Abstract
Deciphering cell-type-specific gene regulatory networks (ctGRNs) is crucial for elucidating fundamental biological processes, such as tissue development and cancer progression. However, accurately inferring ctGRNs from high-dimensional transcriptomic data poses a significant challenge, primarily due to issues like data sparsity, cell heterogeneity, and over-smoothing (i.e. the tendency of node features to become indistinguishable after many graph convolution layers) in deep learning models. To tackle these obstacles, we present GeneLink+, an innovative framework for ctGRN inference leveraging directed graph link prediction (i.e. inferring causal regulator-target edges) tasks. Building upon the robust predictive capabilities of its primary version, GENELink, GeneLink+ incorporates residual-GATv2 blocks, which synergize dynamic attention mechanisms with residual connections. This architecture effectively mitigates information loss during the aggregation process and preserves cell-type-specific gene features, thereby enhancing the identification of regulatory mechanisms as well as the model's interpretability. Furthermore, GeneLink+ uses a modified dot product scheme with learnable weight parameters to adaptively prioritize informative gene pairs when scoring regulatory relationships, thus enabling more precise causal edge attribution. Comprehensive benchmarking across seven datasets demonstrated that GeneLink+ either outperforms or matches the performance of existing state-of-the-art methods in terms of predictive accuracy and biological relevance. Additionally, applications to a wide array of transcriptomic data, encompassing single-cell ribonucleic acid sequencing, small nuclear ribonucleic acid sequencing, and spatially resolved transcriptomics, have unveiled pivotal causal regulatory relationships in blood immune cells, Alzheimer's disease, and breast cancer.