Hierarchical organization of cortical and thalamic connectivity

Abstract

The mammalian cortex is a laminar structure containing many areas and cell types that are densely interconnected in complex ways, and for which generalizable principles of organization remain mostly unknown. Here we describe a major expansion of the Allen Mouse Brain Connectivity Atlas resource¹, involving around a thousand new tracer experiments in the cortex and its main satellite structure, the thalamus. We used Cre driver lines (mice expressing Cre recombinase) to comprehensively and selectively label brain-wide connections by layer and class of projection neuron. Through observations of axon termination patterns, we have derived a set of generalized anatomical rules to describe corticocortical, thalamocortical and corticothalamic projections. We have built a model to assign connection patterns between areas as either feedforward or feedback, and generated testable predictions of hierarchical positions for individual cortical and thalamic areas and for cortical network modules. Our results show that cell-class-specific connections are organized in a shallow hierarchy within the mouse corticothalamic network.

Extended Data Figure 1.. Similarity of connection strengths by distance, virus, hemisphere, and Emx1-IRES-Cre or C57BL/6J mice.

(a-d) Most experiments were done with the Cre-dependent rAAV tracer, rAAV2/1.pCAG.FLEX.EGFP.WPRE. A subset of left hemisphere injections had a duplicate injection of rAAV with a synaptophysin-EGFP fusion transgene in place of the cytoplasmic EGFP (rAAV2/1.pCAG.FLEX.SypEGFP.WPRE). This tracer allowed us to address whether labeling presynaptic terminals would improve accuracy of quantifying target connection strength, particularly in brain regions containing mostly fibers of passage. Data consisted of n=275 experiments (137 EGFP: 138 SypEGFP). These were matched across Cre lines and areas, and represent n=8 Cre lines and n=26 cortical areas. For pairs of spatially-matched experiments, the average projection strength (log10-transformed NPV) measured across the entire brain was lower in SypEGFP vs EGFP experiments (~ 0.8 log unit when <500 μm apart). However, brain-wide projection values were still highly and significantly correlated. Thus, we included the SypEGFP datasets when indicated for analyses of connectivity patterns from given source areas (but only in comparison with other SypEGFP datasets). (a) Spearman correlation coefficients (R) of normalized projection volumes for all possible pairs of injections (different and same tracer, all in the same Cre line) are plotted against the distance between the injection centroids. Linear regressions showed a significant negative slope (p<0.0001) with lower R as distance between injections increases. (b) R is plotted for injections within 500 μm of each other; slopes were not significantly different from zero and means were not significantly different from each other. Average and SD for each group is shown by the large symbols on the left (EGFP v EGFP: 0.81 +/− 0.056, SypEGFP v SypEGFP: 0.79 +/− 0.064, SypEGFP v EGFP: 0.79 +/− 0.071). (c) Quantitative differences in projection strengths measured between replicates with the same virus and between SypEGFP and EGFP (logNPV (EGFP)-logNPV(SypEGFP) injections, all < 500 um apart in the same Cre line (n=133 within virus and 222 between virus comparisons). Boxplots show median, IQR, min, max values, and + indicates mean. (d) Maximum intensity projections from four experiments within 500 μm of each other illustrate overall similarities between replicate injections and tracers (Spearman R is shown for each pair). Injections targeted primary visual cortex (VISp) in Emx1-IRES-Cre mice using either EGFP or SypEGFP tracers as indicated. (e-g) Injections into Emx1-IRES-Cre mice were made into visual areas on the left hemisphere, whereas all C57BL/6J mice received injections into the right hemisphere. Following registration to the CCF, which is a symmetric atlas, we identified three pairs of experiments in which the injection centroids were < 500μm apart after flipping injection site coordinates from the left to the right. Cortical projections were visually similar across both lines and hemispheres, and cortical connectivity strengths (to the 86 cortical targets) from these individual experiments (normalized projection volumes) were positively and strongly correlated as indicated (Spearman rank, r=0.877, 0.904, and 0.965). Thus, in Fig. 2 we merged the Emx1 and C57BL/6J data to represent connection strengths from all layers and classes, and in some of the “anchor” groups we used data from both left and right hemisphere injections.

