Bridging Graph and Pairwise Learning for Gene Regulatory Network Inference from scRNA-seq Data
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: Genomics Radio. Generated commentary on the latest computational biology and genomics papers.
Ines: We're looking at the paper: "Bridging Graph and Pairwise Learning for Gene Regulatory Network Inference from scRNA-seq Data".
Marcus: Here's the gist of the paper. Gene regulatory networks (GRNs) capture the underlying interactions through which transcription factors (TFs) regulate genes The authors propose scGPL, an end-to-end unified framework that jointly models global regulatory topology and local interactions This design bridges the gap between graph-based and pairwise-based methods to achieve state-of-the-art performance on single-cell RNA sequencing data.
Ines: Problem Formulation and Existing Challenges GRN inference aims to recover unknown edges from observational data by representing the network as a directed graph (G, E) where nodes are TFs or genes and edges indicate regulatory relationships Current scRNA-seq datasets present two major challenges for supervised GRN inference, namely severely incomplete edge annotations and a high-dimensional search space where potential interactions vastly outnumber available annotated training samples.
Marcus: Existing methods fall into two categories: graph-based methods, which model the GRN as a whole graph, and pairwise-based methods, which decompose the GRN into individual TF–target gene pairs for modeling.
Ines: Graph-based methods tend to overfit due to their reliance on a single training graph, while pairwise-based methods overlook the global topological structure essential for comprehensively capturing regulatory mechanisms.
Paper summary: Ines: So, we're talking about this paper today, "Bridging Graph and Pairwise Learning for Gene Regulatory Network Inference from scRNA-seq Data." Basically, they're taking gene expression data to figure out how transcription factors regulate genes.
Marcus: That’s the core idea. They look at existing ways to do this, which are split into two main camps: graph-based methods that treat the whole network as one big graph, and pairwise methods that just look at individual pairs of transcription factor and gene interactions.
Ines: And they point out that each approach has a weakness. The graph method tends to overfit because it only trains on one specific training snapshot, but the pairwise method completely misses the big picture, the global structure of how everything connects.
Marcus: Exactly. So this paper proposes something new called scGPL, an end-to-end framework that tries to fix both problems at once by modeling both the global topology and those local interactions simultaneously.
Yuki: From a population perspective, I’m curious if this unified view helps us see regulatory patterns that are more stable across different cell types or even populations, rather than just noise in one specific sample.
Ines: Right. The authors claim that scGPL extracts macro-level features from the global regulatory graph and then uses a dynamic way to sample subgraphs based on importance scores to capture those fine-grained local dependencies.
Marcus: So they fuse these two views: the global representations with those locally enhanced features, which is how they bridge the gap between that old graph method and the pairwise methods.
Yuki: If you can get both the big structure and the small neighborhood context, that should give us a much richer picture of actual biological regulation happening in these single-cell data sets.
Ines: They tested this on seven benchmark datasets, and they report that scGPL consistently achieves state-of-the-art performance, even showing up to a thirteen point seven percent improvement over the strongest baseline method.
Marcus: That number is pretty significant because it validates their integrated design by showing it performs better than what was previously considered the best approach for this problem.
Yuki: It suggests that incorporating that whole graph structure into the pairwise modeling context actually helps support more robust link predictions in these high-dimensional expression spaces.
Ines: Beyond just performance, they showed good robustness too; for instance, under thirty percent edge noise, the average precision recall curve of scGPL only dropped by zero point one five nine percent, which is much less than the thirteen percent drop seen with the baseline method.
Marcus: That kind of stability under noise is important when you're dealing with real single-cell data where sparsity and noise are constant challenges.
Yuki: And I wonder what that means for our understanding of gene regulation in complex systems; does it mean we can trust the predicted connections more when we apply this framework to actual biological samples?
Ines: Well, they also showed that visualizing the latent representations extracted by scGPL using something like t-SNE revealed distinct clustering for transcription factor target gene pairs compared to just looking at the raw input features.
Marcus: So it’s not just a mathematical trick; these latent representations seem to organize themselves in a way that makes sense biologically, separating related interactions from the rest of the noise.
Yuki: That clustering would be very interesting because it could map out functional modules within the network structure itself, which is something we need to understand more about developmental trajectories.
Ines: The paper’s main contribution really comes down to this idea: they propose scGPL, an end-to-end unified framework that enhances pairwise modeling with subgraph derived topological context.
Marcus: It seems like the key is retaining the advantage of having large scale training samples from pairwise methods while giving the model enough context from the whole graph to avoid overfitting.
Yuki: So, in simple terms, what this means for someone just listening to this show? It means we have a new way to map out who talks to whom in genes that uses both the big map and the local neighborhood information at the same time.
Ines: That’s right. It bridges those two different ways of looking at gene regulation so we can make more comprehensive inferences about how transcription factors control their targets.
Marcus: It’s a solid improvement over existing methods because it tackles the limitations of both the pure graph and pure pairwise approaches by combining them into one system.
Yuki: The authors also mention that they successfully predicted associations that weren't even in the original training data, like a case study involving TP53 and EPCAM, which points to its ability to find novel regulatory links.
Ines: That’s a strong signal because it suggests the model isn't just memorizing existing patterns; it’s learning underlying biological relationships that we hadn't explicitly annotated.
Marcus: So, looking at the authors Hang Li and colleagues, this work shows how integrating global structure with local context can yield better predictive power in genomics data.
