Representational geometry, not topography, best characterizes BOLD signals in multimodal brain areas
Poster No:
2129
Submission Type:
Abstract Submission
Authors:
Bogdan Petre1, Martin Lindquist2, Tor Wager1
Institutions:
1Dartmouth College, Hanover, NH, 2Johns Hopkins University, Baltimore, MD
First Author:
Co-Author(s):
Introduction:
Neural representations are characterized both as coarse topographic maps (Sereno, 1995) and embedded population codes (Churchland 2012, Mante 2013). Brain mapping hinges on stereotactic characterizations of brain function, and crucially neglects idiosyncratic fine scale embeddings, but accumulating evidence shows high dimensional embeddings exists at scales relevant to topographic maps (Sengupta, 2017). Here we use BOLD fMRI to systematically evaluate if topographic maps or embedded feature spaces provide a better account of millimeter scale population level brain organization. We show representational geometry offers a richer account of multimodal but not unimodal brain function.
Methods:
We compare two brain encoding models using 3T BOLD fMRI data from the Human Connectome Project (N=610). One model, diffeomorphic multivariate alignment (DMA), preserves local topography. The second is a novel variant of parcel-wise high-dimensional alignment (fixed mean hyperalignment, fmHA) that preserves representational geometry and mean regional evoked responses, but not local topography. Models are fit to functional connectomes estimated from resting-state data (15 min/participant). We first use DMA to align pairs of individuals and then apply fmHA to resulting connectomes. Both models are tested for their ability to predict responses evoked by a battery of 7 cognitive, sensory and motor tasks in independent runs of independent aligned participants (N=305), measured in terms of improvements in pairwise between subject correlations (BSC).
We expect alignment of task evoked responses will improve most in transmodal networks (especially default mode [DMN], frontoparietal [FP], dorsal attention [DA] and salience), and define these as a priori networks of interest. First, resting state alignment is limited by the idiosyncrasies at rest which are greatest in transmodal areas (Gratton, 2018). Second, we expect a unique advantage for fmHA in areas dominated by population code embeddings, while DMA is expected to perform best in areas dominated by coarse scale topographic maps. The latter are undisputed in unimodal areas but topographic organization of multimodal and transmodal brain areas remains largely uncharted.
We expect alignment of task evoked responses will improve most in transmodal networks (especially default mode [DMN], frontoparietal [FP], dorsal attention [DA] and salience), and define these as a priori networks of interest. First, resting state alignment is limited by the idiosyncrasies at rest which are greatest in transmodal areas (Gratton, 2018). Second, we expect a unique advantage for fmHA in areas dominated by population code embeddings, while DMA is expected to perform best in areas dominated by coarse scale topographic maps. The latter are undisputed in unimodal areas but topographic organization of multimodal and transmodal brain areas remains largely uncharted.
Results:
On average across task conditions and throughout the cortex DMA categorically improves BSCs relative to surface anatomical alignment (+7.4% BSC, z-fisher r = +0.023 ± 0.0004, t305=49.0, p < 0.001). fmHA incrementally improves on DMA in transmodal networks (+2% BSC for fmHA-DMA, z-fisher r = +0.005 ± 0.0007, p < 0.001) but DMA was superior to fmHA in unimodal networks (-15% BSC for fmHA-DMA, z-fisher r = -0.07 ± 0.001, p < 0.001). Among transmodal networks, DMN (+15% BSC, z-fisher r = +0.021 ± 0.001, p < 0.001), FP (+14% BSC, z-fisher r = 0.022 ± 0.001, p < 0.001) and DA (+5% BSC, z-fisher r = 0.012 ± 0.001, p < 0.001) networks show statistically significant increases in BSC for fmHA - DMA (mean±sem, Holm-Sidak corrected p for 7 comparisons). This presents a conspicuous profile that casts unimodal and transmodal brain areas in stark relief (Figure).
·Hyperalignment improves between subject correlations in task evoked responses relative to diffeomorphic alignment in multimodal but not unimodal regions.
Conclusions:
Prior efforts to establish shared group level representations in the brain have been limited in scope (Haxby 2011), precision (Bazeille 2021) or are confounded by model averaging (Guntupalli 2016) and transformations which disrupt both topographic maps and evoked response amplitudes simultaneously. We use granular, interpretable models to show support for different representational organizations in different brain areas. While topographic organization offers a useful account throughout the brain, the relative superiority of fmHA in transmodal brain areas shows millimeter scale stereotactic population mapping in these regions may be invalid and recommends measures of shared representational features instead. Conversely, the disruption of functional organization of sensory motor regions by fmHA suggests topographic organization in unimodal areas may be uniquely important.
Higher Cognitive Functions:
Higher Cognitive Functions Other
Modeling and Analysis Methods:
Activation (eg. BOLD task-fMRI) 2
Multivariate Approaches
Neuroanatomy, Physiology, Metabolism and Neurotransmission:
Cortical Anatomy and Brain Mapping 1
Perception, Attention and Motor Behavior:
Perception: Multisensory and Crossmodal
Keywords:
Cognition
Computational Neuroscience
Cortex
FUNCTIONAL MRI
Language
Motor
Multivariate
Somatosensory
Spatial Warping
Vision
1|2Indicates the priority used for review
Provide references using author date format
Churchland, M. M. et al. (2012), Neural population dynamics during reaching. Nature 487, 51–56.
Feilong, M., et al. (2021), The neural basis of intelligence in fine-grained cortical topographies. Elife 10.
Gratton, C. et al. (2018), Functional Brain Networks Are Dominated by Stable Group and Individual Factors, Not Cognitive or Daily Variation. Neuron 98, 439–452.e5.
Guntupalli, J. S. et al. (2016), A Model of Representational Spaces in Human Cortex. Cereb. Cortex 26, 2919–2934.
Haxby, J. V. et al. (2011), A common, high-dimensional model of the representational space in human ventral temporal cortex. Neuron 72, 404–416.
Mante, V., et al. (2013), Context-dependent computation by recurrent dynamics in prefrontal cortex. Nature 503, 78–84.
Sengupta, A., et al. (2017), The effect of acquisition resolution on orientation decoding from V1 BOLD fMRI at 7T. Neuroimage 148, 64–76.
Sereno, M. I. et al. (1995), Borders of multiple visual areas in humans revealed by functional magnetic resonance imaging. Science 268, 889–893.