Organization of cortical and thalamic input to inhibitory neurons in mouse motor cortex

Intracortical inhibition in motor cortex (M1) regulates movement and motor learning. If inhibitory cell types and cortical laminae targeted by cortical and thalamic afferents differ, then these afferents play different roles in regulating M1 output. We quantified input to two classes of M1 interneurons, parvalbumin+ (PV) fast-spiking cells and somatostatin+ (SOM) low-threshold-spiking cells, using monosynaptic rabies tracing. We then compared anatomical connectivity and functional connectivity based on synaptic strength from sensory cortex and thalamus. Functionally, each input innervated M1 interneurons with a unique laminar profile. Different interneuron types were excited in a distinct, complementary manner, suggesting feedforward inhibition proceeds selectively via distinct circuits. Specifically, somatosensory cortex (S1) inputs primarily targeted PV+ neurons in upper layers (L2/3) but SOM+ neurons in middle layers (L5). Somatosensory thalamus (PO) inputs targeted PV+ neurons in middle layers (L5). In contrast to sensory cortical areas, thalamic input to SOM+ neurons was equivalent to PV+ neurons. Thus, long-range excitatory inputs target inhibitory neurons in an area and cell type-specific manner which contrasts with input to neighboring pyramidal cells. In contrast to feedforward inhibition providing generic inhibitory tone in cortex, circuits are selectively organized to recruit inhibition matched to incoming excitatory circuits.

(L5). Somatosensory thalamus (PO) inputs targeted PV+ neurons in middle layers (L5). In 23 contrast to sensory cortical areas, thalamic input to SOM+ neurons was equivalent to PV+ 24 neurons. Thus, long-range excitatory inputs target inhibitory neurons in an area and cell type-25 specific manner which contrasts with input to neighboring pyramidal cells. In contrast to 26 feedforward inhibition providing generic inhibitory tone in cortex, circuits are selectively 27 organized to recruit inhibition matched to incoming excitatory circuits. 28

INTRODUCTION
overall pattern of input to specific interneuron types using monosynaptic rabies tracing. Then, 67 because this method is not layer specific, we measured monosynaptic excitation to M1 68 interneurons functionally using channelrhodopsin-2 (ChR2)-assisted circuit mapping 69 (Cruikshank et al., 2010;Petreanu et al., 2007;Petreanu et al., 2009) to assess synaptic input 70 strength as a function of laminar depth and specific interneuron type across layers in M1. In 71 contrast to the strong preference of thalamocortical input for PV+ interneurons in layer 4 of 72 sensory cortex (Cruikshank et al., 2007;Cruikshank et al., 2010), both methods found thalamic 73 input to SOM+ neurons that was nearly as strong as PV+ neuron input. The laminar pattern, 74 however, differed. Corticocortical afferents from S1 also targeted interneurons differently across 75 cell types: among PV+ neurons, L2/3 cells were most strongly excited, while for SOM+ neurons, 76 input arrived strongly in middle layers (L5). Furthermore, comparing inputs from S1 and 77 thalamus, PV+ neurons were targeted in a complementary fashion, with S1 exciting upper 78 layers (L2/3) and thalamus targeting middle layers (L5). Because S1 and PO targeted excitatory 79 neurons in similar layers (Hooks et al., 2015;Hooks et al., 2013), the complementary pattern of 80 input to PV+ neurons suggests that feedforward inhibition does not simply silence cortex, but 81 integrates into local circuits in a specific fashion. 82

85
