Summary

• Page properties (1 modified, 0 added, 0 removed)
• Attachments (0 modified, 4 added, 0 removed)
  • image-20220103095424-1.png
  • image-20220103095948-1.png
  • image-20220103100841-2.png
  • image-20220103101022-3.png

Details

Page properties

Content

...	...	@@ -12,37 +12,47 @@
12	12	(((
13	13	(% class="col-xs-12 col-sm-8" %)
14	14	(((
15		-== Siibra - Python client for accessing the human atlas ==
	15	+The Showcase is implemented as a series of interactive Jupyter notebooks covering the individual logical steps and can be accessed in a dedicated public EBRAINS collab.
16	16
17		-The multilevel atlas interface called siibra (formerly known as Brainscapes) is designed to allow safe and convenient interaction with brain definitions from different parcellations, facilitating the implementation of reproducible neuroscience workflows on the basis of brain atlases. It allows to work with reference brain templates both at millimeter and micrometer resolutions and provides streamlined access to multimodal data features linked to brain regions. In particular in this demonstrator it is used to retrieve the dataset of receptor density spatial maps.
18	18
19		-The Showcase 1a is built on the [[connectivity data from the 1000BRAINS study>>https://doi.org/10.25493%2F6640-3XH]] available in the Knowledge graph as a dataset with protected access available to EBRAINS users. The access to protected datasets is provided in EBRAINS through the [[Human Data Gateway>>https://wiki.ebrains.eu/bin/view/Collabs/data-proxy/Human%20Data%20Gateway/]] which allows full access once the user has validated the terms of use.
	18	+The EBRAINS collab consists of interlinked Drive, Bucket, Wiki, and Lab. The Drive provides small file storage and contains the notebooks and all supporting code. The Bucket is a large file storage service and holds the pre-computed results of the extensive parameter sweeps and model optimizations to allow skipping the computationally demanding steps. The documentation of the showcase implementation is collected in the Wiki. The Lab service is an instance of JupyterLab—an interactive computing environment where the notebooks can be run and worked with.
20	20
21		-== Brain Network Modeling: TVB ==
22	22
23		-The second major component of the showcase is The Virtual Brain (TVB) simulation platform. TVB is a simulation platform that uses empirical structural and functional data to build whole brain models of individual subjects. For convenient model construction, the system is based on a processing pipeline for structural, functional, and diffusion-weighted magnetic resonance imaging (MRI) data. The pipeline combines several state-of-the-art neuroinformatics tools to generate subject-specific cortical and subcortical parcellations, surface-tessellations, structural and functional connectomes, and region-wise aggregated blood oxygen level-dependent (BOLD) functional MRI (fMRI) time-series.
	21	+The notebooks in this collab will load all required Python modules including Siibra and The Virtual Brain, and the interfaces for launching the computationally demanding parts in the HPC infrastructure. Running the notebooks requires an EBRANS account with permissions to access the Lab and the Knowledge Graph API. In addition, to be able to interact with the HPC infrastructure, the user has to have access to an active allocation on the corresponding FENIX site.
24	24
25		-== Systematic parameter exploration and parameter optimization ==
	23	+==
	24	+Simulation of resting state ==
26	26
27		-The third component is distributed execution of systematic parameter exploration and parameter optimization on the High Performance Computing (HPC) infrastructure available in the FENIX RI. The unified access to the federated infrastructure is enabled by the pyunicore library providing a concise API to the common tasks such as compute job submission and management. That allowed us to create simplified interfaces for the jupyterlab environment for the user to run the parameter explorations. For the Showcase 1a we have implemented custom library for distributed simulations compatible with the controlled data access through the Human data Gateway. The showcase 1b is making use of the Learning to Learn (L2L), which is a gradient-free optimization framework that contains well-documented and tested implementations of various gradient-free optimization algorithms. It also defines an API that makes it easy to optimize (hyper-)parameters for any task (through a construct called an “optimizee”). All the implementations in this package are parallel and can run across different cores and nodes (but equally well on a single core). The basic idea behind “Learning to Learn” is to have an “outer loop” optimizer optimizing the parameters of an “inner loop” optimizee. This particular framework is written for the case where the cycle starts when the outer-loop optimizer generates an instance of a set of parameters and provides it to the inner-loop optimizee. Then, the inner-loop optimizee evaluates how well this set of parameters performs and returns a “fitness” vector for each parameter in the set