Predicting micro-structure from low-dimensional embeddings of brain networks
Poster No:
1585
Submission Type:
Abstract Submission
Authors:
Vincent Bazinet1, Bratislav Misic1
Institutions:
1McGill University, Montreal, Quebec
First Author:
Co-Author:
Introduction:
The human brain is a complex system of heterogeneous regions that interact with each other. At the micro-scale, individual brain regions have unique genetic, cellular, laminar and chemoarchitectural properties, and understanding how this local architecture varies across the cortex is one of the fundamental goals of neuroscience. Their main axes of variation can indeed provide valuable information about the development and function of individual brain regions and ultimately help characterise their role in different psychopathologies [1]. Here, we demonstrate that brain connectivity can be used to characterize variations in a wide range of micro structural properties.
Methods:
In this work, we explored the relationship between micro-structure and brain connectivity by asking (i) which type of brain connectivity (structural or functional) best predict variations in micro-structure, (ii) which type of embedding best summarize these variations, and (iii) what are the main distinctions between the spatial embedding of the brain and network-based embeddings of the brain?
We first reconstructed structural and functional connectomes using diffusion and functional MRI data from the Human Connectome Project (HCP; Fig. 1a) [2], which were parcellated into 800 brain regions [3]. The micro structural attributes studied included T1w/T2w ratio and cortical thickness information from the HCP, transcriptional information from the Allen Human Brain Atlas [4] and regional receptor profiles from publicly available positron emission tomography tracer studies (Fig. 1b) [5].
We next generated three different low-dimensional representations of the structural and functional connectomes using diffusion map embeddings [6], dynamical embeddings [7], and principal components (Fig. 2a). Then, we trained regression models to predict the micro-structural properties of each region of the brain, using the components of our embeddings as features. Half of the nodes in our networks were used to train our models, and the remaining half was used to evaluate their performance.
We first reconstructed structural and functional connectomes using diffusion and functional MRI data from the Human Connectome Project (HCP; Fig. 1a) [2], which were parcellated into 800 brain regions [3]. The micro structural attributes studied included T1w/T2w ratio and cortical thickness information from the HCP, transcriptional information from the Allen Human Brain Atlas [4] and regional receptor profiles from publicly available positron emission tomography tracer studies (Fig. 1b) [5].
We next generated three different low-dimensional representations of the structural and functional connectomes using diffusion map embeddings [6], dynamical embeddings [7], and principal components (Fig. 2a). Then, we trained regression models to predict the micro-structural properties of each region of the brain, using the components of our embeddings as features. Half of the nodes in our networks were used to train our models, and the remaining half was used to evaluate their performance.
Results:
We evaluated the performances of each low-dimensional representation using the coefficient of determination (R2). The best results were obtained, on average, with models relying on diffusion map embeddings of the structural connectome (mean R2=0.59; Fig. 2b, bottom-left panel). This is in line with recent work demonstrating that communication models unfolding on the structural connectome can unveil important features of brain organization [8]. For the functional connectome, models relying on linear representation of connectivity (i.e. principal components) outperformed models relying on diffusion or dynamical embeddings (Fig. 2b, top-left panel)
We also explored whether connectivity information can enhance modelsthat rely on information about the spatial relationship between brain regions (the Euclidean distance between them). We hence trained models using a low-dimensional representation of the distance relationships between brain regions and compared the accuracy of their predictions to those of models combining low-dimensional representations of both the spatial and topological relationships between brain regions. These models combining spatial and topological information outperformed the models relying solely on spatial information by a significant margin (p<0.001; Fig. 2c). This result shows that the information provided in structural and functional brain networks can enhance predictions of micro-structure variations made using the spatial information between brain regions.
We also explored whether connectivity information can enhance modelsthat rely on information about the spatial relationship between brain regions (the Euclidean distance between them). We hence trained models using a low-dimensional representation of the distance relationships between brain regions and compared the accuracy of their predictions to those of models combining low-dimensional representations of both the spatial and topological relationships between brain regions. These models combining spatial and topological information outperformed the models relying solely on spatial information by a significant margin (p<0.001; Fig. 2c). This result shows that the information provided in structural and functional brain networks can enhance predictions of micro-structure variations made using the spatial information between brain regions.
Conclusions:
In summary, the present work shows that brain connectivity provides valuable information about how the local boundaries and non-uniformities in the topographic distribution of micro-structural properties arise. Importantly, it also shows that the information retrieved from brain connectivity is fundamentally different from the background information provided by the spatial relationship between brain regions.
Genetics:
Transcriptomics
Modeling and Analysis Methods:
Connectivity (eg. functional, effective, structural) 1
fMRI Connectivity and Network Modeling 2
Neuroanatomy, Physiology, Metabolism and Neurotransmission:
Cortical Cyto- and Myeloarchitecture
Transmitter Receptors
Keywords:
Computational Neuroscience
1|2Indicates the priority used for review
Provide references using author date format
[2] Van Essen, D. C. (2013). The WU-Minn human connectome project: an overview. Neuroimage, 80, 62-79.
[3] Schaefer, A. (2018). Local-global parcellation of the human cerebral cortex from intrinsic functional connectivity MRI. Cerebral cortex, 28(9), 3095-3114.
[4] Hawrylycz, M. J. (2012). An anatomically comprehensive atlas of the adult human brain transcriptome. Nature, 489(7416), 391-399.
[5] Hansen, J. Y. (2022). Mapping neurotransmitter systems to the structural and functional organization of the human neocortex. Nature Neuroscience, 25, 1569-1581
[6] Coifman, R. R. (2005). Geometric diffusions as a tool for harmonic analysis and structure definition of data: Diffusion maps. Proceedings of the national academy of sciences, 102(21), 7426-7431.
[7] Schaub, M. T. (2019). Multiscale dynamical embeddings of complex networks. Physical Review E, 99(6), 062308.
[8] Avena-Koenigsberger, A. (2018). Communication dynamics in complex brain networks. Nature reviews neuroscience, 19(1), 17-33.