Evidence for two distinct thalamocortical circuits in retrosplenial cortex

Retrosplenial cortex (RSC) lies at the interface between perceptual and memory networks in the brain and mediates between these, although it is not yet known how. It has two distinct subregions, granular (gRSC) and dysgranular (dRSC). The present study investigated how these subregions differ with respect to their electrophysiology and connections, as a step towards understanding their functions. gRSC is more closely connected to the hippocampal system, in which theta-band local field potential oscillations are prominent. We therefore compared theta-rhythmic single-unit activity between the two RSC subregions and found, mostly in gRSC, a subpopulation of non-directional cells with spiking activity strongly entrained by theta oscillations, suggesting a stronger coupling of gRSC to the hippocampal system. We then used retrograde tracers to examine whether differences in neural coding between RSC subregions might reflect differential inputs from the anterior thalamus, which is a prominent source of RSC afferents. We found that gRSC and dRSC differ in their afferents from two AV subfields: dorsomedial (AVDM) and ventrolateral (AVVL). AVVL targets both gRSC and dRSC, while AVDM provides a selective projection to gRSC. These combined results suggest the existence of two distinct but interacting RSC subcircuits: one connecting AVDM to gRSC that may comprise part of the cognitive hippocampal system, and the other connecting AVVL to both RSC regions that may link hippocampal and perceptual regions. We suggest that these subcircuits are distinct to allow for differential weighting during integration of converging sensory and cognitive computations: an integration that may take place in thalamus, RSC or both. Highlights • The two retrosplenial cortex subregions, gRSC and dRSC, differ in their temporal firing characteristics and relation to theta oscillations. • There are differential afferents from the anteroventral thalamic nucleus to gRSC and dRSC, with the dorsomedial subnucleus projecting selectively to gRSC. • The anteroventral thalamus-retrosplenial cortex circuitry thus comprises two functionally and anatomically distinct but connected circuits, differentially connected to the hippocampal system, that may support the integration of cognitive and perceptual information.

subregions, granular (gRSC) and dysgranular (dRSC). The present study investigated how 23 these subregions differ with respect to their electrophysiology and connections, as a step 24 towards understanding their functions. gRSC is more closely connected to the hippocampal 25 system, in which theta-band local field potential oscillations are prominent. We therefore 26 compared theta-rhythmic single-unit activity between the two RSC subregions and found, 27 mostly in gRSC, a subpopulation of non-directional cells with spiking activity strongly 28 entrained by theta oscillations, suggesting a stronger coupling of gRSC to the hippocampal 29 system. We then used retrograde tracers to examine whether differences in neural coding 30 between RSC subregions might reflect differential inputs from the anterior thalamus, which is 31 a prominent source of RSC afferents. We found that gRSC and dRSC differ in their afferents 32 from two AV subfields: dorsomedial (AVDM) and ventrolateral (AVVL). AVVL targets 33 both gRSC and dRSC, while AVDM provides a selective projection to gRSC. These 34 combined results suggest the existence of two distinct but interacting RSC subcircuits: one 35 connecting AVDM to gRSC that may comprise part of the cognitive hippocampal system, 36 and the other connecting AVVL to both RSC regions that may link hippocampal and 37 perceptual regions. We suggest that these subcircuits are distinct to allow for differential 38 weighting during integration of converging sensory and cognitive computations: an 39 integration that may take place in thalamus, RSC or both. 40

Spectral analysis 186
The LFP signal, recorded by low-pass filtering the raw neural signal, was analyzed using a 187 fast Fourier transform (FFT). The power of each frequency component was defined as the 188 square of the magnitude of the output of the FFT. To compare across animals, power 189 measures within each trial were normalized to z-scores and then plotted and smoothed with a 190 Gaussian kernel (bandwidth = 2Hz; standard deviation = 0.5Hz). The power of theta 191 frequency was found as the maximum z-score within the movement-associated (Type-1) theta 192 frequency range (6-12Hz). Average theta frequency was found as the frequency value 193 associated to peak theta power. 194

