Relationship between SC, simulation parameters and behavioral data - conceptual question

[Szymon Tyras]

Hello everyone,

I'm new to whole-brain simulations and currently planning my first study using TVB. I am thinking about the possibility of studying the mechanisms behind behavioral data, in particular about the distinction between structural (understood as the disappearance of white matter) and synaptic mechanisms. 

It seems to me that for this purpose one could correlate the simulation parameters representing the mesoscopic synaptic weight (for example G), the structural weight of the connections (SC) with a behavioral variable and, based on the contribution of a given factor, make a distinction.

However, after reading several publications on the correlations of TVB parameters with clinical outcomes, I have not seen such an approach anywhere, which makes me wonder that I am missing something important. I am aware that the simulation and SC parameters are interrelated (e.g. that the SC limits the possible values of the parameters). 

I'm interested in the opinion of someone more experienced on this subject.

Thank you in advance,

Szymon Tyras

[WOODMAN Michael]

Hi,

Two recent preprints may help provide a starting point for your project,

- https://www.medrxiv.org/content/10.1101/2023.01.11.23284438v1

- https://www.biorxiv.org/content/10.1101/2022.02.17.480902v2

Cheers,

Marmaduke

--
You received this message because you are subscribed to the Google Groups "TVB Users" group.
To unsubscribe from this group and stop receiving emails from it, send an email to tvb-users+...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/tvb-users/066806f5-7c41-4467-a238-cf577d72a9acn%40googlegroups.com.

[Ritter, Petra]

Hi Szymon, 

 

in addition these two publications may be of interest to you as well:

 

https://www.nature.com/articles/s41467-023-38626-y

 

https://elifesciences.org/articles/28927

 

Best,

Petra

 

 

-- 

Prof. Dr. Petra Ritter

BIH Johanna Quandt Professor for Brain Simulation

Director, Brain Simulation Section

Berlin Institute of Health & Dept. of Neurology

Charité University Hospital Berlin

Robert—Koch Platz 4

10115 Berlin

 

EOSC Project VirtualBrainCloud Coordinator

Health Data Cloud Coordinator

Virtual Research Environment Coordinator

Chair Infrastructure & Data Security (IDS) Group of NFDI Section Common Infrastructures

eBRAIN-Health  Coordinator

Testing and Experimentation Facility Health AI and Robotics (TEF-Health) Project Coordinator

 

CONFIDENTIALITY NOTICE: The contents of this email message and any attachments are intended solely for the addressee(s) and may contain confidential and/or privileged information and may be legally protected from disclosure.

To view this discussion on the web visit https://groups.google.com/d/msgid/tvb-users/7ABA643F-7E87-402F-BC2C-2A947F230311%40univ-amu.fr.

[Szymon Tyras]

Dear Dr. Marmaduke, 

Dear Prof. Petra Ritter, 

Thank you for the provided publications. They turned out to be really enlightening. They show how the simulation parameters mediate the effect of SC on brain dynamics. However, one point bothers me - imagine two different brains, one with an SC weight of 1, the other 10. The former could have a G value of 10 and the latter a G value of 1. The total effect on brain dynamics would be indistinguishable. However, two different mechanisms may be responsible for, for example, a certain clinical symptom in these two people (loss of axons in one case and weakening of excitatory synapses in the other). I wonder if regressing out structural data and revealing the correlation between G and e.g. the severity of symptoms would point to the synaptic mechanism and vice versa. Regressed part of variations in ideal case would corresponds to the part of information about influence of SC on G value and left additional information that was gain through fitting procedure (to the FC or FCD).  

Thank you again and very sorry if I am asking about trivial aspects,

Respectfully, 

Szymon Tyras

[Ritter, Petra]

Hi Szymon,

 

indeed the SC values can be arbitrary and need to get normalized before comparing optimized parameters.

 

Even better – compare the actual implications for the state variables you are simulating (e.g. effective connectivity, firing rates etc.) – as done in Schirner et al 2018 and Schirner et al 2023.

 

Best,

Petra

 

-- 

Prof. Dr. Petra Ritter

BIH Johanna Quandt Professor for Brain Simulation

Director, Brain Simulation Section

Berlin Institute of Health & Dept. of Neurology

Charité University Hospital Berlin

Robert—Koch Platz 4

10115 Berlin

 

EOSC Project VirtualBrainCloud Coordinator

Health Data Cloud Coordinator

Virtual Research Environment Coordinator

Chair Infrastructure & Data Security (IDS) Group of NFDI Section Common Infrastructures

eBRAIN-Health  Coordinator

Testing and Experimentation Facility Health AI and Robotics (TEF-Health) Project Coordinator

 

CONFIDENTIALITY NOTICE: The contents of this email message and any attachments are intended solely for the addressee(s) and may contain confidential and/or privileged information and may be legally protected from disclosure.

 

To view this discussion on the web visit https://groups.google.com/d/msgid/tvb-users/2a03c8ac-59d4-45c6-a902-02da513c19c2n%40googlegroups.com.

