratio” (RVR; see Materials and methods) and discovered that 100% of neurons had an RVR below 1 (N = 64 neurons; Fig 5A₂). In fact, for 90.6% of neurons, the variability was more than twice as strong as the signal (RVR < 0.5; Fig 5A₂), indicating that dLGN neurons showed considerable variability in their firing rates that could not be accounted for by the visual stimulus. Similarly, when we repeated the same analysis for pupil size, we found that the movies did not systematically drive pupil size changes (Fig 5B₁), resulting in an overwhelming dominance of trial-to-trial variance in pupil size (Fig 5B₂). Thus, internally driven pupil dynamics remain largely uninterrupted by the naturalistic movie we presented. The analyses above suggest that factors unrelated to the stimulus, such as coupling to arousal-related variables, may be prevalent despite the spiking responses evoked by the movie stimulus.

Download:

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

Fig 5. Coupling of dLGN activity to pupil dynamics is related to differences in visual stimulus encoding.

(A1) Raster plot showing the responses of an example neuron to a movie (black: tonic spikes; red: burst spikes). Bottom: The mean response of the example neuron, centered on the mean across time. Right: The mean firing rate on each trial, centered on the mean across trials. Note that the “Response” and “Variability” axes have the same scale. (A2) The “response variability ratio” (RVR) compares the variance in the mean stimulus response to the mean variance for matching time points across trials (median = 0.16, N = 64 neurons; see Materials and methods). An RVR less than 1 indicates that the trial-to-trial changes in stimulus responses are larger than the change induced by the stimulus. (B1) Trial-by-trial pupil size responses for the same experiment as in (A1). (B2) Distribution of pupil RVRs for all recording sessions with naturalistic movie stimulus presentation (N = 9 recording sessions in 6 mice). An RVR of near zero indicates that stimulus induced fluctuations in pupil size were negligible compared to across-trial fluctuations. (C1) Preferred coupling phase of tonic spikes (blue) and bursts (red) for all neuron-CPD pairs with significant coupling during stimulus presentation. Significant tonic spike coupling was observed in 98.4% of neurons (62/63; mean = 4.3 CPDs per neuron) and burst coupling in 62.7% of neurons (32/51; mean = 1.3 CPDs per neuron). (C2) Preferred phase distribution for tonic spikes and bursts for slower timescales (left: timescale < 0.1 Hz; N = 183 neuron-CPD pairs with tonic spike coupling; N = 37 neuron-CPD pairs with burst coupling; grid lines indicate proportion of 0.25) and faster timescales (right: timescale > 0.1 Hz; N = 90 neuron-CPD pairs with tonic spike coupling; N = 31 neuron-CPD pairs with burst coupling). (C3) Distribution of the preferred phase differences between tonic spiking and bursting for neuron-CPD pairs with significant coupling of both spike types (mean = 2.85, N = 56 neuron-CPD pairs, V-test for non-uniform distribution with a mean of π: V = 16.0, p = 1.6 × 10⁻³). (C4) Coupling strengths (median ± bootstrapped SE) measured during stimulus presentation (bold lines) or without patterned visual stimulation (faded lines, Fig 2E) for tonic spikes (movie median = 0.0050, N = 273 neuron-CPD pairs vs. gray median = 0.0081, N = 681 neuron-CPD pairs; Mann–Whitney U test: U = 1.1 × 10⁵, p = 1.1 × 10⁻³) and bursts (movie median = 0.0332, N = 68 vs. gray median = 0.0528, N = 320 neuron-CPD pairs; Mann–Whitney U test: U = 1.2 × 10⁴, p = 0.16). (D1) Schematic representation of how the data was partitioned in (D) and (E). Each trial of naturalistic stimulus presentation was split into five 1-s segments and a decoder was trained to identify the stimulus segment using spiking activity. Two decoders were trained for each neuron-CPD pair, one for each of 2 CPD phase groupings. Turquoise: Decoders trained on segments during which the CPD was in the phase bin where tonic spiking preferentially occurred. Pink: Decoders trained on segments during which the CPD was in the opposing phase bin. (D2) Schematic of the movie segment decoding, initially performed by training and testing within data partitions. (D3) Decoding accuracy distributions for the 2 phase-groupings for each CPD timescale (black lines: median decoding accuracy; gray line: chance-level performance). (E1) Schematic of the movie segment decoding, now performed by training a decoder on one of the data partitions and testing it on data from the other. The decoding performance across CPD phases was compared to the scores from (D3) to yield a “decoding penalty.” This penalty was compared to the penalty obtained by repeating the same procedure on 10 random partitions of the data segments. A decoding penalty was considered significant if it was greater than the penalties obtained from all 10 random partitions. (E2) Distribution of the ranks of decoding penalties obtained from partitions based on CPD phases among those obtained from 10 random partitions (black lines: median rank). High ranks indicate that the penalty for training and testing across data partitions based on CPD phases were greater than those obtained by random partitions. (E3) Distributions of decoding penalties, only for neuron-CPD pairs with significant penalties (black lines: median decoding penalty). The penalty for training and testing across CPD phases can be as high as 50% across most timescales, suggesting that all timescales of modulation contribute to substantial differences in how stimuli are encoded. CPD, components of pupil dynamics; dLGN, dorsal lateral geniculate nucleus; RVR, response variability ratio.

