Time-resolved effective connectivity in task fMRI: psychophysiological interactions of co-activation patterns

Article

Reference

Time-resolved effective connectivity in task fMRI: psychophysiological interactions of co-activation patterns

FREITAS, Lorena, et al.

Abstract

Investigating context-dependent modulations of Functional Connectivity (FC) with functional magnetic resonance imaging is crucial to reveal the neurological underpinnings of cognitive processing. Most current analysis methods hypothesise sustained FC within the duration of a task, but this assumption has been shown too limiting by recent imaging studies. While several methods have been proposed to study functional dynamics during rest, task-based studies are yet to fully disentangle network modulations. Here, we propose a seed-based method to probe task-dependent modulations of brain activity by revealing Psychophysiological Interactions of Co-activation Patterns (PPI-CAPs). This point process-based approach temporally decomposes task-modulated connectivity into dynamic building blocks which cannot be captured by current methods, such as PPI or Dynamic Causal Modelling. Additionally, it identifies the occurrence of co-activation patterns at single frame resolution as opposed to window-based methods. In a naturalistic setting where participants watched a TV program, we retrieved several patterns of co-activation with a [...]

FREITAS, Lorena, et al . Time-resolved effective connectivity in task fMRI: psychophysiological interactions of co-activation patterns. NeuroImage , 2020, vol. 212, p. 116635

DOI : 10.1016/j.neuroimage.2020.116635

PMID : 32105884

Supplementary Material

* * * * * * * * * * * *

* * * * * * * * * * * *

* * * * * * * * * * * *

* * * * * * * * * * * *

* * * * * * * * * * * *

* * * * * * * * * * * *

Solution:

1

* * * * * * * * * * * *

* * * * * * * * * * * * B)

*

Case 1: When a target voxel is only correlated with the seed, there should be no PPI effect. By adding up the values on the third plot into a Simap, the result will be close to zero, indicating this fact.

Case 2: When a target voxel is only correlated with the task, there should be no PPI effect. By adding up the values on the third plot into a Simap, the result will be close to zero also in this case, indicating this fact.

Case 3: When the target voxel’s correlation with the seed depends on the task, it means that there is a PPI effect. By adding up the values on the third plot into a Simap, the result will be very large, indicating this effect.

*

A)

Figure A.1:An intuitive toy example to illustrate the relationship between the static interaction map (simap) and PPI analysis results. A) Moments where the seed activity is significantly high (red stars) or low (blue stars) are selected. B) Three possible cases are depicted: 1) When the target voxel is correlated only with the seed (meaning that there is no interaction e↵ect), its resulting value in the siMap will be close to zero, as would be expected for its corresponding beta value in a PPI analysis. 2) Similarly, when the target voxel is correlated only with the task, its resulting value in the siMap will be close to zero. 3) when the target voxel is correlated with the PPI regressor, meaning that its relationship with the seed changes according to the context, the averaged value from its suprathreshold frames will yield a high value, indicating a PPI e↵ect.

Group Averaged Regressors

PPI Seed Task

PPI

Seed

Task

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 abs(cos)1

PPI Seed Task PPI Seed Task

PPI Seed Task PPI Seed Task

PPI Seed Task PPI Seed Task

PPI Seed Task PPI Seed Task

PPI Seed Task PPI Seed Task

PPI Seed Task PPI Seed Task

PPI Seed Task PPI Seed Task

PPI Seed Task PPI Seed Task

PPI Seed Task PPI Seed Task

PPI Seed Task PPI Seed Task

PPI Seed Task PPI Seed Task

PPI Seed Task PPI Seed Task

PPI Seed Task PPI Seed Task

PPI Seed Task PPI Seed Task

PPI Seed Task PPI Seed Task Group Averaged Regressors

PPI Seed Task

PPI

Seed

Task

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 abs(cos)1 Orthogonality of regressors per subject

A) B)

C) D)

Figure A.2: Sanity check on the chosen seed signal. Before proceeding with the PPI-CAPs analysis, we checked the orthogonality of the seed, task and ppi regressors using SPM’s Check Orthogonality tool, as well as the skewness of the seed. A) The group-averaged matrix confirms that the regressors are not collinear at the group level. The highest o↵-diagonal value is 0.04. B) Shows the same matrix at the single-subject level. C) The seed signal was not skewed for any of the subjects: the skewness value stayed within the [-0.5 0.5] range for all of them.

D) Shows the distribution of z-scored seed signal values for each subject. Additionally, the magnitude of the seed timecourse was not correlated with the sign of the task: the Pearson’s r coefficient for the subject with the strongest correlation between the magnitude of the seed and the sign of the task was 0.11.

