Changing Connectomes: Evolution, Development, and Dynamics in Network Neuroscience

Changing Connectomes

Changing Connectomes

Evolution, Development, and Dynamics in Network Neuroscience

Marcus Kaiser

The MIT Press

Cambridge, Massachusetts

London, England

© 2020 Massachusetts Institute of Technology

All rights reserved. No part of this book may be reproduced in any form by any electronic or mechanical means (including photocopying, recording, or information storage and retrieval) without permission in writing from the publisher.

This book was set in ITC Stone Serif Std and ITC Stone Sans Std by New Best-set Typesetters Ltd.

Library of Congress Cataloging-in-Publication Data

Names: Kaiser, Marcus, author.

Title: Changing connectomes : evolution, development, and adaptations in network neuroscience / Marcus Kaiser.

Description: Cambridge, Massachusetts : The MIT Press, [2020] | Includes bibliographical references and index.

Identifiers: LCCN 2019059139 | ISBN 9780262044615 (hardcover)

Subjects: LCSH: Neural networks (Neurobiology) | Developmental neurobiology. | Nervous system—Evolution. | Brain—Growth. | Brain—Diseases. | Brain—Degeneration.

Classification: LCC QP363.3 .K34 2020 | DDC 612.6/4018—dc23

LC record available at https://lccn.loc.gov/2019059139

10 9 8 7 6 5 4 3 2 1

d_r0

For Eileen and Felicitas

Contents

Preface

1 Introduction

Part I Connectome Structure

2 Features of Complex Networks

3 Evolution of Neural Systems

4 Organization of Neural Systems

Part II Connectome Maturation

5 Brain Development

6 Layer Formation

7 Axonal Growth

8 Formation of Hubs

9 Module Formation

10 Cortical Folding

Part III Connectome Changes

11 Development and Aging

12 Neurodevelopmental Disorders

13 Neurodegenerative Disorders

14 Recovery from Injury

15 Brain Stimulation Effects

Glossary

References

Index

Preface

The human brain undergoes massive changes during brain development, ranging from early childhood and adolescence to old age. The field of network neuroscience has provided snapshots at different stages, and, through data analysis and computational modeling, mechanisms leading to changing connectomes begin to emerge.

This book aims to provide an overview of network features of the brain, how these features emerge during evolution and brain development, and how alterations ranging from normal aging to developmental and neurodegenerative disorders affect brain networks. The book will give biomedical researchers, who are often only aware of neuroimaging research in humans, information about the knowledge on connectome development in other species. By including an introduction to concepts of network analysis, it will also be accessible to researchers who are new to the field of connectomics. Finally, through Matlab/Octave code examples (available at https://mitpress.mit.edu/changing-connectomes), it will allow computational neuroscience researchers to understand and extend the shown mechanisms of connectome development.

I took my first step into connectomics in the summer of the year 2000, when I did a research internship, while still an undergraduate at Ruhr University Bochum, with Malcolm Young at Newcastle University. In these early days, only connectivity from animal models was available, ranging from the roundworm Caenorhabditis elegans to rat, cat, and rhesus monkey. When I worked with Claus Hilgetag during my PhD studies from 2002 to 2005, knowing about how only a single human brain would be connected seemed like a distant dream. Nowadays, following pioneering studies such as the Human Connectome Project in the United States, large data sets of human brain networks in health and disease are available. Within Europe, the UK Biobank Imaging Study is currently collecting structural and functional brain connectivity data of 100,000 subjects, aiming to complete data collection in 2022. Finally, longitudinal studies of brain development before birth, in early childhood, and during the life span are being performed.

Writing a book related to the fastest growing area of brain research, network neuroscience, is challenging. In particular, given that writing books happens along with other academic commitments, you might ask, "I wonder who they found to pull that off?" I would therefore like to thank the School of Computing at Newcastle University for granting me one semester of research leave to work on this book. I am also indebted to my family, in particular during the final months of finishing this book.

