Identification of Multimodal MRI and EEG Biomarkers using ... · 2 A Taxonomy of Approaches 7 J. Liu and V. Calhoun, "A review of multivariate analyses in imaging genetics," Frontiers

1

1

Univariate - Multivariate Approaches:

Joint Modeling of Imaging & Genetic Data

Vince D. Calhoun, Ph.D.Executive Science Officer &

Director, Image Analysis & MR ResearchThe Mind Research Network

Distinguished Professor, Electrical and Computer Engineering (primary), Biology, Computer Science, Psychiatry, & Neurosciences

The University of New Mexico

Overview

• Introduction

• A Taxonomy of Approaches

• Candidate gene analysis (c1)

• A prior based multivariate analysis (on genetic, c2)

• Data-driven multivariate analysis (on genetic, c2)

• Component-based analysis (on imaging, c3)

• Multivariate analysis across both (c4)

• Multiple variables for brain + gene w/ priors (c4+c2)

• Multivariate analysis across N>2 phenotypes (c4+)

• Integration of additional bioinformatics resources

• Challenges & Future Directions

2

The First Challenge

3 4

Genetic Information

insertiondeletion

• Genetic: single nucleotide polymorphism (SNP)

• Genetic: copy number variation (CNV)

• Epigenetic: methylation

5

Overview

• Introduction











6

2

A Taxonomy of Approaches

7

J. Liu and V. Calhoun, "A review of multivariate analyses in imaging genetics," Frontiers in Neuroinformatics,

vol. 8, pp. 1-11, 2014

Overview

• Introduction











8

Candidate gene approach (c1)

9

Example: APOE4 in Alzheimer’s Disease

10

Properties

11 12

Why Multivariate?

Cross–validation performance using the top M SNP’s selected

via two different methods as the basis for an N-factor principle

component model.

3

Overview

• Introduction











13

A prior based multivariate analysis (on genetic, c2)

14A. Subramanian, et al, Proc Natl Acad Sci U S A, vol. 102, pp. 15545-15550, Oct 25 2005

Gene-set enrichment analysis

A prior based multivariate analysis (on genetic, c2)

• GSEA, Methylation, & Hippocampal Volume

15

Ehrlich, S, “Associations between DNA methylation and schizophrenia-related intermediate

functional and structural imaging phenotypes – a gene set enrichment analysis”, in preparation.

Overview

• Introduction











16

Data-driven multivariate analysis (on genetic, c2)

• Reduction of genetic data

• Penalized regression (LASSO)

• PCA

• ICA

17 18

Methylation Gender Correction (c2)

J. Liu, K. Hutchison, M. Morgan, N. I. Perrone-Bizzozero, J. Sui, and V. D. Calhoun, "Identification of

Genetic and Epigenetic Factors Contributing to Population Structure," PLoS ONE, vol. 5, pp. 1-8, 2010

4

19

• Genomic ~27,000 sites from 23 pairs of chromosomes

• 130 subjects (heavy drinker 33 females, 97 males, age 31.39.7)

Goal: Association with gender, age, BMI, alcohol use, cigarette use, marijuana use, depression, stress, etc.

Results: Methylation correlation

0.490.4

0.45

0.55

0.95

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

Gender Age* BMI* Max_drinks* MJ_use*

*: after gender effect correction

Epigenetic-methylation study



Ancestry effect on methylation



Overview

• Introduction











21

Component-based analysis (on imaging, c3)

22

D. C. Glahn, A. M. Winkler, P. Kochunov, L. Almasy, R. Duggirala, M. A. Carless, J. C. Curran, R. L.

Olvera, A. R. Laird, S. M. Smith, C. F. Beckmann, P. T. Fox, and J. Blangero, "Genetic control over

the resting brain," Proc Natl Acad Sci U S A, vol. 107, pp. 1223-1228, Jan 19 2010

Overview

• Introduction











23

Multiple brain variables & multiple genetic variables (c4)

24

J. Liu and V. Calhoun, "A review of multivariate analyses in imaging genetics," Frontiers in Neuroinformatics,

vol. 8, pp. 1-11, 2014

5

25

fMRI/SNP connection

• fMRI component: specific brain regions with common independent brain function.

• SNP component: a linearly weighted group of SNPs functioning together.

• Relationship assumptionIf an association of SNPs partially define a certain brain function in

specific brain regions, Then, the linked fMRI and SNP components should share a similar pattern of existence across subjects.

fMRI components SNP components

Expression

pattern

Expression

pattern

Subject 1

..

..