Theta analysis 195
To investigate spike-LFP coupling, the LFP signal was filtered offline for Type-1 theta using 196 a 4 th order Butterworth filter with band-pass between 6-12Hz. A zero-phase filtering was 197 applied to prevent the phase shift introduced by the Butterworth filter. The Hilbert transform 198 was then applied to the bandpass-filtered LFP to derive instantaneous theta phases (250Hz 199 sampling rate). Each theta cycle was defined such that peaks occurred at 0 • and 360 • , with a 200 trough at 180 • . To determine at which point to apply the Hilbert Transform, the LFP signal 201 (Hz) was linearly interpolated with spike time (s), with each spike being assigned to the phase 202 of the theta cycle at which it occurred. To visualize the spike-LFP relationship, we then 203 double-plotted a spike-phase histogram for each cell. The probability distribution of spikes 204 relative to theta phases was derived by smoothing the spike-phase histogram using a circular 205 kernel density estimation (KDE) method, with an automatically selected bandwidth parameter 206 using the plug-in rule of Taylor (2008). 207 Theta spike-LFP coherence and rhythmicity were assessed by computing an index of theta 208 phase-coupling (IC) and an index of rhythmicity (IR), respectively. Coupling refers to the 209 phase of theta at which spikes occurred, as spikes may be emitted with a timing coupled to 210 theta (phase-locked) even if there is no overt rhythmicity on the autocorrelogram (Eliav et al., 211 2018). We computed an IC by computing the mean vector length (Rayleigh vector, CircStat 212 toolbox) of the spikes vs theta phase distribution (Climer et al., 2015;Frank et al., 2001). For 213 each cell, spike phases were binned between 0 and 360° with 6° bins. For each cell with IC > 214 99 th percentile of shuffle distribution, the preferred theta phase was estimated as the circular 215 mean of all the spike phases, and a Rayleigh test was conducted to assess the significance of 216 this phase preference across cells (see Supplementary materials, Section 1.2). This value was 217 used to sort cells into theta-modulated vs. not. 218 The IR additionally quantifies the temporal modulation of a cell's firing at theta frequency 219 range (6-12Hz). A theta-rhythmic cell will show peaks at ~125ms on its autocorrelogram. 220 Autocorrelograms of the spike trains were plotted between ± 500ms using 10ms bins, 221 normalized to the maximum value and smoothed (20 bins boxcar). Following methods 222 outlined in Lozano et al. (2017), the IR was calculated as the difference between the expected 223 theta modulation-trough (autocorrelogram value between 60-70ms) and the theta-modulation 224 peak (autocorrelogram value between 120-130ms), divided by their sum. It takes values 225 between -1 and 1. 226

Cell identification and inclusion criteria 227
Head direction and bidirectional cells were identified as described in Jacob et al. (2017), and 228 removed from the current analysis, as they were previously included in Jacob et al. (2017), and they were not theta-modulated. Briefly, to analyze the directional firing characteristics of 230 single units, the spike times and position samples of the rat's facing direction were sorted into 231 bins of 6°. The mean firing rates per angular bin (Hz) were smoothed with a 5-bin (30°) 232 smoothing kernel and polar-plotted. Finally, the angles were doubled and plotted modulo 360° 233 in order to remove any bidirectionality that might be present in the tuning curves of 234 bidirectionally tuned cells. 235 To be considered for analysis, cells needed to have (1) a tuning curve that was not obviously 236 unimodal/bidirectional from visual inspection; (2) Rayleigh vector length of the doubled head 237 direction angles < 0.22, which is the original threshold used by Jacob et al. (2017) and (3)  238 average firing rate > 0.1Hz for the trial. 239 To provide a control for each cell against which to measure its theta-phase coupling we 240 created a time-shifted null distribution in which the entire sequence of spike times was 241 circularly time-shifted by a random amount lying between 20s and the duration of the 242 recording session minus 20s. This preserved the number of spikes and the temporal structure 243 of the spiking sequence, but dissociated the time of spiking from the theta oscillations. This 244 was repeated 1000 times for each cell, and the IC was computed for each repetition. All 245 values were pooled together to derive a control distribution, representing the chance level for 246 which the activity of each RSC cell was locked to theta. A cell was considered as significantly 247 modulated at theta frequency if the IC was >99 th percentile of the control distribution for that 248 neuron. 249 To investigate bursting properties of RSC neurons, inter-spike intervals (ISI) were plotted in a 251 frequency histogram between 0-250ms. Cells with peak ISI <10ms, which defines a burst 252 mode (Fanselow et al., 2001), were classified as bursting and the remainder as tonic. The 253 burst index was calculated as the ratio of spikes with peak ISI <10ms to all the spikes in a trial 254 (Mizuseki et al., 2012). Burst index values closer to 1 indicate the presence of intra-burst 255 spikes, whereas smaller values indicate a prevalence of single spikes. 256

