Cortical glutamatergic projection neuron types contribute to distinct functional subnetworks

The cellular basis of cerebral cortex functional architecture remains not well understood. A major challenge is to monitor and decipher neural network dynamics across broad cortical areas yet with projection-neuron-type resolution in real time during behavior. Combining genetic targeting and wide-field imaging, we monitored activity dynamics of subcortical-projecting (PTFezf2) and intratelencephalic-projecting (ITPlxnD1) types across dorsal cortex of mice during different brain states and behaviors. ITPlxnD1 and PTFezf2 neurons showed distinct activation patterns during wakeful resting, during spontaneous movements and upon sensory stimulation. Distinct ITPlxnD1 and PTFezf2 subnetworks were dynamically tuned to different sensorimotor components of a naturalistic feeding behavior, and optogenetic inhibition of ITsPlxnD1 and PTsFezf2 in subnetwork nodes disrupted distinct components of this behavior. Lastly, ITPlxnD1 and PTFezf2 projection patterns are consistent with their subnetwork activation patterns. Our results show that, in addition to the concept of columnar organization, dynamic areal and projection-neuron-type specific subnetworks are a key feature of cortical functional architecture linking microcircuit components with global brain networks. This study shows that pyramidal tract (PT) and intratelencephalic (IT) projection neurons process information via distinct parallel subnetworks across cortex, each preferentially associated with either motor events or sensorimotor transformation.


Fig. 1 | Distinct activity patterns of IT
and PT Fezf2 neurons during wakeful resting and upon sensory input. a, STP images of GCaMP6f-labeled PT Fezf2 and IT PlxnD1 neurons across dorsal cortex. Arrow indicates anterior-posterior axis. Yellow text indicates approximate location of layer 2/3 (L2/3), layer 5a (L5a), layer 5b (L5b) and layer 6 (L6). Sagittal schematic depicts major projection patterns of IT and PT. b, mRNA in situ images of Fezf2 + (left) and PlexinD1 + (middle) cells. Double in situ overlaid (right) shows Satb2 + (red) and PlexinD1 + (green). PlexinD1 + cells represent a subset of Satb2 + IT cells. c, Example z-scored variance of behavior from video recordings (black trace) and corresponding variance of neural activity from IT PlxnD1 (blue) and PT Fezf2 (green) neurons. Gray blocks indicate active episodes. d, Average variance maps of spontaneous activity during active (right) and quiescent (left) episodes (n = 12 sessions from six mice). e,f, Distribution of percentage of cross-validated IT PlxnD1 and PT Fezf2 activity variance explained by full-frame behavior variance (e) and specific body part (f) from encoding model (n = 12 sessions from six mice). g, Illustration of unimodal sensory stimulation paradigm. h, Mean activity maps of IT PlxnD1 and PT Fezf2 neurons in response to corresponding sensory simulation (average of 12 sessions from six mice). i, Single-trial IT PlxnD1 and PT Fezf2 activity within orofacial (yellow), whisker (red) and visual (purple) areas during orofacial (os), whisker (ws) and visual (vs) stimulation. j, Single-trial heat maps of IT PlxnD1 and PT Fezf2 activity from whisker (bc), orofacial (oc) and visual cortex (vc) in response to corresponding sensory stimulus from one example mouse for each cell type. k, Mean activity of IT PlxnD1 and PT Fezf2 neurons in whisker, orofacial and visual cortex during corresponding sensory stimulus (n = 240 trials in 12 sessions from six mice; shaded region indicates ±2 s.e.m). l, Distribution of IT PlxnD1 and PT Fezf2 activity intensity in whisker, orofacial and visual cortex during corresponding sensory stimulus (n = 240 trials in 12 sessions from six mice). *P < 0.05 and ***P < 0.0005. For box plots, central mark indicates median; bottom and top edges indicate 25th and 75th percentiles; and whiskers extend to extreme points excluding outliers (1.5 times above or below the interquartile range). All statistics are provided in Supplementary Table 1. a.u., arbitrary units; AUC, area under the curve. Article https://doi.org/10.1038/s41593-022-01244-w clusters across both episodes, substantiating distinct cortical activation patterns between IT PlxnD1 and PT Fezf2 (Extended Data Fig. 1f,g). Spatial maps of 75th and 95th percentile activity at each pixel indicated the magnitude of activation to be similar with variance maps across episodes (Extended Data Fig. 1h).
To investigate the correlation between neural activity and spontaneous movements, we built a linear encoding model using the top 200 singular value decomposition (SVD) temporal components of the behavior video as independent variables to explain the top 200 SVD temporal components of neural activity ( Supplementary  Fig. 2a,b). The top 200 components explained more than 85% of neural and behavior variance in both populations with no significant difference ( Supplementary Fig. 2a). Quantifying neural variance explained by spontaneous movements revealed PT Fezf2 activity to be more strongly associated with spontaneous movements compared to IT PlxnD1 (Fig. 1e and Supplementary Methods). Furthermore, forelimb movements contributed significantly more toward PT Fezf2 variance, whereas other movements contributed equivalently between the two populations ( Fig. 1f, Supplementary Fig. 2c and Supplementary Methods).
To evaluate whether difference in PT Fezf2 and IT PlxnD1 responses could be explained by their difference in GCaMP6f expression, we compared the distribution of ΔF/F values from all pixels and peak ΔF/F value at each pixel for every session during spontaneous behavior between the two PNs and found a similar distribution range for both  Fig. 2a,c). To verify if signal correction resulted in similar removal of artifacts, we performed ΔF/F measurements with and without hemodynamic corrections from the hindlimb sensory area aligned to the onset of spontaneous movements (Extended Data Fig. 2d). The peak difference and correlation between corrected and uncorrected signals were similar between IT PlxnD1 and PT Fezf2 (Extended Data Fig. 2e). Together, these results demonstrate distinct activation patterns of IT PlxnD1 and PT Fezf2 during wakeful resting state, with preferential PT Fezf2 activation associated with spontaneous movements.

