Gene regulatory network inference stops choosing sides: scGPL fuses global topology with pairwise samples for up to 13.7% AUPRC gain
Synopsis
Inferring gene regulatory networks from single-cell RNA sequencing has long split into two routes: graph methods train on the whole network as one graph and tend to overfit, while pairwise methods split it into transcription factor–target pairs, gaining abundant samples but losing global topology. scGPL uses a dual-branch design to model both the global regulatory graph and a local subgraph around each candidate pair, reaching the best performance on seven benchmark datasets, with AUPRC up to 13.7% above the strongest baseline.
Interpretation
Gene regulatory network inference can now exploit both the topology of the whole regulatory graph and the abundance of transcription factor–target training pairs inside one end-to-end model, instead of trading one paradigm against the other. Earlier graph methods relied on a single training graph and were prone to overfitting, while pairwise methods had abundant training samples but ignored global network structure and were especially vulnerable to false negatives from incomplete annotation. On seven single-cell datasets curated by the BEELINE framework, scGPL was best on all cell-type-specific networks, with average AUROC 3.4% above scGREAT under the TFs+500 setting and 3.8% above it under TFs+1000.
For each candidate regulatory pair the model builds its own local subgraph, then fuses that local context with the whole-graph representation to predict interactions that reflect both broad pathways and fine-grained regulation. The global branch runs message passing over the full regulatory graph with a multi-layer graph neural network to extract multi-scale node embeddings, while the local branch encodes the pair's cross-cell co-expression with a Transformer and adds structural information about intermediate regulators and co-factors through dynamic subgraph sampling. Subgraph expansion scores neighbors by a weighted combination of node degree and expression cosine similarity and keeps the most informative ones; under TFs+1000 scGPL improves AUPRC by 13.7% over scGREAT, and on the sparsest-positive mDC dataset its EPR is nearly three times that of scGREAT.
Perspective
Evaluation covers seven human and mouse single-cell datasets curated by BEELINE, each in a transcription-factor-plus-500 or plus-1000 highly variable gene setting, compared across cell-type-specific, STRING and non-specific ground-truth networks; the authors propose extending the work to multi-omics data such as paired scRNA-seq and scATAC-seq and to large-scale single-cell foundation models.
The authors note that dynamic subgraph sampling adds computational overhead when the candidate interaction space is very large, that the global branch and subgraph modules still need some prior regulatory interactions to supply structural context, and that current evaluation is mainly within-dataset training and testing, leaving prediction for unseen genes or transfer across biological systems an open problem.