A0435
Title: Graph-linked unified embedding considering label information
Authors: Hiroshi Kobayashi - Doshisha University (Japan) [presenting]
Masaaki Okabe - Doshisha University (Japan)
Hiroshi Yadohisa - Doshisha University (Japan)
Abstract: Graph-Linked Unified Embedding (GLUE) estimates the low-dimensional space shared across datasets derived from multiple sources. GLUE leverages prior knowledge between variables, represented as a knowledge graph and employs graph neural networks to achieve dimension reduction while considering latent structures. This method is particularly effective for analyzing related datasets and is often applied to multi-omics data, which integrates various omics data such as gene expression, proteomics, and DNA methylation. Unlike single-omics analysis, multi-omics data collected through diverse experiments or single-cell measurements contain complementary information that can improve the understanding of complex biological systems and diseases. However, applying GLUE to datasets with staging or cell-type labels may lead to low-dimensional representation that overlooks the label information. To address this, conditional GLUE (CGLUE) is proposed. By employing a conditional variational autoencoder, CGLUE conditions the encoder with labels and integrates multiple datasets into a shared low-dimensional space across sources while preserving label information. This approach promotes the close positioning of data points with the same label in a low-dimensional space, enhancing the interpretability of data based on label-specific information through visualization and analysis.
