EfNST: A composite scaling network of EfficientNet for improving spatial domain identification performance

Abstract

Spatial Transcriptomics (ST) leverages Gene Expression Profiling while preserving Spatial Location and Histological Images, enabling it to provide new insights into tissue structure, tumor microenvironment, and biological development. The identification of spatial domains serves as not only the foundation for ST research but also a crucial step in various downstream analyses. However, accurately identifying spatial domains using computational methods remains a tremendous challenge due to the poor computational performance of many existing algorithms. Here, we propose EfNST, a deep learning algorithm based on a composite scaling network of the EfficientNet Network, designed specifically for the analysis of 10X Visium spatial transcriptomics data. We applied EfNST to three different datasets: human Dorsolateral Prefrontal Cortex, human breast cancer and mouse brain anterior. EfNST outperforms five advanced competing algorithms, achieving the best Adjusted Rand Index (ARI) scores of 0.554, 0.607, and 0.466, respectively. Notably, EfNST demonstrated high accuracy in identifying fine tissue structure and discovering corresponding marker genes with an improved running speed. In conclusion, EfNST offers a novel approach for inferring spatial organization of cells from discrete datapoints, facilitating the exploration of new insights in this field.