GeneRegLink: a deep learning framework for single-cell gene regulatory network inference via graph embedding and end-to-end learning
Abstract
Gene regulatory networks (GRNs) are complex systems of molecular regulators that control the activation and silencing of genes, crucially shaping cellular functions and identity in specific contexts. Deep learning approaches to GRN inference grapple with challenges such as enhancing predictive accuracy amid biological data noise, effectively modeling intricate gene interactions, and innovating architectures for efficient end-to-end learning from raw inputs. Hence, we present GeneRegLink, an innovative end-to-end supervised learning model, leverages graph mining algorithms for GRN inference by predicting potential gene interactions. GeneRegLink addresses key limitations of popular GRN inference tools by improving predictive accuracy in biological data, effectively modeling complex gene interactions through network embeddings, and maintaining computational efficiency via graph sampling and aggregation . It captures both local and global network structures with neighborhood sampling and feature propagation, while offering customizable embeddings tailored to specific biological contexts, such as diseases like triple-negative breast cancer. GeneRegLink outperforms eight state-of-the-art algorithms in predicting scRNA-seq datasets, demonstrated by its superior AUROC and AUPRC scores, which affirm its effectiveness in comparative analysis. These capabilities establish it as a powerful tool for elucidating gene regulatory dynamics.