Preprocessed Consortium for Neuropsychiatric Phenomics dataset

Preprocessing

For scanning parameters and details of the task fMRI paradigms, see 1. The input dataset was acquired from OpenfMRI.org⁶ - accession number ds000030, revision 1.0.3.

Results included in this manuscript come from preprocessing performed using FMRIPREP version 0.4.4 (http://fmriprep.readthedocs.io), a Nipype⁷ based tool. FMRIPREP was run with the following command line arguments:

--participant_label {sid} -w $LOCAL_SCRATCH --output-space T1w fsaverage5 template --nthreads 8 --mem_mb 20000

Where {sid} was the participant label and $LOCAL_SCRATCH was temporary folder for storing intermediate results.

Within the pipeline each T1 weighted volume was corrected for bias field using N4BiasFieldCorrection v2.1.0⁸, skullstripped using antsBrainExtraction.sh v2.1.0 (using OASIS template), and coregistered to skullstripped ICBM 152 Nonlinear Asymmetrical template version 2009c⁹ using nonlinear transformation implemented in ANTs v2.1.0¹⁰. Cortical surface was estimated using FreeSurfer v6.0.0¹¹.

Functional data for each run was motion corrected using MCFLIRT v5.0.9¹². Functional data was skullstripped using combination of BET and 3dAutoMask tools and was coregistered to the corresponding T1 weighted volume using boundary based registration with 9 degrees of freedom - implemented in FreeSurfer v6.0.0¹³. Motion correcting transformations, transformation to T1 weighted space and MNI template warp were applied in a single step using antsApplyTransformations v2.1.0 with Lanczos interpolation.

Three tissue classes were extracted from T1w images using FSL FAST v5.0.9¹⁴. Voxels from cerebrospinal fluid and white matter were used to create a mask in turn used to extract physiological noise regressors using aCompCor¹⁵. The mask was eroded and limited to subcortical regions to limit overlap with grey matter, six principal components were estimated. Framewise displacement and DVARS¹⁶ was calculated for each functional run using Nipype implementation. In addition to those regressors global signal and mean white matter signal was also calculated.

The whole dataset was preprocessed in total three times. After each iteration the decision to modify the preprocessing was purely based on the visual evaluation of the preprocessed data and not based on results of model fitting. First iteration (using FMRIPREP 0.4.2) uncovered inconsistent output image field of view and issues with EPI skullstripping, second iteration (using FMRIPREP 0.4.3) uncovered two cases of failed normalization due to poor initialization. In the final iteration all those issues were resolved. In total the preprocessing consumed ~22,556 single CPU hours.

For more details of the pipeline see http://fmriprep.readthedocs.io/en/0.4.4/workflows.html.

Volume-based task analysis

For a full description of the paradigms for each task, please refer to¹. We analysed the task data using FSL¹⁷ and AFNI¹⁸, implemented using Nipype⁷. Spatial smoothing was applied using AFNI’s 3dBlurInMask with a Gaussian kernel with FWHM=5mm. Activity was estimated using a general linear model (GLM) with FEAT¹⁷. Predictors were convolved with a double-gamma canonical haemodynamic response function¹⁹. Temporal derivatives were added to all task regressors to compensate variability in the haemodynamic response function. Furthermore, the following regressors were added to avoid confounding due to motion: standardised dvars, absolute dvars, the voxelwise standard deviation of dvars, framewise displacement, and the six motion parameters (translation in 3 directions, rotation in 3 directions).

For the Balloon Analog Risk Task (BART), we included 9 task regressors: for each condition (accept, explode, reject), we added a regressor with equal amplitude and durations of 1 second on each trial. Furthermore, we included the same regressors with the amplitude modulated by the number of trials before explosions (perceived as the probability of explosions). The modulator was mean centered to avoid estimation problems due to collinearity. For the conditions that require a response (accept, reject), a regressor was added with equal amplitude, and the duration equal to the reaction time. These regressors were orthogonalised with their fixed-duration counterpart to separate the fixed effect of the trial and the effect covarying with the reaction time. A regressor is added for the control condition.

In the retrieval phase of the Paired-Associate Memory Task (PAMRET), we modelled 4 conditions: true positives, false positives, true negatives, false negatives. For each condition, a regressor is modelled first with fixed durations (3s) and second with reaction time durations, with the latter orthogonalised with the former. With an extra regressor with control trials, there are 9 task regressors in total.

In the Spatial Capacity Task (SCAP), 25 task regressors were included. For each cognitive load (1 - 3 - 5 - 7) and each delay (1.5 - 3 - 4.5) with a correct response, two regressors were added: a regressor with fixed durations of 5 seconds and one with the duration equal to the reaction time, with the second orthogonalised with respect to the first. For both regressors, the onset is after the delay. The last regressor summarises all incorrect trials.

For the Stop-Signal Task (STOPSIGNAL), for each condition (go, stop - succesful, stop - unsuccesful), one task regressor was included with a fixed duration of 1.5s. For the conditions requiring a response (go and stop-unsuccesful), an extra regressor was added with equal amplitude, but the duration equal to the reaction time. Again, these regressors were orthogonalised with respect to the fixed duration regressor of the same condition. A sixth regressor was added with erroneous trials.

In the Task Switching Task (TASKSWITCH), all manipulations were crossed (switch/no switch, congruent/incongruent, CSI delay short/long), resulting in 8 task conditions. As in the SCAP task, we added for each condition two regressors: a regressor with fixed durations of 1 second, and one with the duration equal to the reaction time, with the second orthogonalised with respect to the first. There is a total of 16 regressors.

For subjects who are missing at least one regressor used in the contrasts, the task data are discarded. This is the case for example when no correct answers are registered for a certain condition in the SCAP task. For the SCAP task, we discarded 16 subjects; 14 subjects were removed for TASKSWITCH, 2 subjects for STOPSIGNAL, 2 subjects for BART, and 12 for PAMRET.

All modelled contrasts are listed in the Supplementary material. As is shown, all contrasts are estimated and tested for both a positive and a negative effect.

The total number of subjects modelled in the BART task is 259, while 244 subjects were modelled for the SCAP task. 254 subjects were included the TASKSWITCH task analysis, 197 subjects in the PAMRET task and 255 subjects in the STOPSIGNAL task.

Group level analysis

Subsequent to the single subject analyses, all subjects were entered in a one-sample group level analysis for each task. Three second level analysis strategies were followed: (A) ordinary least squares (OLS) mixed modelling using FLAME¹⁷, (B) generalized least squares (GLS) with a local estimate of random effects variance, using FSL¹⁷, and (C) non-parametric modelling (NP) using RANDOMISE²⁰, with the whole brain first level parameter estimates for each subject as input, and 10,000 permutations. The first two analyses use a group brain mask with voxels that were present in 100% of all subjects. For the permutation tests, a group mask was created where voxels were discarded for further analysis if less than 80% of the subjects have data in those voxels.

In addition to group level statistical maps, activation count maps (ACMs) were generated to show the proportion of participants that show activation, rather than average activation over subjects²¹. These maps indicate whether the effects discovered in the group analyses are consistent over subjects. As in 21, the statistical map for each subject is binarized at z=+/-1.65. For each contrast, the average of these maps is computed over subjects. The average negative map (percentage of subjects showing a negative effect with z < -1.65) is subtracted from the average positive map to indicate the direction of effects.