Animal care and surgical procedures were performed in compliance with the National Institutes of Health (NIH) regulations and with the approval of the Institutional Animal Care and Use Committee of Vanderbilt University. Three male and three female adult squirrel monkeys (Saimiri sciureus weighing 600–800 g, 2–9 years old) were used, which were subjects in our prior studies on the connectivity of distal finger pad representations of areas 3b and 1 (Négyessy et al., 2013; Ashaber et al., 2014, 2020; Pálfi et al., 2018). Here, we provide a brief description of the experimental procedures (for details, see Négyessy et al., 2013; Ashaber et al., 2014).

As previously described, each animal was sedated (ketamine, 15 mg/kg, im), placed in a stereotaxic frame, mechanically ventilated with isoflurane anesthesia, and hydrated with lactated ringers via intravenous infusion. Analgesia during surgery was supplied by buprenorphine (0.01 mg/kg, im). Vital signs [peripheral capillary oxygen saturation, heart rate, electrocardiogram (ECG), end-tidal carbon dioxide, respiratory pattern, and temperature] were monitored. After craniotomy [centered at AP 6 mm and ML 15 mm (Gergen and MacLean, 1962)] and durotomy, areas 3b and 1 were located using the central sulcus and blood vessel landmarks. Following the electrophysiological mapping of hand and finger representations, intrinsic signal optical imaging was used to identify the distal finger pad representations of fingers D2–D4 in areas 3b and 1. Biotinylated dextran amine (BDA, 1:1 mixture of 10% 10K and 10% 3K, Molecular Probes, Inc.) in 0.01 M phosphate buffer (PB, pH 7.4) was then injected into a distal finger pad representation in either area 3b or 1 via iontophoresis (3 μA, 7 s on/off cycle for 20 min) (three cases each). In all cases, the core (≤ 300 μm in diameter) of the BDA injection uptake zone included the upper layers (200–390 μm below the surface), while the lower cortical layers were also included in the core in cases of area 1 injections (160–800 μm below the surface) [for more details, see Table 1 in Négyessy et al. (2013) and Ashaber et al. (2014)]. Upon recovery, heat support was provided for the first 12 h with postoperative analgesia supplied by buprenorphine (0.01 mg/kg, im, two times a day) for 3 days. After a 10–20-day survival period, animals were deeply anesthetized before being transcardially perfused with a fixative composed of 4% paraformaldehyde, 0.1% glutaraldehyde, and 0.2% picric acid in 0.1 M PB (pH 7.3). The brains were immediately removed, and the region of interest was then cut from the cortex, flattened parallel to the cortical surface, and postfixed overnight in 4% paraformaldehyde.

Regularly spaced series (at 130–160 μm except one case with 270 μm) of 50-μm thick tangential sections were cut by vibratome sectioning (see Négyessy et al., 2013; Ashaber et al., 2014). A standard avidin-biotin complex (ABC) protocol (Vectastain Elite ABC kit, Vector Laboratories, Inc.) was used to visualize BDA labeling with nickel-intensified diaminobenzidine (NiDAB) (Sigma-Aldrich) as chromogen (for more details on the procedure, see Négyessy et al., 2000, 2013). First, sections were cryoprotected (30% sucrose in PB) and tissue penetration was enhanced by freezing–thawing. Unbound aldehydes were reduced by borohydride (1% NaBH₄ in PB, 30 min), and intrinsic peroxidase activity was blocked with 1% H₂O₂ in PB (30 min). The sections were then incubated in ABC (1:200 in PB, 0.1M, pH 7.4) overnight at 4°C. After the NiDAB reaction, the sections were osmicated (1% OsO₄ (Electron Microscopy Sciences) in PB (pH 7.4) containing 5% sucrose for 60 min) and flat embedded in resin (Durcupan ACM, Sigma-Aldrich).

Our sample of anterogradely labeled axons consisted of 144 3D reconstructed axons forming connections within somatosensory areas 3b and 1 as well as between these two areas after focal injection of BDA in areas 3b (n = 3) and 1 (n = 3) in six squirrel monkeys (Figure 1; Tables 1, 2). An equal number of patch and no-patch axons were selected for analysis; however, a larger number of axonal segments were reconstructed following area 3b (n = 84) injections compared to area 1 (n = 60) injections (Tables 1, 2). This was mainly due to the low number of axons available in case M, where an oblique cut of the sections prevented the reconstruction of infragranular axons with sufficient length (Tables 1, 2, 4). Tables 1, 2 also show the number of intra-areal and interareal axons reconstructed in the supragranular and infragranular layers of the different cases. Sample images of the BDA label around each injection site and the spacing of bouton varicosities are shown in Figure 1.