Waveform analysis 257
Peak-to-trough waveform width was used to separate narrow from broad spiking cell types 258 (Lewicki, 1998 to the distribution of waveform widths. We defined two cutoffs on this model that divided 262 waveforms into three groups: narrow, broad and in-between. Broad waveforms were those for 263 which the likelihood to be considered broad was larger than 10 times the likelihood to be 264 narrow. Narrow waveforms were those for which the likelihood to be considered narrow was 265 larger than 10 times the likelihood to be broad. Unclassified waveforms were those falling 266 into the dip of the bimodal distribution. The bimodality of the distribution was tested using 267 the calibrated version of the Hartigan dip test (p < 0.01). 268

Statistics 269
For gRSC-dRSC comparisons, if data were found to deviate significantly from a normal 270 distribution (Matlab functions kstest), we used a non-parametric test (Wilcoxon Rank-sum 271 test, WRS, Matlab function ranksum). Otherwise, we used parametric tests. To evaluate 272 simultaneously the effect of two grouping variables (brain region and cell type) on waveform 273 parameters, we used a two-way ANOVA. We also used the Chi-squared goodness-of-fit test 274 (X 2 ) when testing observed numbers of theta-rhythmic cells in gRSC and dRSC against an 275 expected equal proportion, and the number of interneurons against an expected equal 276 proportion. To test the significance between two circular distributions, we employed the 277    Table 1, first row).

377
We then looked at theta-band properties of spiking in single neurons. A total of 642 non-379 directional cells were analyzed from the four rats moving freely in the two-compartment box 380 to determine whether there is differential coupling of gRSC and dRSC with the hippocampal 381 navigational network, reflected in differential theta modulation. Single-cell results were 382 obtained first after pooling cells across animals into a single distribution (pooled analysis), 383 and second, from cells recorded for each animal individually (single subject analysis). Unless 384 otherwise stated, the same results were observed in both analyses and between animals. 385 We first assessed whether cells were theta-modulated by determining their IC. Of the 642 386 non-directional cells identified, 292 cells came from gRSC and 350 from dRSC. Of these 387 cells, half (n = 321/642, 50%) reached the criterion for significant theta modulation, based on 388 the IC being greater than 99 th percentile of the control distribution for that neuron. Table 1  389 reports a summary of recording locations and number of theta-phase-locked cells recorded in 390 each RSC subregion for the four animals. 391

399
Despite the lack of LFP theta differences between gRSC and dRSC recordings discussed 400 above, the distribution of theta-modulated cells varied considerably between RSC subregions, 401 as indicated by the significance of a X 2 test for differences in proportions. Across all animals, 402 67.1% (196/292) of gRSC cells fired in phase with theta, compared to only 35.7% (125/350) 403 of dRSC cells (χ 2 = 62.82; p < 0.0001). Significant differences were confirmed when the data 404 were segregated according to animals (X 2 test, all p < 0.05), except for one animal (R887) for 405 which differences did not reach significance (χ 2 = 0.12; p= 0.73; Table 1). As already noted, 406 for R887 there was under-sampling of gRSC cells (n=15 across 7 trials), possibly due to the 407 tract trajectory passing through white matter tracts before reaching gRSC. We concluded that 408 while the majority of gRSC non-directional cells were theta-modulated, the majority of non-409 directional cells in dRSC fired uncoupled from theta (see pie chart in Figure 6).

Retrograde tracer experiment 564
We next compared, using retrograde tracing, the thalamic afferents to gRSC and dRSC to see 565 if they differ, supporting that these are distinct circuits. The findings have been divided between those tracer injections approximately at the AP levels (-5.5 and -5.8 from bregma) of 567 the electrodes used in the electrophysiological studies and those injections placed more 568 rostrally or caudally within the RSC (see Table 2). Injections were performed in two different 569 laboratories (UCL, n = 6 animals; Cardiff University, n = 17 animals), thus providing an 570 internal replication for our findings (see Supplementary Tables 3-4 for injection 571 coordinates). 572

