Technical Comments

Response to Comment on “Principles of connectivity among morphologically defined cell types in adult neocortex”

See allHide authors and affiliations

Science  09 Sep 2016:
Vol. 353, Issue 6304, pp. 1108
DOI: 10.1126/science.aaf6102


The critique of Barth et al. centers on three points: (i) the completeness of our study is overstated; (ii) the connectivity matrix we describe is biased by technical limitations of our brain-slicing and multipatching methods; and (iii) our cell classification scheme is arbitrary and we have simply renamed previously identified interneuron types. We address these criticisms in our Response.

We will address Barth et al.’s (1) three critiques of our Research Article (2) in turn.

(i) The strength of our study is the number of connections tested (>11,000) in one species, cortical area, and age range, using the same experimental protocol. We agree that our study is not exhaustive, because it is focused on synaptic connectivity between interneurons in layers 1, 2/3, and 5, and we point this out in the abstract of the original paper. However, we do believe that no single study to date has presented a more complete wiring diagram of the adult neocortical microcircuit.

(ii) Severed connections are a well-known limitation of slice recordings, which we discussed explicitly on p. 3 of the supplementary materials (SM) for (2). However, we do not agree that our study is substantially more susceptible to connection loss than previous patching studies, most of which also used 300- to 350-μm slices [we found fewer than half a dozen multipatching studies using 450- to 500-μm slices (35)]. Although cell depth is not reported in many of these studies, the depths at which we recorded (15 to 60 μm) seem typical (6). We patched cells at a wide range of intersomatic distances [see fig. S13A in (2)]; this range reflects the sparsity of labeled interneurons and is not a limitation of octuple patching.

Like many previous studies, we used current-clamp mode to measure connectivity, which suffers less distance-dependent voltage attenuation compared with voltage clamp (7). Using potassium internal solution allowed us to characterize spiking properties and test bidirectional connectivity, whereas cesium-based internal solutions would prevent the patched neuron from firing. We discussed potential problems of voltage attenuation [SM for (2), p. 9], such as the underestimation of synaptic event amplitudes and potential nondetection of small synaptic events on the distal dendrites [e.g., connections from Martinotti cells onto layer 5 (L5) pyramidal neurons].

To directly address Barth et al.’s concern that the connectivity matrix may be distorted due to slice cutting, we simulated the effect of slicing based on interneuron morphology (8) (Fig. 1A) [see also (6)] for the worst-case scenario where two neurons (pre- and postsynaptic) are both located 15-μm deep from the cutting surface. The resulting correction factor was similar across pairs of cell types (Fig. 1B), more or less scaling the entire connectivity matrix by a factor of 1.36 ± 0.10 (mean ± SD) (Fig. 1, C and D). Given these results and the fact that our connectivity principles mostly rely on the presence or absence of connectivity between specific types regardless of its magnitude, we believe that our connectivity principles are valid rules of cortical organization.

Fig. 1 Slice cutting does not distort the connectivity matrix.

(A) Schematic of the correction factor estimation. For each pair of cell types (here, L23 Martinotti cells and L23 BCs), we compute presynaptic axonal densities (blue; here, L23 Martinotti cell) and postsynaptic dendritic densities (violet; here, L23 BC) by rasterizing all neurons of those particular types, randomly rotating the single points around an axis orthogonal to the pial surface, and binning the result in a three-dimensional histogram. Only one neuron per type is shown here for clarity. The bottom contour plots depict the marginal densities for the dendrites (blue) and the axons (violet) of the example cell type pair after radial symmetrization. The neurons of each type were centered on their cell bodies, and the different types were offset by different amounts measured in our experiments. For each offset, we computed the product of the two densities (overlap density, top contour plot) and the relative volume that remains after slice cutting (yellow area under the black marginal density, 1 – q). The inverse of this value is used as a correction factor for the connection probabilities 1/(1 – q). The gray rectangular solid represents a single 300-μm slice. A, D, and L represent the anterior, dorsal and lateral axes, respectively. To mimic the worst-case scenario, the neuron somata were placed 15 μm away from the cutting surface (somata not shown for clarity). (B) Correction factor was weakly correlated with the connection probability. The resulting correction factors had a mean and SD of 1.36 ± 0.10. Pearson correlation between connection probability and the correction factor is 0.02. (C) Original connection probability (p) matrix as shown in (2). (D) Corrected connection probability p/(1 – q) matrix. The structure of the connection probability matrix is very similar to the original one shown in (C). Although cutting does change the connection probabilities, our analysis indicates that it does not affect our connectivity principles because the entire matrix is scaled globally instead of overemphasizing single-connection probabilities.