Axons formed numerous boutons both within and outside the axonal patches, irrespective of the injected area, the pathway (intra-areal vs. inter-areal), or the laminar (supragranular vs. infragranular) location (Figure 2; Tables 3, 4). Figure 2 shows the Neurolucida reconstructions of the population of patch (red) and no-patch (blue) axons from each of the six cases. Major morphological differences appeared between patch and no-patch axons in regard to their length (the distance between the start and end points of the reconstructed segments), direction (how parallel they are), and the number of boutons and bouton density. We found that a larger number of boutons were formed by patch axons compared to no-patch axons (Wilcoxon matched pairs test; p = 0.001, Tables 3, 4). This observation was even more compelling considering that the length of patch axons was significantly shorter than that of no-patch axons (Wilcoxon matched pairs test; p < 0.001, Tables 5, 6). At the level of the supragranular and infragranular layers of the different cases (Supplementary Table S1), the maximum length of patch axons (mean ± SD: 111.42 ± 42.8 μm) was comparable to the minimum length of no-patch axons (mean ± SD: 113.13 ± 43.83 μm) in the sample (Wilcoxon matched pairs test, p = 0.59), i.e., the length distributions overlapped. These comparisons suggested that the different length values of patch and no-patch axons collected from the same sections could not be solely a consequence of the plane of sectioning. Details of the number of boutons and the length of axons in relation to the different categorization of axons (patch, no patch, feedforward, feedback, intra-areal or interareal, and area 3b or 1) and cases are summarized in Tables 3–6.

Only the classification of axons by patch designation (patch or no patch) showed differences in the distribution of the quantitative properties of axons (Figures 3, 4; Supplementary Figures S2–S5). Greater bouton density, convergence, clustering, and more consistent bouton spacing were observed in patch than in no-patch segments. Notably, SD as a measure of consistency of bouton spacing was scaled to the mean and the inter-bouton interval exhibited a similar coefficient of variation (CV = mean/SD) in the case of patch and no-patch axons (mean ± SD, patch: 0.57 ± 0.07; no patch: 0.54 ± 0.03; Wilcoxon matched pairs test, p = 0.45). Axon measures showed thicker axons in no-patch segments and more similar directionality of axons than in patch segments with no differences in tortuosity (Figure 3). These morphological distinctions were independent of whether the projections originated in area 1 or 3b (Figure 4) or whether axons were intra-areal or interareal or feedback (projection from areas 1 to 3b) or feedforward (projection from areas 3b to 1) projections or whether the segments were located in supragranular or infragranular layers (Supplementary Figures S2–S5). Factorial ANOVA supported these observations. After including the categorical variable injection area (areas 3b, 1), patch designation, and pathway type (intra-areal and inter-areal), factorial analysis resulted in a significant major effect only between patch and no-patch axons (Table 7). Interactions were not significant. Post hoc comparisons resulted in significant differences in all variables except the measure of tortuosity (Table 7).

Based on these observations, we determined the relative contribution of the different variables to the patch-like properties of axons. To do this, we utilized PCA to account for variability, followed by stepwise logistic regression to determine the predictors of patch designation across the different variables.

In this study, we explored individual (tortuosity, thickness, bouton density, variability of inter-bouton distance, and clustering tendency of boutons) and collective (distance of the closest boutons of unconnected axonal segments and directional difference of the axonal segments) morphological variables of bouton-forming axonal segments that are relevant to signal transmission and best predict the properties of a patch. We found that, except for tortuosity, all other variables examined differed significantly between axons with varicosities within patches and outside patches. Most notably, we identified proximity of boutons of unconnected axonal segments (bouton-convergence) as the variable with the strongest predicting power of patch designation. In addition to bouton-convergence, bouton density also contributes significantly to the predictive power of an axon's patch designation. As patches are characteristic of mesoscale connectivity in the cerebral cortex of carnivores and primates, an important finding is that in the different pathways (intrinsic, feedforward, or feedback) and layers (supragranular and infragranular) studied axons do not vary in their properties; axons are differentiated exclusively by their patch designation. Taken together, patch and no-patch domains of horizontal, long-distance axons exhibit distinctive morphological properties, which suggest their complementary role in cortical computation and support the modular organization of cortical interactions.

