Interpretable Modeling of Functional Connectivity to Uncover Patterns Predictive of Cognitive Function in Youth

Interpretable predictive modeling of FC-behavior association. Our model learns fine-grained FC patterns predictive of behavioral traits at both regional and participant levels, capturing the overall FC-behavior association. Region-wise predictions are integrated using region-wise relevance scores, yielding a participant level prediction and an interpretable measure of each region’s contribution. These regional predictions and relevance scores are optimized collaboratively using a mean square error (MSE) based loss to enhance prediction performance.

---

🧠 Overview

This repository contains code and resources for uncovering functional connectivity (FC) patterns predictive of behavioral traits using interpretable machine learning models. The project aims to enhance the generalizability and interpretability of brain-wide association studies (BWAS) by learning fine-grained FC patterns at both regional and participant levels.

Features

• Interpretable predictive modeling of functional connectivity data
• Joint learning of region-level relevance scores and prediction functions
• Weighted integration of regional predictions for participant-level trait prediction
• Scalable to large neuroimaging datasets
• Evaluation across multiple cohorts for generalizability

---

📊 Data Sources

This project uses publicly available neuroimaging datasets:

• Adolescent Brain Cognitive Development (ABCD) Study
• Human Connectome Project Development (HCP-D)

Please refer to the respective data portals for access and data usage policies.

---

🚀 Getting Started

System Requirements

• Python (3.9.18)
• PyTorch (1.13.1)
• NumPy (1.25.2)
• pandas (1.5.3)
• SciPy (1.11.3)

Installation

1. Clone the repository and install dependencies:

git clone https://github.com/MLDataAnalytics/interpretable-fc-modeling.git

2. Navigate to the project directory:

cd interpretable-fc-modeling

3. Install the required dependencies. It's recommended to do this in a virtual environment.

pip install -r requirements.txt

---

💻 Usage

Model Traing

To train a model, execute the train_region_interpret_model.py script with the following arguments:

python ./src/train_region_interpret_model.py ${fold_id} ${trait_id} ${data_dir} ${train_list} ${fc_mode} ${output_dir} ${validation_list}

Arguments:

• fold_id: Training fold index (e.g., fold1, ..., fold5 for five-fold cross-validation). This is used in the output file names.

• trait_id: Target trait index (e.g., '1' for 'pc1', which represents general cognition in the ABCD study).

• data_dir: Path to the directory where the functional connectivity (FC) .mat files are saved.

• train_list: A .csv file containing the FC file names and behavioral traits for the training set. Each row corresponds to one individual.

  Example Format (fc_file, cognitive_measure_1, cognitive_measure_2, ...):

  sub-NDARINV5H1AW18P_fc.mat,0.54091746403823,-0.0192227192451836,0.792918923481453
  sub-NDARINVUCUU472H_fc.mat,-1.16157907198747,0.0439085846756245,-0.732536778124967
  sub-NDARINV99NA21PA_fc.mat,1.55319376072347,-0.751235818323214,0.0310268897410735

  The .mat file should contain a variable fc_p2p which is the whole-brain FC matrix (e.g., a 352x352 matrix for 333 cortical regions from Gordon atlas and 19 subcortical regions).

• fc_mode: The type of functional connectivity to use.

  • all: Whole-brain FC (e.g., 352x352).
  • cortex: Cortical-to-cortical FC (e.g., 333x333).
  • subcortical: Subcortical-to-cortical FC (e.g., 19x333).

• output_dir: Path to the directory where the trained model will be saved (e.g., output_dir/model_fold1_mse_pc1_Adam_regAtt_1.0/weights_200.pth).

• validation_list: A .csv file with the same format as train_list for the validation set.

Model Testing

To test the model, execute the test_region_interpret_model.py script:

python ./src/test_region_interpret_model.py ${fold_id} ${trait_id} ${data_dir} ${test_list} ${fc_mode} ${output_dir}

Arguments:

• fold_id: Testing fold index.
• trait_id: Target trait index.
• data_dir: Path to the directory where the FC files are saved.
• test_list: A .csv file containing the FC file names for the testing set.
• fc_mode: Must be one of all, cortex, or subcortical.
• output_dir: Path where testing results will be saved. The script expects the corresponding trained model to be present in a subdirectory (e.g., output_dir/model_fold1_mse_pc1_Adam_regAtt_1.0).

---

📜 Manuscript & Citation

For a comprehensive understanding of the method, please refer to the full manuscript. If you find this method useful or inspiring in your research, please consider citing:

@article{Li2025Uncovering,
  title={{Uncovering functional connectivity patterns predictive of cognition in youth using interpretable predictive modeling}},
  author={Li, Hongming and Cieslak, Matthew and Salo, Taylor and Shinohara, Russell T. and Oathes, Desmond J. and Davatzikos, Christos and Satterthwaite, Theodore D. and Fan, Yong},
  journal={{Proceedings of the National Academy of Sciences}},
  volume={122},
  issue={42},
  pages={e2505600122},
  year={2025},
  publisher={National Academy of Sciences},
  doi={10.1073/pnas.2505600122}
}
@article {Li2025.03.09.642155,
	author = {Li, Hongming and Cieslak, Matthew and Salo, Taylor and Shinohara, Russell T. and Oathes, Desmond J. and Davatzikos, Christos and Satterthwaite, Theodore D. and Fan, Yong},
	title = {Uncovering functional connectivity patterns predictive of cognition in youth using interpretable predictive modeling},
	elocation-id = {2025.03.09.642155},
	year = {2025},
	doi = {10.1101/2025.03.09.642155},
	publisher = {Cold Spring Harbor Laboratory},
	URL = {https://www.biorxiv.org/content/early/2025/03/10/2025.03.09.642155},
	eprint = {https://www.biorxiv.org/content/early/2025/03/10/2025.03.09.642155.full.pdf},
	journal = {bioRxiv}
}