Sensory inputs preferentially activate IT PlxnD1 over PT Fezf2
We next investigated IT PlxnD1 and PT Fezf2 activation after sensory input of the somatosensory and visual system ( Fig. 1g and Methods). Light stimulation and tactile stimulation of whisker and orofacial region strongly activated IT PlxnD1 in primary visual, whisker and mouth-nose somatosensory cortex, respectively, but resulted in no or weak activation of PT Fezf2 in those cortical areas (Fig. 1h). On comparing temporal dynamics from centers of peak activation (Methods), we found strong IT PlxnD1 but weak PT Fezf2 activation in primary sensory cortices in response to whisker, orofacial and visual stimulus (Fig. 1j,k). Comparing activity intensities validated significantly higher activity in IT PlxnD1 compared to PT Fezf2 (Fig. 1l). Peak-normalized maps further revealed activation in corresponding sensory cortices in PT Fezf2 along with broader activation including retrosplenial areas (Extended Data Fig. 2f). Stronger correlations between IT PlxnD1 spatial maps indicate reliability in activation pattern compared to PT Fez2 (Extended Data Fig. 2g). Considering that IT PlxnD1 neurons constitute a major subpopulation of ITs (Fig. 1b), these results provide the first set of in vivo evidence that sensory inputs predominantly activate IT compared to PT PNs, consistent with previous findings that thalamic input predominantly impinges on IT but not PT cells 38 . Notably, IT PlxnD1 activation per se does not lead to significant PT activation at the population level despite demonstrated synaptic connectivity from IT to PT PNs in cortical slice preparations 39 . It is possible that IT PlxnD1 to PT Fezf2 synaptic efficacy is modulated by brain states or that another IT subpopulation might more directly activate PT Fezf2 .
To confirm that wide-field responses reflected calcium dynamics at cellular resolution, we used two-photon imaging to measure responses from single IT PlxnD1 and PT Fezf2 neurons from barrel cortex upon whisker stimulation (Extended Data Fig. 3a and Methods). We recorded from cell bodies of IT PlxnD1 and apical dendrites of PT Fezf2 (dendritic calcium activity in layer 5B is strongly correlated to cell body dynamics 40-44 ; Extended Data Fig. 3b) and used a linear modeling approach to classify neurons as activated, inhibited or unclassified groups (Methods). We first measured the average response for each group of neurons from a single field of view (FOV) and found that, within the activated group, IT PlxnD1 neurons showed significantly higher response compared to the PT Fezf2 group (Extended Data Fig. 3c,d). Average response of all neurons combined from a single FOV resulted in IT PlxnD1 displaying significantly larger response compared to PT Fezf2 (Extended Data Fig. 3e). Combining neuronal responses from all mice across all FOVs resulted in similar response characteristics (Extended Data Fig. 3f-h). These responses followed dynamics very similar to those observed using wide-field imaging during whisker stimulation (Fig. 1k) Fig. 3i). These results show that the response properties in wide-field imaging of IT PlxnD1 and PT Fezf2 neurons closely reflect their cellular resolution dynamics.

PT Fezf2 and IT PlxnD1 neurons are tuned to distinct sensorimotor features
To examine the activation patterns of IT PlxnD1 and PT Fezf2 neurons during sensorimotor processing, we designed a head-fixed mouse feeding behavior. In this setup, mice sense a food pellet approaching on a moving belt, retrieve the pellet into the mouth by licking, recruit both hands to hold the pellet and initiate repeated bouts of handmouth-coordinated eating movements that include bite while handling the pellet; transfer pellet to hands while chewing; and raise hands to bring pellet to mouth, thereby starting the next bout ( Fig. 2a and Supplementary Video 3). We used DeepLabCut 45 to track pellet and body parts in video recordings and wrote custom algorithms to identify major events in successive phases of this behavior (Methods, Fig. 2b, Supplementary Fig. 3a and Supplementary Video 3).
We imaged the spatiotemporal activation patterns of IT PlxnD1 and PT Fezf2 neurons across dorsal cortex while mice engage in the behavior (Supplementary Videos 4 and 5). We calculated average GCaMP6f signals per pixel in frames taken at multiple timepoints centered around when mice retrieve pellet into the mouth (pellet in mouth (PIM)) ( Fig. 2c). Upon sensing the pellet approaching from the right side, mice adjust their postures and hand grip of the support bar while initiating multiple right-directed licks until retrieving the pellet into the mouth. During this period (Fig. 2c), IT PlxnD1 was first activated in left whisker primary sensory cortex (SS bfd ), which then spread to bilateral forelimb and hindlimb sensory areas ( Fig. 2c and Supplementary Video 4). Simultaneously or immediately after, PT Fezf2 was strongly activated in left medial parietal cortex (parietal node) just before right-directed licks, followed by bilateral activation in frontal cortex (medial secondary motor cortex and frontal node) during lick and pellet retrieval into mouth ( Fig. 2c and Supplementary Video 5). During the brief PIM period when mice again adjusted postures and then lifted both hands to hold the pellet (Fig. 2c), IT PlxnD1 was activated in bilateral orofacial primary sensory cortex (frontolateral posterior (FLP) node) and subsequently in an anterior region spanning the lateral primary and secondary motor cortex (frontolateral anterior (FLA) node), whereas PT Fezf2 activation shifted from bilateral frontal to parietal node. In particular, PT Fezf2 activation in parietal cortex reliably preceded hand lift. After initiation of repeated bouts of eating actions, IT PlxnD1 was prominently activated in bilateral FLA and FLP specifically during coordinated oromanual movements, such as biting and handling, whereas PT Fezf2 activation remained minimum throughout the dorsal cortex. For each timepoint, we compared the average ΔF/F values at each pixel to identify regions that were significantly different between the two populations, confirming the differential flow of activation pattern (Fig. 2d). Measuring correlation between maps at each timepoint revealed PT Fezf2 activity maps to be strongly correlated during lick and hand lift, whereas IT PlxnD1 activity maps showed a sharp increase after PIM, which continued during oromanual handling. We found rather weak correlation between activation patterns of the two populations across time ( Supplementary Fig. 3b).
To obtain a spatial map of cortical activation during specific events, we built a linear encoding model using binary timestamps associated with the duration of lick, PIM, hand lift, handling and chewing as independent variables to explain the top 200 SVD temporal components of neural activity. We then transformed the model regression weights to obtain a cortical map of weights associated with each event ( Fig. 2e and Supplementary Methods). This analysis substantiated our initial observations (Fig. 2c). Indeed, during lick onset, IT PlxnD1 was active in left barrel cortex and bilateral forelimb/hindlimb sensory areas, which correlated with sensing the approaching pellet with contralateral whiskers and limb adjustments, respectively. In sharp contrast, PT Fezf2 was active along a medial parietal-frontal axis, which correlated with licking. During PIM, IT PlxnD1 was preferentially active in FLP with lower activity in FLA as well as in forelimb/hindlimb sensory areas, whereas PT Fezf2 was still active along the parietofrontal regions. Before and during hand lift, PT Fezf2 showed strong activation specifically in bilateral parietal areas, whereas IT PlxnD1 showed predominant bilateral activation in both FLA and FLP. During eating and pellet handling, IT PlxnD1 was strongly active in FLA and less active in FLP, whereas PT Fezf2 was only Article https://doi.org/10.1038/s41593-022-01244-w weakly active specifically in the frontal node (Fig. 2e, note scale change in panels). Both IT PlxnD1 and PT Fezf2 showed significantly reduced activity across dorsal cortex during chewing (Fig. 2e).
To capture prominently activated cortical areas associated with onset of various feeding movements, we computed average activity per pixel during the progression from licking, retrieving pellet into b P e l l e t d r o p B e l t s t a r t L i c k P e l l e t i n  Colored lines represent different body parts as indicated (light green shade: handle-and-eat episodes; orange shade: chewing episode). c, Mean IT PlxnD1 and PT Fezf2 sequential activity maps (200-ms steps) during the feeding sequence before and after PIM onset (IT PlxnD1 -23 sessions from six mice; PT Fezf2 -24 sessions from five mice). Note the largely sequential activation of areas and cell types indicated by arrows and numbers: (1) left barrel cortex (IT PlxnD1 ) when right whiskers sensed approaching pellet; (2) parietal node (PT Fezf2 ) while making postural adjustments as pellet arrives; (3) forelimb sensory area (IT PlxnD1 ) with limb movements that adjusted grips of support bar as pellet approaches closer; (4) frontal node (PT Fezf2 ) during lick; (5) orofacial sensory areas (FLP, IT PlxnD1 ) when PIM; (6) parietal node again during hand lift; and (7) FLA-FLP (IT PlxnD1 ) on handling and eating the pellet. d, Difference between IT PlxnD1 and PT Fezf2 average activity maps at each timestep as in c. Only significantly different pixels are displayed (two-sided Wilcoxon rank-sum test with P value adjusted by FDR = 0.05). Blue pixels indicate values significantly larger in IT PlxnD1 compared to PT Fezf2 and vice versa for green pixels. e, Spatial maps of IT PlxnD1 (top) and PT Fezf2 (bottom) regression weights from an encoding model associated with lick, PIM, hand lift, eating and handling and chewing (IT PlxnD1 -23 sessions from six mice; PT Fezf2 -24 sessions from five mice). FDR, false discovery rate. Article https://doi.org/10.1038/s41593-022-01244-w mouth, to hand lift across mice and sessions (from 1 s before to 2 s after PIM; Fig. 3a,d). Comparing the average ΔF/F distribution at each pixel confirmed that PT Fezf2 activation was most prominent along a medial parietal-frontal network, whereas IT PlxnD1 was most strongly engaged along a frontolateral FLA-FLP network (Extended Data Fig. 4a). Activation patterns were much more strongly correlated within each population than between the two populations, with PT Fezf2 maps being more consistent than IT PlxnD1 (Extended Data Fig. 4b). Principal component analysis on combined spatial maps confirmed distinct cortical activation patterns between IT PlxnD1 and PT Fezf2 during various feeding movements (Extended Data Fig. 4c).
To characterize the temporal activation patterns of key cortical nodes during behavior, we extracted temporal traces from the center within each of these four areas and examined their temporal dynamics aligned to the onset of lick, PIM and hand lift (Fig. 3b,c,e,f). PT Fezf2 activity in the frontal node rose sharply before lick, sustained for the duration of licking until PIM and then declined before hand lift; on the other hand, PT Fezf2 activity in the parietal node increased before lick and then declined immediately after, followed by another sharp increase before hand lift and then declined again right after. In contrast, IT PlxnD1 activities in FLA and FLP did not modulate significantly during either licking or hand lift but increased specifically only when mice first retrieved PIM; whereas activation in FLP decreased after a sharp rise after PIM, activities in FLA sustained during biting and handling (Fig. 3g,h). To examine cortical dynamics from both populations within the same region, we measured GCaMP6f signals centered to PIM from all four nodes for each cell type. PT Fezf2 showed strong activation within parietal and frontal nodes specifically during lick and hand lift, whereas IT PlxnD1 showed significantly lower activation within these nodes during these episodes (Fig. 3i,j). In sharp contrast, IT PlxnD1 was preferentially active in FLA and FLP specifically during PIM with sustained activity especially in FLA during biting and handling, but no associated activity was observed in PT Fezf2 within these nodes during the same period (Fig. 3i,j). A similar difference in dynamics was observed from activity aligned to either lick or hand lift onset (Extended Data Fig. 4d).
To validate the differential temporal dynamics among PN types, we projected activity traces onto the top two dimensions identified by linear discriminant analysis (LDA) to visualize the spatial distribution of projected clusters (Methods). The analysis showed that IT PlxnD1 and PT Fezf2 activity clustered independently with little overlap across regions (Fig. 3k). Altogether, these results indicate that IT PlxnD1 and PT Fezf2 operate in distinct and partially parallel subnetworks, which are differentially engaged during specific sensorimotor components of a feeding behavior. It is important to note that the IT class comprises diverse subpopulations beyond IT PlxnD1 ; it is possible that activity of another IT subpopulation might more closely correlate with PT neurons.

Feeding without hand occludes parietal PT Fezf2 activity
To investigate if the observed PN dynamics were causally related to features of the behavior, we developed a variant of the feeding task in which mice lick to retrieve food pellet but eat without using hands (Fig. 4a); this was achieved by using a blocking plate to prevent hand lift until mice no longer attempted to use their hands during eating. We then measured IT PlxnD1 and PT Fezf2 activities aligned to the onset of lick and PIM and compared these to those during normal trials in the same mice ( Fig. 4b-e and Extended Data Fig. 5a,b). Whereas PT Fezf2 activity in the frontal node did not show a difference with or without hand lift, PT Fezf2 activity in the parietal node showed a significant decrease in trials without hand lift specifically during the time when mice would have lifted hands during normal trials ( Fig. 4c and Extended Data Fig. 5a,b). On the other hand, IT PlxnD1 activity in FLA and FLP did not change during the same time with or without hand lift. However, IT PlxnD1 activities in FLA and FLP showed a notable reduction during the eating-handling phase (Fig. 4d,e and Extended Data Fig. 5a,b). Comparing activity intensity during with or without hand lift indicated only PT Fezf2 activities in parietal node to show a significant decline, whereas IT PlxnD1 activities across FLA and FLP did not change during the hand lift phase (Fig. 4f,g).
To visualize cortical regions differentially modulated between trials with or without hand lift, we computed mean pixel-wise activity during a 1-s period after PIM onset from both trial types and subtracted the spatial map of no-hand-lift trials from that of hand-lift trials (Fig. 4h). We computed the difference between the two maps and visualized only significantly different pixels (Fig. 4i). As expected, only the parietal region in PT Fezf2 showed significantly higher activity during hand-lift compared to no-hand-lift trials, whereas no pixels were significantly different within IT PlxnD1 (Fig. 4h,i). These results strengthen the correlation between parietal PT Fezf2 activation and hand lift movement during feeding; they also suggest that IT PlxnD1 activity in FLA and FLP is, in part, related to orofacial sensorimotor components of feeding actions.

