Checking Our Assumptions on Brain Structure and Function

My brain is my brain,
your brain is your brain,
except when it’s not



This is going to be a longish Blog, so you might want to grab a coffee or tea, or equivalent before you continue reading.  It’s okay, we’ll wait.

The motivation for this blog came from a recent paper from our lab, which Joelle took the lead on. It caused some ripples in the lab, with colleagues, and on social media. It’s really about checking simple assumptions, in this case about brain connectivity and how we measure it.

Mapping brain connectivity has become a massive scientific industry. The epitome of this is the Human Connectome Project (HCP), which was spawned by this seminal paper from Sporns, Tononi and Kotter - an attempt to create a population level map of the structural and functional connectivity of the human brain. For those who are not immersed in this stuff, structural connectivity refers to the physical connections between parts of the brain. In the ideal case, this would be synapses between cells, but this is hard to do in living humans, so we estimate connections using an MRI sequence called diffusion tensor imaging (DTI). By analogy, this gives us the wiring diagram of the brain. However, that’s only an indication of the potential for communication, not whether it has been used. This is what functional connectivity is meant to reveal – what collections of regions are actually communicating. We aren’t going to spend time on the nuances of these two, as there are a lot of arguments about how each is measured and what each means. Functional connectivity, at its essence, is really only telling you that there is a statistical association between two parts of the brain, and not that they are, in fact, actually communicating. Other methods, generally labelled as effective connectivity, are better suited for that, but I don’t plan to talk about that here (you can get more info on this from Scholarpedia and a recent review. So, structural connectivity tells you about the potential for two brain areas to talk while functional connectivity gives you a hint that two areas may be talking. Because of this uncertainty, it’s not surprising then, that knowing one doesn’t perfectly predict the other.

But if we go back to the goal of things like the connectome project, you might expect that if we knew someone’s structural connectivity, we should be able to predict their specific pattern of functional connectivity better than someone else’s. For example, the structural connectome for Randy’s brain should predict Randy’s functional connectome better than Joelle’s structural connectome should. This seems so obvious.

So, Joelle tested this.  Not with Randy and Joelle’s brains, but rather across six different datasets, with sample sizes ranging from 48 to 766, which included data from the Human Connectome Project. The result was very surprising. In all but one dataset, the correlation between a person’s structural and functional connectome was no higher than the correlation between a given person’s structural connectome and someone else’s functional connectome. This is shown in the plot below where the blue histogram is self-with-self connectomes and the red for self-with-other. The only dataset where self-with-self was higher was the highly resolved Glasser parcellation from the Human connectome project. Other parcellation schemes from the Human Connectome data set did not show this differentiation.

So, this was a bit concerning. For the connectomics field, and particularly human connectomics, what was to be a reasonable assumption turns out to false, unless you have a very high-quality dataset, which few of us can acquire. Even then, the separation of the self vs. other is quite small. For my lab, this is a problem because we have been assuming that if we combine structural and functional connectomes, we will improve subject specificity, especially in the clinical work. This is one of the goals for TheVirtualBrain consortium.

There are a few things to note here before we all pack up and go home. First, the variance of structural connectomes across individuals is much lower than functional connectomes. This is probably related to the metrics that we derive to estimate structural connectivity, which are heavily processed to minimize noise measurement. Comparable noise suppression is done for functional connectivity, but the metric relies on signal fluctuations, which gives a wider dynamic range.

There is variance in structural connectomes, no question, as you can relate this to other measures such as cognition (an example). But, what our little experiment suggests is that much of the unique variance may be muted. The Glasser parcellation tries to preserve as much of the individual information on structure as possible, and yet even there the difference between self-to-self and self-to-other was really small. This does suggest that parcellation schemes that better preserve individual anatomy will boost structural variance.

Rather than being discouraged by this, we see it as motivation for further exploration. One route could be to re-evaluate the metrics that we are using to assess the fit between structure and function. The study we published uses the “traditional” functional connectivity measure across a very long time window – what’s been referred to as static connectivity. We do know that if you look at shorter time-steps, the functional connectivity pattern changes. There is reason to believe that this dynamic functional connectivity may be more sensitive to individual differences. The “static” functional connectome only captures one pattern of functional connectivity. Insofar as structural connectomes reflect potential for functional networks, perhaps using dynamic functional connectivity estimates will yield more specific correspondence. There is some great work led by Mac Shine [@jmacshine] on this (example 1; example 2).

A second route to explore could be building a biophysical model to bridge these data rather than an analytic one. We are doing this with TheVirtualBrain and have some promising outcomes. Combining structure and function in a model distills the features of the empirical measures to those that best produce the brain dynamics of interest. For instance, dynamic functional connectivity patterns can be reproduced in TVB models when there are nonlinearities in the regional activity. We recently showed that TVB model parameters are a better predictor of cognitive status in dementia than structural or functional connectivity. This may seem paradoxical since the model is based on structural and functional connectivity, but it’s likely this reflects the distillation process we mentioned. Modeling forces you to focus on data features that you believe are most relevant for explaining the phenomena of interest. It also gives you an explicit way to test assumptions. 

So let this be our PSA : Don’t ever be reluctant to do that.


McLab Icon_C8A698 mini.png