One of the methodological limitations is that sampling of the serial sections used in this study does not allow tracing back to the origin of the axon. Therefore, it is possible that separately reconstructed axonal segments belong to the same neuron. While this possibility cannot be ruled out, single-cell reconstructions show that recurrent horizontal axonal branches of cortical spiny neurons are oriented in different rostro-caudal and medio-lateral directions, instead of raising multiple parallel collaterals (Buzás et al., 2006; Binzegger et al., 2007). Based on this observation, we expect our sample to include at most a negligible number of linear no-patch axonal segments, which belong to the same neuron. Within patches, the chance that the sampled segments belonging to the same axonal arborization may be higher. However, the distal branches of fully reconstructed neurons typically branch in a Y-shape and are usually divergent, unlike our sample, in which patch axons are overlapping (Buzás et al., 2006; Binzegger et al., 2007). Studies show a strong laminar preference of horizontal connections of neurons in different layers of the visual cortex (Kisvárday, 2016; Martin et al., 2017). In this study, with a single exception, we reconstructed patch and no-patch axons from more than one section with wide inter-section gaps (i.e., in widely separated layers), and in different horizontal locations, which also decreases the possibility of sampling axons from the same neuron.

Our approach does not allow the unequivocal identification of the type of parent neuron of the axonal segments studied. Although, as described in section Introduction, long-distance horizontal connections are known to be mostly formed by pyramidal neurons, there are transcolumnar and projective GABAergic neurons that can be anterogradely labeled in our experiments (DeFelipe et al., 2013). Even if this is the case, the number of labeled GABAergic axonal segments probably forms a small fraction of the total number of axonal segments studied. The population of GABAergic neurons with a long-distance projection is a subfraction of the total number of GABAergic cells, which form 10–30% of the neurons in the cerebral cortex (DeFelipe et al., 2013). Furthermore, qualitative and quantitative morphological observations are in agreement with the horizontal distribution of axons of individual pyramidal cells. As shown previously, the axonal boutons of pyramidal neurons mostly appear in the form of varicosities similar to the one found here (Kisvárday et al., 1997). Furthermore, in cat visual cortex, single-cell labeling revealed that intrinsic horizontal connections of pyramidal neurons consist of linear segments that terminate in a rich arborization of short, distal branches (Buzás et al., 2006; Muir and Douglas, 2011). Also in cat visual cortex, the horizontal pattern of boutons formed by nearby pyramidal neurons is consistent with the distribution of patch and no-patch domains described here (Kisvárday et al., 1997; Buzás et al., 2006). Together, these observations suggest that our sample consisted of axons originating from pyramidal neurons. Specifically, no-patch axons probably represent the linear segments whereas patch axons can correspond to the terminal arborizations of the horizontal axonal connections.

The chance of including labeled axons by retrograde backfilling of neurons projecting to the injection site has been discussed in our previous studies (Négyessy et al., 2013). Our observations about the apparent lack of thalamocortical labeling and secondary labeling of patches formed by neurons after probable backfilling suggest that these factors did not significantly influence our results. Horizontal sectioning may also have resulted in inconsistencies in the laminar localization of the axonal segments, as discussed earlier (Pálfi et al., 2018). In that study, we found a laminar preference of connectivity similar to that as shown before (Burton and Fabri, 1995). In our samples, the majority of patches and boutons were localized in the supragranular layer (Pálfi et al., 2018), which is consistent with the prevalence of patchy organization of horizontal connections in layers 2/3 (Douglas and Martin, 2004; Kisvárday, 2016).

It should also be noted that our thickness measurements of the axons probably indicate larger values than in reality due to the relatively low resolution of the digital camera attached to the Neurolucida setup and also to the relatively larger scattering of light through the osmicated and resin-embedded sections. Therefore, these values are not comparable to those published in the literature (Anderson et al., 2002; Innocenti and Caminiti, 2017; Koestinger et al., 2017). However, this systematic effect of measurement would not affect the results of the comparisons made in this study.