A key concern of Barth et al. is our finding of much lower connectivity between excitatory cells compared with previous studies, which we believe could reflect true differences in connectivity between adult and juvenile animals. Ours is the first large-scale slice electrophysiology study on adult mice (≥2 months). Most others were typically done on rodents 14 to 21 days postnatal (P14 to P21). [Based on the weight range, the study that Barth et al. cite (9) appears to use juveniles as well.] When we tested connectivity among layer 5 pyramidal neurons in juvenile (P15 to P20) slices, we found it to be around 10% for nearby neurons (18 of 186 pairs in <50 μm radius), comparable to previous reports (10, 11) and significantly higher than the connectivity we measured in adult slices using the same technique (0 of 44 pairs in <50 μm radius; P < 0.02, Mann-Whitney U test). However, synaptic weights in adult pyramidal cells could be smaller in amplitude, or synapses could be electrotonically further from the soma, both of which might put them below the detection limit. Slicing could also affect adult tissue differently than juvenile tissue. For example, pyramidal cells may be more likely to survive in juvenile slices. Therefore, we generally agree with Barth et al. that a detailed comparison of connectivity in juvenile and adult animals is needed. Improved in vivo approaches, such as rabies virus circuit mapping or large-scale electron microscopy reconstruction (1214), will be crucial to resolve the issues mentioned above, including the rare occurrence of gap junctions in our adult slices.

(iii) The lack of a unified taxonomy of cell types has been a major challenge in neuroscience (15, 16). Our interneuron classification was based on morphological features (1719) and was validated by an automated method using cross-validated logistic regression. With random class labels, the classification accuracy dropped to ~48% on average (from ~97%), suggesting that our morphological classification scheme is robust. When we projected the morphological features on the weight vector of the classifier, we found a clear clustering of cell types, further corroborating our method. Finally, our classification was confirmed independently for L1 interneurons by a recent single-cell transcriptomics study (20). Some of the morphological types we found might be further subdivided, but more data are needed. All fully reconstructed neurons from our study (n = 298) have been submitted to and will soon be publicly available.

Barth et al. are correct that we are not the first to study L5 interneurons. For example, previous studies (21, 22) recovered a few L5 parvalbumin-expressing (PV) neurons that resemble horizontally elongated cells (HECs). However, the validation of a morphological cell type requires a much larger amount of data. By recovering the morphology of many L5 neurons, we found that L5 PV neurons have three distinct morphologies (axon confined locally, axon extending horizontally, and axon projecting to both L5 and L2/3) with distinct connectivity profiles, suggesting that L5 PV neurons form a heterogeneous group. The L5 basket cells (BCs) were similar to L2/3 BCs, whereas shrub cells (SC) and HECs were different in every respect. It is unlikely that the fast-spiking SCs reported in our study are identical to the small and nest basket cells reported in (23), because small and nest basket cells often display an adapting regular-spiking firing pattern (unlike SCs) and were observed in neocortical L2/3 and L4 (unlike SCs) (23). In adult visual cortex, we found that L5 vasointestinal peptide (VIP-cre/Ai9 mice) (24) cells were very sparse, and the few we studied did not exhibit obvious bipolar morphology as previously reported in somatosensory cortex (25). Therefore, more work is needed to study the L5 VIP types.

Barth et al. point out that we did not find certain previously described interneuron subtypes [namely, cholecystokinin–expressing (CCK+) BCs and long-range γ-aminobutyric acid–releasing (GABAergic) interneurons] in our study. CCK+ BCs have been well documented in the hippocampus (26) but are less well characterized in neocortex. CCK+ BCs in the neocortex have been reported in rats (27, 28) and CCK+ cells are thought to be rare [only 5% of all GABAergic interneurons in adult mouse visual cortex (29)] and heterogeneous. None of the BCs in our data set appear to fit the electrophysiological profile of CCK+ BCs. If they exist in adult visual cortex, they might have been missed in our study because of their low prevalence. Long-range GABAergic cells in the neocortex are also rare (<0.5% of all interneurons), and 60 to 90% of them are located in L6 and the white matter (3032), neither of which were targeted in our study. If they do exist in L1 to L5, they could be easily missed in our study because their main axon may be cut during slicing.

In summary, we believe that many of the potential confounds raised by Barth et al. apply quite generally to in vitro patching experiments and do not especially bias our study. Beyond their specific technical concerns, the main message of the Barth et al. Comment seems to be that more work is needed to validate and extend the cortical connectivity matrix. On this point, we respond with enthusiastic agreement. By continued application of the multipatching approach and the addition of complementary approaches in the intact brain (1214, 3336), we hope to asymptotically approach the true underlying average connectivity matrix and to identify variations linked to species, ages, and cortical areas. These are exciting times, and we look forward to participating in this common concerted effort.


View Abstract

Stay Connected to Science

Navigate This Article