Group SiMap:

k=3

500 1000 1500 500

1000 1500

k=4

500 1000 1500 500

1000 1500

k=5

500 1000 1500 500

1000 1500

k=6

500 1000 1500 500

1000 1500

k=7

500 1000 1500 500

1000 1500

k=8

500 1000 1500 500

1000 1500 k=3

500 1000 1500 500

1000 1500

k=4

500 1000 1500 500

1000 1500

k=5

500 1000 1500 500

1000 1500

k=6

500 1000 1500 500

1000 1500

k=7

500 1000 1500 500

1000 1500

k=8

500 1000 1500 500

1000 1500

Consensus Matrices

k=3

500 1500 500

1000 1500

k=4

500 1500 500

1000 1500

k=5

500 1500 500

1000 1500

k=6

500 1500 500

1000 1500

k=7

500 1500 500

1000 1500

k=8

500 500 1000

1500

0.5 1

1

0.5

0

Figure A.4: An analysis of the number of occurrences of each PPI-CAP per subject combined with consensus clustering indicates the optimal number of groups in which to cluster the suprathreshold frames. (Top) The PPI-CAPs distributions fork2[3,8] showed thatkvalues of 6 and above yielded PPI-CAPs that never occurred for some subjects, while values between 4 and 5 had the most homogeneous distribution of pattern occurrence per subject. (Bottom) Visual inspection of the consensus matrices showed that the most stable values fork(that is, those for which any two frames would most consistently be clustered together or separately) werek= 3 andk= 4. Since 4 represents a better trade-o↵in terms of variety and robustness, we chose this value for further analysis (see its clear pattern highlighted in green). The colormap represents the proportion of time when two frames were consistently clustered either together or separately.

-0.2 0 0.2 0

0.2

0.4

Null Seed Distribution p = 0.007992

-0.2 0 0.2

0 0.2

0.4 Null Task Distribution p = 0.01998

-0.4 -0.2 0 0.2

0 0.2

0.4 Null PPI Distribution p = 0.15884

-0.5 0 0.5

0 0.2

0.4 Null Seed Distribution p = 0.000999

-0.2 0 0.2

0 0.2

0.4 Null Task Distribution p = 0.11988

-0.2 0 0.2 0.4

0 0.2

0.4 Null PPI Distribution p = 0.15584

-0.5 0 0.5 1

0 0.2

0.4 Null Seed Distribution p = 0.000999

-0.2 0 0.2

0 0.2

0.4 Null Task Distribution p = 0.008991

-0.2 0 0.2

0 0.2

0.4 Null PPI Distribution p = 0.11888

-0.5 0 0.5 1

0 0.2

0.4 Null Seed Distribution p = 0.000999

-0.2 0 0.2

0 0.2

0.4 Null Task Distribution p = 0.31668

-0.2 0 0.2

0 0.2

0.4 Null PPI Distribution

p = 0.000999

Occurrence in Time

% of subjects% of subjects% of subjects% of subjects

1 2 3 4

1 2 3 4

1 2 3 4

1 2 3 4

PPI-CAP 1PPI-CAP 2PPI-CAP 3PPI-CAP 4

Corresponding movie frame

00’00’

’ 00’13’

’ 00’17’

’ 00’25’

’ 01’24’

’ 02’09’

’ 02’57’

’ 03’48’

’ 04’43’

’ 05’06’

’

00’00’

’ 00’13’

’ 00’17’

’ 00’25’

’ 01’24’

’ 02’09’

’ 02’57’

’ 03’48’

’ 04’43’

’ 05’06’

’

00’00’

’ 00’13’

’ 00’17’

’ 00’25’

’ 01’24’

’ 02’09’

’ 02’57’

’ 03’48’

’ 04’43’

’ 05’06’

’

00’00’

’ 00’13’

’ 00’17’

’ 00’25’

’ 01’24’

’ 02’09’

’ 02’57’

’ 03’48’

’ 04’43’

’ 05’06’

’ 00’00’’ 00’40’’ 01’20’’ 02’00’’ 02’40’’ 03’20’’ 04’00’’ 04’40’’ 05’20’’

Time (m’s’’)

00’00’’ 00’40’’ 01’20’’ 02’00’’ 02’40’’ 03’20’’ 04’00’’ 04’40’’ 05’20’’

Time (m’s’’)

00’00’’ 00’40’’ 01’20’’ 02’00’’ 02’40’’ 03’20’’ 04’00’’ 04’40’’ 05’20’’

Time (m’s’’)

00’00’’ 00’40’’ 01’20’’ 02’00’’ 02’40’’ 03’20’’ 04’00’’ 04’40’’ 05’20’’

Time (m’s’’)

1’38’’ 1’51’’ 2’17’’ 4’22’’

0’01’’ 0’12’’ 3’23’’ 3’54’’

0’37’’ 0’48’’ 3’08’’ 5’05’’

1’11’’ 2’12’’ 3’35’’ 3’50’’

53 40 27 13 0 13 27 40

53 40 27 13 0 13 27 40

53 40 27 13 0 13 27 40 53 40 27 13 0 13 27 40

2 3

1 4

1

2 4

3

1 2

3

4

2

1

3

4

Figure A.6: The high temporal resolution of the PPI-CAPs approach allows us to link brain activity to stimuli presented at specific time points. Here, we show moments when each PPI-CAP was highly consistent across participants (left). The y-axis shows the percentage of subjects, out of the total 15 participants, who expressed the PPI-CAP at each given time point. The histogram on the right hand side of the panel depicts the distribution of random consistency across subjects. This was calculated by randomly permuting the time points on which the PPI-CAP appeared for each subject, then re-calculating the subject consistency across time (the same plot as the ones on the left). Finally, we plotted the distribution of possible consistency values. The dashed line represents the value of the 99^thpercentile from the random distribution, indicating that a subject consistency above this threshold is significant. Selected time points of high consistency across subjects for each PPI-CAP are indicated by numbers on the left plots, and the corresponding video frames are presented for those times on the rightmost side of the figure.