A multivariate distance-based analytic framework for connectome-wide association studies - PubMed (original) (raw)

A multivariate distance-based analytic framework for connectome-wide association studies

Zarrar Shehzad et al. Neuroimage. 2014 Jun.

Abstract

The identification of phenotypic associations in high-dimensional brain connectivity data represents the next frontier in the neuroimaging connectomics era. Exploration of brain-phenotype relationships remains limited by statistical approaches that are computationally intensive, depend on a priori hypotheses, or require stringent correction for multiple comparisons. Here, we propose a computationally efficient, data-driven technique for connectome-wide association studies (CWAS) that provides a comprehensive voxel-wise survey of brain-behavior relationships across the connectome; the approach identifies voxels whose whole-brain connectivity patterns vary significantly with a phenotypic variable. Using resting state fMRI data, we demonstrate the utility of our analytic framework by identifying significant connectivity-phenotype relationships for full-scale IQ and assessing their overlap with existent neuroimaging findings, as synthesized by openly available automated meta-analysis (www.neurosynth.org). The results appeared to be robust to the removal of nuisance covariates (i.e., mean connectivity, global signal, and motion) and varying brain resolution (i.e., voxelwise results are highly similar to results using 800 parcellations). We show that CWAS findings can be used to guide subsequent seed-based correlation analyses. Finally, we demonstrate the applicability of the approach by examining CWAS for three additional datasets, each encompassing a distinct phenotypic variable: neurotypical development, Attention-Deficit/Hyperactivity Disorder diagnostic status, and L-DOPA pharmacological manipulation. For each phenotype, our approach to CWAS identified distinct connectome-wide association profiles, not previously attainable in a single study utilizing traditional univariate approaches. As a computationally efficient, extensible, and scalable method, our CWAS framework can accelerate the discovery of brain-behavior relationships in the connectome.

Keywords: Brain–behavior relationships; Connectome; Discovery; Functional connectivity; Phenotype; Resting-state.

Copyright © 2014 Elsevier Inc. All rights reserved.

PubMed Disclaimer

Figures

Figure 1

Figure 1. Schematic of the CWAS Analytic Framework

The proposed framework is carried out for each grey matter voxel separately. Starting with (1) a single seed voxel, (2) maps of temporal correlations (i.e., functional connectivity) between the seed voxel and every other grey matter voxel are generated (yellow-red = positive correlations, blue-cyan = negative correlations). Next, (3) the distance (one minus the correlation) between every pair of participants' connectivity maps is computed (red-white = positive correlation or distance of 0-1, white-blue = negative correlation or distance of 1-2), and (4) an analysis of distance, analogous to an analysis of variance, is performed. The sum of squared distances between typically developing children (TDC) and children with ADHD is compared to the residual sum of squared distances (the error term) to produce a pseudo-F statistic. An accompanying p-value is obtained by permutation. This workflow (steps 1-4) is then repeated for every grey matter voxel to produce a whole-brain map of _p_-values values, which is then thresholded at p<0.05 and corrected for multiple comparisons by setting a minimum cluster size (p < 0.05) based on Monte Carlo simulations. All flat maps are of the right hemisphere and are rendered on the PALS-B12 atlas in caret.

Figure 2

Figure 2. Overview of Analyses

Goals or objectives are given in the first column, followed by the particular analyses used in the second column, the datasets examined in the third column, and the associated figures in the final column. Supplementary analyses provide further applications and details of MDMR-based CWAS, including false positive and negative rates using simulations and the potential impact of selecting difference distance measures.

Figure 3

Figure 3. Connectome-Wide Associations for IQ Revealed with MDMR

(a) Significant associations between whole-brain functional connectivity and intelligence (full-scale IQ) are shown with p < 0.05 (cluster-corrected using Gaussian random field theory [GRF]) for resting-state Scans 1 and 2. (b) Overlap of significant associations related to IQ between the two scans is shown in red.

Figure 4

Figure 4. Large-Scale Network Membership of IQ CWAS

(a) Overlap of significant associations related to IQ between two scans (same as Fig 3b) is shown in red and overlaid on seven functional networks of interest presented by Yeo et al. (2011). The networks are displayed as semi-transparent and correspond to the colored labels given in (b). (b) The proportion of significant associations related to IQ in each of the seven networks is depicted graphically.

Figure 5

Figure 5. Reproducibility of IQ CWAS

Associations related to IQ were calculated foe each of 500 bootstrap subsamples.(a), Based upon all possible pairings of boostrap subsamples, the histograms display the overlap (Dice coefficient) between significant associations (p < 0.05, uncorrected) and the similarity (Spearman rho) between unthresholded associations. (b) Four random subsamples were selected and the significant associations related to IQ are shown with p < 0.05 (cluster-corrected using GRF).