of parameters. Lastly, the outer-loop optimizer generates a new set of parameters using the fitness vector it got back from the inner-loop optimizee. On the whole, what this means is that the outer-loop Optimizer works only with parameters and fitness values and doesn’t have access to the actual underlying model of the optimizee. The only thing the optimizee does is to evaluate the fitness of the given parameter. In our implementation, this fitness function is defined to minimize the distance between empirical and simulated BOLD signals through the swFCD observable, as described above. Once properly configured, L2L allows running the simulation for a “mean” subject signal built as an average of all individual subjects in the dataset, or for each such individual subject in particular. This execution is distributed on top of the computing capabilities of the Jülich supercomputing center, which allows computing in a matter of minutes which otherwise would take several days to compute. In the current implementation, a systematic parameter sweep was used to guarantee accurately finding a global minimum of the parameters to fit.
	26	+The fist notebook in the inter-individual variability workflow explores the resting-state simulation for a subject of the 1000BRAINS dataset. Functional data are simulated by means of a brain network model implemented in TVB, which is an ensemble of neural mass models linked via the weights of the structural connectivity (SC) matrix. Following topics are covered:
28	28
29		-In addition to the systematic parameter exploration, we have integrated a Bayesian framework allowing for inference of the full posterior values of the parameters as a fourth component for the Showcase 1a. For this we have employed the Simulation Based Inference (SBI) in which a deep neural estimator is trained to provide a relationship between the parameters of a model (black box simulator), and selected descriptive statistics of the observed data. As numerous repeated simulations are at the core of the training phase of SBI (sampling of the prior parameter distributions), the implementation can reuse the infrastructure for the systematic parameter sweeps on the HPC infrastructure. Specifically, for each subject an estimator was trained on 2000 simulations to retrieve the global coupling parameter G using the properties of functional connectivity and functional connectivity dynamics derived from empirical fMRI data. The fMRI time series data hasn't been made publicly available yet, and the procedure is at this stage demonstrated on simulated data.
30		-
	28	+1. The neural mass model by Montbrio, Pazo and Roxin
	29	+1. Construction of the TVB model for a particular subject
	30	+1. Dynamics of the model, and execution of a parameter study
	31	+1. Summary of the simulation results for the whole cohort
31	31
32		-== Virtual ageing trajectories: 1000BRAINS cohort ==
	33	+[[image:image-20220103100841-2.png]]
33	33
34		-(% style="color:#000000; font-family:Arial; font-size:11pt; font-style:normal; font-variant:normal; font-weight:400; text-decoration:none; white-space:pre-wrap" %)Beyond the general common architecture, the human brain is characterized by a high interindividual variability with regard to its structure and functional abilities, including measurable cognitive outcome. Disentangling the sources and consequences of this variability is key to understanding the capabilities of our brains and any pathologies occurring over the lifespan. This also includes implementing measures of variability into simulations and modelling approaches to make them biologically plausible and as realistic as possible.
35	35
36		-(% style="line-height:1.38; margin-top:13px" %)
37		-(% style="color:#000000; font-family:Arial; font-size:11pt; font-style:normal; font-variant:normal; font-weight:400; text-decoration:none; white-space:pre-wrap" %)The aging brain is a particularly remarkable use case in this respect as the interindividual variability of brains and cognitive abilities of older subjects from the general population is very pronounced. The relevant factors influencing this variability, be it genetics or environmental and lifestyle factors, are manifold and barely understood yet. Similarly, the different levels within the organization of the brain, from the molecular, cellular to the systems level, contribute to these effects to varying degrees. The aim of the first branch of Showcase 1 is thus to model brain aging with biologically plausible assumptions about the architecture of the brain, using the EBRAINS platform and integrating information from the different organizational levels of the brain, and to create a virtually aged cohort of brain models which allow testing of hypotheses and causal linkage of potential sources and effects of the empirically observed variability.
	36	+== Regional variability data ==
38	38
39		-== Brain region variance: incorporating receptor densities ==