IT PlxnD1 and PT Fezf2 inhibition differentially disrupts feeding
To first investigate if the cortical regions active during feeding behavior were necessary for its proper execution, we optogenetically inhibited bilateral regions of dorsal cortex using vGat-ChR2 mice expressing channelrhodopsin-2 in GABAergic neurons (Extended Data Fig. 6a). We examined the effects of bilateral inhibition of parietal, frontal, FLP and FLA nodes on different components of the behavior, including pellet retrieval by licking, hand lift after PIM and mouth-hand mediated eating bouts. Our results show that the parietofrontal and frontolateral regions are necessary to orchestrate orofacial and forelimb movements that enable pellet retrieval and mouth-hand coordinated eating behavior (Extended Data Fig. 6).
We then investigated if IT PlxnD1 and PT Fezf2 neurons within the same cortical region were causally associated with distinct sensorimotor components of the feeding behavior. We expressed light-activated anion channelrhodopsin (GtACR1) in IT PlxnD1 or PT Fezf2 neurons in frontal and frontolateral regions of the same mouse and examined the effects of bilaterally inhibiting either population during specific phases of feeding (Methods and Fig. 5a). During pellet retrieval phase, PT Fezf2 inhibition in both frontal and frontolateral nodes resulted in a sharp decrease in tongue length throughout the inhibition, whereas IT PlxnD1 inhibition resulted in a momentary decrease at inhibition onset, after within frontolateral and parietofrontal nodes show distinct temporal dynamics during feeding behavior. a,d, Mean activity map of PT Fezf2 (a, 24 sessions from five mice) and IT PlxnD1 (d, 23 sessions from six mice) during feeding from 1 s before to 2 s after PIM. b,e, Example PT Fezf2 (b) and IT PlxnD1 (e) activity from FLA (magenta), FLP (orange), frontal (dark brown) and parietal (light brown) nodes during feeding behavior; vertical bars indicate behavioral events. c,f, Single-trial heat maps of PT Fezf2 activity from frontal and parietal (c) and IT PlxnD1 from FLA and FLP nodes (f) centered to lick, PIM and hand lift onset (five sessions from one example mouse). g, Mean PT Fezf2 activity within frontal and parietal nodes centered to lick, PIM and hand lift onset. Gray dashed lines indicate median onset times of other events relative to centered event (five sessions from one example mouse; shaded region indicates ±2 s.e.m). Gray shade indicates eating-handling episode. h, Mean IT PlxnD1 activity within FLA and FLP centered to lick, PIM and hand lift onset (five sessions from one example mouse; shaded region indicates ±2 s.e.m). i, Single-trial heat maps of IT PlxnD1 and PT Fezf2 activities within parietal, frontal, FLP and FLA centered to PIM (IT PlxnD1 -23 sessions from six mice; PT Fezf2 -24 sessions from five mice). j, Mean IT PlxnD1 and PT Fezf2 activity within parietal (light brown), frontal (dark brown), FLP (orange) and FLA (magenta) centered to PIM (sample size as in i; shaded region indicates ±2 s.e.m). Left inset: overlaid activity maps of IT PlxnD1 (blue) and PT Fezf2 (green) after thresholding indicating distinct nodes preferentially active during the feeding sequence. k, Distribution of IT PlxnD1 and PT Fezf2 activities centered to PIM onset from parietal, frontal, FLP and FLA projected to the subspace spanned by the first two LDA dimensions (IT PlxnD1 -23 sessions from six mice; PT Fezf2 -24 sessions from five mice). Handl., handling. Handl. . This resulted in a significant decrease in the mean tongue length compared to control trials only when disrupting PT Fezf2 but not IT PlxnD1 neurons (Fig. 5c), which was substantiated by comparing the relative tongue length difference between the two populations (Extended Data Fig. 7a). PT Fezf2 inhibition in both frontal and frontolateral regions, after the mouse picked the pellet, disrupted the ability to bring hands to the mouth to hold the pellet, resulting in a sharp decrease in the proportion of hand lift episodes, whereas only a small effect was observed on IT PlxnD1 inhibition ( Fig. 5d and Supplementary Video 8). This was validated by comparing the proportion of hand lifts during inhibition between the two populations ( Fig. 5d). Disrupting PT Fezf2 in both frontal and frontolateral regions after retrieving the pellet and bringing hands to mouth (during food handling) strongly affected the gross mobility of hands, such that mice were unable to properly bring the pellet toward the mouth ( Fig. 5e and Supplementary Video 9), resulting in an increase in hand-mouth distance and decrease in velocity ( Fig. 5e-g). This was validated by comparing the relative hand-to-mouth distance and hand velocity difference between the two populations (Extended Data Fig. 7b,c). Time (s) Although no such gross deficits were observed on IT PlxnD1 disruption ( Fig. 5e-g), inhibiting both frontal and frontolateral resulted in a more subtle effect wherein mice had difficulty in using fingers to grasp the pellet properly, resulting in decreased agility and spending significantly longer time handling the pellet close to the mouth ( Fig. 5e and Supplementary Video 10). Indeed, we found that, during IT PlxnD1 inhibition, the hands were closer to the mouth for a significantly longer time than control trials (Fig. 5h). The decreased agility was accompanied by a drop in the number of grasps and rigid finger movements, resulting in a decrease in the distance between fingers during food handling (Supplementary Video 10, Fig. 5i,j and Methods). These results provide causal evidence that IT PlxnD1 and PT Fezf2 neurons within the same cortical regions differentially contribute to controlling distinct motor actions of feeding. Whereas PT Fezf2 is associated with controlling major oral and forelimb movements, including lick and hand lift, IT PlxnD1 is likely involved in finer-scale coordination, such as finger movements during food handling.

Distinct projections of IT PlxnD1 and PT Fezf2 subnetworks
To explore the anatomical basis of IT PlxnD1 and PT Fezf2 subnetworks revealed by wide-field calcium imaging, we examined their projection patterns by anterograde tracing using recombinase-dependent AAV in driver lines. Using serial two-photon tomography across the whole mouse brain 46 , we extracted the brain-wide axonal projections and registered them to the Allen Mouse Common Coordinate Framework (Methods and refs. 47,48 ) and quantified and projected axonal traces within specific regions across multiple planes. With an isocortex mask, we extracted axonal traces specifically within the neocortex and projected signals to the dorsal cortical surface (Fig. 6a).
As expected, PT Fezf2 in parietal and frontal regions show very few intracortical projections (Fig. 6a, Extended Data Fig. 8a and Supplementary Videos 13 and 14). PT Fezf2 in frontal node predominantly project to dorsal striatum, pallidum (PAL), sensorimotor and polymodal thalamus (THsm and THpm), hypothalamus (HY), motor and behavior-state-related midbrain regions (MBmot and MBsat) and motor and behavior-state-related Pons within the hindbrain (P-mot and P-sat). PT Fezf2 in parietal node projected to a similar set of subcortical regions as those of frontal node but often at topographically different locations within each target region ( Fig. 6b and Extended Data Fig.  8b-e). To analyze the projection patterns, we projected axonal traces within the three-dimensional (3D) masks for each region across its coronal and sagittal plane. PT Fezf2 in parietal and frontal nodes both projected to the medial regions of caudate putamen (CP), with frontal node to ventromedial region and parietal node to dorsomedial region (Extended Data Fig. 8b); they did not project to ventral striatum. Within the thalamus, frontal PT Fezf2 neurons project predominantly to ventromedial regions both in primary and association thalamus, whereas parietal PT Fezf2 preferentially targeted dorsolateral regions in both subregions (Extended Data Fig. 8c). Within the midbrain, frontal and parietal PT Fezf2 specifically targeted the motor superior colliculus (SCm) with no projections to sensory superior colliculus or inferior colliculus. Within SCm, frontal PT Fezf2 preferentially targeted ventrolateral regions, whereas parietal PT Fezf2 projected to dorsomedial areas (Extended Data Fig. 8d). Together, the large set of PT Fezf2 subcortical targets may mediate the intention and preparation and coordinate the execution of tongue and forelimb movements during pellet retrieval and handling. In particular, the thalamic targets of parietofrontal nodes might project back to corresponding cortical regions and support PT Fezf2 -mediated cortico-thalamic-cortical pathways, including parietal-frontal communications 38 .
In contrast to PT Fezf2 , IT PlxnD1 formed extensive projections within cerebral cortex and striatum (Fig. 6b, Extended Data Fig. 8b and Supplementary Videos 11 and 12). Within the dorsal cortex, IT PlxnD1 in FLA projected strongly to FLP and to contralateral FLA, and IT PlxnD1 in FLP projected strongly to FLA and to contralateral FLP. Therefore, IT PlxnD1 neurons mediate reciprocal connections between ipsilateral FLA-FLP and between bilateral homotypic FLA and FLP (Fig. 6a). In addition, IT PlxnD1 neurons from FLA predominantly project to FLP (MOp), lateral secondary motor cortex (MOs), forelimb and nose primary sensory cortex (SSp-ul and SSp-n), secondary sensory cortex (SSs) and visceral areas (VISCs). Interestingly, IT PlxnD1 neurons in FLP also projected to other similar regions targeted by FLA (Fig. 6b). Beyond cortex, IT PlxnD1 neurons in FLA and FLP projected strongly to the ventrolateral and mediolateral division of the striatum, respectively (STRd; Fig. 6b and Extended Data Fig. 8b). These reciprocal connections between FLA and FLP and their projections to other cortical and striatal targets likely contribute to the concerted activation of bilateral FLA-FLP subnetwork during pellet eating bouts involving coordinated mouth-hand sensorimotor actions. As driver lines allow integrated physiological and anatomical analysis of the same PN types, our results begin to uncover the anatomical and connectional basis of functional IT PlxnD1 and PT Fezf2 subnetworks.

