Toward Deep Learning Techniques for Cell-Type and Spatial Resolution Estimation of Regulatory Networks

Understanding and decoding the intricacies of gene regulatory networks is crucial in genomics for insights into gene expression and cellular functions. Traditionally, research in this field has heavily relied on transcriptome data and machine learning to infer these networks, but this approach has mostly used bulk tissue samples. This

method overlooks the nuances of individual cells and their microenvironments, limiting our understanding to a broader, macroscopic level. The advent of spatial transcriptomics marks a significant shift, promising to unravel these networks at a single- cell and spatial level. This technology allows for the exploration of gene expression in relation to spatial dynamics,

enhancing our understanding of tissue organization and cellular functions. However, adapting machine learning to spatial genomics faces challenges. One major issue is the scarcity of spatial transcriptome data, which hampers the effectiveness of deep learning methods known for their superiority in network estimation. Another challenge is the need for models that account for the physical positions of cells,

as traditional methods treat data as independent and identically distributed, ignoring spatial relationships. To address these challenges, this proposal outlines two main objectives: Aim 1: Developing deep learning methods for cell-type resolution regulatory network estimation capable of transferring between scRNA-seq and spatial transcriptomics data. We will develop machine learning

mdoels that can integrate components that explicitly model the regulatory network, distinguishing cell types based on transcriptomic data. The approach will use domain-invariant regularization to adapt from scRNA-seq to spatial transcriptomics, employing GTEx and HuBMap data sets. Aim 2: Developing deep learning methods with spatial regularization for estimating regulatory networks

at spatial resolution within spatial transcriptomics. We will develop techniques that factor in the spatial positioning of cells during the learning process. The hypothesis is that cells in close spatial proximity have similar regulatory structures. This aim will also use GTEx and HuBMap data, along with collaborative efforts on spatial

transcriptome data of the human dorsolateral prefrontal cortex. Overall, this proposal seeks to lead the development of advanced deep learning models, integrating cell-type resolution and spatial dimensions to revolutionize our understanding of regulatory networks in genomics to both the cell-type and spatial resolution.