CaHoT-GRN: context-aware high-order topology learning for robust single-cell gene regulatory network inference

Area of Science:

• Computational Biology
• Systems Biology
• Genomics

Background:

• Gene regulatory networks (GRNs) control cellular functions via transcription factors (TFs).
• Single-cell sequencing advanced GRN studies, but current inference methods often miss sequence context and TF properties, leading to unreliable networks.
• Existing methods predominantly use expression data, neglecting crucial sequence semantic and physicochemical information.

Purpose of the Study:

• To develop a robust framework, CaHoT-GRN, for accurate single-cell GRN inference.
• To incorporate sequence semantics and high-order topological information for improved network reconstruction.
• To address limitations of existing methods by integrating diverse biological data types.

Main Methods:

• Leveraging pretrained biological large language models for semantic embeddings from gene/protein sequences.
• Constructing a heterogeneous information network (HIN) using meta-path generation and protein-protein interactions to model cooperative regulation.
• Employing a similarity co-attention module to capture topological consistency and long-range gene associations.

Main Results:

• CaHoT-GRN achieved an average AUC of 0.846 and AUPR of 0.420 on single-cell transcriptomic datasets.
• The framework matched or outperformed existing GRN inference methods.
• Downstream analyses, including pathway and motif analyses, confirmed the biological relevance of the inferred networks.

Conclusions:

• CaHoT-GRN offers a robust and context-aware approach for single-cell GRN inference.
• The integration of sequence semantics and network topology enhances the reliability and biological interpretability of GRNs.
• The method demonstrates significant improvements over existing techniques, paving the way for more accurate understanding of gene regulation.