Figure 6

Figure 6. Correspondence between CWAS and Neurosynth (Automated Meta-Analyses)

Using our approach, significant associations with intelligence based on MDMR are shown in blue-white for Scans 1 (a) and 2 (b). Using the automated meta-analysis provided with Neurosynth, significant brain activity associated with key terms that were related to IQ (intelligence, reasoning, and WM) are shown in purple-white. The overlap between our approach and the automated meta-analysis is shown in yellow.(c) Only the overlap is shown between the combined MDMR results for Scans 1 and 2 compared to the neurosynth meta-analytic findings.

Figure 7

Figure 7. Effects of Global Signal and Mean Connectivity on IQ CWAS

Significant associations for intelligence are shown with p < 0.05 (GRF corrected) when (a) removing the global signal at the individual level, (b) adding the mean connectivity (over all voxelwise functional connections) as a covariate at the group level, or (c) for comparison when using no global-based correction (same as Figure 3a). All results are shown for both resting-state scans.

Figure 8

Figure 8

Comparison of MDMR analysis and GLM seed-based connectivity analysis (SCA). (a) For each scan, a voxel's significance based on the MDMR analysis (x-axis) is plotted against the percent of individual connections exhibiting a significant connectivity-IQ relationship for the same voxel in a GLM analysis (y-axis).(b) For each scan, brain regions ranking among the top 15% based on the MDMR analysis are shown in red, those ranking in the same percentiles based on the number of significant GLM findings are shown in green, and the overlap between the two is shown in yellow. (c) The same plot as (b) is shown except mean connectivity is added as a covariate in the MDMR and GLM-based analyses.

Figure 9

Figure 9. MDMR as a Guide to SCA

(a) From Scan 1, 100 regions-of-interest (individual voxels) were selected from associations related to IQ and divided into four groups (25 each) including global maxima significant (maximally significant voxels with p < 0.05; red), significant (random voxels with _p_ < 0.05; green), non-significant (random voxels with _p_ > 0.05; cyan), and global minima (minimally significant voxels with p > 0.05; purple). Bar plots show the percent of individual functional connections that significantly vary with IQ based on a GLM analysis (y-axis), which are averaged within each ROI group (x-axis) for resting-state Scans 1 and 2. (b) As in plot (a), but for 100 ROIs selected from Scan 2.

Figure 10

Figure 10. Applicability of MDMR-based CWAS

(a) Significant (p < 0.05, cluster-corrected using GRF) connectome-wide associations for development (age), ADHD diagnosis (children with ADHD vs. typically developing controls), and L-DOPA administration (placebo vs. L-DOPA).(b) For each dataset, the percent of significant associations within each of seven functional networks of interest (Yeo et al., 2011).

Figure 11

Figure 11. Connectome-Wide Associations for IQ Across Different Parcellations

Significant connectome-wide associations for intelligence are shown using 50, 200, 800, and 3200 parcellations as well as the original voxelwise results. All analyses use resting-state Scan 1 and are thresholded at p < 0.05 (cluster-corrected using GRF).

Figure 12

Figure 12. Applicability of Parcellation-Based CWAS

Significant connectome-wide associations for ADHD diagnosis (children with ADHD vs. typically developing controls), development (age), full-scale intelligence (Scan 1), and L-DOPA administration (placebo vs. L-DOPA) are shown based on voxelwise data (left) and 800 parcellation units (right). All datasets are threshold voxelwise at p< 0.05, then cluster-level corrected using GRF at p < 0.05.

Similar articles

Cited by

References

    1. Akil H, Martone ME, Essen DCV. Challenges and Opportunities in Mining Neuroscience Data. Science. 2011;331:708–712. - PMC - PubMed
    1. Anderson MJ. A new method for non-parametric multivariate analysis of variance. Austral Ecol. 2001;26:32–46.
    1. Andersson JLR, Jenkinson M, Smith S. Non-Linear Registration, aka Spatial Normalisation (No TR07JA2), FMRIB Technical Report. FMRIB Centre; Oxford, UK: 2007.
    1. Andrews-Hanna JR, Reidler JS, Sepulcre J, Poulin R, Buckner RL. Functional-Anatomic Fractionation of the Brain's Default Network. Neuron. 2010;65:550–562. - PMC - PubMed
    1. Beckmann CF. Modelling with independent components. NeuroImage. 2012;62:891–901. - PubMed

Publication types

MeSH terms

Grants and funding

LinkOut - more resources