Anatomical tracing of long-range inputs to PV+ and SOM+ neurons 86 Incoming cortical and thalamic excitation to M1 directly excites pyramidal neurons 87 EGFP-B19G to label potential starter neurons and make them express both TVA receptor (to 100 become infected with EnvA pseudotyped rabies virus) as well as the rabies coat protein B19G 101 (Fig. 1f). The starter cells in these injections spanned cortical laminae (L2-6). After two weeks to 102 6 119 brains and that the retrograde tracing method was reliable. To minimize excessive comparisons, 120 we limited comparisons to brain regions where at least one area exceeded a detection threshold 121 (dotted line in Fig. 1o-p). Overall, the pattern was quite similar for inputs to PV+ and SOM+ 122 neurons in M1. Most presynaptic inputs were found in cortex (Fig. 1o), in particular frontal and 123 motor areas on the dorsal surface of cortex, as well as somatosensory cortex, which is strongly 124 reciprocally connected to mouse M1 (Hooks, 2017;Mao et al., 2011). About 20% of inputs for 125 both interneuron classes were localized to thalamus, consistent with similar corticocortical and 126 thalamocortical input to these cell types. We further examined the thalamocortical data by 127 subdividing labeled neurons by thalamic nuclei (Fig. 1p). Although the data showed that the 128 thalamic nuclei associated with M1 based on prior anatomic studies (Deschenes et al., 1998;129 Kuramoto et al., 2009;Kuramoto et al., 2015;Ohno et al., 2012) were robustly represented, 130 including VAVL, VM, and PO, we did not detect significant differences between number of 131 neurons projecting to PV+ or SOM+ neurons. The ability to detect such differences was limited 132 by correction for false detection rate (Benjamini et al., 2009), with non-significant trends in VM 133 and two medial nuclei (PC and PF). 134 We extended the retrograde viral tracing study to two other cortical areas strongly 135 reciprocally connected with motor areas, primary somatosensory cortex and frontal cortex (M2). 136 Experimental procedures were identical, except for the targeting of viral injections to different 137 cortical areas in PV-Cre and SOM-Cre mouse lines. This gave a richer dataset for comparison. 138 We first made comparisons across cortical areas of starter cell within a given interneuron line 139 ( Fig. 2b-d). For SOM+ neurons, the major areas labeled were motor cortex and frontal cortex 140 (greater in fM1 and M2 starter cases than in vS1 starter cases) and somatosensory cortex 141 (greater in vS1 starter cases), with a relatively uniform fraction of thalamic input (~20%) across 142 neocortical areas). PV+ input was similar, with the additional note of significantly greater input 143 from limbic areas (mostly orbital cortex) to M2 than to other areas. Comparison of input from 144 specific thalamic nuclei was considerably different across injection sites. In general, the pattern 145 of thalamic label could be used to infer the cortical target, with a reasonably large number of 146 nuclei providing a specific pattern of cortical input. For example, in cases using PV+ starter 147 neurons, vS1 injections could be identified by more robust VPM and PO labeling. fM1 injections 148 had strong input from VAVL and VM as well as PO and midline nuclei. M2 input was similar to 149 fM1 in receiving strong VAVL and VM input, but also received significant MD input (which vS1 150 and fM1 did not) as well as input from midline nuclei including PC and PF. This is consistent 151 with substantial differences in thalamocortical targeting across cortical areas. However, in 152 comparing input to different cell types within a given cortical area, few significant differences 153 were detected. For M2, PV+ interneurons labeled more thalamic cells in VM and PC compared 154 to SOM+ neurons in M2. In vS1, SOM+ interneurons labeled more VAVL neurons than PV+ 155 neurons. For most nuclei, there were no detectable differences. In general, this was consistent 156 with much larger differences in thalamic label for cortical areas than for specific cell types within 157 a cortical area. 158 Synaptic mapping of long-range excitatory inputs from S1 and thalamus to M1 160 Because SOM+ neurons comprise a diverse range of neurons, including Martinotti and 267 non-Martinotti subsets (Tremblay et al., 2016), we were curious whether the input pattern to the 268 SOM+ neurons labeled in our SOM-Cre x Ai14 strategy would be able to be replicated in a more 269 targeted subset of cells (Huang et al., 2016;Tasic et al., 2018). We labeled the Tacr1+ subset 270 of interneurons using a NK1R-CreER mouse crossed to a tdTomato reporter. Like SOM+ and 