Joseph Callicott MD fMRIMRI Summer Course 62014 Introduction The Age of B ig Data Lohr GOOD with numbers Fascinated by data The sound you hear is opportunity knocking NY Times 2222012 ID: 931431
Download Presentation The PPT/PDF document "Multivariate analyses in clinical popula..." is the property of its rightful owner. Permission is granted to download and print the materials on this web site for personal, non-commercial use only, and to display it on your personal computer provided you do not modify the materials and that you retain all copyright notices contained in the materials. By downloading content from our website, you accept the terms of this agreement.
Slide1
Multivariate analyses in clinical populations: General factors & neuroimaging
Joseph Callicott, MD
fMRI/MRI Summer Course 6/20/14
Slide2IntroductionThe ‘Age of Big Data’
Lohr
, “GOOD with numbers? Fascinated by data? The sound you hear is opportunity
knocking…” (NY Times, 2/22/2012)We routinely collect ‘multimodal data’ E.g., mood rating scale and structural MRI Compile or compare, but typically without multimodal analysisProjects classified as ‘geno-,’ ‘proteo-,’ or ‘pheno-’ already connotes ‘big data:’Each fMRI image presents ~20K analysesGWAS model = strict correction for multiple comparisonsCurrent model = parallel correlation/association per datasetProposed model = multivariate approach w/data reductionSimplified analysesSmaller statistical ‘cost’Some current theoretical approaches become testable’ (RDoC)
Slide3OutlineA Tale of Two Lectures:
Imaging genetics & schizophrenia
Relevant issues for
clinical populationsWhy imaging genetics?Imaging genetics 101Multivariate analyses of fMRI: within experimental datasetGeneral vs specific factors in datag“i” : factor analytic solution of general factors in fMRI task data Multivariate analyses of fMRI redux: across experimental dataset
Slide4Issues of special interest to clinical studies…BOLD fMRI in clinical populations:
BOLD fMRI is not, strictly speaking, a clinically informative measure
No pathognomic findings, to date
Performance likely to differ fundamentally in tasks that HVs will perform near ceilingI.e., ~100% accuracy and faster RT than patientsBOLD fMRI in healthy subjects‘predictive relationships’ between behavior and BOLD implicit in design, perhaps not strongly correlated with group map activationGenetic associations to BOLD (the bulk of the imaging genetics literature) do not necessarily connote a ‘real’ effect of a given polymorphismIs the phenotype heritable? In past, twin or sibling studiesCurrently, within ‘only’ HV = GCTA (Visscher)
(Callicott et. al. 2000;
Manoach
et al., 2000; Callicott et al. 2003; others)
(Van Snellenberg et al. 2006)
Slide5So, then, why
imaging genetics?
Crass commercial message:
Simple plan, high impactHave an fMRI task in a relatively large sample?Healthy controls preferableN > 40Draw blood or swab cheek Genotyping at most resolutions fast & cheap*In SPM: ANCOVA or regression suffice and seem reasonably poweredCOMT led from primate to human imaging, and then to drugs targeting cognitive impairments
Slide6Seriously, though, why imaging genetics
?
Few routes to neural mechanism using
in vivo human dataAnimal model Drug studyBOLD fMRI (MRSI, MEG, EEG)Genes do not code for mental illness, per seGenes code for heritable aspects of brain function, intermediate- or endo-phenotypesGenetic risk for illnesses like schizophrenia is polygenic, heterogeneousGene interact with each other and the environment BOLD fMRI, as an alternate metric of specific or general cognitive systems, offers ‘real world validation’ Putative genetic mutations (including private mutations (CNVs)In spite of growing sample sizes, association studies risk false positivesRDoC domains and constructsIf these do not correspond to brain systems we can map, then may be as doomed as DSM
Slide7Take home message…Larger samples needed (GWAS), typically via collaboration across centers (ENIGMA)
BIG DATA is here
Multivariate or non-hypothesis-driven analyses offer the potential for novel, highly informative findings
CNV & cognitionVery good software often freely available:PLINK AFNI GingerALE-SLEUTH-MANGO R (many)
(Stefansson et al., Nature, 2014)
Slide8(*
Visscher
et al., 2010; Nan et al. 2012**;
Postuma et al. 2002***; McGue & Bouchard, 1998#;^Burmeister et al., 2008)
Slide9Interest in imaging genetics predicated on heritability of phenotypes…
Callicott
et al
. Cereb Cortex 2000Patients > Controls (N=13) (N=18)Callicott et al. Am J Psychiatry 2003Healthy Siblings > Controls (N=48) (N=33)
PFC BOLD during our Nback h2 = 0.4-0.5
Blokland
e
t
al.
Biol
Psych
2009,
J
Neurosci
2011;
Koten
et al.
