Graph Masked Autoencoder With Multi-Head Latent Attention Mechanism for Gene Regulatory Link Prediction

Reconstructing gene regulatory networks (GRNs) from single-cell RNA sequencing (scRNA-seq) data remains challenging due to high dimensionality, noise, and cellular heterogeneity. The performance limitations of unsupervised methods necessitate more advanced supervised approaches. To address this, we present GMAMLink, an effective two-stage framework that integrates graph masked autoencoding with multi-head latent attention (MLA) for GRN inference. Central to our approach is a decoupled representation-topology learning paradigm: the model first performs self-supervised pre-training using a graph masked autoencoder (GraphMAE) to learn robust gene representations, followed by supervised fine-tuning, where MLA adaptively fuses these embeddings with the graph structure for precise link prediction. By separating gene representation learning from topological inference, this design enhances robustness and predictive accuracy. Extensive evaluations across seven scRNA-seq datasets and four benchmark networks demonstrate that GMAMLink consistently achieves superior or competitive performance compared to eight state-of-the-art methods. In a case study on triple-negative breast cancer (TNBC), GMAMLink reconstructed metastasis-specific regulatory networks and identified NUPR1 as a central hub gene in glycolytic pathways, with its elevated expression significantly correlated with metastatic progression and poor patient survival. These findings highlight GMAMLink's potential to uncover context-specific regulatory programs and prioritize high-value biological targets for combating cancer metastasis.