271 PV+ crosses, this resulted in labeled neurons across the cortical thickness. We repeated our 272 input mapping experiments in these mice and presented the data in a similar format (Fig. 8h). In 273 contrast to all SOM+ and PV+ neurons, Tacr1+ cells responded well to S1 inputs across 274 multiple layers. Of note, the strongest responses were in in L2/3 and L6, with weaker input to L5 275 ( Fig. 8f-h). The input from PO also did not correspond to the SOM+ pattern, with strongly 276 responding neurons in L5A and L5B. 277 Because our inputs to PV+ and SOM+ neurons were quantified as normalized input 278 strength, we wondered whether the absolute strength of input to PV+ and SOM+ differed across 279 pathways. We compared input across pathways and cell types by plotting the absolute input 280 strength for all cells on the same scale, as other investigators have done (Kinnischtzke et al., 281 2014). This was done by summing the EPSC amplitudes across the points of the input map (as 282 in Fig. 3l). Cell vector plots were presented as before (Fig. 8g) using the same logarithmic scale 283 for all plots to improve presentation of more frequent recordings with weaker inputs. In general, 284 excitatory input from S1 and PO to either interneuron type (PV+ or SOM+, top and middle row) 285 was of similar magnitude. Averaging input by layer confirmed that monosynaptic input was of 286 comparable strength (Fig. 8i). The overall pattern is similar to our normalized data. The 287 distribution of S1 and PO input to PV+ is nearly overlapping, with the exception of input to L2/3 288 PV+ neurons. Input to SOM+ neurons is of similar strength from S1 and PO except for L5A, 289 where the absence of thalamic input is pronounced. Corticocortical input is similar or stronger to 290 SOM+ neurons than PV+ neurons at each laminar depth except L2/3. Thalamocortical input is 291 stronger to L5A PV+ neurons than to SOM+ cells. Outside of this layer, however, input strength 292 to SOM+ cells is equal to PV+ neurons. 293

Direct comparison of synaptic and anatomical connectivity 294
By performing comparable and anatomical experiments, we sought to make a 295 quantitative connection between input to PV+ and SOM+ neurons as measured by these distinct 296 methods. Differences in how the data are quantified presented challenges. First, because the 297 presynaptic neurons labeled by monosynaptic tracing might connect to a starter cell in any 298 cortical layer, we considered our anatomical tracing data to potentially present an averaged 299 view of connectivity to specific interneurons across all cortical layers. To make this data 300 comparable to our physiological data where target PV+ and SOM+ neurons were more explicitly 301 identified by laminar position, we sought to derive an average connection strength to PV+ or 302 SOM+ neurons by converting the layer-by-layer strength of synaptic connectivity (Fig. 9a,   These measurements might not reflect the relative contribution of the two interneuronal 325 populations to a train of stimuli, which might be more physiological. Since PV+ and SOM+ 326 neurons show differences in short term plasticity of local excitatory inputs (Beierlein et al.,327 2003), we were interested to test whether later synaptic inputs in a train might alter which 328 interneuron population was more likely to carry feedforward inhibition.
In contrast to these results in PV+ neurons, SOM+ cells showed some degree of 342 facilitation for both classes of input. S1 input was more facilitating at 20 Hz than 10 Hz. 343 Similarly, PO input was also more facilitating at 20 and 40 Hz. These experiments were 344 performed using the same viral vector and recording conditions for both mouse lines, 345 suggesting that the short-term plasticity observed differed between postsynaptic targets and 346 was not due exclusively to an artifact of measuring short term plasticity with optical methods 347 inhibition proceeds through distinct interneurons (of either the same or different type) for inputs 408 recruiting motor cortex. Our data indicates that there is space for selectivity in inhibitory circuitry.