Extended Data Figure 2.. Characterization of cortical projection neuron classes and layer selectivity across mouse lines.

(a) Brain-wide projection patterns were visually inspected for every experiment and manually classified into one of six categories based on projections to ipsi- and contra- lateral cortex, striatum, thalamus, and midbrain/pons/medulla structures as described for IT, PT, and CT classes. (b-d) Unsupervised hierarchical clustering (using Euclidean distance and average linkage) of projection weights validates and reveals major classes of cortical projection neurons. (b) Each column of the heat map shows one of the 1,081 injection experiments. Colors in the “manual PN class” are coded as in (c) for projection class. Rows show selected major brain regions which distinguish known classes of projection neurons. Values in each cell are the fraction of total brain projection volume in the given region. The dendrogram was split into 9 clusters, with two subclusters identified post-hoc for cluster 5. The numbers of experiments per cluster were (1: n=24, 2: n=4, 3: n=204, 4: n=158, 5a: n=148, 5b: n=230, 6: n=174, 7: n=12, 8: n=16, 9: n=111. The numbers of experiments per projection class were CT: n=119, IT: n=342, IT PT: n=158, IT PT CT: n=189, local: n=100, PT: n=173. (c) The relative frequency of experiments from manually-assigned projection classes within each cluster is shown. There was significant enrichment of 1, or 2 related, classes in each cluster (dots; Fisher’s exact t-test, p<0.01). (d) Maximum intensity projections of GFP-labeled axons across the brain from one example per cluster. (e) Characterization of layer-selectivity in wild type and 14 Cre lines derived from injection experiments. Number of experiments per line is listed in Supplementary Table 1. For every injection and line, we assessed layer-selectivity based on the manually annotated injection sites. Polygons were drawn around every injection site so that, after registration to the CCF, injection volume in each layer could be informatically-derived. A layer-selectivity index was calculated for each experiment (the fraction of the total injection volume contained in each layer, scaled by the relative volume of each layer in the injection source region, because layer volumes differ by area). Plots show individual data points and the average layer selectivity index +/− 95% confidence intervals (in black) for the set of 15 mouse lines. Red lines in each Cre graph show average values from C57BL/6J experiments. Red lines in the C57BL/6J graph are averages from the Emx1-IRES-Cre experiments, which also labels cells across all layers. There is a bias toward L5 neuron infection in both C57BL/6J and Emx1-IRES-Cre, highlighting the importance of using layer selective Cre lines for better coverage of cortical outputs.

Extended Data Figure 3.. Computationally removing the distance-dependence of connection weights alters the modular structure of the cortex.

To test the degree to which spatial proximity of regions affects modularity analysis, we used a power-law to fit the distance component of our ipsilateral corticocortical connectivity matrix (Knox et al., 2019). Then, we repeated our modularity analysis on the “distance-subtracted” matrix built from these residuals. (a) Weighted connectivity matrix for 43 cortical areas showing the value of the residuals from a power-law to fit the distance component. Rows are sources, columns are targets. Colors on the rows indicate distance-subtracted community structure with varying levels of resolution (γ = 0.5–1.5 on the y axis, γ = 0.8 only on the top portion of the x-axis). Columns are colored by their module affiliation in the distance-subtracted matrix above their module affiliation in the original matrix (Figure 1e). The inset in the top left corner shows the modularity metric (Q) for each level of γ, along with the Q value for a shuffled network containing the same weights. The Q values for modularity in the distance-subtracted matrix were smaller than for the original cortical matrix (e.g., 0.2754 vs 0.4638 at γ = 0.8) and the range of values for which Q was greater than Q_shuffled was narrower (0.7 ≤ γ ≤ 1.7), but some modules were still present in the distance-subtracted cortical connectivity matrix. The difference between Q and Q_shuffled was greatest for γ = 0.8. The first distance-subtracted module was comprised of the entire somatomotor module, most of the lateral module, and two regions from the prefrontal module. The second distance-subtracted module contained the visual, auditory, and medial modules, plus most of the prefrontal module and one region from the lateral module (temporal association area). Notably, these modules were like those reported by Rubinov et al. (2015). As γ increased past 1.0, regions began to split from the two large modules in small groups that generally did not reflect the original divisions, except for the auditory areas. (b) Diagram shows the ipsilateral cortical network in 2D using a force-directed layout algorithm. Nodes are color coded by module. Edge thickness shows residual values and edges between modules are colored as a blend of the module colors. (c) Cortical regions color-coded by their distance-subtracted community affiliation at γ = 0.8 show spatial relationships.