Subject N

RRR: one (of many) possible frameworks

Spatial ICA

ICA Components

fMRI

PPI database

Genetic module extraction

Modules

Two sample t-test

Collaborative

sparse reduced

rank regression

GWAS

Module, Gene, SNP

ranking

Functional enrichment

analysis

Imaging Endo-phenotype

ExtractionGenetic Network

construction

Association analysis

Canonical Correlation Analysis (CCA) model

Illustration of combining both imaging and SNP data with the CCA

model to identify correlated genes and voxels.

H. Cao, J. Duan, D. Lin, Y. Shugart, V. D. Calhoun, and Y.-P. Wang, "Sparse Representation Based Biomarker Selection for

Schizophrenia with Integrated Analysis of fMRI and SNPs," NeuroImage, in press,28

A1

Parallel ICA: Two Goals

X1

S2S1

W2W1

X2

Data1: (fMRI) Data2: (SNP)

A2

Identify

Hidden

Features

Identify

Hidden

Features

Identify Linked Features

J. Liu, O. Demirci, and V. D. Calhoun, "A Parallel Independent Component Analysis Approach to Investigate

Genomic Influence on Brain Function," IEEE Signal Proc. Letters, vol. 15, pp. 413-416, 2008.

MAX : {H(Y1) + H(Y2) }, <Infomax> 2

1i 2j2

1 2

1i 2j

Cov(a ,a )g( )=Correlation(A ,A ) =

Var(a )×Var(a )Subject to: arg max g{W1,W2 ŝ1,ŝ2 },

Simulation

29

Simulation results suggest that parallel ICA, in general, is able to extract more accurately the components and connections

than a correlation test, in particular for weak linkages. Results also indicate that the ratio of sample size to SNP size should

be at least 0.02. However, when the data have a low odds ratio or cases vs. controls ratio, the correlation test provides

results reliably, though with lower accuracy.

Liu, J, et al, IEEE Bioinformatics and Biomedicine. 2008: Philadelphia, PA.

Simulation: Designed to provide a more complete understanding of Parallel ICA while applied to genomic SNP array

studies. We specified the parameters for each component and input them into PLINK, an open-source whole genome

association analysis toolset [http://pngu.mgh.harvard.edu/purcell/plink/].

Conditions: sample size effect. case to control ratio, SNP array size effect, case-related SNP’s vs. total SNP’s, odds ratio,

connection strength between genotype and phenotype effects

Initial Proof of Concept: SNP/fMRI Fusion

30

J. Liu, G. D. Pearlson, A. Windemuth, G. Ruano, N. I. Perrone-Bizzozero, and V. D. Calhoun, "Combining

fMRI and SNP data to investigate connections between brain function and genetics using parallel ICA,"

Hum.Brain Map., vol. 30, pp. 241-255, 2009

SNP Z score Gene

Rs1466163 -4.08 AADC: aromatic L-amino acid decarboxylase

Rs2429511 3.97 ADRA2A: alpha-2A adrenergic receptor gene

Rs3087454 -3.09 CHRNA7: alpha 7 nicotinic cholinergic receptor

Rs821616 2.96 DISC1: disrupted in schizophrenia 1

Rs885834 -2.78 CHAT: choline acetyltransferase

Rs1355920 -2.77 CHRNA7: cholinergic receptor, nicotinic, alpha 7

R4765623 2.73 SCARB1: scavenger receptor class B, member 1

Rs4784642 -2.71 GNAO1: guanine nucleotide binding protein (G protein),

alpha activating activity polypeptide O

Rs2071521 2.58 APOC3: apolipoprotein C-III

Rs7520974 2.55 CHRM3: muscarinic-3 cholinergic receptor

Data Description: 20 Sz & 43 Healthy controls

fMRI: one image per subject (Target activation in AOD task)

SNP: one array per subject (384 SNP genotypes - -> 367 SNPs)

Control vs Patient

p<0.001

6

SNP Selection and Approach• Schizophrenia patients and healthy controls

• MCIC data: Boston, Iowa, Minnesota and New Mexico

• Genome-wide 1M SNP data - [biallelic coding (AA, AB, or BB]

• fMRI sensorimotor task- Block design motor response to auditory

stimulation

• 208 subjects with SNP (777365 SNPs) and fMRI data (52322 voxels)

• SNP selection

• SNPs differentiating schizophrenia patients and healthy controls are

included (p-value < 0.005, 3318 SNPs)

• SNPs related to schizophrenia risk genes are included (1843 SNPs

selected, related to DISC1, COMT, etc.)