For a first step toward neuroscience and connectome research, I would like to thank Klaus-Peter Hoffmann and Malcolm Young, respectively. Much of this work relies on fruitful discussions and collaborations with my colleagues, mainly Claus Hilgetag, but also Rolf Kötter, Olaf Sporns, Herbert Jaeger, Miles Whittington, and Arjen van Ooyen. I also benefited from discussions with current and former members of my research team. For providing feedback on parts of this book, I would like to thank Alexandros Goulas, Ann-Shyn Chiang, Arjen van Ooyen, Bruno Mota, Cheol Han, Cornelis Stam, David Willshaw, Georg Striedter, Hongkui Zeng, John-Paul Taylor, Larry Swanson, Lisa Ronan, Markus Butz-Ostendorf, Michael Nitsche, Olaf Sporns, Petra Vertes, Roman Bauer, Sven Bestmann, Thiebaut de Schotten, and Yujiang Wang.

Finally, I would also like to thank my editor at the MIT Press, Bob Prior, for his support in getting this book project off the ground and landing safely now.

1 Introduction

How does the brain change over time? While researchers are beginning to understand more about how the set of connections between neurons and regions, the connectome, is organized, this book describes how such an organization arises during development and evolution, how it changes in health and disease, and how external interventions can alter its architecture. For this aim, we describe mechanistic insights, based on computational models and experimental studies, that can link the changes at the local level of individual neurons to the observed large-scale alterations in connectivity between brain regions.

Over the past several years, connectome information about large cohorts of human subjects and patients at different stages of brain development or disease progression has become available. There are continuing efforts to increase data quality, to enable data sharing, and to analyze brain network architecture. There is increasing availability of longitudinal data, measuring the same subjects at multiple time points. Along with connectome data of the early stages, before and shortly after birth, this allows us for the first time to observe the development of connectomes.

Given the availability of both data on brain networks and tools to describe the organization and behavior of these networks, the field of network neuroscience can inform how we look at connectome changes. There are different approaches that one can take to analyze connectome changes. First, one may look at changes in the network organization at different time points using network science approaches. This analysis of network features gives a first insight into which network functions—for example, integration and segregation of information flows—might be altered. Second, one may simulate neural activity within the network to understand how changes in network structure influence network behavior. Finally, one may use computational models to evaluate the mechanisms that lead to the observed changes in network structure. Such approaches can help us to understand how connectome changes arise and change brain function, and they could suggest hypotheses that can be tested experimentally in the future.

There are also more direct applications of network neuroscience in understanding connectome changes. As will be described, there are distinct network changes for brain disorders. Computer models will be essential to inform diagnosis and treatment of individual patients. As each case is different, it is impossible to have an experimental animal or clinical human study with exactly the same condition as found in an individual. In addition, given dozens of variables that play a role, and with interactions between variables influencing each other, it will be impossible to provide mathematical equations that describe the relationships between all variables. The only solution for better diagnosis and treatment of individual patients is a personalized computational representation, based on connectome, physiome, genome, and other available data. With such a system, the plausibility of different disease origins and the outcomes of different interventions can be tested in silico to find the most suitable option for an individual patient.

The study of mechanisms of network changes can also be helpful for designing experimental studies, for understanding the link between brain network structure and performance, and for improving the design and development of artificial neural networks. The range of applications that can benefit from an understanding of connectome changes includes, for example, the following:

• The design of artificial neural networks—for example, for deep learning—includes features of connectomes such as layers but only includes a small subset of mechanisms for network growth and development. This particularly limits the ability of such systems to provide more complex behavior such as multimodal integration, adaptivity to new environments, and learning from small training data sets. Building networks that grow based on mechanisms identified in biological neural networks might provide new breakthroughs in this field.

• Knowing about the link between connectome structure and function will help us to understand why network features arise during individual brain development and during the evolution of neural systems. Not all network features might have a direct or strong link to cognitive processing.

• Knowing the developmental origin of brain diseases provides another biomarker that can be used to help with diagnosis, the stratification of the patient cohort, and treatment planning. Understanding the developmental pathways of network changes could, given the current brain network of a patient, help to predict which factors played a role in the genesis of these connectome changes.

• Understanding the mechanisms that lead to network changes can help to improve brain function following brain injury by facilitating the design of rehabilitation interventions that increase positive network changes while trying to prevent network changes that have a negative effect on brain function.

• For brain stimulation, while there are some models about immediate stimulation effects, during and shortly after stimulation, it will be crucial to understand long-term effects in order to predict effects and minimize negative side effects.

