Network R01 Functional Alignment Analyses

Authors: Chris Iyer

Poldrack Lab, Stanford University Updated: 8/7/2024

This respository contains code + results for functional alignment, task decoding, and conjunction analysis for our Network R01 rest data (discovery sample). To replicate the analyses, run the following scripts:

1. connectivity.py: loads rest data from /oak/stanford/groups/russpold/data/network_grant/discovery_BIDS_21.0.1/derivatives/glm_data_MNI. Derives a voxel-to-parcel connectivity matrix for each session of rest data (each of 5 subjects have 10-12 rest sessions). This means each voxel is correlated to each parcel of the 400-dimension Schaefer parcellation (found in data/templates; alternatively can pass atlast='DiFuMo to the get_parcellation() function). Saves connectomes to outputs/connectomes.
2. connectome_average.py: averages each subject's 12 session-wise connectomes into 1 per subject. We first used reliability.ipynb to visualize the connectomes and verify that we see within-subject, across-session shared structure--this was clear, so we averaged within-subject.
3. srm.py: derives connectivity-based Shared Response Models (SRM; Chen et al. 2015; Nastase et al. 2020; Guntupalli et al. 2018). This derives one SRM per parcel (as an anatomical constraint), and concatenates across all parcels. The derivation is done on all subjects' connectomes at once (none are left out--later, in contrast_map_analyses/conjunction_analysis.py, this is done with one subject left out at a time). It saves a parcelwise shared response, and subject-specific transformation matrices (concatenated across parcels, from each subject's native voxel space into the shared space). This is our functional alignment method of choice.
4. task_beta_mapping.py: models trial-level activation during our battery of cognitive control tasks. These trial maps will then be decoded in the next script. NOTE: the failure of our task decoding to get reliable signal may be because our trials are too fast to get reliable trial-level activation estimates in this script.
5. task_decoding.py: decodes trial labels of the trial activation maps produced above, comparing linear-SVC-based decoding performance on only-anatomically-aligned data to SRM-transformed (i.e. functionally aligned) data. Outputs (a .pkl file of classifier accuracy values and a .png plot of classifier ROC-AUC on each task) are saved to outputs/decoding.
6. visualize_results.ipynb: knits confusion matrices from decoding into a PDF and runs statistical tests on classifier performance data. Not strictly necessary but you can take a look!
7. ALTERNATE ANALYSIS: the purpose of the task decoding analysis is to quantify the additional signal within functionally-aligned task data by decoding task states and comparing to un-aligned data. As a total alternative, I implemented a conjunction analysis of aligned and un-aligned task contrast maps in contrast_map_analyses/conjunction_analysis.py. This script re-derives SRM transformations leaving one subject out of the initial derivation. It then transforms all other subjects' task contrast maps into the left-out subject's native space, and computes the Dice coefficient and Pearson R of all pairs of subjects' contrast maps. It also computes these for MNI-only (non-SRMd) data. This offers another way of assessing the benefit of SRM. Results are saved to outputs/conjunction_analysis--both the raw coefficients and a plot.

Additional files/folders:

• environment.yml contains my conda environment.
• run.sbatch contains a sample sbatch script to run all or some of the python scripts on Sherlock, with the memory parameters I used.
• archive contains old code and sanity checks I ran, including some contrast map reliability checks and a script performing within-subject decoding.

Google doc explaining steps in more detail

Please contact c.iyer@columbia.edu for all questions!