Our analyses indicate that the tortuosity of bouton-forming axonal segments of pyramidal cells is an invariable property and, therefore, neurons save material and use other strategies in finding their specific targets (Anderson et al., 2002; Budd et al., 2010). We found that all variables related to bouton placement differed significantly between patch and no-patch axons. It should be noted that the CV of inter-bouton intervals observed here is in agreement with the 0.5 reported by Shepherd et al. (2002) and indicates that bouton spacing is neither completely random (CV = 1) nor regular (CV = 0). Our findings support that the subrandom distribution of interbouton intervals is a feature of unmyelinated axonal branches in the brain (Hellwig et al., 1994; Shepherd et al., 2002). Even though bouton spacing variability is a constant fraction of the mean spacing along axons, irrespective of the type of neuron or branch order, only boutons on the terminal branches of lateral connections form clusters (Anderson et al., 2002; Shepherd et al., 2002; Binzegger et al., 2007). These observations are supported by our results showing larger bouton densities, clustering tendency, and smaller variability of interbouton intervals along the axonal segments within patches compared to no-patch segments. Furthermore, boutons of unconnected axonal segments are significantly closer to each other within patches than outside of it. These findings are in line with observations showing the overlapping distribution of terminal axon arborizations of a population of nearby pyramidal neurons in cat visual cortex (Kisvárday et al., 1997; Buzás et al., 2006). By showing higher bouton densities within patches compared to no-patch domains, we confirmed our previous observations in primate somatosensory cortex (Négyessy et al., 2013; Ashaber et al., 2014). In these previous studies, we showed that bouton density is significantly higher in axonal patches than in any other locations including regions with BDA-labeled tracks of long-range horizontal axons.

In addition to bouton spacing, no-patch axons were significantly thicker than patch axons, which is in accordance with axon branching order, given that patches are formed by the most distal branches (Buzás et al., 2006; Binzegger et al., 2007). The direction of axonal segments also differs between patch and no-patch domains. The highly varying path of an axon within patches can promote convergence in contrast to the nearly parallel direction of no-patch axons, especially considering that axons were reconstructed in ~50-μm thick sections. Viewing from a larger scale, no-patch axons extend radially from an injection site similar to the distribution of patches around the injection site (Négyessy et al., 2013; Ashaber et al., 2014; Pálfi et al., 2018).

Our application of multivariate data analysis techniques further revealed what makes patch and no-patch domains fundamentally different cortical sites. PCA resulted in a clear clustering of patch and no-patch axons based solely on the first PC, which explained more than half of the variance in the data. Although, with the exception of tortuosity, all variables exhibited a relatively high load on the first PC, the analysis showed that bouton density and bouton-convergence have the largest distinguishing power. To see which of these factors has the highest predicting power of patch designation, we applied stepwise logistic regression. The model identified bouton-convergence with the highest predicting power. Bouton density also contributed significantly to the model fit; however, this parameter had far less predicting power, as shown by the observed probabilities. Altogether, these findings indicate that patches are special cortical loci that can attract axons of a population of nearby pyramidal neurons, which results in the convergence of bouton clusters formed by the distal terminal arborizations (Buzás et al., 2006; Binzegger et al., 2007). Similar evidence for bouton distributions obtained by labeling a small population of neurons and pooling single neuron reconstructions also suggest a highly overlapping distribution of the final terminal arborizations in the cat visual cortex (Kisvárday et al., 1997; Buzás et al., 2006; Martin et al., 2014, 2017).

One of the major principles governing patch formation is probably selectivity for different molecular cues. Reaction-diffusion models can explain patch formation (Bauer et al., 2014). However, pyramidal neurons exhibit a large variability on the basis of molecular composition (Lodato and Arlotta, 2015; Zeng and Sanes, 2017). Considering the multitude of component molecules, some degree of neuronal similarity can be dispersed throughout the cortical mantle and result in the diversity of target sites found in the case of no-patch axons. In such a scenario, the strength of molecular similarity can determine the specificity of target sites of a population of cortical neurons. Another probable factor determining the target neurons is the similarity of functional properties with the presynaptic neuron (Kaas, 2012). Similar rules seem to govern the formation of long-distance interareal, feedforward, and feedback connections. The understanding of how chemical and activity dependent mechanisms interact in the formation of the different cortical circuits requires further investigation.

In the sensory cortex, feedforward interarea input is responsible for activation hotspots or “imprint” (Chavane et al., 2011) at the population level and the formation of the classical receptive field (RF) of neurons (somatosensory cortex: Favorov and Whitsel, 1988; Chen et al., 2003; Friedman et al., 2008; and the visual cortex: Kisvárday et al., 1997; Angelucci et al., 2017). On the other hand, lateral intra-areal connections are thought to transmit contextual information and are responsible for the configuration of extra-classical RF together with feedback interarea connections (Kisvárday, 2016; Angelucci et al., 2017; Chavane et al., 2022). However, how synaptic computations bring the function of these pathways into effect is not completely understood (Douglas and Martin, 2007; Boucsein et al., 2011; Harris and Mrsic-Flogel, 2013; Muller et al., 2018; Rockland, 2018; Chavane et al., 2022). In this study, we made a distinction between the efferents of a population of nearby projection neurons, which is independent of whether the projection pathway is intra-areal, feedforward, or feedback. We found a major difference between patch and no-patch axonal segments. Patches are identified as convergence spots of efferents formed by a population of neighboring neurons, in contrast to the rather parallel spread of neighboring no-patch axons. Notably, this distinction applies to all connections studied, specifically intra-areal, feedforward, and feedback. Also, patch and no-patch characteristics of the axons are independent of the laminar localization. These observations suggest a fundamental similarity in the computation performed by the different pathways. In contrast to the structural properties of the feedforward thalamocortical synaptic boutons, which support their driving functions (Negyessy and Goldman-Rakic, 2005; Petrof and Sherman, 2013), the ultrastructural characteristics of the cortico-cortical synaptic boutons are more homogenous (Ashaber et al., 2020). These electron microscopy findings provide further hints to the importance of convergence in cortical computation.

The convergence of afferents is recognized as the substrate for increasing the effectiveness of signal transmission through the neuronal network. This recognition is supported by the observation that cortical microstimulation can induce a patchy pattern of activation (Roe et al., 2015; Xu et al., 2019; Friedman et al., 2020). Patches can provide the functional bias found in the case of orientation tuning in the visual cortex (Kisvárday, 2016). In contrast, no-patch axons can diversify the input via conveying contextual information to postsynaptic neurons (Kisvárday, 2016; Chavane et al., 2022). Consequently, neurons with overlapping RFs can significantly shape the response specificity of sensory cortical patches. The functional similarity of the neurons forming the patchy projection argues against the lack of specificity of sensory cortical interactions (Chavane et al., 2022). Techniques sensitive to subthreshold changes in membrane potential can reveal input diversity provided by non-convergent projections of no-patch axons (Chavane et al., 2022). The recognition that no-patch regions of a population of neurons probably interact with different functional neuronal populations can explain the formation of cortical circuits that can support both functional specificity and variability. This observation is consistent with the role of patch domains in the emergence of signal correlations of the neuronal responses whereas the no-patch domain can contribute to noise correlations (Panzeri et al., 2015; Bányai and Orbán, 2019).

Neural activity can spread at different velocities inside and outside the patches due to the different axon thicknesses. However, given the short length of the distal branches, it is reasonable to assume that larger bouton densities compensate for the slower conduction speed within the patches. Such compensation in conduction velocity would critically depend on the relative diameter sizes between the axon and the bouton, where a larger difference results in a longer lag in signal propagation (Segev and Schneidman, 1999; Alcami and El Hady, 2019). In addition, a bouton density higher than a critical value can result in a nonlinear increase in conduction delay (Segev and Schneidman, 1999). Further uncertainties in conduction speed can be introduced by activity-dependent changes in axon thickness and bouton size (Alcami and El Hady, 2019). How these factors affect the spread of activity along patch axons remains to be tested. Until these issues are clarified, as a working hypothesis, one can assume a similar dynamic in the two bouton-forming axonal domains of the cerebral cortex.

In our previous studies, we explored the direction of information flow within the circuitry of areas 3b and 1 (Wang et al., 2013; Pálfi et al., 2018). The picture that emerged from that data is consistent with the findings of Polack and Contreras (2012) who found that in the mouse visual cortex information is rapidly forwarded to numerous cortical areas, simultaneously triggering slower serial processing within areas, which can be influenced by the later feedback component in population activity (Semedo et al., 2022). Patchy organization is a characteristic feature of the wiring motif of the mesoscale network of the cerebral cortex of higher order mammals (Roe, 2019). Patches provide distributed modules of axonal convergence that are well positioned to function as sites for the synchronization of neural activity proceeding at rapid and slower time scales. In contrast, due to horizontal dispersion, no-patch axons are more suitable to modulate large-scale dynamics of cortical synchronization.