https://doi.org/10.1371/journal.pbio.3002614.g005

To investigate the coupling of spiking activity to pupil dynamics during naturalistic stimulus viewing, we decomposed the pupil size signal as before and characterized the phase coupling of bursts and tonic spikes (N = 9 recording sessions in 6 mice). We found significant tonic spike coupling in 98.4% of neurons (62/63; mean = 4.3 CPDs per neuron) and burst coupling in 62.7% of neurons (32/51; mean = 1.3 CPDs per neuron; Fig 5C₁), with opposite phase preferences at slower CPDs (timescales < 0.1 Hz; Fig 5C₂). Although stimulus presentation appeared to disrupt the preferred phase relationship between bursts and tonic spikes for faster CPDs (timescales > 0.1 Hz; Fig 5C₂), the opposing-phase relationship was still present overall (mean tonic spike—burst phase difference = 2.85, N = 56 neuron-CPD pairs with significant coupling to both spike types; V-test for non-uniform distribution with a mean of π: V = 16.0, p = 1.6 × 10⁻³; Fig 5C₃). The disruption at faster timescales is likely due to the presence of stimulus-driven bursts [35,36], and also demonstrates that the opposing phase relationship is not guaranteed based on intrinsic biophysical properties of dLGN neurons, such as the burst generation mechanism. Coupling strengths tended to be lower for tonic spikes during movies (median = 0.0050, N = 273 neuron-CPD pairs) compared to gray screen conditions (median = 0.0081, N = 681 neuron-CPD pairs; Mann–Whitney U test: U = 1.1 × 10⁵, p = 1.1 × 10⁻³; Fig 5C₄). In contrast, coupling strengths remained similar for bursts (movie median = 0.0332, N = 68 versus gray screen median = 0.0528, N = 320 neuron-CPD pairs; Mann–Whitney U test: U = 1.2 × 10⁴, p = 0.16; Fig 5C₄).

To examine the implications of these findings for the encoding of arousal-related variables by dLGN spiking activity, we leveraged the repeating patterns of activity induced by the stimulus and next asked if pupil dynamics could be decoded from patterns of spiking activity. We first split each 5 s trial into 1 s segments, and, for each CPD, used the mean phase during each segment to assign one of 2 phase labels. We then used spiking activity from each individual neuron with significant phase coupling to decode the phase label. We found that the phase labels could be decoded above chance level for every timescale of pupil dynamics, both when using tonic spikes (S6A Fig) or bursts (S6B Fig). Together, these results show that dLGN neurons also represent multi-scale aspects of pupil dynamics during encoding of a naturalistic visual stimulus.

Given the different stimulus-response functions of bursts and tonic spikes [35–37], and changes in feature selectivity that can occur with arousal [38–40], we next asked whether the movie stimulus is differently represented by spiking activity during different phases of pupil dynamics. To this end, we used spiking activity to decode which of the five 1-s stimulus segments was presented. For each neuron-CPD pair, we sorted the stimulus segments into 2 groups according to the phase of the component at the time when the segment was on screen (Fig 5D₁). These 2 phase groupings were designed such that one phase bin was centered on the preferred phase of tonic spiking, while the other phase bin was centered 180° opposite. Critically, we used all spikes to decode the identity of the stimulus segment, regardless of whether they were tonic spikes or part of a burst, to simulate the perspective of a downstream neuron in the cortex that would be blind to this classification.

