Activity flows over task-evoked networks shape cognitive

Activity flows over task-evoked networks shape
cognitive task activations across task switches
Cole
Neurocognition
Lab
Michael W. Cole, Takuya Ito, Douglas H. Schultz, Ravi D. Mill
Center for Molecular & Behavioral Neuroscience, Rutgers University-Newark, Newark, NJ
Validation with computational model:
Does task-evoked FC emerge?
i=n-1
j’s predicted activity = ∑ (i’s activity × Connectivity i-with-j)
Visual
50
Motor/tactile
Regions
100
150
Cingulo-opercular
Premotor (FP)
200
Default-mode
250
Frontoparietal
Auditory (primary)
300
Dorsal attention
Hippocampal
350
50
100
150
200
Regions
250
300
350
Resting-state fMRI, Pearson r
-0.2
0
0.2
0.4
0.6
Task vs. rest dynamics
Output (population firing rate)
i≠j
Task state
Mean
(+ 4 SDs)
Linear range
Input (sum of population inputs)
BOTH
VERTICAL
LEFT INDEX
Instructions
3925 ms
+
Delay
1570 - 6280 ms


Rule set 1 description:
If BOTH stimuli are VERTICAL,
press your LEFT INDEX finger
Answer: TRUE
(Left index finger)
Trial 1
2355 ms
Time
Task 64
Logic rule 4
Sensory rule 1
Motor rule 1
NEITHER
RED
LEFT INDEX
Instructions
3925 ms
+

