Tu213 - Floren: A Combination of Heterogeneous Graph Transformer and Graph Self-supervised Learning to Analyze Highly Specific Patterns in Single-cell Transcriptomic Data

Abstract Text: Single-cell transcriptomics has emerged as a powerful technique, with increasing data collection and complex experimental designs necessitating more flexible and accurate analytical methods. Graph Neural Networks (GNNs) have shown promise in this domain, with over 40 tools published in the past four years. However, most focus on homogeneous graph learning for cell annotation or clustering. To our knowledge, only DeepMAPS employs heterogeneous graph learning, and only Cellograph analyzes single-cell transcriptomic data at the differential gene expression (DGE) level.
We present FloREN, a GNN pipeline that extends DeepMAPS to multiple samples and feature connections, enhancing the Cellograph approach to address not only DGE but also transcription factors (TFs), cell states, regulatory pathways, and more. Our method integrates an autoencoder (AE) to establish a common representation for cells and genes, applies Heterogeneous Graph Transformers (HGT) to learn informative networks per sample, and utilizes a modified Graph Contrastive Learning (graphCL) technique to extract group-relevant patterns. We demonstrate FloREN's learning capabilities and flexibility on approximately 2 million cells from around 400 patients across four autoimmune conditions.