Repository For Neuroimaging Data To Facilitate Cross-Disciplinary Research
By Deborah Borfitz
November 16, 2022 | Scientists at the Montreal Neurological Institute of McGill University (Montreal, Canada) have created an open-source database, called “neuromaps,” housing more than 40 of the uncountably high number of brain maps in existence. The brain maps selected for aggregation were those thought to be most useful to the neuroimaging research community in contextualizing their results against structural and functional phenotypes, according to Justine Hansen, a Ph.D. student in the Integrated Program of Neuroscience.
The neuromaps toolbox was introduced in an article appearing recently in Nature Methods (DOI: 10.1038/s41592-022-01625-w). “Ultimately, we hope that neuromaps will add a spark to the analysis of human brain maps and increase accessibility of data and software tools to people with diverse research interests,” says Hansen.
With time, the data repository will grow more comprehensive and extensive as others in the field add in brain maps, pending approval and quality checks, which they find interesting and useful, she says. The starting point was neuroimaging data students in the lab found themselves repeatedly using, including microarchitecture (cortical thickness and myelination), receptor distributions, gene expression, and functional hierarchy maps delineating lower-order regions responsible for functions like sensing and moving from higher-order ones involved in cognitive processes such as language and planning.
Each map associates every brain region with a value that represents some measure, Hansen explains, such as a number corresponding to cortical thickness. New maps are routinely being produced and are sometimes specific to a particular experiment—e.g., “which areas of the brain light up in the fMRI when someone watches minute 30:00 of The Grand Budapest Hotel?”—and other times come out of analysis pipelines.
In an article she published last year in Nature Human Behavior (DOI: 10.1038/s41562-021-01082-z), she describes a map of how gene expression and functional activity covaries in the brain. The map is the output of a statistical procedure called partial least squares analysis.
Researchers often don’t make their original brain imaging data available either to protect patient privacy or because no one asks them to share it, says Hansen, noting that neuromaps contains group averages rather than individual-level data. When people were asked about having their data added to neuromaps, and given the proper credit, “the responses were overwhelmingly positive.”
Lots of other brain data are “hidden away” in a GitHub repository or in an outdated platform, she adds. “Neuromaps it just making it easier to find this data and fetch it.”
Other open-access platforms such as NeuroVault and BALSA contain a lot of surface and volumetric data. “Neuromaps is first and foremost a tool for contextualization,” says Hansen. “One aspect of this is the data repository, but equally important is the transformation and significance testing functionalities.”
In other words, “neuromaps isn't a space for just any brain map to reside,” she continues. “It tries to be a bit restrictive and collect maps that will help researchers get an idea of how their brain map relates to features of brain structure and function.”
Data from the Human Connectome Project (HCP) of the National Institutes of Health was used to generate the transformation functions, says Hansen, adding that brain data are generally available in one of four “coordinate systems.” Two brain maps can’t be immediately compared since they need to be in the same space.
Neuromaps implements functions for transforming data between different spaces and the HCP data was used to generate “robust transforms,” she says. She went into detail in a recent blog post, but to summarize the concept of transforms she explained, “imagine you want to compare two maps of the earth, but one is provided as a globe and the other as a 2-D flat map. How can you make quantitative judgements? Neuromaps effectively transforms data from the globe to the flat map.”
Neuromaps developed “very organically,” says Hansen. The idea of putting brain maps together to create a central resource within the lab quickly gave way to the larger plan to develop a public repository. “But then Ross Markello [co-first author of the Nature Methods paper and key developer] had the brilliant idea of including the transformations. This was a game changer for neuromaps... [making it] a tool that makes it much easier to compare your brain data to all sorts of other brain data.”
Thanks to Markello, who has “a knack for writing good code,” neuromaps is intuitive to use, Hansen says. The main limitation is that it is written in Python and not currently available in other coding languages.
The addition of the transforms solved one of the early holdups in developing neuromaps—namely, deciding how to make the data available, she continues. “In what space? High or low resolution? in multiple spaces [i.e., have repeats]? Instead, we just make data available in their original space and then make it possible to transform between spaces however you'd like.”
Technically speaking, “the transformation functions make it possible to directly compare data, and the spatial nulls are there to generate statistical null models that account for the spatial autocorrelation of brain data,” says Hansen. The brain is spatially embedded and, “regions that are close together tend to be more similar than regions that are far apart.”
Neuromaps has three parts—the library, transformations, and spatial nulls—and each of them can be used alone, she adds.
Using The Tool
Researchers who plan to use neuromaps will need to do their own interpretation to make sense of the outputs or develop new hypotheses from these findings, says Hansen. “Just because neuromaps lets you compare brain data to other brain data doesn't mean that every association is ‘real’ [e.g., biologically supported].”
Additionally, “researchers will need to know which maps they want to compare their data against,” she says, although an integrated one-function analytic workflow is “in the works.” This requires that researchers make their own list of the annotations/brain maps they want to contextualize their work against.
While the data in neuromaps isn't subject-specific, it is possible for researchers to compare their subject-level data to the aggregated data, Hansen points out. For example, each of 10 brain maps from 10 individuals could be contextualized against the 20 maps in neuromaps, giving those 10 people a "fingerprint" based on how much their brain data looks like normative group-averaged features of brain structure and function.
“But it's important to remember that neuromaps data is group-averaged and normative,” she emphasizes. “An individual's cortical thickness map might look different from a group-average from many individuals, and this might tell us something about the individual.”
Her expectation is that neuromaps will “bridge gaps between scientific fields,” Hansen says. Researchers who study electrophysiological data, for instance, can now relate their research with genetic data or receptor distributions.
“The brain is a single integrated organ, so it makes sense to integrate all the layers of brain organization when we analyze data,” says Hansen. “Presumably, each level of organization—like gene expression versus functional hierarchy verses cortical thickness—are interacting with one another and related.”