• Combine selected SNPs: 5157 SNPs as final input

• Parallel ICA

• Use subject type (SZ patients. vs. healthy control) as reference

• Apply reference PCA to SNP and fMRI data to reduce dimension and

select components of interest

• Apply parallel ICA to identify linked components optimized to the

correlation between the two modalities

31

J. Chen, V. D. Calhoun, G. D. Pearlson, S. Ehrlich, J. Turner, B. C. Ho, T. Wassink, A. Michael, and J. Liu, "Multifaceted

genomic risk for brain function in schizophrenia," NeuroImage, vol. 61, pp. 866-875, 2012.

Number of Components

• SNP data component number selection

• Component number estimated to be 8 for fMRI data (MDL)

• Increase component number from 2 to 60 for SNP data

• Compare the identified fMRI and SNP components obtained with

different SNP component numbers

• Sliding window covering 5 consecutive component numbers

• Locate a region in which identified components remain stable

(select 5% most contributing SNPs from the identified component

and evaluate the overlapping ratios)

32

SNP component number Overlapping ratio Correlation

6 0.843 0.982

7 0.927 0.984

8 0.851 0.957

9 0.970 0.951

Reference SNP component number = 5

J. Chen, V. D. Calhoun, and J. Liu, "ICA Order Selection Based on Consistency: Application to Genotype Data," in Proc.

EMBS, San Diego, CA, 2012.

Consistency

• Leave-one-out validation

• Divide subjects into 10 sets

• Apply parallel-ICA to 90% of the data in each validation run

• Evaluate the consistency of the components based on correlations

between components identified in each validation run and the reference

components identified with the full dataset

33

……

1 2 …… 10Subjects

Resulting Linked Component• fMRI component number = 8, SNP component number = 5

• One pair of linked components is identified, with p-value passing

Bonferroni correction

• Bootstrap: Multiple runs of parallel-ICA with 5157 randomly selected

SNPs. The median correlation was 0.16

34

fMRI component index

SNP component index

rfMRI-SNP P-value

1 3 -0.065 3.49E-01

2 2 0.042 5.48E-01

3 1 0.099 1.51E-01

4 2 -0.138 4.54E-02

5 1 0.141 4.16E-02

6 5 -0.128 6.44E-02

7 1 0.178 9.95E-03

8 4 0.282 3.39E-05



Validation Results

• For each validation run, identify

the replicated fMRI/SNP

component exhibiting most similar

pattern to the reference fMRI/SNP

component identified with full

dataset

• Calculate the fMRI-SNP

correlation between the replicated

components obtained in the

validation run

• Determine if the replicated

components are the highest linked

pair (highest correlation)

• A successful validation: both SNP

and fMRI components show

similar patterns to the reference

SNP and fMRI components,

respectively; and they are linked

35

Index of validation run

rSNP-fMRIHighest linked

pair?

1 0.301 √

2 -0.295 √

3 -0.334 0.339

4 -0.301 √

5 0.246 √

6 -0.249 0.259

7 -0.306 √

8 0.347 √

9 0.387 √

10 0.304 √



Identified fMRI component

36

Brain region Brodmann area Volume Max z-score

Postcentral Gyrus 1, 2, 3, 5, 7, 40, 43 9.7/4.4 6.23(42,-32,60)/5.70(-39,-32,62)

Precentral Gyrus 4, 6 8.2/1.7 6.03(42,-12,56)/4.40(-59,-12,42)

Inferior Parietal Lobule 40 3.0/0.5 6.03(45,-35,57)/4.88(-45,-32,57)

Medial Frontal Gyrus :6, 32 1.0/1.6 4.82(3,-3,53)/5.27(0,-3,53)

Superior Temporal Gyrus 21, 22, 38 1.4/0.5 3.24(62,-12,1)/2.93(-50,-3,-5)



7

Summary of SNP Results• Conduct pathway analysis and functional annotation clustering based

on identified 94 genes

• IPA (Ingenuity Pathway Analysis) identifies “Schizophrenia of humans”

as one of the top biofunctions, involving 11 genes

• IPA also identifies a number of significant canonical pathways, four of

which are related to neurotransmitter signaling

• David’s Bioinformatics Resource reports the most significant cluster to be

functionally related to synapse. A cluster annotated as “cell projection” is

also identified

37

Disease and disorder Gene p-valueSchizophrenia of humans BDNF, COMT, DISC1, DRD3, ERBB4, GAD1,

