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

Gene set enrichment analysis (GSEA) is widely used to interpret genome-wide expression data, but pathway inference from differentially expressed gene (DEG) signatures remains limited by data scarcity, model incompatibility with tabular data, and limited interpretability. We aimed to develop an efficient and explainable model that predict gene-pathway associations from a large curated DEG signature compendium and supports potential therapeutic target discovery in complex diseases such as osteoarthritis (OA) and other disease contexts. We developed SaintGSE, a supervised prediction framework combining an autoencoder for dimensionality reduction and the SAINT transformer model for modeling tabular DEG signatures. The model was trained on a large public DEG compendium with pathway labels derived from EnrichR and interpreted using Integrated Gradients (IG) to prioritize driver genes. Experimental validation was conducted using mouse (n=5) and human (n=3) chondrocyte cultures, and in vivo OA-induced mouse models (n=5 per group). SaintGSE analysis identified key OA-related signaling pathways potentially modulated by natural extract from Senna obtusifolia, including NF-&#x3ba;B and p38. In vitro, the extract reduced OA catabolic factors compared with IL-1&#x3b2;-treated controls, including Mmp3 (mean difference 52.06, 95% CI 32.8-71.3, p<0.0001), Mmp13 (mean difference 5.13, 95% CI 2.9-7.4, p<0.0001), and Cox2 (mean difference 1.39, 95% CI 1.1-1.7, p<0.0001). In vivo, the extract significantly reduced cartilage damage compared with the DMM-operated PBS-treated group (effect estimate [mean rank difference] 13.3, p=0.0009). IG-based driver genes provided compact, condition-specific candidates supporting each pathway prediction. SaintGSE offers scalable, explainable pathway prediction from DEG signatures and enables mechanism- and target-oriented follow-up for disease studies and drug candidate screening.