VaeSSC: Enhanced GRN Inference with Structural Similarity Constrained Beta-VAE

Hongmin Zhang, Ming Shi, Zhongwei Huang, Zhi Yang, Ran Zhou, Haitao Gan

Abstract: Gene regulatory network (GRN) encodes the intricate molecular interactions that govern the regulation of cell identity, thereby controlling the functions and characteristics of cells. With the emergence of single-cell transcriptomics, single-cell RNA sequencing has provided a powerful data foundation for the reconstruction of GRN. Consequently, the reconstruction of GRN has garnered significant attention. In recent years, deep learning has demonstrated remarkable performance across various domains, leading some researchers to apply deep learning models to the reconstruction of GRN. However, often overlooked is the correlation that exists among different cell types at different stages, resulting in ample room for improvement in the performance of GRN reconstruction. To address the need for models to capture the correlation information between cells, we propose a method called VaeSSC, which effectively captures the structural information of adjacent cells. By fully integrating the structural information of adjacent cells’ GRN, our method ensures that the reconstructed GRN conform more closely to objective principles, thereby enhancing the performance of GRN reconstruction. Extensive experiments conducted against challenging GRN reconstruction methods from the past have demonstrated the effectiveness of our proposed method.