GRIN2B, HTR7, NOTCH4, NRG1, NRG36.49E-09

Neurotransmitter signaling pathway Gene p-valueGABA receptor signaling GABRA4, GABRG3, GAD1 2.13E-03

Dopamine receptor signaling COMT, DRD3, PPP2R2C 7.66E-03

Neuregulin signaling NRG1, NRG3, ERBB4 1.15E-02

Glutamate receptor signaling GRIN2B, GRID2 3.74E-02

Functional annotation cluster Gene p-valueSynapse GABRA4, GABRG3, GAD1, GRIN2B, GRID2, ERBB4,

SHC4, OTOF, PSD3, CTBP25.20E-03

Cell projection DRD3, GAD1, GRIN2B, MYCBP2, DNAH11, WNT2, ESR1, CDH13, ALCAM, MYO5A

9.70E-02



ADNI Data

• European-American ADNI subjects (N=757)

• 209 HC (Mean/SD Age = 76.05/4.94; 113 Males) with

no past history of neurological or psychiatric disorder,

• 367 subjects with MCI (Mean/SD Age = 74.95/7.37;

239 Males)

• 181 subjects (Mean/SD Age = 75.57/7.48; 100 Males)

with clinically-assessed AD

• SNP data

• 533,872 SNPs (after QA)

• sMRI data

• Baseline 1.5T MRI scans

• Segmented into GM maps

38

S. Meda, B. Narayanan, J. Liu, N. I. Perrone-Bizzozero, M. Stevens, V. D. Calhoun, D. C. Glahn, L. Shen, S. L.

Risacher, A. J. Sayking, and G. D. Pearlson, "A large scale multivariate parallel ICA method reveals novel imaging-

genetic relationships for Alzheimer's Disease in the ADNI cohort," NeuroImage, vol. 60, pp. 1608-1621, 2012

ADNI Data

39




Genetic Components by Group

40




Alzheimer’s disease functional interaction pathway

41modified from Sleegers K et al., TIGS 2009

42

sMRI/SNP

A. sMRI component –A (group difference)

B. sMRI component –B (linked, but no group difference)

K. Jagannathan, V. D. Calhoun, J. Gelernter, M. Stevens, J. Liu, F. Bolognani, A. Windemuth, G. Ruano, and G. D.

Pearlson, "Genetic associations of brain structural networks in schizophrenia: a preliminary study using parallel ICA,"

Biological Psychiatry, vol. 68, pp. 657-666, 2010

Structural deficits in brain regions consistently implicated in previous schizophrenia reports, including frontal and

temporal lobes and thalamus were related to SNPs from 16 genes, several previously associated with schizophrenia

risk and/or involved in normal CNS development, including AKT, PI3k, SLC6A4, DRD2, CHRM2 and ADORA2A.

8

Genetics and P3 ERP generation

• Subjects: 41 healthy subjects (24 female, 17 male)

• EEG collected during AOD task,

target/novel ERPs extracted

• Blood sample collected, genotyped 384 SNPs from 222

genes 6 physiological systems.

43

J. Liu, K. A. Kiehl, G. D. Pearlson, N. I. Perrone-Bizzozero, and V. D. Calhoun, "Genetic Determinants of

Target and Novelty Processing," NeuroImage, vol. 46, pp. 809-816, 2009. 44

ERP Topography & SNP Associations

Tar

get

Sti

muli

Novel

Sti

muli

0.55

0.47


Target and Novelty Processing," NeuroImage, vol. 46, pp. 809-816, 2009.

SNPs Genes

rs1800545

rs7412

rs1128503

rs6578993

rs1045642

rs2278718

rs4784642

rs521674

ADRA2A

APOE

ABCB1

TH

ABCB1

MDH1

GNAO1

ADRA2A

SNPs Genes

rs1800545

rs7412

rs6578993

rs2278718

rs1128503

rs429358

rs3813065

rs4121817

rs521674

ADRA2A

APOE

TH

MDH1

ABCB1

APOE

PIK3C3

PIK3C3

ADRA2A

45

Pathway Analysis


Target and Novelty Processing," NeuroImage, vol. 46, pp. 809-816, 2009. 46

Dataset

Number of

Training

Subjects

Number of

Testing

Subjects

1 4 36

2 12 28

3 20 20

4 28 12

5 36 4

Subjects 20 Patients

20 Healthy Controls

fMRI data Auditory Oddball Task

Gene data 367 SNPs

Classification with SNP & fMRI

SVM Classification

fMRI + Gene

fMRI

Gene

H. Yang, J. Liu, J. Sui, G. Pearlson, and V. D. Calhoun, "A Hybrid Machine Learning Method for Fusing fMRI and

Genetic Data to Classify Schizophrenia," Frontiers in Human Neuroscience, vol. 4, pp. 1-9, 2010.

