Integrating Machine Learning and Biological Context for Single-Cell Gene Regulatory Network Inference

Single-cell RNA sequencing (scRNA-seq) data is a powerful approach for aiding in the uncovering of intricate regulatory networks that drive cellular processes. These gene regulatory networks (GRNs) play a pivotal role in understanding how transcription factors and target genes interact to control gene expression in diverse biological contexts. However, existing methods for GRN inference often rely heavily on mathematical models or correlation-based techniques, frequently neglecting biological knowledge. Alternatively, some approaches incorporate biological knowledge with strict constraints, limiting their ability to predict novel interactions. This oversight can result in reduced biological interpretability and the inclusion of false-positive regulatory interactions. To address these limitations, we propose a novel framework that integrates dataset-specific information with GO term annotations to enhance GRN inference. By combining mathematical modeling with biologically informed criteria, our approach prioritizes genes based on their functional relevance and improves the interpretability of inferred networks. Using scRNA-seq data from donors with type 2 diabetes, we demonstrate that our method significantly reduces false positives and can potentially identify biologically relevant transcription factors that other approaches may miss. Our framework offers a comprehensive and grounded approach. This integration of biological knowledge with computational models not only enhances the precision of GRN predictions but also provides a deeper understanding of the regulatory mechanisms underlying complex cellular processes and disease pathogenesis.