Extended Data Figure 4.. Whole brain single neuron reconstructions reveal L4 IT projections.

(a) L4 neurons are classified into at least 3 morphological types as shown. (b) Image shows sparse labeling of L2/3 and L4 neurons in the tamoxifen-inducible Cux2-IRES-CreERT2 driver crossed with the Ai166 reporter and using a low-dose of tamoxifen via oral gavage for 1 day. L4 neurons were identified based on their apical dendrite and local axons, using additional anatomical context when possible. Reconstruction was performed using Vaa3D-TeraVR on the high resolution whole brain image stack (composed of more than 10K images, resolution XYZ: 0.3 × 0.3 × 1 μm) acquired with a two-photon fluorescence micro-optical sectioning tomography system (2p-fMOST). (c) We identified 25 total L4 neurons for complete morphological reconstruction of dendrite and axon for three cell types and three cortical areas. In this Cre line at least, SSCs were most frequently identified. (d) Dorsal surface view shows the corticocortical projection patterns from three anterograde tracer experiments into the predominantly L4 Cre lines for somatosensory cortex (SS), visual cortex (VIS) and auditory cortex (AUD). (e-k) Each panel shows two examples of reconstructed cells of the same L4 type in SS, VIS, or AUD. Local morphology for each cell is shown in the inset. Arrowheads indicate axon clusters outside local region. Red=axon, blue=basal dendrite, black=apical dendrite. Consistent with canonical descriptions, we found SSCs in the somatosensory cortex that had only local axon clusters (e). However, even in these cases, we frequently observed what appeared as an aborted axon branch (no terminal cluster found; long arrow). We also found SSCs in SS that did have clear axon clusters in nearby areas (g), and, in AUD cortex, SSCs projected even to the opposite hemisphere (f). (h-k) Although we identified fewer TPC and UPC cell types in this experiment, for both types we still found cells with near and long-range projections.

Extended Data Figure 5.. Locations and cortical projection patterns from thalamic tracer experiments.

(a) Locations of the thalamic tracer injection centroids (blue dots) are shown mapped onto virtual 2D coronal planes from the Allen CCFv3. To minimize the number of sections shown, all centroids are mapped within 200 μm of their original location. See Supplementary Table 1 (thalamus tab) for more details on Cre lines and coverage. (b) Example TC projections are shown in a flat map view of the ipsilateral cortical hemisphere for different thalamic nuclei arranged by the clusters identified in Fig. 3 and related to cortical modules. Most thalamic clusters projected primarily to a single module (Fig. 3c), but some thalamic regions projected across multiple modules (e.g., AV, VAL, PF, CL), or projected strongly to both prefrontal and another module; e.g., somatomotor (MD-1, VM), lateral (PVT, MD-2, PT) or medial regions (RE, AM).

Extended Data Figure 6.. Comparison of corticothalamic projection strengths derived from EGFP and SypEGFP tracer experiments.

