MyPLS 2.0 - Partial least squares analysis for multivariate brain-behavior associations
Poster No:
1111
Submission Type:
Abstract Submission
Authors:
Daniela Zöller1,2,3, Valeria Kebets4, Thomas Bolton1,2, Dimitri Van De Ville1,2
Institutions:
1Institute of Bioengineering, Ecole Polytechnique Fédérale de Lausanne (EPFL), Geneva, Switzerland, 2Department of Radiology and Medical Informatics, University of Geneva, Geneva, Switzerland, 3Department of Psychiatry, University of Geneva, Geneva, Switzerland, 4Clinical Imaging Research Centre, National University of Singapore, Singapore
First Author:
Daniela Zöller
Institute of Bioengineering, Ecole Polytechnique Fédérale de Lausanne (EPFL)|Department of Radiology and Medical Informatics, University of Geneva|Department of Psychiatry, University of Geneva
Geneva, Switzerland|Geneva, Switzerland|Geneva, Switzerland
Institute of Bioengineering, Ecole Polytechnique Fédérale de Lausanne (EPFL)|Department of Radiology and Medical Informatics, University of Geneva|Department of Psychiatry, University of Geneva
Geneva, Switzerland|Geneva, Switzerland|Geneva, Switzerland
Co-Author(s):
Thomas Bolton
Institute of Bioengineering, Ecole Polytechnique Fédérale de Lausanne (EPFL)|Department of Radiology and Medical Informatics, University of Geneva
Geneva, Switzerland|Geneva, Switzerland
Institute of Bioengineering, Ecole Polytechnique Fédérale de Lausanne (EPFL)|Department of Radiology and Medical Informatics, University of Geneva
Geneva, Switzerland|Geneva, Switzerland
Dimitri Van De Ville
Institute of Bioengineering, Ecole Polytechnique Fédérale de Lausanne (EPFL)|Department of Radiology and Medical Informatics, University of Geneva
Geneva, Switzerland|Geneva, Switzerland
Institute of Bioengineering, Ecole Polytechnique Fédérale de Lausanne (EPFL)|Department of Radiology and Medical Informatics, University of Geneva
Geneva, Switzerland|Geneva, Switzerland
Introduction:
Unsupervised learning methods such as Partial Least Squares (PLS) can allow to overcome the limitations that arise with classification when classes are not well defined. PLS is a data-driven multivariate statistical technique that aims to extract relationships between two data matrices (McIntosh et al., 2004). PLS has previously been used to link neural variability with age (Garrett et al., 2010), or atrophy to symptoms in Parkinson's disease (Zeighami et al., 2019).
Here, we present a toolbox that deploys Behavior PLS, which aims to maximize the covariance between neuroimaging and behavioral data by deriving latent components (LCs) that are optimally weighted linear combinations of the original variables.
Here, we present a toolbox that deploys Behavior PLS, which aims to maximize the covariance between neuroimaging and behavioral data by deriving latent components (LCs) that are optimally weighted linear combinations of the original variables.
Methods:
myPLS is a Matlab-based analysis pipeline, openly available on GitHub (https://github.com/danizoeller/myPLS). The code is based on PLS correlation (PLSC) described in Krishnan et al. (2011).
Our command-line toolbox builds on a previous version of myPLS, which was presented at OHBM 2018. It incorporates a higher flexibility of possible input data formats (1D, 2D, or volumetric brain data). Besides 3D volumetric brain data (e.g., voxelwise cortical volume), one- and two-dimensional data formats are widely used by the neuroimaging community; e.g., when looking at region-wise graph metrics (1D) or correlation matrices (2D). Further, the user can choose between several options for the processing steps that are to be conducted, and between different alternatives for the visualization of the results.
The PLS analysis is described in Fig. 1.
Our command-line toolbox builds on a previous version of myPLS, which was presented at OHBM 2018. It incorporates a higher flexibility of possible input data formats (1D, 2D, or volumetric brain data). Besides 3D volumetric brain data (e.g., voxelwise cortical volume), one- and two-dimensional data formats are widely used by the neuroimaging community; e.g., when looking at region-wise graph metrics (1D) or correlation matrices (2D). Further, the user can choose between several options for the processing steps that are to be conducted, and between different alternatives for the visualization of the results.
The PLS analysis is described in Fig. 1.
Results:
We illustrate the toolbox with two datasets linking (1) resting-state functional connectivity (RSFC) to clinical, cognitive and personality measures, and (2) resting-state coupling of amygdala and hippocampus to anxiety.
In the first example (taken from Kebets et al., 2019), we use 2D imaging data, and compute PLSC without considering grouping information. Our analysis comprises data from 224 subjects (Fig. 2A), including healthy controls (HCs) as well as individuals with psychiatric disorders. PLSC was computed between whole-brain RSFC and behavior, resulting in four significant LCs (Fig. 2A). For the sake of brevity, we only describe LC1 here (full results can be found in Kebets et al., 2019). LC1 (r=0.78, Fig. 2C) accounts for 20% of RSFC-behavior covariance (Fig. 2B), and is characterized by high loadings on clinical measures (e.g., mood lability), and low loadings on cognitive scores (Fig. 2D). The associated RSFC pattern (Fig. 2E) showed decreased RSFC within the somatomotor network, while greater RSFC between sensory-motor and dorsal attention networks showed greater RSFC with the salience network and with subcortical regions.