Science
2009
Slide10Finding genes for highly heritable, but complex diseases
affected person
unaffected
“nonpenetrant”(Goldman et al. Nat Rev Genet 2006)Remains difficult, even when n=100KCaused by many (100-1000s) of genesThe effects of a mutation vary between peopleHas all the genes (note this doesn’t mean the exact same set)
May still carry some genes (like a parent of a sick person)
Has all of the genes but is NOT sick for reasons we can’t explain
Slide11Catechol O-methyltransferase (COMT): NIMH Intramural Success Story
(Apud et al. 2006)
(Egan et al. 2001)
(Meyer-Lindenberg et al. 2006)
Slide12Functional impact COMT Val105/158Met
val
/met
rs4680
5’
Now validated at multiple levels:
Animal models:
Reduced enzymatic activity
Altered synaptic dopamine levels
Human data:
Reduced enzymatic activity in vitro
lymphoblastoid
cell lines
Altered transcription/reduced activity post mortem
Altered D1 but not D2 receptor density in PFC
PFC efficiency in BOLD fMRI
Combined:
S
ex effects mostly in males ((
Papaleo
et al., 2014)
striatum
mammalian
PFC
Slide13genotype effect
F=5.41, df= 2, 449;
p<.005.
Executive cognitionEffect of rs4680 on frontal lobe function (Egan et al
PNAS
2001)
n = 218
n = 181
n = 58
vv>vm>mm
, SPM 99, p<.005
Physiological efficiency
Circa 2014: How have these findings held up?
Replicated but n’s ~20
Slide14BOLD phenotypes in simple association: COMT and PFC14
(
Mier
, Kirsch, Meyer-Lindenberg; 2009)
Slide15BOLD phenotypes in simple association: 5-HTTP and Amygdala15
(
Munafo
, Brown, and Hariri; 2007))
Slide16BOLD phenotypes in simple association: Power?16
(Barnett et al, 2008)
Slide171st generation imaging genetics: simple association
Candidate genes
KIBRA impaired memory & expressed in hippocampus (
Papassotiropoulos et al., Science 2006)Replication in 3 independent populations in behavioral memory measuresIn 30 healthy subjects, KIBRA associated with reduced hippo activation
Slide18Genetic mutations modeled in cell culture or animalsAssociation based on disease GWAS (ZNF804A)Esslinger et al. 2009 (Science)
Rasetti et al. (Arch Gen Psychiatry)
2
nd generation imaging genetics: GWAS era
Slide19PFC neuronal function: ‘optimized’ by dopamine & GABA interactions
(Goldman-Rakic & Selemon 1997)
(Seamans et al., 2001)
2nd generation imaging genetics: Epistasis and pathways
Slide20(Straub et al., 2007)
2
nd
generation imaging genetics: Epistasis
Slide21COMT
:
V/V
V/MM/MBray Hap:-/-
-/-
-/-
+/-
+/-
+/-
V/V
V/M
M/M
V/V
V/M
M/M
+/+
+/+
+/+
BOLD
fMRI
Left DLPFC (
a.u
.)
COMT x Dysbindin interaction
Epistasis (gene-gene interaction)
Initially based on candidate-by-candidate
Buckholtz et al.,
Mol
Psychiatry 2007
Data-driven (machine learning)
Nicodemus et al., Hum Genet 2010
Now predicated on detailed cellular or animal modeling
COMT x DTNBP1 (
Papaleo
et al., 2013)
DISC1 x NKCC1 (Kim et al., Cell 2012 & Callicott et al. J
Clin
Invest 2013)
2
nd
generation
imaging genetics: Translational neuroscience
Slide2222
More of the same (‘sophisticated
univariate
’)? Network/connectivity? Hypothesis-free?Hypothesis-free pattern detection (random forest)ICA/PCA/CPCA networksNext generation imaging genetics?Novel phenotypes (processing speed)
Slide23Sophisticated univariate: Imaging GWAS
BOLD fMRI GWAS
Nback (n = 364)
Illumina 650K chip genotypingAutomated extraction of AAL ROIsFirst GWAS + using BOLD fMRI (Callicott, Spencer, et al., in prep)
Slide24As a heritable trait, BOLD fMRI phenotypes show other sensitivities…..