(a-d) Maximum intensity projections from four experiments within 500 μm of each other targeting VISp (same experiment labeled VISp-3 below) using either EGFP or SypEGFP tracers in the Rbp4-Cre_KL100 (L5) or Ntsr1_Cre_GN220 (L6) line as indicated. (a’-d’) Coronal STPT images near the center of the densest terminal zone in LGd show axon and presynaptic terminal labeling in LGd and other thalamic targets, including the ventral lateral geniculate (LGd, LGv), the intergeniculate leaflet (IGL) and the lateral posterior nucleus (LP). The anterior pretectal nucleus (APN) in the midbrain is also indicated. SypEGFP labeling is more punctate and has less fluorescence in axons and fiber tracts. (a”-d”) Coronal STPT images near the center of one of the densest terminal zones in the middle of LP. (a”’-d”’) Coronal STPT images near the center of the second densest terminal zone in the anterior part of LP. This image also contains a portion of the terminal zone in LD. (e-h) Directed, weighted, connectivity matrices (11 × 44) showing log10-transformed normalized projection volumes for the Cre lines representing CT projections labeled from Layer 5 (e,f) or Layer 6 (g,h) with EGFP or SypEGFP tracer as indicated. True negatives (including passing fibers) at the regional level were masked and colored dark grey. The color map is the same as in Fig. 4. The matrix shows relative differences for connections originating from L5 vs. L6 (L5−L6/L5+L6) for EGFP-based measures (i) and SypEGFP-based measures (j). (k) Normalized projection strengths for corticothalamic targets (n=484) were significantly correlated from matched cortical locations between EGFP and SypEGFP tracers for both Cre lines (Spearman r=0.71, 0.73, p<0.0001). On average, EGFP CT NPVs were ~ 0.5 log unit larger than SypEGFP for Rbp4 experiments, but were not different for the Ntsr1 line. (l) Normalized projection strengths for corticothalamic targets (n=484) contacted by L5 or L6 cortical neurons in matched injection locations were also significantly correlated for both EGFP and SypEGFP tracers (Spearman r=0.51, 0.60, p<0.0001), although more weakly than for the same line between viruses. Specific connections with different fiber to terminal ratios are colored by source module (light blue = from VISp, orange = from SSp, dark blue = from RSPagl). (m) Relative differences in projection strength to LP and LGd are plotted from n=6 VISp injection experiments (VISp-1 to VISp-6 in matrix rows above) for each Cre line and viral tracer. (n) Relative difference ratios calculated for L5 to L6 using EGFP are plotted against those obtained using SypEGFP (n=484 CT connections, n=278 above threshold). There is a significant correlation (Spearman r=0.68, p<0.0001). Specific connections are colored by source module (from panel l) and labeled with the target.

Extended Data Figure 7.. Validation of informatics-processing steps: CCF registration and quantitation from segmentation.

(a-c) To determine how precise the registration process is which we rely on here for quantification of signal by layer in the cortex, we manually delineated layers 1 to 6b, using background fluorescence in coronal STPT images, for n=9 cortical areas (ACAd, ORBvl, AId, PERI, SSp-bfd, MOp, VISp, RSPd, and AUDp) in n=4 mice per region. We then quantified the percentage of voxels within each manually annotated layer that were assigned to all cortical layers following automated registration to the CCFv3. (a) A confusion matrix show the mean % of overlapping voxel labels averaged across these areas (individual region data in Supplementary Table 6). (b,c) Boxplots show the median and mean (indicated with “+”); whiskers show the min:max range for the % overlap for individual experiments (b) or cortical areas (c, colored dots). Across these cortical areas, the average % overlap ranged from 86–96% of voxels appropriately registered for all layers, except for L6b, which was not included in subsequent layer quantifications. For some areas and layers, the precision was worse than others, e.g., while 66% of voxels were appropriately assigned to L2/3 in ACAd, the remaining 34% were assigned to neighboring L5. In ORBvl, only 51% of voxels were appropriately labeled for L6a. We want to note, however, that delineating layer 5 from L6a in ORBvl in coronal sections using just background fluorescence was very difficult even for experienced anatomists, so some of the imprecision may in fact come from the manual drawing. Even with these exceptions noted, in all cases a large majority of voxels were registered and assigned correctly. (d-e) Frequency distributions of informatically-derived quantification for manually verified true negative and positive targets. (d) The numbers of Log10-transformed normalized projection values are plotted for all corticocortical and thalamocortical targets manually verified as true negative (n=24,272) or true positive (n=12,921). Most true positive values were between log₁₀=−4 and log₁₀=1. At log₁₀=−1.5 (red arrow), 639 true negatives remained (2.6%), while 7,100 true positives were still included (54.9%), resulting in a false positive rate of 8.3% at this threshold level. (e) Numbers of Log10-transformed normalized projection values are plotted for all corticocortical and thalamocortical targets manually verified as true negative (n=15,789) or true positive (n=4,503). At log₁₀=−2.5 (red arrow), 362 true negatives remained (2.3%), while 3,335 true positives were still included (74.1%), resulting in a false positive rate of 9.8% at this threshold level.