[Szymon Tyras]

Dear Prof. Petra Ritter, 

Thank you again for fast response. If I understand your answer correctly it indeed means that when one is interested in simulations parameters directly as synaptic weights values it should be done only under the assumption (statistical or empirical) that SC underlying them were similar. Am I getting this correctly? 

As you suggested it obviously would be the best option to test impact of SC directly (same simulations parameters, different SCs). However, I am primarily interested in outcome that is not directly simulated (clinical symptoms). 

Thank you again,

Respectfully, 

Szymon Tyras

[Ritter, Petra]

Hi Szymon,

 

we can relate simulated dynamics / simulation inferred patient specific state variable features to clinical scores.

 

The behavior of the model’s state variables as a consequence of altered SC or altered model parameters may reveal the altered neuronal processes underlying the clinical symptoms.

 

Using TVB as computational microscope to observe / infer the neuronal dynamics that are not easily accessible empirically  – and relate them to behavior or symptoms.

To view this discussion on the web visit https://groups.google.com/d/msgid/tvb-users/e2d1a058-3026-4234-a1e8-49f2145e8d25n%40googlegroups.com.

[Randy McIntosh]

Hi Szymon – to jump into this conversation, it seems you’re interested in requires combining the simulation parameters (e.g. G or E/I) and empirical data (e.g, SC) to ascertain unique contributions to prediction of clinical symptoms.  That per se has been done in a pub from the TVB group looking at dementia here. 

 

It is a good idea, but I would advise not regressing out variations but rather include both simulation and empirical estimates to determine joint and unique prediction.  Looking at the variation across the sample would then allow you to differentiate gradations of contributions to predicting clinical symptoms – such as more white matter (SC) vs more local E/I (NB: assuming it’s a continuous effect).

To view this discussion on the web visit https://groups.google.com/d/msgid/tvb-users/2a03c8ac-59d4-45c6-a902-02da513c19c2n%40googlegroups.com.

[Szymon Tyras]

Dear Prof. Petra Ritter, 

Thank you very much for answer. Indeed, it seems that your proposal is best option to determine causality. However, it requires knowing/finding meaningful enough relation between outcome of interest and dynamics. Sometimes it will be difficult. If I understand the whole process correctly one can assume that somewhere in simulated dynamics there is information about outcome of interest - because we generally assume that empirical dynamics stands behind it and fitting procedure reproduce it as accurately as possible, and use optimal simulation parameters as model output - hidden state that would not be seen otherwise (for example synaptic weights). Then this output can be use to correlate it with outcome of interest. If I am not wrong this should be able to reveal some informations about synaptic underpinnings as done for example in Good et. al. (https://www.eneuro.org/content/9/1/ENEURO.0075-21.2022). Obviously it would not be as valuable as finding direct relations parameters-brain dynamics-outcome that you have mentioned but still informative enough as first approximation.  

Dear Dr. Randy McIntosh, 

Thank you for answering. Indeed it is what I was planning to do. In particular (instead of using ML as compared to paper you cite) I was planning to put simulations parameters (G, Ji's) and SC weights 

into single PLS correlation model to correlate it with clinical outcome. If single latent variable would be significant it would mean that (linear) combination of this data is correlated with clinical outcome. Then I could simply check contributions of SC weights and simulations parameters. If, for example SC weights would not be different from 0 one can assume that synaptic mechanisms is leading or vice versa. If both contributions will be significant than the true mechanism probably depend on combination of both. 

Is this, in principle make sense? 

Thank you both again,

Respectfully, 

Szymon Tyras

[Randy McIntosh]

Hi Szymon – your plan sounds like a good one. I look forward to seeing the results!

To view this discussion on the web visit https://groups.google.com/d/msgid/tvb-users/e22966cf-2809-4223-9367-1e2af7116dc5n%40googlegroups.com.

[Szymon Tyras]

Dear Dr. Randy McIntosh, 

Thank you for help and kind words. 

Respectfully, 

Szymon Tyras

[Viktor Jirsa]

Hi Szymon, 

you posted a conceptual question to this forum, then let me add a conceptual comment. The link to clinical outcome has been answered. Regarding the disentangling of mechanisms (such as G and SC), you may want to study the proposed papers from the perspective of correlation vs causality, having the following illustrative toy model in mind: you seek to understand the solutions for x^2=4. Your mathematical high-school training teaches you the solutions x = +2 and -2. An optimisation approach will estimate one of them, let us say x=+2, and adding more sophisticated analysis will systematically confirm that the square of +2 equals 4. Taking second order statistics (correlation, FC, etc) will estimate both solutions and approximate the solution to be x = 0. Inference using MC will sample a probability density distribution for two modes around +2 and -2.  Obviously, there are many variants to this example, but critical issues are already illustrated in this simple setup. 

Best regards, Viktor

To view this discussion on the web visit https://groups.google.com/d/msgid/tvb-users/448410cb-e2a7-4c29-879a-6da75d271002n%40googlegroups.com.

[Szymon Tyras]

Dear Dr. Viktor Jirsa,

Thank you for your answer regarding statistical methods. Indeed, Bayesian approach can be useful in such case. 

Respectfully, 

Szymon Tyras