SaintGSE: Transformer-based efficient and explainable gene set enrichment analysis

SaintGSE: Transformer-based efficient and explainable gene set enrichment analysis SaintGSE is an artificial intelligence model designed to predict human gene–pathway relationships using large-scale differentially expressed gene (DEG) datasets. By leveraging an autoencoder and the SAINT transformer model, SaintGSE addresses practical challenges in gene expression analysis, including data scarcity, model compatibility, and interpretability. This project uses and modifies code from the SAINT project (https://github.com/somepago/saint), licensed under the Apache License 2.0. Key Features * AI-driven pathway prediction: Uses an autoencoder and the SAINT model to analyze DEG signatures and predict associated signaling pathways. * Osteoarthritis study: Applied to osteoarthritis (OA) to identify key pathways and potential therapeutic targets. * Explainability (Integrated Gradients): Produces gene-level attributions via Integrated Gradients (IG) to identify influential genes supporting each pathway prediction. Installation Before installation, we recommend building a conda environment from the attached YAML file and activating it. Our code has been tested with python=3.8 on Linux. ``` $ cd /path/to/SaintGSE $ conda env create -f saintgse_env.yml $ conda activate saintgse_env ``` Clone the repository and install the dependencies: ``` $ git clone https://github.com/MSjeon27/saintgse.git $ cd saintgse ``` The code in this dataset is also accessible via GitHub. You can find the GitHub repository at the following link: https://github.com/MSjeon27/SaintGSE Usage Step 0. Preprocessing the input DEG (from pyDESeq2 result) Currently, SaintGSE has the function of converting mouse genes into human genes. The preprocessing code serves to change the human or mouse DEG data into the format used for SaintGSE. * human DEGs ``` $ preprocessing.py --query_fc /path/to/your/DEGs.tsv --out Preprocessed_fc.tsv ``` * mouse DEGs ``` $ preprocessing.py --query_fc /path/to/your/DEGs.tsv --org mouse --out Preprocessed_fc.tsv ``` Step 1. Training SaintGSE for a target pathway Available pathway names can be found in: /datasets/pathway_list_in_DEG.txt Training: ```bash $ python SaintGSE.py --pathway "Proteins Involved in Osteoarthritis" --pretrain ``` Step 2. Prediction through SaintGSE (Interpretation) Run prediction for a target pathway using a pre-trained or previously trained checkpoint: ```bash $ python SaintGSE.py --pathway "Proteins Involved in Osteoarthritis" --predict Preprocessed_fc.tsv ``` If IG is enabled, SaintGSE also produces gene-level attributions per sample. Typical outputs include: - pathway_IG_sample_scores.tsv (sample-level pathway calibrated score; closer to 1 indicates stronger association, closer to 0 indicates weaker association) - *_IG_gene_contributions.tsv (gene-level IG contributions per sample) Recommended driver gene definition (cumulative |IG|): In our OA analyses, we observed that a compact subset of DEGs (approximately 7–20% of the DEG signature) often explained 50% of the total cumulative |IG|. We recommend defining informative genes as the smallest set of genes that cumulatively accounts for the top 50% of |IG| (top 50% of cumulative |IG|). Note You can control IG computation via optional arguments: - ig_steps (e.g., 10/20/50/100/256) (default: 64) - ig_baseline (zero or mean) (default: zero) How to Cite If you use this model or repository in your research, please cite it as follows: ``` Jeon, MS & Nam, JH et al., "SaintGSE: Transformer-based efficient and explainable gene set enrichment analysis," 2024. GitHub repository. Available at: https://github.com/MSjeon27/SaintGSE ``` For more information or any questions regarding citation, feel free to contact us (msjeon27@cau.ac.kr).