Extended Data Figure 8.. CC projection patterns by layer and class between reciprocally connected areas with known hierarchy.

(a) In the visual module, VISp and VISal are reciprocally connected (black line). VISp is the de facto bottom of visual cortex hierarchy. The output to VISal from VISp is feedforward (FF). The reciprocal connection (VISal to VISp) is feedback (FB). In the FF direction (top), VISp projections from L2/3, L4, and L5 IT projections were densest in L2/3-L5 of VISal, and relatively sparse in L1 and L6 (cluster 4). Rbp4 projections from VISp to VISal were densest in L4 and L6, with moderate levels in L2/3 (cluster 8). L5 PT and L6 CT cells projected, albeit sparsely, to L1 and L5 (cluster 2). In the FB direction (bottom), L2/3 IT axons were broadly distributed across layers, with a sparser region in L5 (cluster 6). VISal L4 IT cells project noticeably more weakly to VISp (as opposed to the panel above), and terminate with a different pattern (L1 and L5/6, cluster 6). L5 IT cells project densely to superficial layers in VISp (cluster 1). Rbp4 axons were dense in L1 and deep layers (cluster 6). Projections from L5 PT and L6 CT cells were also sparse, but present in L1 and L6 (cluster 6). (b) In the somatomotor module, SSp-bfd and SSs cortex are reciprocally connected. SSp-bfd to SSs is FF; the reverse is FB. In the FF direction (top), L2/3 and L4 IT cells preferentially innervate L2/3-L5, with relatively fewer terminals in L1 and L6 (clusters 3 and 4). L5 IT projections densely innervate L1 and L2/3 (cluster 1). Rbp4 projections were densest in L4 and L6, with moderate levels in L2/3 (cluster 8). L5 PT and L6 CT cell projections were sparse, and to L1 and/or deep layers (cluster 2 and 6). In the FB direction (bottom), the patterns looked remarkably like FB projections from VISal to VISp. Note again the strong connection originating from L4 cells only in the FF direction. (c) VISp (in the visual module) and ACAd (in the prefrontal module) are reciprocally connected. ACAd exerts top-down control of VISp activity (FB); the reverse (VISp to ACAd) is considered FF. In the FF direction (top), L2/3, L4, and L5 cells all preferentially innervate L1 (cluster 1). In the FB direction (bottom), L2/3 cells also predominantly terminate in L1, but L5 cells project to both L1 and deep layers (L5 and L6, cluster 6). Note also there is a potentially significant sub-layer distinction; axons from VISp to ACAd are relatively deeper in L1 (or at the border of L1 and L2/3) of ACAd, compared to the more superficial termination of ACAd axons in L1 of VISp. (all panels) Overall, FF projections are more often in clusters 1, 4, and 8, and FB projections in cluster 6. Cluster assignments are indicated in each panel; n/a indicates the connection was either absent or below threshold for clustering. Areas in each module are shown in a top down cortex view and the network as a force-directed layout (edges = normalized connection density from Fig. 1e). STPT images in the approximate center of each target region show the laminar distribution of axons arising from labeled neurons in the different Cre lines. Images were rotated so that the pial surface is always at the top of each panel.

Extended Data Figure 9.. TC and CT projection patterns and rules between reciprocally connected areas.