We observed that decoders trained and tested on data from the same phase group (Fig 5D₂) achieved nearly perfect performance (Fig 5D₃; phase bin 1 decoding accuracy median = 99.6%, phase bin 2 decoding accuracy median = 99.6%, N = 399 neuron-CPD pairs), indicating that spiking activity during either phase was informative about the stimulus segment. However, when we tested the decoder trained using spiking activity from one phase group with activity occurring during the opposite phase grouping (Fig 5E₁), we found a marked decrease in decoding performance (Fig 5E₃; median penalty for train 1 → test 2 decoding = −33.6%, train 2 → test 1 = −34.5%). Critically, this decoding penalty was only considered significant if it was greater than the penalties obtained by performing the same cross-group decoding analysis on random partitions of the data (Fig 5E₂). The overall negative decoding penalty implies that encoding of the stimulus into firing rates differs between the phases of pupil dynamics. Overall, high within-state decoding accuracy and prominent decoding penalties across states suggest that the trial-to-trial variability in Fig 5A is not random but may stem from arousal-related changes in the stimulus-response properties of dLGN neurons. Importantly, these results were not limited to a specific temporal scale of pupil dynamics but were observed across the entire range of timescales examined, indicating that multiple temporal scales of modulation affect stimulus encoding.

Discussion

Our results establish that activity in dLGN, the primary visual thalamic nucleus, is coupled to distinct CPDs that occur over multiple timescales, ranging from several seconds to several minutes. Throughout these timescales, bursts and tonic spikes exhibited robust phase preferences, occurring during opposite phases of pupil dynamics across all timescales. Individual neurons were coupled to multiple timescales and were diverse in the timescale that drove the strongest modulation. The coupling between dLGN spiking and pupil dynamics extended beyond effects related to differences in pupil size per se and could not be attributed to transitions between quiescence and locomotion, or to saccadic eye movements. Furthermore, the coupling persisted even during viewing of a rich naturalistic movie, where we observed differences in how visual stimuli were encoded across phases of pupil dynamics. Together, our findings support the notion that arousal-related modulation of visual thalamus in the waking state results from a combination of changes in dLGN spiking activity linked to diverse temporal scales.

Our data-driven decomposition of the pupil signal moves beyond a binary classification of behavioral state and reveals CPDs which together predict dLGN activity modulations across multiple timescales. It is important to note that the components we recovered here do not necessarily map directly onto the time-courses of independent internal processes or behavioral sequences. Yet, our results support the idea that modulations of both the pupil signal and dLGN neural activity result from a combination of multiple intrinsic and/or extrinsic influences, possibly separated by the temporal scale over which they occur. Multi-scale influences have been described in other contexts, including the presence of diverse activity timescales within and between neurons [41–43], the action of several behavioral state-related neuromodulators with distinct timescales [24,44] or the largely non-overlapping activation dynamics of different ensembles of neurons within the same neuromodulatory system [45]. Furthermore, quantification of behavior in freely moving animals has revealed a multi-scale organization, where long-lasting “idle” states can be punctuated by rapid “active” behaviors (and vice-versa; [46]), mirroring the modulations we observed here, where slow and fast modulations were superimposed in dLGN activity. Finally, in the context of task performance, cognitive processes orchestrated by arousal-related neuromodulators can change on fast and slow timescales [21], such as quick reorienting to surprising stimuli or the choice of specific behavioral strategies over longer timescales. In humans, variations in stimulus detection ability have been linked to the phase of slow components (0.01 to 0.1 Hz) of EEG signals [47]. Thus, the multi-scale modulation we observed in the dLGN could reflect sensory processes adapting to the various timescales of behavioral organization of the organism.

While providing a quantitative account of dLGN modulations linked to multi-scale pupil dynamics, locomotion, and saccades, a limitation of our study is that it has not addressed these modulations on a mechanistic level. However, the relationship to pupil size strongly implicates the neuromodulators norepinepherine (NE) [20,24,48] and acetylcholine (ACh) [11,24] in the present findings. NE is provided to the thalamus by the locus coeruleus (LC) [49,50], while ACh in the rodent thalamus comes from several brainstem nuclei (PPN/LDT and PBG) [51–53] and the basal forebrain (BF) [51,54]. Consistent with a potential role for the LC-NE system in the present findings, stimulation of the LC in vivo has been shown to suppress bursting in the dLGN [55] and primary somatosensory thalamus (VPM) [56], while causing pupil dilation [48]. Additionally, NE levels in the mouse thalamus during NREM sleep have been found to fluctuate with a frequency of 0.02 Hz and are linked to correlates of sensory arousability [57]. Given the prevalence of high power of CPDs around 0.02 Hz we observed, it also seems likely that the LC-NE system drives some of the coupling between dLGN spiking and pupil dynamics at such timescales during wakefulness. In addition, stimulation of the cholinergic PPT/LDT increases firing rates in the VPM [58], and switches dLGN neurons from burst to tonic mode [59], supporting the role of cholinergic nuclei in the modulations observed here. Although the activity of the PPN/LDT and PBG have not been directly linked to pupil size fluctuations, they are involved in coordinating locomotor patterns and eye movements [60,61], meaning they could specifically drive the subset of modulations we observed linked to these behaviors. It is important to note that neuromodulatory influences in the thalamocortical system extend beyond direct effects on thalamic neurons, and also involve the thalamic reticular nucleus [51,54,62–64], corticothalamic L6 neurons [65,66], and retinal boutons [40,67,68], all of which could contribute to additional arousal-related influences on thalamic neurons. Furthermore, although our analyses in darkness control for changes in retinal illumination linked to pupil size per se, there are known effects of pupil size on feed-forward signaling. Changes in retinal irradiation linked to pupil size can shift color selectivity in the visual system [39], and there is evidence for interactions between luminance and behavioral modulation [69] that warrants further exploration. Future studies directly tracking and manipulating the activity of these modulatory systems in the thalamus promise to yield further insight into the relationship between internal states and feed-forward sensory signals across diverse time scales.