Some specific examples of differences stand out in our quantitative maps. Our data 410 suggests that the incoming sensory information from S1 excites PV+ neurons in L2/3 but SOM+ 411 neurons in L5. Since these same afferents from S1 also strongly target pyramidal neurons in 412 L2/3 and L5A (Mao et al., 2011), it suggests that, in part, the feedforward inhibition recruited to 413 L2/3 pyramids and L5A pyramids can proceed through different neural populations, as cortical 414 interneurons generally, though not exclusively, inhibit pyramidal neurons in the same layer 415 (Katzel et al., 2011). Since SOM+ neurons may also inhibit PV+ cells, helping to disinhibit cortex 416 (Pfeffer et al., 2013), it is possible that S1 input can have differential inputs on the excitability of 417 L2/3 and L5A pyramids, despite similarity in monosynaptic excitation to these targets. This 418 suggests that differences in cell types play a role in determining the efficacy of feedforward input 419 and the rules of long-term plasticity in these two target layers. The pattern of thalamic input to 420 PV+ interneurons also differed from S1 input. This suggests that where cortical and thalamic 421 afferents target the same excitatory network, to the extent S1 and PO target partially non- thalamic input strength to SOM+ interneurons was marginally higher than that to PV+ neurons 449 except for L5A (Fig. 8h middle right). However, there is considerable variability between 450 individual cells (Fig. 8g). Similarly, corticocortical input from S1 was higher to SOM+ neurons in 451 L5A and L5B (Fig. 8h top right). This collectively suggests that there is no uniform rule that 452 thalamic input will be selective for PV+ neurons across all cortical areas and layers. 453 In addition to monosynaptic input strength, it is also worth considering that input from S1 454 and PO will not occur in isolation but as part of trains of action potentials, especially should 455 thalamus fire in burst mode (Sherman, 2001;Steriade et al., 1993). In the case of PO, short 456 term depression of input to PV+ neurons is the strongest short term plasticity we observed (Fig.  457 10). In contrast, these inputs were facilitating to SOM+ neurons. This suggests that as rapidly 458 as the second or third pulse in a train, the recruitment of feedforward inhibition could shift from 459 PV+ to SOM+ neurons (Tan et al., 2008). The difference in short term facilitation and 460 depression to SOM+ and PV+ neurons respectively is not as large for S1 afferents as for PO 461 inputs, though the general pattern is the same. This effect is qualified by differences in the 462 intrinsic excitability of the particular cells involved, as well as their own depression onto 463 pyramidal neuron targets. Inhibition from SOM+ neurons is hypothesized to be facilitating, in 464 contrast to the depression observed in fast spiking neurons (Beierlein et al., 2003). If this is the 465 case in M1, then this would serve to further shift the balance of disynaptic feedforward inhibition 466 from PV+ to SOM+ neurons, giving SOM+ neurons the potential to play a role in regulating 467 excitability and plasticity (Chen et al., 2015). 468

Comparison of anatomical and functional methods for mapping microcircuitry 469
The emergence of techniques for recording, imaging, and reconstructing the whole brain help clarify how neural connectivity contributes to brain function. 480 Here we directly compared synaptic input and retrograde anatomical tracing of 481 presynaptic input to M1 interneurons. Based on prior work, we hypothesized that 482 thalamocortical input would favor PV+ interneurons. In S1, thalamic inputs to fast spiking 483 (presumed PV+) neurons were more likely to be connected (Bruno and Simons, 2002) and were 484 vastly stronger by EPSC amplitude, as much as 10-20x in L4 and L5/6 ( Fig. 6, (Cruikshank et  485   al., 2010)) as well as later 2/3 (Fig. 2, (Naskar et al., 2021)). In M1 by contrast, both synaptic 486 connectivity mapping experiments and retrograde viral tracing data showed strong 487 thalamocortical connections to PV+ and SOM+ neurons in M1 (Figs. 1 and 6-9). Indeed, input to 488 PV+ neurons was only ~10-30% stronger than that to SOM+ neurons (Fig. 9). This difference 489 was not significant in the anatomy data, though non-significant trends (Fig. 1) were present in 490 three thalamic nuclei not assessed functionally (VM, PC, and PF). Thus, both measures of 491 anatomical and functional connectivity were consistent in demonstrating distinct thalamocortical 492 connectivity rules for primary motor cortex compared to sensory cortex. Similarly, our measures 493 of corticocortical input to SOM+ neurons were comparable between anatomical and functional 494 methods (Fig. 9), though the functional method showed a slight preference for SOM+ neurons 495 (because of stronger input to the more numerous interneuron population in L5A and L5B), while 496 the anatomical method showed a nonsignificant trend for larger S1 input to PV+ neurons (Fig.  497 1). This is in contrast to corticocortical input in S1, where long range input to PV+ neurons was 498 stronger in L2/3 for multiple pathways (Naskar et al., 2021). 499 However, there are some limitations to the comparison. With anatomical methods, it was 500 not possible to label SOM+ or PV+ interneurons in specific layers and show differences in 501 connectivity as was the case in functional mapping experiments. Furthermore, quantifying 502 differences in synaptic strength is not precisely the same as labeled presynaptic neurons. 503 Although we hypothesize that stronger synaptic connections would result in more points of 504 contact between corticocortical or thalamocortical axons and rabies-infected starter cells, there 505 are not control experiments to show that increased synaptic strength results in an increase in 506 labeled presynaptic neurons. Furthermore, while we can quantify synaptic strength to PV+ or 507 SOM+ neurons from S1 or PO inputs in a given layer, it is not certain that deriving a single value 508 for connection strength for the whole population is the strongest means to make such a 509 comparison.