(a) Schematic summarizes observed projection patterns between core thalamic nuclei (blue circle) and their reciprocally connected cortical targets (L1–L6 color coded). Laminar patterns are from Fig. 5g. STPT images of labeled axon terminals between 3 pairs of core nuclei and primary sensory cortex that perfectly follow rules in both directions. In the FF direction (LGd to VISp, VPL to SSp-ll, VPM to SSp-n) projections are dense in L4 or L4 and L6 (clusters 4, 8). In the FB direction, CT projections predominantly arise from L6. (b) Schematic summarizes observed projection patterns between matrix-focal thalamic nuclei (orange circle) and their reciprocally connected cortical targets. STPT images of reciprocal connections between PT and ILA, MD and ORBl, and MD and AId illustrate the schematized rules. Projections from these thalamic nuclei belong to clusters with relatively less L1 axon (FF-like, clusters 3, 7, 9). The reciprocal CT input is also stronger from L6 (FB), like the core nuclei above. (c) Three schematics are shown to summarize observed projection patterns between matrix-multiareal thalamic nuclei (red circle) and their reciprocally connected cortical targets. The top schematic shows dense TC projections to L1 (FB) with CT projections originating from L5 (FF). The middle schematic (with relevant example images boxed) shows reciprocal connection patterns in which TC projections target mid-layers (FF-like) and the reciprocal CT input is stronger from L6 (FB). The bottom schematic shows the same TC projection pattern as the top schematic, but with CT projections originating ~equally from L5 and L6. STPT images show reciprocal connections between multiarea-matrix thalamic regions LP, PO, RE, and VM to 3 cortical targets each. Some regions have target-specific projections that are either FF or FB. For example, different from the LP-to-VISp projection (FB), axons from LP to VISam and ACAd target mid-layers as opposed to L1 (clusters 8 and 5, FF), and the reciprocal connection arises more from L6 (typical for FB). Projections from PO, RE, and VM to all three cortical targets are consistent with a FB projection (denser terminations in L1 and either L5 or L6 (clusters 2 and 6). Reciprocal CT projections originate from L5 or, both L5 and L6. We did not see CT input arising equally from both layers or more from L5 when the reciprocal TC projection was considered FF, consistent with the “no-strong-loops” hypothesis. (all panels) Overall, FF projections from core thalamic regions are in clusters 4 and 8. FB from matrix-multiareal thalamic regions are in clusters 2 and 6, like CC FB. The matrix-focal results support the notion that patterns with relatively less L1 involvement (3, 5, 7, 9) are FF, particularly given the strong reciprocal input observed from L6. STPT images are from the approximate center of the axon termination field for each target region. Cortex images were rotated so that the pial surface is at the top. Cluster assignments (for TC) are indicated in each panel. Text labels above image show FF and FB direction based on relative position in Fig. 6. Dashed lines indicate region borders.

Extended Data Figure 10.. Robustness of the hierarchical organization results.

We constructed multiple hierarchies using only C57BL6/J and Emx1-IRES-Cre experiments (WT) or Cre data without the Cre line confidence measure to compare with results in Fig. 6. The hierarchical position of each area $H_i^0$ and the global hierarchy score h_g are defined as in Equations (4) and (5) in Methods, but with the same confidence for all lines, i.e, conf(T)=1 for all Cre lines (T). (a,b) In both cases, connection types 2 and 6 are assigned to one direction (feedback), while other clusters are grouped to the opposite direction (feedforward). Cluster 7 was not identified in the WT dataset. (c) Corticothalamic connections were also classified as in Fig. 6b for the Cre data. CT connections were not included for WT as these are exclusively defined by Cre lines. (d,e) Global hierarchy scores from the original, observed data, and the distributions of hierarchy scores obtained from shuffled datasets (n=100) are shown for CC connections only (green), compared to scores obtained when TC connections are sequentially included (pink). The upper bound scores for an artificially-perfect hierarchy using the WT datasets (e) are: 0.630 for CC and 0.601 for CC+TC connections. (f) Z-scores were calculated for the global hierarchy scores compared to shuffled data for each of the three versions of cortical hierarchy (CC, CC+TC, CC+TC+CT). The highest z-scores were observed when using Cre line confidence weighting (compared to those with no confidence weighting or wild type data only). (g) Predicted hierarchical positions of 37 cortical and 24 thalamic areas based on CC, CC+TC, or CC+TC+CT connections. Areas are ordered in each panel by the scores obtained using Cre line data with confidence weighting (Cre conf, black circles). Scores from Cre line data without confidence weighting (gray circles) and scores from wild type/Emx1-IRES-Cre data (open circles) are plotted for direct comparison. Y-axis labels are color coded by module assignment (for cortical areas). (h) Robustness of the cortical hierarchy (w/ Cre conf) against individual Cre lines and projection classes. The left panel shows Spearman rank correlation coefficients between the CC and CC+TC hierarchy with n=13 layer-/class-specific Cre lines included vs. each of the Cre lines removed. The right panel shows results when data from Cre lines with the same layer and class were removed together. Removal of these lines and classes produced relatively minor deviations from the overall hierarchy determined with all data. Note that in both panels the y-axis starts at R=0.85. For all lines and classes, the correlation with the hierarchy using the complete dataset is very high. The lowest correlations occurred following removal of Cux2-IRES-Cre, Rbp4-Cre_KL100, and Tlx3-Cre_PL56.