Distinct IT PlxnD1 and PT Fezf2 dynamics under ketamine
Given the distinct spatiotemporal activation patterns of IT PlxnD1 and PT Fezf2 during spontaneous and goal-directed behavior, we further explored whether they differ in network dynamics in a dissociation-like brain state under ketamine-xylazine anesthesia 35,49 . We found significant differences in both the temporal dynamics and spatial propagation of activities between the two cell types. IT PlxnD1 oscillated at a higher frequency compared to PT Fezf2 . Whereas IT PlxnD1 activities spread multi-directions across most of the dorsal cortex, PT Fezf2 activities mainly propagated from retrosplenial toward the frontolateral regions (Extended Data Figs. 9 and 10 and Supplementary Videos 15 and 16). These results show that, even under an unconstrained brain state, IT PlxnD1 and PT Fezf2 subnetworks operate with distinct spatiotemporal dynamics and spectral properties, likely reflecting their differences in biophysical, physiological 50 and connectional properties (for example, Fig. 6). and PT Fezf2 frontal (inh (n = 167), control (n = 176)) and FLA (inh (n = 202), control (n = 204)) nodes. Control (black). f,g, Distribution of mean-normalized handmouth distance (f) and mean absolute hand velocity (g) for 5 s during control and inhibition of IT PlxnD1 and PT Fezf2 frontal and FLA nodes. h, Distribution of mean hand-near-mouth duration for 5 s during control and inhibition of IT PlxnD1 frontal nodes. i, Mean trace of inter-finger 1-2 distance during inhibition of IT PlxnD1 frontal (brown) and FLA (magenta) nodes. Control (black). Red line over fingers 1 and 2 illustrates the variable measured. j, Distribution of mean inter-finger 1-2 distance for 5 s during control and inhibition of IT PlxnD1 frontal and FLA nodes. Sample size for f-j as in e. All data pooled from four mice for IT PlxnD1 and three for PT Fezf2 . *P < 0.05, **P < 0.005 and ***P < 0. 0005

Discussion
Whereas early cytoarchitectonic analyses of cell distribution patterns identified numerous cortical areas 51 and their characteristic laminar organization 52,53 , single-cell recording revealed the vertical groupings of neuronal receptive field properties 3,4 . Since its formulation, columnar configuration as the basic units of cortical organization has been a foundational concept 5 , yet, to date, the anatomic basis and functional significance of 'cortical columns' remain contentious [54][55][56] . Multi-cellular recordings and computational simulation led to the hypothesis of a 'canonical circuit' template, which may perform similar operations across cortical areas [6][7][8]57 ; but its cellular basis and relationship to global cortical networks remain unsolved. An enduring challenge for understanding cortical architecture is its neuronal diversity and wiring complexity 10 . Meeting this challenge requires methods to monitor and interpret neural activity patterns across cortical layers and areas with cell type resolution in real time and in behaving animals. Wide-field calcium imaging in rodent cortex provides an opportunity to bridge cellular and cortex-wide measurement of neural activity 22 . Among diverse PNs, IT and PT represent two major top-level classes that mediate intracortical processing and subcortical output channels, respectively, with distinct gene expression 12,58 , developmental trajectories 59 , morphological and connectivity features 39,50 , biophysical properties 50 and functional specializations in specific cortical areas 60,61 and behavior [62][63][64] . Here we demonstrate that IT PlxnD1 and PT Fezf2 neurons operate in separate and partially parallel subnetworks during a range of brain states and sensorimotor behaviors and control distinct aspects of feeding movements. These results suggest a revision of the concept of cortical architecture predominantly shaped by the notion of columnar organization 55 ; they indicate that dynamic areal and PN-type-specific subnetworks are a key feature of cortical functional architecture that integrates microcircuit components and global brain networks. It is possible that columnar information flow between IT and PT, and, thus, the functional integration of corresponding subnetworks, might be dynamically gated by inhibitory and modulatory mechanisms according to brain states and behavioral demand.
Modeling and experimental studies have suggested that the source of signals measured by wide-field imaging from cortical surface differs depending on the depth of cell body layer 65   Article https://doi.org/10.1038/s41593-022-01244-w fluorescence originating across the cortical depth 65,66 . Although a large proportion of signal originates from extra somatic layers, especially for deep-layer neurons, a significant amount also arises from the cell body layer 65 . Additionally, the high correlation between calcium dynamics in cell body and apical dendrites suggests that dendritic signals closely reflect cell body dynamics [40][41][42][43][44] . Furthermore, GCaMP wide-field signals are strongly associated with neuronal action potentials both at single-cell resolution 67 and across cortical depth in a local region 36, 68 . Along with these limitations, it is important to note that GCaMP6f signals have relatively slow temporal dynamics (hundreds of milliseconds); complementary methods with better temporal resolution for spiking activities (for example, electrophysiological recordings 69 ) are necessary to decipher information flow and neural circuit operation. The posterior parietal cortex (PPC) is an associational hub receiving inputs from virtually all sensory modalities and frontal motor areas, and, it supports a variety of functions, including sensorimotor transformation, decision-making and movement planning [70][71][72] . PPC subdivisions are strongly connected with frontal secondary motor cortex in a topographically organized manner 73,74 , and this reciprocally connected network has been implicated in movement intension, planning and the conversion of sensory information to motor commands 75 . The cellular basis of parietal-frontal network is not well understood 71,76 .
Here we found that sequential co-activation of parietal-frontal PT Fezf2 neurons are the most prominent and prevalent activity signatures that precede and correlate with tongue, forelimb and other body part movements. During the feeding task, training mice to eat without hands specifically occluded PT Fezf2 parietal activation that normally precedes hand lift. Furthermore, optogenetic inhibition of PT Fezf2 neurons within the frontal node disrupts licking and hand lift, whereas inhibition of the parietal node disrupts hand-to-mouth movement trajectory. Together, these results suggest PT Fezf2 neurons as a key component of the parietal-frontal network implicated in sensorimotor transformation and action control. As PT Fezf2 neurons do not extend significant intracortical projections, their co-activation in the parietal-frontal network might result from coordinated pre-synaptic inputs from, for example, a set of IT PNs that communicate between the two areas or from cortico-thalamic-cortical pathways 38,77,78 linking these two areas. As the topographic connections between parietal and frontal subdivisions appear to correlate with multiple sensory modalities and body axis 73,74,76 , cellular resolution analysis using two-photon imaging and optogenetic recordings may resolve these topographically organized circuits that mediate different forms of sensorimotor transformation and action control.
Although IT PlxnD1 neurons show broad and complex activity patterns during several brain states and numerous episodes of sensorimotor behaviors, we discovered a prominent FLP-FLA subnetwork that correlates with coordinated mouth and hand movements during feeding. Notably, this subnetwork is weakly correlated with pellet retrieval and hand lift to mouth, when PT Fezf2 neurons in the parietal-frontal subnetwork showed strong activation. Whereas FLP mostly comprises primary sensory areas of the orofacial and forelimb regions, FLA comprises frontolateral regions of primary and secondary motor areas. The prominent reciprocal IT PlxnD1 projections between these two areas and across bilateral FLP-FLA suggest an anatomical basis underlying the concerted activity dynamics of this functional subnetwork. Furthermore, optogenetic inhibition of IT PlxnD1 neurons within frontolateral nodes resulted in finger movement deficits during pellet handling. Together, these results suggest a significant role of FLP-FLA subnetwork in the sensorimotor coordination of orofacial and forelimb movement during feeding.
Our focus on IT PlxnD1 and PT Fezf2 populations in the current study does not yet achieve a full description of cortical network operations. Indeed, top-level classes further include cortico-thalamic, near-projecting and layer 6b populations 12 ; and the IT class alone comprises diverse transcriptomic 12 and projection 17 types that mediate myriad cortical processing streams 79 . Although IT PlxnD1 represents a major subset, other IT subpopulations remain to be recognized and analyzed using similar approaches. It is possible, for example, that another IT type might feature a direct pre-synaptic connection to PT Fezf2 (for example, refs. 7,39 ) and share a more similar spatiotemporal activity pattern and closer relationship to the PT Fezf2 subnetwork. Finer-resolution genetic tools for examining additional PN types will achieve an increasingly more comprehensive view of functional cortical networks. Furthermore, simultaneous analysis of two or more PN types in the same animal will be particularly informative in revealing their functional interactions underlying cortical processing.

