3743 - Cell-Cell Interaction-Informed Deep Deciphering of Gene Regulatory Networks

Purpose/Objective(s): Deciphering gene regulatory networks (GRNs) is critical for understanding the molecular determinants of radiation response in cancer. Single-cell RNA sequencing (scRNA-seq) offers unprecedented resolution of cellular heterogeneity, yet traditional GRN inference methods often fail to capture the complexity of cell-cell interactions. We developed Cell-Cell Interaction Informed Deep Inference (C-CDI), which leverages explicit cell–cell interactions by transforming scRNA-seq data into a structured image format. A novel deep learning framework is applied to capture both local and global dependencies, thereby enhancing GRN inference. This approach aims to elucidate molecular determinants of radiation response and identify potential biomarkers for personalized radiotherapy.

Materials/Methods: C-CDI involves the construction of two-dimensional “CelloGraphs” by optimally arranging cells on a spatial grid according to system entropy, effectively encoding cell-cell relationships. By creating spatially meaningful CelloGraphs for each gene, clearer expression patterns of the gene across all cells emerge, enhancing the ability of the convolutional neural network (CNN) to comprehend gene co-expression. These image representations are then processed using a CNN to extract deep features and infer regulatory interactions between gene pairs. We benchmarked the method on multiple scRNA-seq datasets, comparing its performance against established techniques (e.g., CNNC, MLP, GRNBoost2, GENIE3) using metrics such as the area under the receiver operating characteristic curve (AUC) and precision-recall analyses.

Results: C-CDI consistently outperformed conventional methods, yielding mean AUC improvements of 8.2%–13.2% as compared to the second-best method CNNC in tasks including transcription factor–gene interaction, pathway prediction, and causality inference. The enhanced spatial depiction of gene expression patterns facilitated the identification of distinct regulatory modules. Preliminary application to tumor-derived scRNA-seq data further revealed candidate biomarkers associated with radioresistance.

Conclusion: C-CDI framework combines deep learning with innovative image representation of cell-cell interactions to advance GRN inference. Beyond its robust performance, this method shows promise for uncovering molecular signatures relevant to radiation response, potentially guiding the design of personalized radiation therapy.