In part I, before we can look at connectome changes, we first need to describe how brain networks can be measured and how we can analyze their features. Chapter 2 provides an overview of network reconstruction and of the analysis of topological and spatial features. Furthermore, I show how activity in these networks can be modeled by giving a brief overview of dynamic features of brain networks. Chapter 3 shows how topological features arise during brain evolution, starting with simple nerve nets and moving on to modular and hierarchical networks. Chapter 4 gives an overview of the architecture of brain networks for the organisms for which we already have full or partial information about their connectomes: Caenorhabditis elegans, fruit fly, pigeon, mouse, rat, ferret, cat, rhesus monkey, marmoset monkey, and human. Part II discusses the maturation of network features during individual brain development. Chapter 5 shows how regional patterns such as cortical maps can be formed and how genetic factors, competition, and homeostasis can induce these patterns. Chapter 6 shows how layers can form; it includes experimental and computational results indicating the roles of cell growth, cell migration, and cell death. Chapter 7 looks at axon growth and the formation of synaptic connections determining principles of initial connection establishment. Chapter 8 looks at network hubs, outlining different hub types and different mechanisms that can generate hubs during brain development. Chapter 9 describes how modules, enabling segregated information processing, can arise due to developmental time windows and genetic factors. Chapter 10 describes how cortical folding changes during development and what principles and mechanisms might cause these changes. Part III looks at how connectomes change during the life span in health and disease and how interventions can interfere with these processes. Chapter 11 is about healthy brain development, outlining changes until adulthood. Chapter 12 talks about changes due to neurodevelopmental disorders such as schizophrenia, autism spectrum disorders, major depression, epilepsy, and Tourette’s syndrome as well as about underlying mechanisms for these changes. Chapter 13 deals with age-related disorders such as Alzheimer’s disease, Lewy body dementia, and Parkinson’s disease as well as with models of disease progression. Chapter 14 describes how connectomes react to lesions, caused by stroke, traumatic brain injury, or loss of peripheral input, and how computational models can be used to test underlying mechanisms for these changes. Chapter 15 highlights the emerging role of brain stimulation, being used on patients but also on healthy subjects, in altering the dynamics but also the topology of brain networks.

How can we determine brain connectivity and how can we analyze brain networks? The next chapter will show how structural and functional connectivity can be determined and how topological, spatial, and dynamic characteristics of the network can be analyzed.

I Connectome Structure

The connectome architecture can be studied in several ways. It is important to keep in mind that scope, resolution, and parcellation of the nervous system influence the observed network features. The scope of the network could range from the cortex and subcortical structures to a network that also includes the peripheral nervous system. Therefore, when comparing between different studies in humans or between different species, some changes in the network are already due to such differences.

While we have a complete network of the nervous system for C. elegans, studies of the human brain are usually limited to the cortex and some subcortical structures. Even for the central nervous system in humans, including the cerebellum, which has 10 times as many neurons as the telencephalon (Herculano-Houzel, 2009), leads to a completely different network. The resolution could be at the level of fiber tract connectivity between regions, neuronal connectivity between smaller patches of brain tissue such as cortical columns, or axonal connectivity between individual neurons.

With higher resolution, the number of network nodes increases drastically: at the regional resolution a human cortical network might consist of 100 regions, whereas at the resolution of individual neurons it consists of 10 billion nodes. Correspondingly, the edge density, the proportion of existing connections relative to all possible connections, changes from 10% at the regional resolution to 0.0001% at the neuronal resolution.

Finally, even within a resolution level, parcellation schemes lead to a wide range of network nodes. For the human brain, regional parcellations of the cortex vary from 68 regions (Hagmann et al., 2008) to 360 regions (Glasser et al., 2016) with a wide range of different anatomical, functional, or multimodal parcellation approaches (Eickhoff, Constable, and Yeo, 2018; Eickhoff, Yeo, and Genon, 2018).

2 Features of Complex Networks

The set of connections in neural systems, now called the connectome (Sporns et al., 2005), has been the focus of neuroanatomy for more than a hundred years (Ramón y Cajal, 1892; His, 1888). However, it has attracted recent interest due to the increasing availability of network information at the global (Felleman and Van Essen, 1991; Scannell et al., 1995; Burns and Young, 2000; Tuch et al., 2003) and local levels (White et al., 1986; Denk and Horstmann, 2004; Lichtman et al., 2008; Seung, 2009) as well as the availability of network analysis tools that can elucidate the link between structure and function of neural systems. Within the neuroanatomical network (structural connectivity), the nonlinear dynamics of neurons and neuronal populations result in patterns of statistical dependencies (functional connectivity) and causal interactions (effective connectivity), defining three major modalities of complex neural systems (Sporns et al., 2004). How is the network structure related to its function, and what effect does changing network components have (Kaiser, 2007)? Since 1992 (Achacoso and Yamamoto, 1992; Young, 1992), tools from network analysis (Costa, Rodrigues, et al., 2007) have been applied to study these questions in neural systems.