Long history within animal literature showing significant effects of environment on brain structure &
function
Beneficial effects of ‘enrichment’ (toys, limit isolation) (Hebb, Am J Psychiatry, 1955) fMRI during social stress task influenced by environmentUrban upbringing or urbanicity linked to increased risk for mental illness (Van Os et al., Nature, 2010)(Lederbogen et al., Nature, 2012)Sophisticated univariate
:
N
ovel questions
Slide25As a heritable trait, BOLD fMRI phenotypes show other sensitivities…..fMRI during WM ( 3 cohorts (USA1 = 124; USA2 =92; Italy1=226 )
Sensitivity to childhood environment (Urbanicity) (
Ihne
et al., in submission)Sophisticated univariate: Imaging G x E
Slide26fMRI during WM ( 3 cohorts (USA1 = 124; USA2 =112; Italy1=226 )Gene-by-environment interaction (COMT x Urbanicity) (Ihne et al., in submission)
Ihne et al., in preparation
Sophisticated
univariate: Imaging G x E
Slide27Constrained principle component analysis (CPCA) (David AA Baranger
– Wash U)
27http://www.nitrc.org/projects/fmricpcaTodd Woodward and colleagues, University of British Columbia:CPCA provides a “unified framework [for]… regression analysis and principal component analysis .” To identify functional systems using from singular-value decomposition of BOLD time series, These systems are imaged by constraining analyzed BOLD signal from a particular interval of time against all other scans (i.e., all others are baseline) Multivariate network analysis: ICA/PCA/CPCA banish ‘blob-ology’
Slide28CPCA28
Z or ‘activation’ matrix = individual time series for all subjects (rows) for all voxels in the brain (columns)
Our standard SPM5 via XNAT first level processing of 0B alternating with 2B
G or ‘design’ matrix = a model to predict BOLD signal changes (columns) over all fMRI scans (rows)SPM5 often uses a canonical hemodynamic response function (HRF) to deconvolve signal, fMRI-CPCA uses finite impulse response function (FIR)http://www.nitrc.org/projects/fmricpca
Slide29CPCA29
N-back model not complicated:
Simply provide onset and offset of 0B and 2B task epochs
Components = extracted components represent networks Component loadings= loosely, correlation coefficients between component scores and BOLD signal that was predicted from imposed constraints (design)http://www.nitrc.org/projects/fmricpca
Slide3030
Identify and then display components
using MRICon for anatomical localization
(http://www.nitrc.org/projects/mricron)In this case, not really using estimated hemodynamicsRather, we wish to compare effect of diagnosis or genotype using component scores and predictor weightsPredictor weight = contribution of G matrix to changes in components over the fMRI time series (~ correlation of component score and g) CPCA
Slide31CPCA: Confusing Problematic Conflicting Agonizing
Unspecified error required recalculation of component weights
Same networks found with addition of a fourth DMN
Differentiates NC and SIBs from SCZNo longer appears to be identifying intermediate phenotype
Slide32CPCA: Nback systems
Anti-task Network
resembles
cingulate from DMN+ hippocampus
WM Network
DLPFC
+parietal
Motor system
Anti-task #2
resembles parietal
From DMN
+ cerebellum
Slide33CPCA: Factors sensitive to disease, not genetics
p
<0.05
p<0.05Unspecified error required recalculation of component weightsSame networks found with addition of a fourth DMNDifferentiates NC and SIBs from SCZ
No longer appeared to identify intermediate phenotype
Slide34CPCA: Not particularly sensitive in general420 HV CPCA (2back) = 4 factors
Neuro-
= 6 cognitive factors
2B as measured in labg estimates
Slide35Big data benefits reproducibility…
Slide36ENIGMA: first
GWAS+
sMRI
(Stein et al. 2009, 2010; Thompson et al., 2014)Big data benefits reproducibility…Heritability for novel phenesReplication on large scale
Slide37OutlineA Tale of Two Lectures:
Imaging genetics & schizophrenia
Why imaging genetics?
Imaging genetics 101Multivariate analyses of fMRI: within experimental datasetGeneral vs specific factors in datag“i” : factor analytic solution of general factors in fMRI task data Multivariate analyses of fMRI redux: across experimental dataset
Slide38Are phenotypes independent?
Slide39Are phenotypes independent?
Slide40Are phenotypes independent?
Slide41The general cognitive factor (Spearman’s g)
(Dickinson et al., 2008)
(Jensen, 1998)
Slide42Where is g?
‘Lesion maps’ from 241 patients w/ focal brain damage and
g
(Gläscher et al. PNAS 2010)
.
Barbey et al. (Brain 2012)
found similar results in 182 focal brain lesion patients
Various conceptual, functional and structural support for PFC and PAR (at minimum)
Slide43Is
g
associated with fMRI activation?
