Machine Learning for Science: Mathematics at the Interface of Data-driven and Mechanistic Modelling
24 August 2023
Researchers using computational methods to study MRI scans say they can produce a more detailed picture of the structural architecture of the human brain than was previously possible.
Their improved roadmap of how brain regions are related, they say, might help us understand both ‘normal’ brain development in healthy humans and how it differs in patients with brain disorders like dementia and psychiatric illnesses like schizophrenia.
In a paper published this week in Nature Neuroscience – they add that their method of quantifying the brain’s structural network also mapped closely onto brain networks derived from gene expression data.
This could be used to help researchers identify which genes are expressed in regions of the brain affected by mental health conditions – and which genes, therefore, may be implicated in the development of these conditions. The work was carried out by a team of researchers from this Department and from the University’s Department of Psychiatry. The team includes Accelerate Research Fellow Dr Sarah Morgan and Isaac Sebenius, Accelerate PhD student currently studying for his PhD in Psychiatry under Dr Morgan’s supervision.
“What I’m really excited about,” says Isaac , the first author on the study, “is that this new method connects imaging data with more fundamental biological processes like gene expression. Brain scans are relatively easy to acquire, so it’s great if we can use them to say something about what is really going with someone’s neurobiology.”
“We hope to use this new method to study illnesses like schizophrenia and conditions like Alzheimer’s in more detail,” says Dr Sarah Morgan.
“In fact, we could use this method to study any sort of mental health condition, including psychotic disorders and depression. We could also use it to study the brain’s development from its early stages to its ageing, and to find out which connections within the brain are important for different cognitive processes.”
The ‘robust estimation’ approach described in Nature Neuroscience takes data from over 11,000 MRI scans and uses a computational approach to find structural similarities and connections between different regions of the brain.
It builds on work previously carried out in the Department of Psychiatry that used certain values for each brain region – such as cortical thickness, surface area and grey matter volume – to look for correlations between them (Seidlitz et al, Neuron 2018).
“This worked really well,” Sarah (right) says. “But using a more advanced computational approach to estimate the similarity between brain regions gives results which are both statistically more robust and more consistent with what we already know about the brain’s structure. Our method doesn’t just use a set of summary values per brain region, but instead draws on the underlying distributions of those values. You build up a multivariate distribution and then calculate the similarity of those distributions between brain regions and that seems to work very effectively.”
As well as building up a robust picture of structural connectivity within the brain, the new method was also able to estimate brain age from MRI scans.
“And that’s really important,” Sarah says, “because we know with mental health conditions that people have different risks of different conditions depending on their age. So we used our method to estimate age and found we could do it more accurately with this method than with others.”
She is also excited about their finding that their measurements of structural similarity between regions of the brain is very strongly coupled to the measurement of gene expression similarities between regions of the brain.
“We’d like to use this new method to look at psychotic disorders and see, for example, if we can predict symptoms from patterns of brain connectivity and relate those to the gene expression data – which might give us some ideas about which biological mechanisms underlie which symptoms. This method could be very useful for that.”
Robust estimation of cortical similarity networks from brain MRI. Isaac Sebenius, Jakob Seidlitz, Varun Warrier, Richard A. I. Bethlehem, Aaron Alexander-Bloch, Travis T. Mallard, Rafael Romero Garcia, Edward T. Bullmore & Sarah E. Morgan. Nature Neuroscience (2023).