The modulations in burst and tonic spiking we observed indicate that, throughout periods of wakefulness, and even during stimulus viewing, the dLGN is in constant alternation between 2 distinct information processing modes. One prominent view is that tonic spikes encode information about a stimulus linearly, while stimulus-driven bursts provide an all-or-none “wake-up call” to the cortex [14], by providing augmented retinogeniculate [37] and geniculocortical [18] communication, and improved stimulus detectability [70] (for related work in the somatosensory system, see e.g., [71,72]). Burst spikes, in comparison with tonic spikes, also have different feature selectivity: they track low frequency stimulus content [9,73,74] with an earlier response phase [32,36,75], integrate input over longer timescales with biphasic response kernels [35,36,76–78], and prefer smaller stimuli [79] with a stronger suppressive surround [36]. Such differences likely contribute to the decoding penalty we observed when training and testing decoders on spiking activity from opposing phases of the modulations we observed. However, it is also possible that neuromodulatory mechanisms change the stimulus response properties of dLGN neurons, without eliciting a switch in firing mode from tonic spiking to bursting. It remains to be seen whether spatiotemporal receptive fields change across phases of pupil dynamics, and whether these changes can be attributed to membrane hyperpolarization and bursting.

Our results focused on the commonalities of multi-scale modulation across the recorded population of dLGN neurons. However, we also observed a substantial diversity across neurons in the strength of coupling within a timescale, the timescale with strongest coupling, and the degree to which stimulus decoding was influenced. The mouse dLGN consists of various cell types [80] and functional subtypes [81,82], some of which are known to be differentially affected by arousal-linked variables [38,68]. Moving forward, considering the specific functional role of a given neuron, along with its temporal structure of modulation, will lead to a better understanding of dynamic stimulus processing in the visual system.

In this study, revealing the multi-scale nature of arousal-related activity modulations in dLGN was possible by applying a decomposition approach to the pupil signal. This approach builds on previous studies using raw pupil size as a marker for internal arousal processes in sensory circuits [7,40,66,83–87] or the pupil’s dynamics within a restricted frequency range [6,24,43]. Importantly, however, our analysis of pupil components and phases in relation to spiking activity is not intended to undermine the usefulness of pupil size itself, which remains an accessible signal that offers a relevant index into internal state. Indeed, our analyses show that modulations indexed by pupil size may be viewed as a specific case of phase coupling to pupil dynamics: coupling to high amplitude peaks or troughs of a component. We therefore view these approaches as complimentary and considering both allows for a more nuanced and comprehensive account of the modulations of sensory processing.

While our study has expanded the use of the pupil signal to reveal arousal-related modulations to multiple temporal scales, recent work has emphasized that a larger array of behaviors may also help to better explain activity in sensory systems [19,88]. Individual behaviors, such as locomotion onset [8,12], eye movements [89], head movements [69], or particular postures in freely moving conditions [90], modulate activity in the rodent early visual system. In the present study, locomotion and eye movements were shown to relate to a subset of pupil-linked modulations. However, given that a larger set of behavioral components has been linked to activity modulation in the visual cortex, this subset may expand as more behaviors are measured. These behavior-related modulations often occur in time-windows of seconds or less, which may be explained by the need to account for their immediate influence on sensory inputs [34,91]. Our present work shows that sensory neurons are modulated not only at these faster timescales, but also at longer timescales. In the future, a more complete characterization of the organization of spontaneous behaviors across temporal scales [46] promises to advance our understanding of multi-scale modulations of sensory processing by behavioral states.

Materials and methods

Supporting information

Acknowledgments

We would like to thank M. Sotgia for lab management and support with animal handling and histology, Y. Bauer, G. Born, and A. H. Kotkat for their contributions to the data collections and spike sorting, S. Schörnich for IT support, and B. Grothe for providing excellent research infrastructure.