Nback (n= 161) higher g greater efficiencyReplication (n= 582) higher g greater efficiencyExact overlap
Notes:
Analysis: SPM5 multiple regression controlling for age, sex
2B accuracy
g
(r = 0.3, p < 0.001)
Replication…cont’dReplication 3 (n= 211)
Areas within replication exactly overlapping discovery…
discovery
Replication 5 (n= 306)
Replication 4 (n= 393)
Replication 4
Replication 3
Replication 5
Slide45DSVT
v g
But…Faces v gMTL v g
Slide46g
correlates with similar areas
across 4 tasks in same 161 HVs
Nback (n= 161)Notes:Analysis: SPM5 multiple regression controlling for age, sex161 with QC+ NB, MTL, Faces, DSVTNB as discovery ROI, others queried at p < 0.05
Slide47Is there a general solution for fMRI?161 HVs with QC+ Nback, MTL (incidental encoding), Faces (response to aversive faces), and DSVT (processing speed)Individual 1st level maps created for each task
Sue Tong: automated script to extract parameter estimates in Automated Anatomical Labeling (AAL) ROIs
Mean fMRI ‘signal” transformed to Z score
Factor analysis:Principle component extraction Orthogonal and oblique rotationsFactor scores estimated i (fMRI g) = sum of factor scores Comparison across task and against cognitive measures (big g)
Slide48fMRI (Nback) i
(161 HVs, max likelihood extraction w/ varimax rotation, 60.1% total variance explained, goodness-of-fit p < 1e-5)
.50
.41.40.42.26.42.30.13.11
.13
.16
.10
Slide49fMRI (Nback) i
(161 HVs, max likelihood extraction w/ varimax rotation, 60.1% total variance explained, goodness-of-fit p < 1e-5)
.79
.90.59.54.7
.36
.43
.56
.52
.27
Slide50fMRI (Nback) i
Slide51fMRI (Faces) i
(161 HVs, max likelihood extraction w/ varimax rotation, 58.1 % total variance explained, goodness-of-fit p < 1e-5)
.44
.40.45.40.39.41
.10
.10
.10
.10
.10
Slide52fMRI (Faces) i
Slide53fMRI (DSVT) i
(161 HVs, max likelihood extraction w/ varimax rotation, 51.6 % total variance explained, goodness-of-fit p < 1e-5)
.40
.41.42.41.40.41.10
.10
.10
.10
.10
Slide54fMRI (DSVT) i
Slide55fMRI (MTL) i
(161 HVs, max likelihood extraction w/ varimax rotation, 51.6 % total variance explained, goodness-of-fit p < 1e-5)
.40
.41.42.41.40.41.10
.10
.10
.10
.10
Slide56fMRI (MTL) i
Slide57An i by any other name…
gFaces i
Nback i
DSVT i
MTL i
g
Pearson Correlation
1
-.060
.042
-.025
-.210
**
Sig. (1-tailed)
.226
.300
.378
.004
Faces i
Pearson Correlation
-.060
1
-.182
*
-.027
.014
Sig. (1-tailed)
.226
.010
.365
.428
Nback i
Pearson Correlation
.042
-.182
*
1
.056
-.049
Sig. (1-tailed)
.300
.010
.241
.267
DSVT i
Pearson Correlation
-.025
-.027
.056
1
.124
Sig. (1-tailed)
.378
.365
.241
.059
MTL i
Pearson Correlation
-.210
**
.014
-.049
.124
1
Sig. (1-tailed)
.004
.428
.267
.059
**. Correlation is significant at the 0.01 level (1-tailed).
*. Correlation is significant at the 0.05 level (1-tailed).
Structural MRI
(351 HVs, max likelihood extraction w/ varimax rotation, 48 % total variance explained, goodness-of-fit p < 1e-5)
Slide59Summary:
A Tale of Two Lectures:
Imaging genetics & schizophrenia
Imaging genetics easier in an age of data sharing and public databases
BOLD fMRI (IMHO) has never been about diagnosis = hello
RDoC
!
Multivariate analyses of fMRI: novel findings, novel questions
General vs specific factors in data
g inspires a straight-forward, replicable multivariate analysis of fMRI (Factor analytic approach) (Dickinson et al.,
Biol
Psych 2008; JAMA Psych 2014, numerous)
“
i
” : factor analytic solution of general factors in fMRI task data
Multivariate
analyses of fMRI
redux
:
Data reduction writ large
Replication across tasks, labs, designs?
Slide60Further musing…Multimodal data, multimodal analysis
fMRI phenotypes are not independent
Aspects within each task representing individual ‘positive manifold’
Is heritability about this general shared variance or specific task aspects?FMRI i as a data reduction methodNot gComplicated, but factor solution may be informed by other dataStructural MRI factor solution, no relationship to g or other cognitive factors g holds special relationship to fMRI dataTest whether reduced factor structure more related to genes, other MRI, clinical measures
Slide61Thanks:
Dwight Dickinson
Sue Tong
Jessica IhneKaren BermanBarbara SpencerGraham BaumMorgan BartholomewAmanda ZheutlinCTNB clinical staff