Online content
Any methods, additional references, Nature Portfolio reporting summaries, source data, extended data, supplementary information, acknowledgements, peer review information; details of author contributions and competing interests; and statements of data and code availability are available at https://doi.org/10.1038/s41593-022-01244-w.

Animals
All experimental procedures were carried out in accordance with National Institutes of Health (NIH) guidelines and approved by the Institutional Animal Care and Use Committees (IACUCs) of Cold Spring Harbor Laboratory (CSHL) and Duke University. Fifty-seven male and female mice were included as part of the study. All mice were housed in groups of at least 2-5 on a 12-hour light/dark cycle. To express GCaMP6f within a specific projection neuron population, 14 FezF2-CreER and 16 PlexinD1-CreER knockin mouse lines generated in the lab were crossed with Ai148 ( Jackson Laboratory, 030328), a GCaMP6f reporter line. Three VGAT-ChR2-EYFP mice ( Jackson Laboratory, 014548) that express the blue-light-activated opsin ChR2 in GABAergic interneuron population were used for optogenetic manipulation. Six PlexinD1-CreER mice and four FezF2-CreER mice crossed with a reporter line expressing LSL-Flp were used for viral expression of flp-dependent anterograde tracing. Four PlexinD1-CreER mice and three FezF2-CreER mice were used for cell-type-specific inhibition experiments. Four PlexinD1-CreER mice and three FezF2-CreER mice crossed with Ai148 were used for two-photon imaging experiments. Expression of reporters were controlled via the intraperitoneal injection of tamoxifen (20 mg ml −1 , dissolved in corn oil) between 1 month and 2 months postnatal. All mouse colonies at CSHL were maintained in accordance with husbandry protocols approved by the IACUC and housed by gender in groups of 2-4 with access to food and water ad libitum and on a 12-hour light/dark cycle.

Surgical procedures
For wide-field calcium imaging and optogenetic manipulation, adult mice older than 6 weeks were anesthetized by inhalation of isoflurane maintained between 1% and 2%. Ketoprofen (5 mg kg −1 ) was administered intraperitonially as analgesia before and after surgery, and lidocaine (2-4 mg kg −1 ) was applied subcutaneously under the scalp before surgery. Mice were mounted on a stereotaxic headframe (Kopf Instruments, 940 series, or Leica Biosystems, Angle Two). An incision was made over the scalp to expose the dorsal surface of the skull, and the skin was pushed aside and fixed in position with tissue adhesive (Vetbond, 3M). The surface was cleared using saline, and an outer wall was created using dental cement (C&B Metabond, Parkell; Ortho-Jet, Lang Dental), keeping most of the skull exposed. A custom-designed circular head plate was implanted using the dental cement to hold it in place. After cleaning the exposed skull thoroughly, a layer of cyanoacrylate (Zap-A-Gap CA+, Pacer Technology) was applied to clear the bone and provide a smooth surface to image calcium activity or for optogenetic stimulation 24 . For viral injections, we followed the same anesthesia procedure. Under anesthesia, an incision was made over the scalp; a small burr hole was drilled in the skull; and brain surface was exposed. A pulled glass pipette tip of 20-30 μm containing the viral suspension was lowered into the brain; a 300-400-nl volume was delivered at a rate of 30 nl min −1 using a Picospritzer (General Valve Corporation); the pipette remained in place for 10 minutes, preventing backflow, before retraction, after which the incision was closed with 5/0 nylon suture thread (Ethilon Nylon Suture, Ethicon) or TissueGlue (Vetbond, 3M); and mice were kept warm on a heating pad until complete recovery 37 . For cell-type-specific optogenetic manipulations, we first drilled through the skull using a 0.5-mm bur bilaterally over the frontal and frontolateral anterior areas in each mouse, followed by viral injection (GtACR1) as described earlier. We then implanted fiber-optic cannulae (outer diameter 1.25-mm ceramic ferrule, 400-μm core, 0.39 NA, R-FOC-L400C-39A, RWD), placing them on the surface of the brain without penetrating into tissue and sealed them to the skull using dental cement (Tetric EvoFlow, Ivoclar Vivadent), followed by head bar implantation.

Whole-brain serial two-photon tomography and image analysis
Whole-brain serial two-photon (STP) imaging was performed as described previously 46 . In brief, perfused and post-fixed brains of adult mice were embedded, cross-linked and imaged across coronal sections with a Chameleon Ultrafast-2 Ti:Sapphire laser. Images were further processed using ImageJ/Fiji and Adobe Photoshop before analysis. To analyze GCaMP distribution and projection patterns of IT PlxnD1 and PT Fezf2 , each frame was background-subtracted and aligned to 3D Allen map 80 , after which projection intensity in each brain region was computed. A more detailed description of imaging and analysis is provided in Supplementary Methods.

