STDSN: Domain Separation Network for Transfer Learning in Spatial Transcriptomics Deconvolution

Résumé fourni par la source

Spatial transcriptomics (ST) preserves spatial information of gene expression within tissue architecture. However, its resolution limitations often necessitate deconvolution methods to infer cell type composition. Most existing methods rely on singlecell RNA sequencing (scRNA-seq) data as a reference but fail to fully account for the differences in unique features between ST and scRNA-seq data, which limits deconvolution performance. To address this issue, we propose a novel deconvolution method called STDSN based on transfer learning. STDSN adopts the concept of domain separation networks(DSN) and employs a shared-private encoder architecture. The shared encoder extracts common features from both simulated ST data generated from scRNA-seq and real ST data, while the private encoders capture their private features. This design minimizes the features differences and enables the model trained on simulated data to generalize effectively to real ST data for accurate deconvolution. We evaluated STDSN on 32 simulated datasets and 2 real spatial datasets, demonstrating that it significantly outperforms existing deconvolution methods.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.