	38	+The first step of the regional variability workflow loads the data from the Knowledge Graph and defines the regional bias on the model. In this case we require: 1) structural connectivity matrices, 2) GABA and AMPA receptor densities at each region, and 3) empirical resting-state fMRI data for fitting and validation of the simulations. The three datasets shall be characterised in the same parcellation.
40	40
41		-(% style="color:#000000; font-family:Arial; font-size:11pt; font-style:normal; font-variant:normal; font-weight:400; text-decoration:none; white-space:pre-wrap" %)The second branch of the Showcase 1 focuses on the impact of variability of the properties of brain regions on the expressed dynamics. Over the last two decades, whole-brain network models have proven useful instruments to describe the brain's resting-state activity: from explaining the intrinsic dynamical behavior of the brain to the classification of brain states in disorders of consciousness. These whole-brain models are built at the resolution of interconnected cortical regions and subcortical brain areas, with each region simulated at the population levels. So far, a major limitation of these models lies in the assumption that all brain regions are of identical characteristics while it is well-known that brain areas differ in a variety of properties, e.g., neuronal densities, local cytoarchitecture, and types of neuroreceptors [Deco2021]. Accounting for these regional variances is crucial for translational purposes as for example differences in neuroreceptor densities determine the impact of pharmacological agents on particular brain regions and thus their potential to modulate the whole-brain dynamics. The purpose of Showcase #1(b) is, therefore, to illustrate how whole-brain network models can be adapted to include such regional variances and to develop workflows to run the simulations taking full advantage of the HBP's EBRAINS infrastructure, leading to workflows that are accessible and re-usable by the scientific community.
	40	+Link to the notebook:
42	42
43		-(% style="color:#000000; font-family:Arial; font-size:11pt; font-style:normal; font-variant:normal; font-weight:400; text-decoration:none; white-space:pre-wrap" %)In implementation of Showcase 1b, the simulation is based on a network of nodes that represent brain ROI, each running the (% style="color:#000000; font-family:Arial; font-size:11pt; font-style:normal; font-variant:normal; font-weight:700; text-decoration:none; white-space:pre-wrap" %)**BEI**(% style="color:#000000; font-family:Arial; font-size:11pt; font-style:normal; font-variant:normal; font-weight:400; text-decoration:none; white-space:pre-wrap" %) (Balanced Excitation-Inhibition) model [Deco2014], which defines two neuronal subpopulations, an excitatory and an inhibitory one, both interacting through neurochemical (e.g., AMPA, GABA, NMDA, etc.) currents. The objective of the Showcase 1b is to demonstrate and find out the more precise mechanism through which regional neurotransmitter density affects this excitation-inhibition balance. Numerically, the goal is to minimize the difference in sliding window Functional Connectivity Dynamics ((% style="color:#000000; font-family:Arial; font-size:11pt; font-style:normal; font-variant:normal; font-weight:700; text-decoration:none; white-space:pre-wrap" %)**swFCD**(% style="color:#000000; font-family:Arial; font-size:11pt; font-style:normal; font-variant:normal; font-weight:400; text-decoration:none; white-space:pre-wrap" %)) between a reference empirical signal and a simulated one. The swFCD subdivides the time-series into successive windows and, for each one, computes its corresponding Functional Connectivity through the usual Pearson correlation, resulting in a series of NxN matrices. Two swFCD can be compared by means of the Kolmogorov-Smirnov statistic.
	42	+* [[regional_variability/notebooks/1_data_setup.ipynb>>https://lab.ch.ebrains.eu/user-redirect/lab/tree/shared/SGA3%20D1.2%20Showcase%201/regional_variability/notebooks/1_data_setup.ipynb]]
44	44
45		-
	44	+(% style="text-align:center" %)
	45	+[[image:image-20220103095424-1.png]]
	46	+
	47	+== Fitting model parameters for models with regional bias ==
	48	+
	49	+A series of simulations of the whole-brain network model is launched in the EBRAINS HPC facilities in order to identify the optimal model parameters leading to simulated resting-state brain activity that best resembles the empirically observed activity.
	50	+
	51	+Link to the notebook:
	52	+
	53	+* [[regional_variability/notebooks/2_parameter_swep_proto.ipynb>>https://lab.ch.ebrains.eu/user-redirect/lab/tree/shared/SGA3%20D1.2%20Showcase%201/regional_variability/notebooks/2_parameter_swep_proto.ipynb]]
	54	+
	55	+[[image:image-20220103095948-1.png]]
46	46	)))
47	47
48	48

image-20220103095424-1.png

Author

...	...	@@ -1,0 +1,1 @@
	1	+XWiki.fousekjan

Size

...	...	@@ -1,0 +1,1 @@
	1	+33.7 KB

Content

image-20220103095948-1.png

Author

...	...	@@ -1,0 +1,1 @@
	1	+XWiki.fousekjan

Size

...	...	@@ -1,0 +1,1 @@
	1	+4.1 KB

Content

image-20220103100841-2.png

Author

...	...	@@ -1,0 +1,1 @@
	1	+XWiki.fousekjan

Size

...	...	@@ -1,0 +1,1 @@
	1	+116.4 KB

Content

image-20220103101022-3.png

Author

...	...	@@ -1,0 +1,1 @@
	1	+XWiki.fousekjan

Size

...	...	@@ -1,0 +1,1 @@
	1	+87.1 KB

Content