In situ dybridization
Hybridization chain reaction (HCR) in situ was performed as previously described 81 . Probes were ordered from Molecular Instruments. Mouse brain was sliced into 50-μm-thick slices after paraformaldehyde (PFA) perfusion fixation and sucrose protection. HCR in situ was performed via the free-floating method in a 24-well plate. First, brain slices were exposed to probe hybridization buffer with HCR Probe Set at 37 °C for 24 hours. Brain slices were washed with probe wash buffer, incubated with amplification buffer and amplified at 25 °C for 24 hours. On day 3, brain slices were washed, counter-stained with DAPI and mounted. PlexinD1 (546 nm), Fezf2 (546 nm) and Satb2 (647 nm) probes were used to examine overlaps between these markers.

Feeding behavior paradigm
We developed a novel behavior paradigm wherein mice use an ethological behavioral sequence to capture, handle and feed on food pellets while being head-fixed. In brief, a food pellet is automatically dispensed onto a conveyer belt that delivers it to the head-fixed mouse. The mouse then picks the pellet with its tongue to its mouth, followed by bringing its hand to the mouth to manipulate and eat it. At the end of trial, the belt moves back to its starting position to initiate the delivery of a new pellet. Most mice can perform the task within 2 weeks of initiating training. A detailed description of the task and training protocol is provideed in Supplementary Methods.

Behavior tracking and classification
Using two high-speed cameras (FL3-U3-13S2C-CS, Teledyne FLIR) fitted with varifocal lens (COT10Z0513CS, B&H), we recorded behavior from both the front and left side of the mouse at 100 frames per second (fps) as they performed the task under IR illumination. We used DeepLabCut (version 2.0.8) 45  To identify hand position during optogenetic manipulation, we tracked the location of the first finger (Fig. 5e) of one of the hands. Instantaneous hand velocity (speed) was quantified as the absolute value of the first derivative of the hand position with respect to time. To quantify deficits in hand lift trajectory, we measured the absolute velocity of 15-Hz low-pass-filtered (to remove high-frequency noise) hand trajectory during the lift episode and compared its integral between 0 and 1 s after hand lift between inhibition and control trials. Hands were considered close to mouth if the distance between finger and mouth was below a custom-defined threshold. Inter-finger distance was quantified by extracting the instantaneous Euclidean distance between the position of the first and second finger (Fig. 5j). Control trials used for comparison were extracted from the closest non-inhibition trial preceding each inhibition trial. To compare the effect of inhibition between IT PlxnD1 and PT Fezf2 (Extended Data Fig. 7b,c), the first n control trials to match the sample size of inhibition trials were used to compute the distribution of difference between control and inhibition trials.

Wide-field calcium imaging
We used wide-field imaging to simultaneously measure GCaMP6f activity across the dorsal cortex. The imaging system used was as described previously 24 . In brief, the cortical surface was illuminated with alternating blue (470 nm) and violet (405 nm) LEDs at 60 Hz. Images were acquired with an sCMOS (edge 5.5, PCO) camera. We used the 405-nm excitation signal to regress out hemodynamic signal from 470-nm excitation and to obtain calcium-dependent ΔF/F. For spontaneous and ketamine-anesthetized measurements, because activity was measured for 180 s continuously, signal was first de-trended by fitting and subtracting a 7th-order polynomial to the raw signal associated with 405-nm and 470-nm excitation (Supplementary Fig. 1c) before regressing out non-calcium-dependent signal as described before. This resulting imaging rate of 30 fps after hemodynamic correction was used for all subsequent analysis of calcium activity. All wide-field data were rigidly aligned to the Allen CCFv3 dorsal map using four anatomical landmarks-the left, center and right edges of the anterior ridge between the frontal cortex and the olfactory bulb and the lambda 24,80 -thereby allowing data to be combined across mice and sessions. A detailed description of imaging components and correction is provided in Supplementary Methods.

Optogenetic manipulation
To disrupt cortical activity in VGAT-ChR2 mice during behavior, we built a laser scanning system that can direct laser stimulation unilaterally or bilaterally across the whole dorsal cortex surface. A collimated beam of blue light (470 nm) from a laser (SSL-473-0100-10TN-D-LED, Sanctity Laser) was fed into a two-dimensional (2D) galvo system (GVS 002, Thorlabs) that was directed onto cortical surface using custom-written software. The system contained an additional path to simultaneously visualize the cortical surface using a camera (BFS-U3-16S2C-CS, Teledyne FLIR). Using this system, we directed blue light with a beam diameter of 400 μm (full width half maximum) bilaterally at 30 Hz. Laser power at the stimulation site on the cortical surface was set between 10 mW and 15 mW. We bilaterally inhibited cortical areas identified from regions active during the feeding behavior task: FLA (1.6 mm anterior, 2.3 mm lateral), FLP (0.5 mm anterior, 3.5 mm lateral), frontal (2 mm anterior, 1 mm lateral) and parietal (−1.2 mm posterior, 1.2 mm lateral). Using median onset times associated with lick, PIM and hand lift from previous behavior trials, we turned on the stimulation before median lick onset time or during licking before median PIM onset time or after PIM but before median hand lift onset time or after hand lift onset time during manipulation, bilaterally inhibiting each of the four regions of interest (ROIs) for durations ranging from 5 to 7 s. The inhibition was randomly turned on between control trials where we did not provide any laser stimulation. For cell-type-specific manipulation, we used splitter branching fiber-optic patch cords (400 μm core, SBP(2) 1 m FCM-2xZF2.5, Doric Lenses) attached to the head of a 532-nm laser (GL532T3-100FC, SLOC Lasers). The output fibers were attached bilaterally to either the frontal or frontolateral anterior optic fiber implants during behavior for manipulation of either IT PlxnD1 or PT Fezf2 neurons. Laser power was set between 5 mW and 8 mW. The inhibition protocol was as described earlier.

Neural and behavior data analysis
All neural and behavior analysis was performed on MATLAB version 2018b and Python 3.8/3.9.
Wakeful resting state analysis. Mice were first habituated to head fixation in the setup as described earlier (Supplementary Methods). Calcium dynamics were recorded for 3 minutes at 30 fps with simultaneous behavior video recording at 20 fps. Both calcium activity and behavior videos were band-pass filtered between 0.01 Hz and 5 Hz. Variance from behavior video recordings was used to identify active and quiescent episodes. To quantify the amount of neural activity variance explained by behavior, we computed the SVD of both neural and behavior data. We then used a linear model to explain the top 200 temporal components of neural data using the top 200 temporal components of behavior video data as independent variables. We performed five-fold cross-validation of the model to obtain the cross-validated R 2 (ref. 82 ). To quantify neural activity variance explained by each body part, we defined a window around each body part and extracted the average motion energy amplitude within each window. We then used a linear model to explain the top 200 SVD temporal components of neural activity data using the signal per body part as independent variables. We performed five-fold cross-validation of the model to obtain the cross-validated R 2 (ref. 82 ). A detailed description of the analysis is provided in Supplementary Methods.

