GENELink

GENELink infers gene regulatory networks (GRNs) from single-cell RNA sequencing (scRNA-seq) data using supervised graph-based methods that integrate transcription factor–DNA binding data such as ChIP-seq to improve regulatory interaction prediction.

Key Features:

• Supervised GRN inference (graph-based link prediction): Employs a graph-based link prediction approach to infer latent interactions between transcription factors (TFs) and target genes from scRNA-seq data.
• Integration of TF–DNA binding data (ChIP-seq): Leverages observed TF–gene pairs from transcription factor–DNA binding assays such as ChIP-seq to supervise network inference.
• Graph Attention Network: Uses a graph attention network to project single-cell gene expression into a low-dimensional space for downstream representation learning.
• Gene representation learning: Learns optimized gene embeddings for similarity measurement and causal inference between gene pairs.
• Performance and validation: Was benchmarked against eight existing GRN reconstruction methods across seven scRNA-seq datasets with four types of ground-truth networks and achieved comparable or superior performance.
• Application to breast cancer metastasis: Applied to human breast cancer metastasis data to reveal regulatory heterogeneity between primary tumors and lung metastases, including differential roles of Notch and Wnt signaling and the importance of mitochondrial oxidative phosphorylation (OXPHOS) during the seeding step.
• Ontology enrichment and experimental validation: Identifies ontology enrichment in unique lung metastasis GRNs that were validated through pharmacological assays.

Scientific Applications:

• Developmental biology: Dissects cell-specific regulatory mechanisms and dynamic regulatory landscapes at single-cell resolution.
• Oncology: Analyzes tumor regulatory heterogeneity and metastatic processes, exemplified by studies of breast cancer primary tumors versus lung metastases and pathways such as Notch, Wnt, and OXPHOS.
• Method benchmarking: Facilitates comparison and evaluation of GRN reconstruction methods on diverse scRNA-seq datasets and ground-truth networks.

Methodology:

Projects single-cell gene expression onto a low-dimensional space using a graph attention network with observed TF–gene pairs; learns gene representations optimized for similarity measurement or causal inference; and performs graph-based link prediction supervised by TF–DNA binding data such as ChIP-seq.