In the second example (taken from Zöller et al., 2019), we use 1D imaging data and include grouping information in the PLSC analysis. Our analysis comprises data from 152 subjects (Fig. 2F), including 74 HCs and 78 patients with 22q11.2 deletion syndrome, a genetic disorder with an elevated risk for developing psychiatric disorders. We investigated the resting-state coupling between an amygdala and hippocampus network and six other brain networks. PLSC was computed between these six coupling measures and anxiety, resulting in one significant LC (Fig. 2F) with r=0.31 (Fig. 2H), and accounting for 80% of coupling-anxiety covariance (Fig. 2G). Behavior saliences indicate that LC1 reflects a relationship that is present only in patients (Fig. 2I). Imaging saliences (Fig. 2J) show that higher coupling of amygdala/hippocampus with two networks (language network and dorsal anterior cingulate cortex/dorsolateral prefrontal cortex), and lower coupling with the anterior default mode network, are associated with higher anxiety.
In the first example (taken from Kebets et al., 2019), we use 2D imaging data, and compute PLSC without considering grouping information. Our analysis comprises data from 224 subjects (Fig. 2A), including healthy controls (HCs) as well as individuals with psychiatric disorders. PLSC was computed between whole-brain RSFC and behavior, resulting in four significant LCs (Fig. 2A). For the sake of brevity, we only describe LC1 here (full results can be found in Kebets et al., 2019). LC1 (r=0.78, Fig. 2C) accounts for 20% of RSFC-behavior covariance (Fig. 2B), and is characterized by high loadings on clinical measures (e.g., mood lability), and low loadings on cognitive scores (Fig. 2D). The associated RSFC pattern (Fig. 2E) showed decreased RSFC within the somatomotor network, while greater RSFC between sensory-motor and dorsal attention networks showed greater RSFC with the salience network and with subcortical regions.
In the second example (taken from Zöller et al., 2019), we use 1D imaging data and include grouping information in the PLSC analysis. Our analysis comprises data from 152 subjects (Fig. 2F), including 74 HCs and 78 patients with 22q11.2 deletion syndrome, a genetic disorder with an elevated risk for developing psychiatric disorders. We investigated the resting-state coupling between an amygdala and hippocampus network and six other brain networks. PLSC was computed between these six coupling measures and anxiety, resulting in one significant LC (Fig. 2F) with r=0.31 (Fig. 2H), and accounting for 80% of coupling-anxiety covariance (Fig. 2G). Behavior saliences indicate that LC1 reflects a relationship that is present only in patients (Fig. 2I). Imaging saliences (Fig. 2J) show that higher coupling of amygdala/hippocampus with two networks (language network and dorsal anterior cingulate cortex/dorsolateral prefrontal cortex), and lower coupling with the anterior default mode network, are associated with higher anxiety.
Conclusions:
PLSC is a powerful approach for linking data from two modalities. Here, we present a user-friendly set of command line tools that incorporate a multitude of options for the application of PLS correlation to different types of neuroimaging data.
Modeling and Analysis Methods:
Activation (eg. BOLD task-fMRI)
Connectivity (eg. functional, effective, structural)
Methods Development 1
Multivariate Approaches 2
Neuroinformatics and Data Sharing:
Workflows
Keywords:
Data analysis
FUNCTIONAL MRI
Machine Learning
Modeling
Multivariate
Statistical Methods
Other - Partial Least Squares
1|2Indicates the priority used for review
My abstract is being submitted as a Software Demonstration.
Please indicate below if your study was a "resting state" or "task-activation” study.
Healthy subjects only or patients (note that patient studies may also involve healthy subjects):
Was any human subjects research approved by the relevant Institutional Review Board or ethics panel? NOTE: Any human subjects studies without IRB approval will be automatically rejected.
Was any animal research approved by the relevant IACUC or other animal research panel? NOTE: Any animal studies without IACUC approval will be automatically rejected.
Please indicate which methods were used in your research:
For human MRI, what field strength scanner do you use?
Which processing packages did you use for your study?
Provide references using author date format
Garrett D.D., Kovacevic N., McIntosh A.R., Grady C.L. (2010), ‘Blood oxygen level-dependent signal variability is more than just noise’, Journal of Neuroscience, vol. 30, issue 14, pp. 4914-21.
Kebets V., Holmes A.J., Orban C., Tang S., Li J., Sun N., Kong R., Poldrack R.A., Yeo B.T.T. (2019), ‘Somatosensory-Motor Dysconnectivity Spans Multiple Transdiagnostic Dimensions of Psychopathology’, Biological Psychiatry, vol. 86, issue 10, pp. 779-791.
Krishnan, A., Williams, L.J., McIntosh, A.R., Abdi, H. (2011), ‘Partial Least Squares (PLS) methods for neuroimaging: A tutorial and review’, Neuroimage, vol. 56, pp. 455–475.
McIntosh, A.R., Lobaugh, N.J. (2004), ‘Partial least squares analysis of neuroimaging data: Applications and advances’, Neuroimage, vol. 23, pp. 250–263.
Zeighami Y., Fereshtehnejad S.M., Dadar M., Collins D.L., Postuma R.B., Mišić B., Dagher A. (2019), ‘A clinical-anatomical signature of Parkinson's disease identified with partial least squares and magnetic resonance imaging’, Neuroimage, vol. 190, pp. 69-78.
Zöller, D., Sandini, C., Karahanoğlu, F. I., Padula, M. C., Schaer, M., Eliez, S., & Van De Ville, D. (2019), ‘Large-scale brain network dynamics provide a measure of psychosis and anxiety in 22q11.2 deletion syndrome’, Biological Psychiatry: Cognitive Neuroscience and Neuroimaging, vol. 4, issue 10, pp. 881-892.