Sensory stimulation analysis.
Before stimulation, mice were injected with chlorprothixene (1 mg kg −1 intraperitoneally) and maintained under light isoflurane anesthesia (0.8-1% with O 2 ). We then placed a custom-designed cardboard attached to two piezo actuators (BA5010, PiezoDrive) close to the left whisker pad between whiskers and just below the upper and lower lip. We also placed an orange LED close to the dorsal region of the left eye. We used an Arduino Uno Rev3 (A00006, Arduino) to drive the piezo and LED. A single trial consisted of 3 s of baseline followed by whisker stimulation at 25 Hz for 1 s, 3-s delay, orofacial stimulation at 25 Hz for 1 s, 3-s delay, blinking visual stimulus at about 16 Hz for 1 s, followed by 3-s delay before starting the next trial. We recorded one session per day consisting of 20 trials. To extract temporal traces, we used spatial maps obtained by averaging IT PlxnD1 activity per pixel during the 1-s stimulation period in response to each sensory stimulation. We identified centers of peak activity in each map and used a circular window of 560-μm diameter to extract signals within the circular mask and average them per frame. To compute activity intensity during sensory stimulation, we computed the integral of signals extracted from each ROI for 1 s during the stimulation. Two-photon imaging and analysis. We used a Sutter movable objective microscope to measure single-neuron calcium dynamics at 30.9 Hz over the left whisker somatosensory cortex. The location was identified using the peak activity after whisker stimulation from wide-field imaging experiments. Each trial consisted of 3 s of baseline followed by 1-s whisker stimulation (as described previously), followed by another 3 s of post-stimulus measurement. For each FOV, we measured responses across 20 trials. We recorded from cell bodies in IT PlxnD1 and apical dendrites of PT Fezf2 (200-500-μm dorsoventral). We did not record from cell bodies in PT Fezf2 because they were relatively dim due to the depth. Dendritic calcium activity in layer 5B neurons has shown to be strongly correlated to cell body dynamics 40-44 . We used suite2p https://doi.org/10.1038/s41593-022-01244-w (https://www.suite2p.org/) to identify neurons and extract calcium dynamics, followed by removal of neuropil activity and z-score computation for each neuron. To classify neurons, we used linear modeling to fit the response of each cell to a predictor variable containing ones during whisker stimulation and zeros otherwise. We used the statsmodels module in Python to model the fit and obtain regression weights along with the associated statistical significance. Neurons with significantly positive regression weights (P < 0.05) were classified as activated neurons, whereas those with significantly negative weights were classified as inhibited neurons. All other neurons were grouped as unclassified.
Feeding behavior analysis. To identify sequential activation pattern during feeding behavior, we extracted frames 1 s before and 1 s after PIM onset for all trials across mice and sessions. Because each frame is registered to the Allen CCFv3, we computed mean for each pixel at every sampling point to obtain an average activation map at each timepoint centered around PIM.
To identify activation maps associated with specific behavior event, we used a linear modeling approach. We used binary timestamps associated with each behavior event as independent variables to explain the top 200 SVD temporal components associated with neural activity. Spatial maps associated with each behavior event were obtained by computing the dot product between regression weights and spatial components of SVD. A detailed analysis is described in Supplementary Methods.
To identify the center of activation so as to extract temporal traces, we first calculated the average activity per pixel for 1 s before to 2 s after PIM onset across mice and sessions (Fig. 3a,d). We then applied a mask containing mouth and nose primary sensory dorsal cortex region (as labeled by the Allen CCF v3) over the IT PlxnD1 activation map and identified the center of peak activation and used it as the center of frontolateral posterior node (Fig. 3d, orange). Similarly, we used MOs and MOp masks over IT PlxnD1 activation map to identify the center of the frontolateral anterior node (Fig. 3d, magenta). We used MOs mask over the PT Fezf2 activation map to identify the center of the frontal node (Fig. 3a, dark brown) and a few cortical regions in the posterior area (RSP agl, VIS am , VIS a , SS p-tr , SS p-ll , SS p-ul , SS p-un , VIS rl and SS p-bfd ) to identify the center of the parietal node (Fig. 3a, light brown). We used a circular window mask of 560-μm diameter around these centers to extract signals within these masks and averaged them per frame to obtain temporal dynamics from each node. The Allen masks were used only to help identify the centers of peak activation and were not used to parcellate the cortex for any analysis. To identify distinct activation clusters using LDA, we first combined temporal activity centered to PIM onset from all trials within an ROI from both PNs along the temporal dimension. We then concatenated the PN type class labels associated with each trial and performed LDA on the activity matrix and class labels using the LDA toolbox.
(LDA: https://www.mathworks.com/matlabcentral/fileexchange/2 9673-lda-linear-discriminant-analysis, MATLAB Central File Exchange, retrieved on 28 December 2021). We then projected the temporal activity matrix on the first two dimensions identified by the analysis and colored them based on PN type to visualize clusters.

Reporting summary
Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.

Data availability
Sample data are available at https://doi.org/10.6084/m9.figshare. 21437604.v1. All additional data will be made available upon reasonable request. Source data are provided with this paper.

Nature Neuroscience
Article https://doi.org/10.1038/s41593-022-01244-w Extended Data Fig. 8 | Axonal projection of IT PlxnD1 and PT Fezf2 in subcortical structures. a Three dimensional rendering of axonal projections of IT PlxnD1 from FLA and FLP and PT Fezf2 from frontal and parietal node. Yellow circle indicates injection site. b Spatial distribution of axonal projections of IT PlxnD1 from FLA and FLP (top) and PT Fezf2 parietal and frontal nodes (bottom) within the striatum projected onto the coronal and sagittal plane. c Spatial distribution of axonal projections of PT Fezf2 from parietal and frontal nodes within the primary and association thalamus projected onto the coronal and sagittal plane. d Spatial distribution of axonal projections of PT Fezf2 from parietal and frontal nodes within the motor Superior colliculus (SCm, magenta), sensory superior colliculus (SCs, yellow) and inferior colliculus (IC, brown) projected onto the coronal and sagittal plane. e Spatial distribution of axonal projections of PT Fezf2 from parietal and frontal nodes (bottom) within the hindbrain projected onto the coronal and sagittal plane. f Brain-wide volume and peak normalized projection intensity maps of IT PlxnD1 from FLA and FLP and PT Fezf2 from frontal and parietal nodes from two mice. Black font indicates injection site; larger gray font indicates regions with significant projections; smaller gray font indicates regions analyzed.  and PT Fezf2 (green) spatial power maps for each frequency band projected to the subspace spanned by the top two principal components (n = 18 maps in each group). IT PlxnD1 and PT Fezf2 both clustered independently with further segregation between IT PlxnD1 0.6-0.9 Hz and 1-1.4 Hz frequency bands, substantiating the distinct activation patterns between the two populations. d Activation sequence of the most dominant pattern (1 st dimension) identified by seqNMF from IT PlxnD1 (top) and PT Fezf2 (bottom) activity combined across mice and sessions.