Figure 1.. Cortical tracer experiments and network modularity.

(a) Top-down view of the right cortical hemisphere in CCFv3. (b) a virtual cortical flat map shows all 43 annotated areas. The white dotted line indicates the boundaries of what is visible in a. (c) Cortical injection locations plotted on the flat map. (d) Key summarizes layer and projection class selectively for 15 mouse lines. The color code is also used in c; experiments in lines not listed are colored dark gray. (e) Matrix shows ipsilateral normalized connection densities between 43 cortical areas. Top left corner: the modularity metric (Q) and Q for a shuffled network are plotted for each γ level. Colors to the left of each row indicate community structure at γ = 0–2.5. Community structure was determined independently for each value of γ, but colors were matched to show how communities split as γ increased. Columns are colored by the six modules identified at γ = 1.3. (f) Cortical regions on the flat map color-coded by module affiliation at γ = 1.3. (g) Network diagram shows ipsilateral corticocortical connections using a force-directed layout algorithm. Nodes are color coded by module. Edge thickness shows relative normalized connection density. Edges between modules are colored as a blend of the connected node colors.

Figure 2.. Corticocortical projection patterns by layer and class.

(a) 43 groups of experiments spatially-matched to one Rbp4 anchor (green dots). Most group members were < 500 μm from the anchor (median = 296 μm). Green circles indicate the variance in distance to Rbp4 for each group. (b-e) Data from three groups are shown. (b-d) STPT images at the center of each injection site per Cre line were manually overlaid by finding the best match between the pial surface (top) and white matter boundary (bottom), then pseudocolored by line. Scale bar = 250 μm. (e) Top down views of CC projections for spatially-matched experiments. (f) Directed, weighted connectivity matrices (27 × 86) for seven mouse lines: WT and the six Cre lines in (e). Each row contains the log10-transformed normalized projection volumes (NPV) from a single experiment in one of 27 source areas. Columns show cortical target regions. Rows and columns follow the same order in each matrix. White boxes highlight regions in the same module. True negatives and passing fibers were masked out (dark grey). Rows for which an experiment was missing (often because of low Cre expression) are light grey. The color map ranges from 10^−3.5 to 10^0.5 Log NPV. It is truncated at both ends. (g) Average out-degrees (+/− SEM) across all sources for each Cre line are plotted for ipsilateral and contralateral cortex. (h) The fraction of true positive targets shared by each line with its Rbp4 anchor is shown in the box plot (gray). The fraction of positive targets unique to Rbp4 (green) or to the line indicated (white) are also shown. Box plots show median and IQR. Whiskers show min and max values.

Figure 3.. Thalamocortical projection patterns by region and class.

(a) Left, flat map views show TC projections labeled from the region indicated. Right, STPT images from the center of a cortical target (* on left) show example axon lamination patterns associated with each projection class. (b) Key summarizes projection class assigned for 29 thalamic nuclei. (c) The TC connectivity matrix (70 × 43) for individual viral tracer injection experiments with verified cortical projections. Each row shows log10-transformed NPVs from one experiment to the 43 ipsilateral cortical targets (columns). Cre line names for each row are in Supplementary Table 5). Unsupervised hierarchical clustering, using Spearman correlation and average linkages, revealed 7 clusters containing thalamic regions with cortical projection patterns resembling the cortical modules. Matrix color map is identical to Fig. 2.