Delay
1570 - 6280 ms Trial 1
2355 ms

Rule set 64 description:
If NEITHER stimulus is RED,
press your LEFT INDEX finger
[other finger, same hand if false]
Answer: FALSE
(Left middle finger)
Sensory Rules
1. Red
2. Vertical
3. Hi Pitch
4. Constant
Motor Rules
1. Left Index
2. Left Middle
3. Right Index
4. Right Middle
• New cognitive paradigm with 12 task rules (4 rules with 3 rule domains) per
subject, recombined into 64 task contexts
• Multiband fMRI with N=100 subjects, 15 minutes resting-state fMRI, 60 minutes
of task fMRI, TR=0.785 s, 2 mm cubic voxels
fMRI task-evoked FC,
No task regression
Pre-fMRI task-evoked FC
fMRI task-evoked FC,
Perfect (correct HRF) task regression
HRF shape (correct)
5
150
200
250
1.21%
300
350
Basis set
model
-30
-20
0
-10
100
0.15
0.1
0.05
Regions
10
0
0.05
150
200
250
-0.05
300
-0.1
50 100 150 200
Left motor cortex (right hand)
Right motor cortex (left hand)
350
250 300 350
Regions
0.04
Pearson r difference
100
0.03
0.02
0.01
0
-0.01
-0.02
-0.03
50
100 150 200 250 300 350
Regions
10.1% of all connections significant
(p<0.05, FDR corrected)
0.8% of all connections significant
(p<0.05, FDR corrected)
Prediction of task activation patterns
enhanced by task-evoked FC
Actual task activations
R2=0.40
50
50
50
100
100
100
150
150
150
200
200
200
250
250
250
300
300
300
350
Region
(left hand > right hand)
(Arb. units)
50
50
No
Canonical
FIR
regression HRF
model
10
fMRI activations
15
0
20
30
19.89%
Left vs. right hand task FC
Basis set task regression
• Task-evoked FC increases are substantially attenuated with basis-set-based
task regression, suggesting other approaches leave false positives
• Left vs. right hand button press FC differences match expected hemispheres,
but with likely inflated responses without task regression
Model: Task-evoked FC is inflated by fMRI
350
BOTH
NOTBOTH EITHER NEITHER HIPITCH CONSTANT
Logic rules
RED
VERTICAL L.INDEX R.INDEX
Semantic rules
L.MID.
R.MID.
Task-evoked FC predictions
350
BOTH
NOTBOTH EITHER NEITHER HIPITCH CONSTANT
Motor rules
Logic rules
Motor prediction
RED
VERTICAL L.INDEX R.INDEX
Semantic rules
L.MID.
R.MID.
BOTH
NOTBOTH EITHER NEITHER HIPITCH CONSTANT
Motor rules
Logic rules
Motor prediction
RED
VERTICAL L.INDEX R.INDEX
Semantic rules
L.MID.
R.MID.
Motor rules
fMRI activity (z-scored)
r=-0.59
r=0.56
-2.5
Region
fMRI task-evoked FC,
fMRI task-evoked (”background”) FC ,
fMRI task-evoked FC,
Wrong HRF task regression
FIR-modeled task regression HRF-basis-set task regression
FIR modeling
HRF shape (wrong)
Regressor
...
Logic rule 1
Sensory rule 2
Motor rule 1
20.05%
Resting-state FC predictions R2=0.26
-2
-1.5
-1
-0.5
0
0.5
1
1.5
2
2.5
• Task activations better predicted using task-evoked FC (p<0.00001), suggesting
task-evoked FC meaningfully affects the flow of activity through networks
• Task-evoked FC was especially important for the motor/tactile network activations for the motor rules (negative vs. positive correlation)
HRF shape (basis set)
2
4
6
8
10
...
Task 1
Logic Rules
1. Both
2. Not Both
3. Either
4. Neither
2% of time points
Mean
(+ 4 SDs)
Task stimulation
(5 regions)
C-PRO Cognitive Paradigm
8% of time points
Resting state
0.8
Concrete-Permuted Rule Operations
(C-PRO) cognitive paradigm
• Simulated resting-state fMRI FC reflects structural + synaptic connectivity patterns (see
Cole et al. 2016; Nat Neuro)
• Task-evoked FC was enhanced by local stimulation, respecting the network organization
• Task-evoked FC (and activity flow) only enhanced when baseline activity well below the
linear range (bias=-20), with task increasing
activity into linear range (4x more often)
20.06%
20
Region
Regions
...
i=2
Task simulation
25
Pearson r difference
i=1
fMRI simulation
Left vs. right hand task FC
No task regression
Regions
Region
j
Task stimulation
(5 regions)
Correlation change from rest
w
o
fl
y
t
i
v
i
t
c
A
Task-evoked FC
Pearson correlation
Prediction of
held-out activity in j
Resting-state FC
Population synaptic weight
• Resting-state networks are present during tasks (Cole et al.,
2014) and shape task activations
(Cole et al., 2016)
• Do task functional connectivity
(FC) changes to resting-state networks influence task activations?
• Novel “activity flow mapping”
method with fMRI
• Predicts activity patterns; better
prediction = larger influence of FC
architecture
Synaptic connectivity
Correcting task FC inflation in empirical fMRI data
% increased from rest FC
How do functional networks shape
task-evoked activations?
2
4
6
Time
8
281 regressors
(task block duration)
10 ...
5 regressors
(99.5% of variance
among 1000 HRF shapes)
• Task-evoked FC (task - resting-state FC) is inflated by fMRI, driven by convolution with a hemodynamic response function (HRF) (top-middle)
• Task-evoked FC inflation can be corrected if HRF shape is known (top-right)
• Using an incorrect (bottom-left) HRF only partially corrects inflation, while an
over-parameterized (bottom-middle) HRF model can make it worse
• Using an HRF basis set approach (Woolrich et al., 2004) consistently corrects
the inflation across a variety of unknown HRF shapes (bottom-right)
Summary & Conclusions
• Does fMRI inflate task-evoked FC and can it be corrected?
Yes – Basis-set-based regression correctly fits task-evoked activation HRF shape, removing mean task responses and allowing
proper estimates of task-evoked FC with fMRI.
• Does task-evoked FC contribute to shaping task activations?
Yes – Task activation patterns could be predicted much better
(14% more variance) when mapping activity flow with task FC.
Address correspondence to [email protected]
A copy of this poster can be found at www.colelab.org