Overview

• Introduction











47

Constrained ICA

48

CDK14 (7q21)

ZNF879

GRM6 (5q35)

LOC 100506971

UGDH

SLC5

ZNF354C

ADAMTS2

J. Liu, M. Ghassemi, A. Michael, D. Boutte, W. Wells, N. I. Perrone-Bizzozero, F. Macciardi, D. Mathalon, J. Ford, S. Potkin,

J. Turner, FBIRN, and V. D. Calhoun, "An ICA with reference approach in identification of genetic variation and associated

brain networks," Frontiers in Human Neuroscience, vol. 6, pp. 1-10, 2012.

9

Parallel ICA w/ Reference (pICA-R)

49

ji, 2j1i

2

2j1i

ji,

2

2j1i

2

2k222y2k

2

2

1y1

10/201/21/2U1/2

1/21/21/2

AW

1/21/21/2

AvarAvar

A,AcovmaxA,AcorrmaxF3

)X(Wr Ylnf EmaxSr,distY H maxF2

Ylnf EmaxY H maxF1

WXWU,e1

1Y

XWSSAX

2

1

1/2

11/21/2

λ

J. Chen, V. D. Calhoun, G. D. Pearlson, N. Perrone-Bizzozero, J. Sui, J. A. Turner, J. Bustillo, S. Ehrlich, S. Sponheim,

J. Canive, B. C. Ho, and J. Liu, "Guided Exploration of Genomic Risk for Gray Matter Abnormalities in Schizophrenia

Using Parallel Independent Component Analysis with Reference," NeuroImage, vol. 83, pp. 384-396, 2013.

pICA-R: Schizophrenia & ANK pathway

50

J. Chen, V. D. Calhoun, G. D. Pearlson, N. Perrone-Bizzozero, J. Sui, J. A. Turner, J. Bustillo, S. Ehrlich, S. Sponheim,

J. Canive, B. C. Ho, and J. Liu, "Guided Exploration of Genomic Risk for Gray Matter Abnormalities in Schizophrenia

Using Parallel Independent Component Analysis with Reference," NeuroImage, in press

Overview

• Introduction











51

3-way parallel ICA

52

Results (N=112, fMRI, sMRI, 65K SNPs in coding genes)

Genetics

Structural

Variation

Functional

Activity

Corr = 0.44

Pval = 1.5e-5

Corr = 0.34

Pval = 2.0e-3

Corr = 0.39

Pval = 9.0e-4

(a) sMRI Component (b) fMRI Component

(c) Genetic Component (d) Pairwise ConnectionsV. Vergara, A. Ulloa, V. D. Calhoun, D. Boutte, J. Chen, and J. Liu, "A Three-way Parallel ICA Approach to

Analyze Links among Genetics, Brain Structure and Brain Function," NeuroImage, in press,

Overview

• Introduction











54

10

55 56

Overview

• Introduction











57

Challenges & Issues

• Overfitting (cross-validation, prior information)

• Replication

• Interpretation (e.g. GSEA does not model exact

interaction among SNPs, latent component may

not have a biological purpose).

• Can be partially addressed by incorporating

known biological, cellular, or behavioral specific

information.

• Lots of information not fully incorporated

(proteomic, gene expression, epigenetic,

behavioral and environmental variables)

• CNV (low incidence)

58

• http://mialab.mrn.org/software

• freeware, written in MATLAB

• Group ICA of fMRI Toolbox (GIFT)• Single subject/Group ICA

• Many algorithms (e.g. ICA, cICA, IVA)

• MANCOVA testing framework

• Source Based Morphometry

• Model order estimation

• ICASSO (clustering/stability)

• Dynamic connectivity (dFNC)

• Fusion ICA Toolbox (FIT)• Parallel ICA, jICA

• mCCA+jICA & much more!

• Simulation Toolbox (SimTB)• Flexible generation of fMRI-like data

• Task & rest fMRI

• Space * Time * Amplitude

Left Hemisphere

Visual Stimuli Onset

Left Hemisphere

Visual Stimuli Onset

Software

60

Documents

Identification of Multimodal MRI and EEG Biomarkers using ... · 2 A Taxonomy of Approaches 7 J. Liu and V. Calhoun, "A review of multivariate analyses in imaging genetics," Frontiers