What are the benefits of using network analysis in neuroimaging research? First, networks provide an abstraction that can reduce the complexity when dealing with neural networks. Human brains show a large variability in size and surface shape (Van Essen and Drury, 1997). Network analysis, by hiding these features, can help to identify similarities and differences in the organization of neural networks. Second, the overall organization of brain networks has been proven reliable in that features such as small worldness and modularity, present but varying to some degree, are found in all human brain networks (and those of other species, too). Third, using the same frame of reference, given by the identity of network nodes as representing brain regions, both comparisons between subjects as well as comparisons of different kinds of networks (e.g., structural vs. functional) are feasible (Rubinov and Sporns, 2010).

The analysis of networks originated from the mathematical field of graph theory (Diestel, 1997), later leading to percolation theory (Stauffer and Aharony, 2003) or social network analysis (Wasserman and Faust, 1994). In 1736, Leonhard Euler worked on the problem of crossing all bridges over the river Pregel in Königsberg (now Kaliningrad) exactly once and returning to the origin, a path now called an Euler tour. These and other problems can be studied by using graph representations. Graphs are sets of nodes and edges. Edges can either be undirected, going in both directions, or directed (arcs or arrows) in that one can go from one node to the other but not in the reverse direction. A path is a walk through the graph where each node is only visited once. A cycle is a closed walk, meaning a path that returns back to the first node. A graph could also contain loops that are edges that connect a node to itself; however, for analysis purposes we only observe simple graphs without loops. In engineering, graphs are called networks if there is a source and sink of flow in the system and a capacity for flow through each edge (e.g., flow of water or electricity). However, following conventions in the field of network science, I will denote all brain connectivity graphs as networks.

For brain networks, nodes could be neurons or cortical areas and edges could be axons or fiber tracts. Thus, edges could refer to the structural connectivity of a neural network. Alternatively, edges could signify correlations between the activity patterns of nodes forming functional connectivity. Finally, a directed edge between two nodes could exist if activity in one node modulates activity in the other node, forming effective connectivity (Sporns et al., 2004). Network representations are an abstract way to look at neural systems. Among the factors missing from network models of nodes, say brain areas, are the location, the size, and the functional properties of the nodes. In contrast, geographical or spatial networks also give information about the spatial location of a node. Two- or three-dimensional Cartesian coordinates in a metric space indicate the location of neurons or areas.

Network analysis techniques can be applied to the analysis of brain connectivity, and I will discuss structural connectivity as an example. Using neuroanatomical or neuroimaging techniques, one can test which nodes of a network are connected, that is, whether projections in one or both directions exist between a pair of nodes (see figure 2.1A). How can information about brain connectivity be represented? If a projection between two nodes is found, the value one is entered in the adjacency matrix; the value zero defines absent connections or cases where the existence of connections was not tested (see figure 2.1B). The memory demands for storing the matrix can be prohibitive for large networks as N² elements are stored for a network of N nodes. As most neuronal networks are sparse, storing only information about existing edges can save storage space. Using a list of edges, the adjacency list (see figure 2.1C) stores each edge in one row listing the source node, the target node, and—for networks with variable connection weight—the strength of a connection.

Figure 2.1

Representations of networks. (A) Directed graph with two directed edges or arcs (AC and DC) and one undirected edge being equivalent to a pair of directed edges in both directions (BC). (B) The same graph can be represented in a computer using an adjacency matrix where a value of 1 denotes the existence of an edge and 0 the absence of an edge. In this example, rows show outgoing connections of a node and columns show incoming connections. (C) Sparse matrices (few edges) can also be represented as adjacency lists to save memory. Each edge is represented by the source node, the target node, and the weight of the edge (here: uniform value of 1).