How reliable is retrograde labeling for anatomically assessing connectivity? The data 511 presented here suggest anatomical and functional methods can predict similar connectivity. 512 First, the retrograde tracing data provides strong support for retrograde tracing as an effective 513 measure of area-to-area connectivity. Tracing from starter cells in each cortical area specifically 514 labeled cortical areas and a pattern of thalamic nuclei unique to that cortical area and 515 reproducibly across replicates (Fig. 2b-e). For example, M2 gets input from limbic areas while 516 M1 and S1 do not. Furthermore, even small brain areas, such as individual thalamic nuclei, 517 have distinct patterns of label for each of the three cortical areas studies. However, because the 518 specific thalamic labeling was almost identical for input to PV+ and SOM+ interneurons for the 519 regions that we studied, evidence that retrograde tracing varies strongly with postsynaptic target 520 would be better to test in cell types with vastly different connectivity, especially where afferents 521 avoid innervating one cell type while connecting strongly to another intermingled type. 522

Subcellular organization of input to interneurons 523
Our circuit mapping approach gives some insight into the subcellular localization of 524 synaptic input in the dendritic arbor of postsynaptic neurons. This insight is constrained by 525 methodological limitations of patch clamp which limit our ability to study distal dendrites. 526 Interneurons, however, are generally more electronically compact than pyramidal neurons. 527 There is some organization of the input around the PV+ and SOM+ soma, suggesting some 528 mechanisms for afferent targeting exist and may differ between interneuron subtypes. S1 input 529 to PV+ neurons is evenly distributed perisomatically, while these same afferents tend to target 530 dendrites ~50-150 µm below the soma of SOM+ cells. Similarly, PO input to PV+ cells is also 531 shifted deep to the soma center by ~50-150 µm. This offset is interesting in contrast to thalamic 532 inputs to pyramidal neurons (Petreanu et al., 2009). In these cells, PO input targets the apical 533 dendrites, and especially the L1 arbors of L3 pyramidal neurons but is soma-centered (and quite 534 strong) for L5A pyramidal neurons. This is consistent with each cell type directing input from a 535 defined presynaptic pathway to some extent within its arbor. 536

Different SOM+ subtypes have different connectivity 537
One difficulty in making a general statement about SOM+ connectivity in M1 was the 538 variability within this dataset, especially for thalamic inputs (Fig. 7). In our assessment of 539 monosynaptic input strength and short-term plasticity, PV+ neurons behaved in a much more When synaptic strength was mapped, this presented a pattern of input strength quite in 555 contrast to that of the SOM-Cre x Ai14 mice (Fig. 8h-i). In particular, S1 input had been 556 distributed across L2/3, L5A, and L5B to SOM+ neurons, but most strongly to L5A. In contrast, 557 this subset had a quite distinct pattern, with substantial L2/3 and L6 input. PO input was noisy 558 to SOM+ neurons in general. This emphasized the likelihood that each molecular subtype of 559 interneuron had distinct mechanisms to govern synaptic strength. 560

Conclusion 561
Overall, these results point towards circuit mechanisms by which different pathways of 562 incoming excitation may excite specific subcircuits, including specific networks for inhibition in 563 primary motor cortex. Having some means by which recruitment of feedforward inhibition from 564 each pathway does not generally silence all M1 circuits is perhaps necessary for control of 565 movement when M1 is receiving competing inputs from different brain regions. 566 567

Stereotactic injections. 569
