Baycrest Research Division - Enriching Care, Enhancing Knowledge, Enlightening Minds

Rotman > Resources > Support + Services

back-project brain scores behavioral PLS

I'm Online

nicolette
Posted on 07/12/17 15:25:09
Number of posts: 5

nicolette posts:

Dear PLS experts,

I'm conducting commandline behavioural PLS analyses comprising cortical thickness (CT; FreeSurfer) vertex-wise (149955 vertices) and 6 physiological variables across 53 participants. I'd like to examine whole-brain associations between CT and physio data and back-project brainscores to visualize on the brain.

My questions:

1) Do the brainscores (boot_result.compare_u) resemble the order of the original order of CT data, i.e., can I use their vertex coordinates to back-project brainscores onto the hemispheres? It seems to be ok (https://www.rotman-baycrest.on.ca/index.php?action=view_thread&id=848&module=bbmodule&src=%40random45c35fcb17881).

If so, the back-projected brainscore maps overlaps quite well with two other analyses: a) clusters derived from a between-group whole-brain GLM and b) the related effect-size map (gamma.mgh/rstd.mgh, i.e., the between-group contrast/pooled residual sd, created in FreeSurfer). The latter shows a particularly good overlap with the PLS result (e.g., at ES=0.4-0.7). I think that's pretty exciting as two different analyses overlap quite well (in my eyes).

2) I want to extract average CT in some of the PLS clusters to plot against physio data. However, there's quite a number of smaller clusters for which I don't want to do that. One could e.g., choose those clusters where analyses overlap but perhaps you have a better suggestion?

Thanks a ton in advance!

Sincerely,

Nic

PS: I included some (a lot of) detail below but thought it may help catch potential mistakes and perhaps help others looking into similar analyses.

### ### ### ANALYSES ### ### ###

### GET CORTICAL THICKNESS (CT) VERTEX-WISE (from the "fsaverage-normalized" map so that everyone has the same number of vertices) ###

mris_convert -c $SUBJECTS_DIR/$sub/surf/lh.thickness.fwhm10.fsaverage.mgh $SUBJECTS_DIR/fsaverage/surf/lh.sphere.reg $output_dir/$sub.lh.thickness.fwhm10.asc

- Isolate the CT column from each participant's ascii file, paste those into one big file, and delete the rows that have 0s for everyone (medial surface)

- load this file into MATLAB

### PLS (6.15.1) ANALYSIS IN MATLAB (R2015-b) - one for each hemisphere ###

lh=load('nozero.lh.all.txt'); % Left-hemisphere CT data

lhtransp=lh.'; % Transpose to get rows=participants and columns=vertices

nsubjects=53;

ncolumns=149955;

ncolumns=6;

in2=fopen('physio_all.txt');

physiodata=fscanf(in2,'%f',[ncolumns,inf])';

fclose(in2);

save 'physiodata.mat' 'physiodata'

opt.method=3;

opt.num_perm=10000;

opt.num_boot=10000;

opt.stacked_behavdata=physiodata;

resulsvertexlh=pls_analysis({lhtransp},[nsubjects], 1, opt);

save 'resulsvertexlh.mat' 'resulsvertexlh'

resulsvertexlh.boot_result.compare_u % Brain scores

resulsvertexlh.perm_result.sprob % Significance of LVs

pct=(resulsvertexlh.s.^2/sum(resulsvertexlh.s.^2)) % Percent covariance accounted

### CREATE LABEL TO VIEW PLS RESULTS ###

# Create text file (keep '"' before label; vertices faceX faceY faceZ brainscore) and save as "lh.brainscore.label"

#label , from subject lh.pls

29112

0 0 0 100 -2.79

12 -42.53 30.9 85.07 -2.13

15 -42.53 -30.9 85.07 -2.26

27 -95.11 30.9 0 -2.56

31 -68.82 -50 -52.57 -2.67

32 0 -100 0 -2.5

40 -16.25 -50 -85.07 -2.27

52 44.92 86.27 23.25 -2.9

...

Replies:

Untitled Post

rmcintosh
Posted on 07/13/17 07:33:40
Number of posts: 394

rmcintosh replies:

Hi Nic - thanks for the quesitons and it looks like some nice results are emerging. To answer your questions:

1) the compare_u are bootstrap ratios (singular vector weight/standard error) technically - the variable usc are brain scores, but anyway, the ordering of elements is same as whatever you did to extract the CT measures, there shouldnt be any reordering when you read the data into matlab

2) the strategy of which peaks to include in your CT averages (i.e, overlap with GLM) is probably okay. You can also plot the brainscores themselves (usc) against the physiological measure as the scores are the weighted sum of the CT measures (using the singular vector weights). The variable "orig_corr" is this correlation

back-project brain scores behavioral PLS

I'm Online

nicolette
Posted on 07/14/17 12:04:39
Number of posts: 5

nicolette replies:

Hi Randy,

Thank you for your quick answer and the good information! I went back to my original PLS analysis and actually only report the BSR while using the usc for plotting against physio data and orig_corr for barplots showing the correlations for the significant LV.

The new whole-brain PLS results are exciting and I'm very happy to use the method :)

Have a great day,

Nic

back-project brain scores behavioral PLS

I'm Online

nicolette
Posted on 08/07/17 13:08:27
Number of posts: 5

nicolette replies:

Dear Randy,

I have a couple more questions and would be really grateful for your help.

1) How can I cluster-correct the whole-brain behavioral PLS analysis on the command line?

2) The whole-brain PLS shows a similar correlation pattern (boot_results.orig_corr) to the cluster PLS but directions are perfectly reversed. Could this be related to the PLS method or should I interpret the opposite relationships indeed as opposite relationships?

I thank you greatly in advance,

Nic

Untitled Post

rmcintosh
Posted on 08/07/17 18:32:39
Number of posts: 394

rmcintosh replies:

quote:

Dear Randy,

I have a couple more questions and would be really grateful for your help.

1) How can I cluster-correct the whole-brain behavioral PLS analysis on the command line?

I thank you greatly in advance,

Nic

Hi Nic

1) I am not sure you need to do cluster correction. The inferences are made on the whole image, not the clusters

2) the sign flip is probably arbitrary but let me be sure. In the cluster analysis, did you get something like area1 if postiive and area2 is negative with a positive weight for behavior 1 and negative for behavior 2 (e.g., +area1 * -area2 vs. +beh * -beh2) and in the whole brain *both* sides are flippped? i.e., -area1 * +area2 vs -behav1 * +behav2 ?

if thats the case you can simply flip the signs on both sides and should be ok

Baycrest Research Division - Enriching Care, Enhancing Knowledge, Enlightening Minds

Replies:

Contact Us: