Abstract 4366672: PAH-former: Transfer Learning for Efficient Discovery of Pulmonary Arterial Hypertension-Associated Genes

Introduction: Pulmonary arterial hypertension (PAH) is a severe disease characterized by pulmonary vascular remodeling. Current research is restricted by its rarity and lack of accurate in vivo models, making therapeutic target identification challenging. While single-cell RNA sequencing (scRNA-seq) offers high resolution, prioritizing candidate genes from its vast datasets remains difficult and often subjective, lacking causal validation. Research Questions: This study aimed to develop and validate a novel Geneformer-based platform, PAH-former, to overcome limitations of traditional differential gene expression analysis and robustly identify disease-associated genes in PAH. We hypothesized that fine-tuning Geneformer with PAH-specific scRNA-seq data would enable superior identification of pathogenic genes. Methods: PAH-former was developed by fine-tuning the Geneformer foundational model using public human PAH and control lung scRNA-seq datasets. Three training approaches were compared for cell classification accuracy. In silico gene deletion and overexpression analyses were performed using the fine-tuned model to predict genes driving PAH-like state transitions. Candidate genes were validated in vitro by siRNA knockdown in human pulmonary arterial endothelial cells (HPAECs), assessing changes in SOX18 mRNA expression, a known PAH-related transcription factor. Results: One of the three models achieved the highest cell classification accuracy (0.847) and F1 score (0.748). In silico deletion analysis with this PAH-former identified 134 candidate genes predicted to shift cells from a healthy to an PAH state; these included known disease related genes and many novel candidates, with minimal overlap with differential gene expression analysis from the original study. Gene Ontology enrichment of these 134 genes revealed relevant pathways, including TNF-α/NF-κB signaling, oxidative stress response, and VEGFA-VEGFR2 signaling. Knockdown of three novel candidate genes in HPAECs resulted in statistically significant increases in SOX18 mRNA expression. Conclusion: Our fine-tuned Geneformer platform, PAH-former, effectively identified both known and novel candidate genes associated with IPAH pathogenesis. This approach demonstrates the potential of context-specific fine-tuning of large-scale transformer models for unbiased gene discovery in complex diseases like IPAH, paving the way for new therapeutic strategies.