Hierarchical organization of cortical and thalamic connectivity
Nature, 575, pages195–202(2019)
Hierarchical organization of cortical and thalamic connectivity
Julie A. Harris, et.al.
Allen Institute for Brain Science, Seattle, WA, USA
University of Washington, Department of Applied Mathematics, Seattle, WA, USA
Laboratory of Systems Neuroscience, National Institute of Mental Health, Bethesda, MD, USA
Wenzhou Medical University, Wenzhou, P. R. China
[paraphrase]
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 resource, 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.
Cognitive processes and voluntary control of behaviour originate in the cortex. To
understand how incoming sensory information is processed and integrated with past
experiences and current states in order to generate appropriate behaviour requires
knowledge of the anatomical patterns and rules of connectivity between cortical areas.
Connectomes—\complete descriptions of brain wiring—\exist at different levels of spatial
granularity, from single cells to populations of cells and entire areas (micro-, meso-, and
macro-scale). Common organizational features of macro- and meso-scale cortical
connectivity have been distilled across data sets, often using graph theory approaches to
describe network architecture. For example, cortical areas have unique patterns of
connections (“fingerprints”), connection strengths follow a log-normal distribution
spanning more than four orders of magnitude, and the organization of cortical areas is
modular, with distinct modules corresponding to specific functions.
The concept of hierarchical organization is important for understanding the cortex, and has
inspired the development of neural network methods in deep machine learning. A hierarchy
of cortical areas was first derived by mapping anatomical patterns of corticocortical (CC)
connections onto feedforward and feedback directions. In primate, feedforward connections
were characterized by dense axon terminations in layer (L)4 of the target area, and feedback
connections as dense terminals in superficial and deep layers (avoiding L4). Differences in
the layers of origin are also associated with feedforward and feedback connections. It is still
unclear whether the concept of a cortical hierarchy, which has been derived largely from
sensory systems, can be applied globally across the entire cortex, and how it arises from
connections made by different classes of neurons. Each cortical region comprises distinct
types of excitatory neurons that are largely organized by layers, but also by long-distance
projection patterns: intratelencephalic (IT) in L2–L6, pyramidal tract (PT) in L5, and
corticothalamic (CT) in L6.
Thalamic nuclei make important contributions to cortical function. They serve as a
“relay” for primary sensory information, and are well positioned to influence cortical
information processing through reciprocal or transthalamic loops. Thalamocortical (TC)
projection neurons are classified into three major classes: core, intralaminar, and matrix.
Like CC projections, feedforward and feedback rules have been proposed for TC and CT
projections. Core projections (to L4) are described as “driver” (feedforward) and matrix
projections (to L1) as “modulator” (feedback). For CT connections, input from L6 is
considered feedback, and from L5 feedforward.
We hypothesize that a unifying hierarchical organization across the entire cortex and its
major input structure, the thalamus, is governed by a set of anatomical rules for CC, CT
and TC connections. By using diverse Cre driver mouse lines to selectively label cells from
different cortical layers and classes, we have substantially expanded the Allen Mouse Brain
Connectivity Atlas resource (http://connectivity.brain-map.org1), adding 1,256 new tracing
experiments. Our findings follow analyses of projection patterns spanning nearly the entire
mouse cortex and thalamus, and show how these patterns relate to layer and cell class. We
test the above hypothesis by building a computational hierarchical model using
anatomical rules derived from observations of axon termination patterns. Our results
show that the mouse cortex and thalamus form an integrated hierarchical organization.
We used a genetic viral tracing approach, building on our previously established whole-
brain imaging and informatics pipeline, to map projections originating from unique cell
populations in the same cortical area, and from distinct projection classes in the thalamus.
Our study represents a big step towards a true mesoscale connectome. It will be
informative for future connectome studies with more refined cell types and single cells,
which will no doubt reveal additional principles of cell-type-specific brain connectivity.
With these mesoscale data, we derived several generalizable anatomical rules of cortical
and thalamic connections, and tested whether the organizing principle of a hierarchy
applies to mouse cortex and thalamus.
The cortex is organized as a modular network, which provides a structural view of possible
paths of information flow, but does not impose direction or order onto that flow. By
contrast, a hierarchy implies that interareal connections belong to at least two general types:
feedforward or feedback. Specific anatomical projection patterns were previously
associated with information transmission in these directions in primate and rodent visual
cortex. In our data, we observed many similar patterns. Two patterns that differed were the
superficial layer projections (cluster 1) and the deep layer projections (cluster 9). Felleman
and van Essen noted the occasional superficial-only pattern, but they called it feedback
because it did not involve L4. Our results suggest this pattern is associated with
feedforward. The strength and presence of projections between areas from the
predominantly L4 Cre lines was also unexpected, given canonical circuit diagrams, and
might be explained by varying degrees of layer selectivity. However, by reconstructing the
complete dendritic and axonal morphology of single cells, we directly show that L4
neurons, even spiny stellate cells, can in fact have long-range projections.
The hierarchy that we find is shallower than might have been expected, even with inclusion
of thalamic regions. The difference between the lowest and the highest areas is less than
two full levels, and the all-area hierarchy global score is at 19% between random and
perfectly hierarchical. This might be characteristic of the mouse cortex, given its high
connection density, particularly when considering all non-zero connection strengths. We did
not explicitly include strengths in computing hierarchy, except that weak connections were
removed. Notably, hierarchical position alone does not explain all of the connections of a
given area. This complexity may be why some have argued that the concept of a hierarchy
is overly simplistic for describing functional properties. Given the number of different
connection types that arise from a single area, future computational models that incorporate
more than feedforward and feedback labels will enable further insights into the organization
of brain networks.
Cortical hierarchies were previously derived from classic anterograde or retrograde tracing
without cell-class resolution. Using Cre lines, we have mapped both layer of origin and
target lamination pattern in the same experiment. We found that L2/3 and L4 neurons have
predominantly feedforward layer projection patterns, whereas L5 and L6 neurons have both
feedforward and feedback patterns. However, these general relationships depend on the
specific source–target connection and Cre line. The Cre data set, with all this detail,
produced the most robust hierarchy. However, our results from wild-type mice provide a
solid benchmark for others interested in applying these hierarchical model algorithms to
classic tracing data. The calculation of global hierarchy scores for other data sets will
enable direct comparisons between species and quantitative assessments of how
development or disease might affect hierarchical organization.
Links to Consciousness Discussion