Injections at the level of recording 581
Thirteen retrograde tracer injections (n=10 animals) targeted RSC at approximately the same 582 AP levels as the recording sites (for representative injections see Figure 9; all injection sites 583 reported in Table 2). In three cases (730, 731 and 878), dual injections were made, one in 584 dRSC (n = 6) and the other in gRSC (n = 7; Table 2 In all cases with injections largely confined to dRSC, retrograde cell label was virtually 593 restricted to the ventrolateral (AVVL) subfield of the AV nucleus (e.g., cases 222#10, 222#9). 594 In contrast, in four of the six cases that targeted gRSC (cases 878, 731, 730, 682), retrograde 595 cell label was observed in both AV subfields, i.e., AVVL and dorsomedial AV (AVDM). The 596 two remaining gRSC cases (077941#13, 077942#14) contained label in both AVVL and 597 AVDM, but the preponderance was in AVVL (Figure 9). 598 The overall pattern of differences (AVVL to dRSC; AVVL and AVDM to gRSC) can be 599 observed in Figure 9, which provides a schematic depicting the AV area with the densest cell 600 body labelling across the various injection cases (label in AD is not shown). In further support 601 of this difference in AV innervation, Figure 10

637
>0.5 mm rostral or caudal to the electrophysiology recording sites ( Table 2; for representative  640 injections see Supplementary Figures 2-3). Of these, six injections were from three cases 641 with dual tracer injections at different AP levels (Supplementary Figure 2). Nine tracer 642 injections seemed restricted to the dRSC, four to the gRSC and three included both RSC 643 subregions (Supplementary Figure 3). These dRSC injections included both superficial and 644 deep layers, except for two cases where the injections appeared restricted to deep dRSC 645 (rostral 227#16, caudal 227#22). Irrespective of AP level, we observed consistent labelling 646 differences after dRSC and gRSC injections (Supplementary Figure 2) that closely matched 647 those differences seen when the injections were placed at the level of the electrodes. Once 648 again, dRSC injections were consistently associated with AV label in AVVL (only 649 occasionally extending marginally into AVDM towards its dorsal border), while gRSC 650 injections resulted in label in both AVVL and AVDM (Figure 9).

651
Further examples of these differential patterns of AV label can be see in Figure 11. This 652 figure contrasts when a tracer injection involved both dRSC and gRSC, versus when it was 653 located in just the dRSC subregion. An injection of CTB involving both RSC subregions 654 resulted in extensive cell labelling across both AVDM and AVVL (case 227#202; Figure 11 227#22 and 227#24, 227#16; Figures 11-12) also supported a 675 previously described topographical pattern (Shibata, 1993): namely, that ventral AV 676 preferentially projects to the rostral RSC whereas dorsal AV targets caudal RSC (see Figure  677 11 C, D). This same gRSC injection (227#24) showed a concentration of AVVL cell label 678 (Figure 12), whereas three other cases with gRSC injections resulted in combined 679 AVVL/AVDM cell label (cases 227#204, 227#202,187#9; as also described above for the 680 gRSC injections close to the AP of the recording sites see Figures 8-9).  projections from one of the main inputs to RSC, the anterior thalamus, to see if there might be 716 differential projections from this region; we found that gRSC was distinguished by a 717 projection from the dorsomedial subfield of the anteroventral thalamus (AVDM) that was 718 much weaker or missing following the dRSC injections. Below, we examine these findings, 719 and then conclude with some speculations about the possible functional role of these two 720 thalamocortical subcircuits. 721

Increased theta coupling in gRSC 722
Comparison of the two subregions (in both pooled and single subject analysis) revealed that 723 gRSC had a much higher mean IC (index of theta phase-coupling), firing frequency and 724 spike-bursting propensity, and theta-modulated cells in this region exhibited considerably larger theta-modulation depth. However, we found no difference in local field potential (LFP) 726 theta oscillations between the two subregions, highlighting a dissociation between theta 727 rhythms in the LFP and oscillations present at the single cell level (i.e., theta rhythms are not 728 generated uniquely from local ensemble activity within gRSC, in line with results of Young and 729 McNaughton, 2009;Borst et al., 1987, Talk et al., 2004. These observations raise two 730 questions: what causes these differences and what, if any, are their functional implications? 731 One possibility is that these signals are a consequence of differential network connectivity 732 between gRSC/dRSC and the medial septal theta "clock." Septal GABAergic neurons express 733 theta-modulated firing and innervate both CA1 and gRSC -but dRSC much less so -with a 734 common projection pattern (Unal et al., 2015). They selectively target and inhibit GABAergic 735 interneurons, resulting in a rhythmic disinhibition of the principal neurons, the activity of 736 which thus becomes phase-locked to septo-hippocampal theta. The second major question arising from the differential theta coupling observed between 763 gRSC and dRSC is what, if any, might be the functional consequences. Theta is widely 764 thought to have the function of synchronizing the excitability of brain regions throughout the 765 hippocampal formation (Korotkova et al., 2018), perhaps to facilitate information transfer or 766 synaptic plasticity. Within hippocampus itself, each theta cycle contains a sequential 767 activation of hippocampal neurons (usually manifesting as place cells) and may function to 768 organize sequential information as part of a spatio-temporal memory system. It may be that 769 gRSC with its greater hippocampal connectivity has the role of mediating between incoming 770 sensory information and ongoing spatial computations, with theta facilitating the transfer and 771 maintaining the spatio-temporal organization. We will return to this point later. 772