Figure 4.. Corticothalamic projections from layers 5 and 6.

(a,b) CT connectivity matrices (27 × 44) for L5 (a, Rbp4) and L6 (b, average of Ntsr1 and Syt6). Each row shows log10-transformed NPVs from one of the 27 cortical source areas in Fig. 2 to the 44 ipsilateral thalamic target regions (columns). (c) The fraction of true positive CT targets shared by WT (black circle) and each L6 line (yellow) with its Rbp4 anchor is plotted in the box plot (gray). The fraction of positive targets unique to Rbp4 (green) or unique to the L6 line (white) are also shown. Box plots show median and IQR. Whiskers show min and max values. (d) Log NPVs for thalamic targets shared by Ntsr1 and Syt6 were significantly correlated (Spearman r=0.77, p<0.0001). (e) Log NPVs for thalamic targets shared by replicate experiments in the same Cre line < 500 μm apart were significantly correlated (Spearman r=0.84, p<0.0001). (f) The average log NPVs originating from L6 are plotted against L5 for all spatially-matched experiments (Spearman r=0.65, p<0.0001). (g) The matrix shows the relative difference for each source x target connection originating from L5 vs. L6 (L5−L6/L5+L6).

Figure 5.. Corticocortical and thalamocortical target lamination patterns.

(a) Unsupervised hierarchical clustering on relative projection density per layer. Each column is a unique combination of mouse line, cortical or thalamic source area, and cortical target. Connections to agranular (no L4) regions are colored gray for L4. The dotted line indicates where the dendrogram was cut into 9 clusters. (b) Median relative density by layer for each cluster. (c) Number of cortical or thalamic connections in each cluster, plotted on the left and right y-axis, respectively. (d) The frequency of cortical and thalamic targets assigned to each cluster. The dotted line indicates the overall frequency of CC targets in the entire dataset (90.53%). (e) Representative STPT images show axonal lamination patterns from a connection assigned to each cluster from cortex or thalamus. In panels 4, 8, and 9, thalamic axons passing through superficial corpus callosum are indicated (*cc). (f) The relative frequency with which each cortical Cre line and TC projection class appears in the clusters. The fraction of experiments in a cluster belonging to each Cre line/class was divided by the overall frequency of experiments from that Cre line/class. A relative frequency value of “1” (white) indicates that Cre line appeared in that cluster with the same frequency as in the entire dataset. Values <1 (green) indicate lower and >1 (pink) indicate higher than expected frequency in a cluster. Dots indicate significant positive enrichment in that cluster (Fisher’s exact test, p<0.0001). (g) Schematic diagram shows significantly enriched axon lamination patterns associated with each layer and/or class of origin in the source area.

Figure 6.. A hierarchical organization of areas and modules.

(a) Direction mapping results for CC and TC terminal layer patterns. Median relative density by layer for each cluster shown from Fig. 5b. (b) Direction mapping for CT connections. Scatterplot shows the log10-transformed NPV for every CT connection from L5 and L6. Points are color coded by the mapping (FF or FB) predicted from the CC+TC hierarchy. Linear discriminant analysis (red line) assigned connections below = FF and above = FB. (c) Global hierarchy scores for CC connections only (green), compared to the scores when TC and CT connections are sequentially included (pink, blue). Scores for the original, observed, data are shown as single outlined bars. Distributions of hierarchy scores were obtained from shuffled datasets (n=100). The medians of the shuffled distributions estimate the lower bound (0.001, 0.044, −0.002). (d) 37 cortical areas and 24 thalamic nuclei rank ordered by their CC+TC+CT hierarchy scores. Scores for each area using only CC or CC+TC connections are also plotted. Y-axis labels are color coded by module assignment for cortical areas. (e) Network diagram showing interconnections of all cortical visual areas (visual module = light blue, medial module = dark blue). Edge width = relative connection density (from Fig. 1e). The curved lines show outputs (left) and inputs (right) to each node. Nodes are positioned along a single axis based on hierarchical score. (f) Intermodule network diagram. Edge width = sum of connection densities from Fig. 1e.