Animal protocols were approved by Institutional Animal Care and Use Committee at University 570 of Pittsburgh (Protocol #20118160) and followed the recommendations in the Guide for the Care 571 and Use of Laboratory Animals of the National Institutes of Health. Mice of either sex were 572 used. Experimental procedures were similar to previous studies (Hooks et al., 2013). Cre driver 573 mouse lines (PV-Cre, SOM-Cre, and NK1R-CreER) were used in conjunction with the lsl-574 tdTomato reporter line, Ai14, to label specific interneuron populations (see Table 2). Tacr1 575 experiments using NK1R-CreER, required tamoxifen injections to drive recombination. Animals 576 were anesthetized using isoflurane and placed in a custom stereotactic apparatus. Juvenile included CAG-ChR2-mVenus and hSyn-Chronos-GFP. AAV-CaMKIIa-hChR2(H134R)-EYFP 579 was used in Tacr1 mapping experiments (see Table 3 for details and serotypes). Injections were 580 made with glass pipettes (Drummond) using a custom-made positive displacement system 581 (Narashige). Stereotactic coordinates are listed in Table 1. A pair of injections (50-100 nL; 500 582 and 800 µm depth from pia) was made in the cortex. One location (50-100 nL) was injected into 583 thalamic targets. using NeuroInfo software (MBF Bioscience). Somata were detected by NeuroInfo using an 597 artificial neural network. Their coordinates in the Allen CCF used to quantify in which structure 598 the neuron was detected using custom Matlab software. Presynaptic labeled (tdTomato+) 599 neurons in each structure were quantified as a fraction of total labeled neurons in that brain to 600 normalized for differences between mice and injections. For comparisons between starter cell 601 types (PV+ or Cre+) or injection sites (M2, fM1, vS1), t-tests were performed. To reduce the 602 number of comparisons to a manageable size, brain regions were only included in the analysis if 603 the fraction of presynaptic neurons labeled exceeded a threshold value: 2% for 24 large brain 604 areas ( Fig. 1o and Fig. 2bd) or 0.5% for 33 smaller thalamic nuclei (Fig. 1p and 2c and 2e-h). 605 The two-stage procedure of Benjamini, Krieger, & Yekutieli (Benjamini et al., 2009) was used to 606 control for the false discovery rate (FDR). Black * (with p-values noted) mark significant 607 differences between presynaptic labeling. Gray * used for comparisons not significant after 608 correcting for FDR. 609 Electrophysiology and photostimulation. Brain slices were prepared >14 days after viral chloride, 3.1 sodium pyruvate, 11.6 sodium ascorbate, 25 NaHCO3, 25 D-glucose, 7 MgCl2, 2.5 613 KCl, 1.25 NaH2PO4, 0.5 CaCl2). Off-coronal sections (300 µm) of M1 were cut using a vibratome 614 (VT1200S, Leica). Additional sections were cut to confirm injection location. Slices were 615 incubated at 37 ºC in oxygenated ACSF (in mM: 127 NaCl, 25 NaHCO3, 25 D-glucose, 2.5 KCl, 616 2 CaCl2, 1 MgCl2, 1.25 NaH2PO4,) for >30 min and maintained at room temperature (22 ºC) 617 thereafter. Electrophysiology data were low pass filtered (1 kHz). EPSCs were detected with a threshold of 647 >6x standard deviation from baseline). Mean EPSC amplitude for input maps was averaged 648 over 75 ms post-stimulus (reported in pA). For comparison between cells in the same slice, total 649 input was computed by summing supratheshold pixels. Paired comparisons across cells use the 650 nonparametric Wilcoxon signed-rank test. Mean ratio of excitation were calculated as geometric 651 means, with the strongest input layer shown as 1.0. In these comparisons, the bar graph was 652 overlaid with a line showing the bootstrap mean and SD (10000 replicates) resampling the 653 paired data. To compare the axonal profile of S1 and PO axons (Fig. 1), we quantified the 654 fluorescence as before (Hooks et al., 2013)    showing retrograde label in S1 and thalamus from M1 starter cells in a SOM-Cre mouse. Note 985 S1 label and thalamic label.