2.1 Reconstructing Connectomes

How can one get information about brain connectivity between regions, the macroscopic connectome (Akil et al., 2011)? The classical way to find out about structural connectivity is to inject dyes into a brain region. The dye is then taken up by dendrites and cell bodies and travels within a neuron either in an anterograde (from soma to synapse) or a retrograde (from synapse to soma) direction. Typical dyes are horseradish peroxidase, fluorescent microspheres, Phaseolus vulgaris-leucoagglutinin, Fluoro-Gold, Cholera toxin subunit B, DiI, and tritiated amino acids. Allowing some time for the tracers to travel, which could be several weeks for the large human brain, the neural tissue can be sliced up, and dyes can indicate the origin and target of cortical fiber tracts. Whereas this approach yields high-resolution information about structural connectivity, it is an invasive technique usually unsuitable for human subjects (however, there are some postmortem studies). In the following, I will therefore present noninvasive neuroimaging solutions to yield structural and functional connectivity.

The workflow for yielding connectivity data starts with anatomical magnetic resonance imaging (MRI) scans with high resolution (see figure 2.2). These scans are later used to register the location of brain regions. For establishing functional connectivity, a time series of brain activity in different voxels or regions can be derived. The correlation between the time series of different voxels or, using aggregated measures, brain regions can be detected and represented as a correlation matrix (with values ranging from –1 to 1). This matrix either can directly be interpreted as a weighted network or can be transformed into a binary matrix in that only values above a threshold lead to a network connection.

For establishing structural connectivity, diffusion tensor imaging (DTI) or diffusion spectrum imaging (DSI) can be applied. Using deterministic tracking, for example, the number of streamlines between brain regions can be represented in a matrix. For probabilistic tracking, matrix elements would represent the probability to reach a target node starting from a source node. In both cases, the weighted matrix can either be analyzed directly or be thresholded so that connections are only formed if a minimum number of streamlines or a minimum probability has been reached.

The choice of nodes and edges can be influenced by the anatomical parcellation schemes and measures for determining connectivity (Rubinov and Sporns, 2010; Eickhoff, Yeo, and Genon, 2018). This choice must be carefully considered as different choices might not only change the topology by removing or adding a few nodes or connections but might alter the local and global network features that will be discussed in the following sections.

Connections of a network can be binarized or weighted. Binary connections only report the absence or presence of a connection. Weighted links can also show the strength of a connection. For structural connectivity, weights can indicate the number of fibers between brain regions (e.g., the streamline count of deterministic tracking), the degree of myelination, the probability that a node can be reached from another node (e.g., probabilistic tracking), or the amount of dye traveling from one node to another (traditional tract-tracing studies). For functional connectivity, weights can indicate the correlation in the time course of signals of different nodes.

2.2 Topological Features

Local Scale—Single Node Features

Networks can be characterized at different levels, ranging from properties characterizing a whole network at the global scale to properties of network components at the local scale. Starting from the local scale, components of a network are its nodes and edges. Edges can be weighted, taking continuous (metric) or discrete (ordinal) values indicating the strength of a connection. Alternatively, they could just have binary values with zero for absent and one for existing connections.

Figure 2.2

Workflow for structural and functional connectivity analysis. High-resolution anatomical magnetic resonance imaging scans of each subject are used as references for further measurements (1). For establishing functional connectivity, a time series of brain activity in different voxels or regions can be derived (3). The correlation between the time series of different voxels or, using aggregated measures, brain regions can be detected and represented as a correlation matrix (5). This matrix can either directly be interpreted as a weighted network (6) or can be binarized in that only values above a threshold lead to a network connection (7). For establishing structural connectivity, diffusion tensor imaging or diffusion spectrum imaging can be applied (2). Using deterministic tracking, for example, the number of streamlines between brain regions can be represented in a matrix (4). This weighted matrix can either be analyzed directly (6) or be thresholded so that connections are only formed if a minimum number of streamlines has been reached (7).

When one thinks about neural systems, there could also be multiple edges between two nodes—for example, a fiber bundle connecting two brain regions. However, such multigraph networks are usually simplified in that the number of fibers is either neglected (binary values) or included in the strength of a connection. In addition to fiber count, one might also think of other properties of connections such as delays for signal propagation or degree of myelination. Whereas such properties likely have significant impact on network function, they are currently not part of the analysis of network topology but are increasingly part of models for network dynamics.

The other component at the local scale is a network node. A node could be a single neuron but, as for edges, could also be an aggregate unit of neurons such as