4.2
Differential AV projections to gRSC and dRSC 773 Our dataset of retrograde tracer injections adds to previous anatomical descriptions of anterior 774 thalamic inputs to RSC. Overall, in agreement with previous studies (Shibata, 1993), a 775 topography emerged from our tracing studies such that ventral AV projects to rostral RSC 776 while dorsal AV projects to caudal RSC. Within the AV projections, we found a clear 777 differentiation of AVDM and AVVL projections into the gRSC and dRSC. The AVVL 778 subfield targeted both retrosplenial subregions, providing moderate fiber input to both 779 superficial and deep layers (see also Van Groen and Wyss, 1990a, 1992Shibata, 1993). 780 In contrast, the AVDM subfield provided a selective projection to the gRSC; mainly to the 781 superficial layers (Shibata, 1993). Overall, the present results argue for the presence of two 782 related but distinct AV-RSC subcircuits: AVVL to dRSC, AVVL and AVDM to gRSC. 783 Two anatomical facts are relevant for clarifying the observed variations in the 784 electrophysiology data between gRSC and dRSC. First, the difference in the overall 785 projection strength to the two RSC subregions should be noted, in addition to the 786 topographical organization of AV projections from within AV to each of these subregions 787 (Shibata et al., 1993). In relation to the first point, our data demonstrate stronger AV 788 projections to gRSC compared to dRSC (see also Van Groen and Wyss, 2003). Second, there 789 are differences in the source of these projections from different AV subfields. Superficial 790 gRSC layers receive a distinct projection from AVDM, which sends no or only moderate 791 projections to dRSC. In contrast, dRSC only receives widespread AVVL projections. Hence, variations in AV fiber density (stronger AV to gRSC) but may also be explained by 794 considering anatomically distinct cell populations within the AV projecting to the two RSC 795 subregions. 796

4.3
Two interacting thalamocortical subcircuits 797 Previous behavioral, connectivity and neuroanatomical evidence, together with our current 798 findings, combine to indicate that the RSC is not a unitary structure (see Aggleton et al., this 799 issue, for review). gRSC is more strongly coupled with the hippocampal formation, as 800 evidenced by both anatomical connectivity and theta modulation: the latter of which accords 801 with the fact that theta is a distinguishing feature of neurons in brain regions that are linked to 802 the hippocampal spatial system (Burgess et al., 2002;Buzsáki, 2002Buzsáki, , 2005 interacting head direction signal. The question thus arises: why does the brain maintain these 813 separable subsystems? 814 that may agree or differ in their content, and thus allow for comparison and appropriate error 816 correction. The streams in RSC comprise interoceptive signals arising from subcortical 817 regions and exteroceptive signals coming in from the sensory periphery. A specific example 818 of these converging information streams is provided by the RSC head direction signal, which 819 arises from convergence of self-motion inputs (primarily interoceptive cues such as vestibular 820 signals, motor efference and proprioception) and static environmental information from the 821 sensory cortices (Yoder and Taube

Conclusion and future directions 832
Our results detail differences in theta-related spiking between gRSC and dRSC along with the 833 existence of two interdependent AV-RSC subnetworks, suggesting two distinct streams of 834 information to the RSC. We propose that these conform to a more interoceptive stream that 835 conveys the results of internal processing and a more exteroceptive stream that conveys the state of the external world as assessed by the senses. The cognitive stream is characterized by 837 close connectivity with the hippocampal system, as well as a unique input from the AVDM.