Yuki: It pushes the idea that for complex biological systems, you need more than just looking at individual connections; you need the whole system context to really grasp what's happening.
Conclusion: Ines: So we’ve been looking at scGPL, this new framework that tries to combine the best of both graph and pairwise learning for figuring out gene regulatory networks from single-cell data.
Marcus: Yeah, the authors titled it "Bridging Graph and Pairwise Learning for Gene Regulatory Network Inference from scRNA-seq Data". They’re basically trying to use a unified system.
Yuki: From my side, what this means is they’ve built something that should handle the messy nature of single-cell data better than either method alone.
Ines: It does try to model both the big picture global structure and those fine local interactions at the same time.
Marcus: And the results are pretty solid. They show state-of-the-art performance, with some improvements over existing methods on seven different datasets.
Yuki: That’s interesting because it means we might get more reliable network structures to study across different cell types or even populations.
Ines: It suggests that this integrated approach gives us a much more comprehensive view of how transcription factors control their targets in these complex systems.
Hang Li, Jiaxin Qi, Yan Cui, Yuhua Zheng, Jianqiang Huang
Hangzhou Institute for Advanced Study, University of Chinese Academy of Sciences · Computer Network Information Center, Chinese Academy of Sciences
q-bio.GN, cs.AI
Submitted: 2026-11-10
Updated: 2026-11-10
Comments: Accepted to Bioinformatics
Code: https://github.com/talent-hang/scGPL
License: http://creativecommons.org/licenses/by/4.0/
Importance score: 92/100
The gist: Gene regulatory networks (GRNs) capture the underlying interactions through which transcription factors (TFs) regulate genes The authors propose scGPL…
Key concepts
- Graph-based Methods
- These methods treat the entire gene regulatory network as one big graph. They are good at capturing the overall structure but often struggle with overfitting when trained on limited data, as they rely too heavily on a single training graph.
- Pairwise-based Methods
- These methods focus on modeling individual interactions between a transcription factor and its target gene. They benefit from having many training examples but lose sight of the larger, global network structure that governs how these local interactions fit together.
- scGPL Architecture
- scGPL is an end-to-end system with two main parts: a Global Branch using a Graph Neural Network to find macro features of the whole network, and a Pairwise Branch featuring dynamic subgraph sampling to find fine-grained local dependencies.
Summary
Gene regulatory networks (GRNs) capture the underlying interactions through which transcription factors (TFs) regulate genes The authors propose scGPL, an end-to-end unified framework that jointly models global regulatory topology and local interactions This design bridges the gap between graph-based and pairwise-based methods to achieve state-of-the-art performance on single-cell RNA sequencing data. p. 0 cn
Abstract
Motivation: Gene regulatory networks (GRNs) capture the underlying interactions through which transcription factors (TFs)
regulate genes. Based on gene expression data, existing GRN inference approaches generally fall into two categories: graphbased methods, which model the GRN as a whole graph, and pairwise-based methods, which decompose the GRN into
individual TF–target gene pairs for modeling.Where the paper says it
Problem Formulation and Existing Challenges
GRN inference aims to recover unknown edges from observational data by representing the network as a directed graph (G, E) where nodes are TFs or genes and edges indicate regulatory relationships Current scRNA-seq datasets present two major challenges for supervised GRN inference, namely severely incomplete edge annotations and a high-dimensional search space where potential interactions vastly outnumber available annotated training samples. Existing methods fall into two categories: graph-based methods, which model the GRN as a whole graph, and pairwise-based methods, which decompose the GRN into individual TF–target gene pairs for modeling. p. 0 cn
Abstract
Motivation: Gene regulatory networks (GRNs) capture the underlying interactions through which transcription factors (TFs)
regulate genes. Based on gene expression data, existing GRN inference approaches generally fall into two categories: graphbased methods, which model the GRN as a whole graph, and pairwise-based methods, which decompose the GRN into
individual TF–target gene pairs for modeling. Graph-based methods tend to overfit due to their reliance on a single training graph, whereas pairwise-based
methods benefit from abundant TF–target training pairs. In contrast, pairwise-based methods overlook the global topological
structure, which is essential for comprehensively capturing regulatory mechanisms.Where the paper says it
Where the paper says it
How the Training Signal is Built
scGPL employs a dual-branch architecture that synergizes global and local views to extract features The Global Branch uses a multi-layer Graph Neural Network to extract macro-level gene embeddings from the entire expression matrix and prior network, capturing hierarchical topological features. The Pairwise Branch consists of two sub-modules: a Transformer Module that captures co-expression dependencies from pairwise expression profiles, and a dynamic Subgraph Module that extracts local topological features. p. 2 Overview
To overcome the limitations of graph-based and pairwisebased methods, we propose scGPL, an end-to-end unified
framework that seamlessly blends global regulatory topology
with fine-grained local interactions. As illustrated in Figure 2,
scGPL consists of three integrated components: (1) Global
Branch, which extracts multi-scale topological embeddings
from the entire regulatory graph to capture broad regulatory
pathways; (2) Pairwise Branch, which serves as the core
of our localized modeling and comprises two sub-modules:
a Transformer Module for capturing non-linear co-expression
dependencies, and a dynamic Subgraph Module for encoding
local regulatory contexts; (3) Multi-scale Fusion Module,
which hierarchically integrates these global and local biological
perspectives to yield comprehensive and robust link predictions. Starting from the initial gene pair, the module iteratively expands the subgraph by
selecting top-k neighbors based on a hybrid importance score, providing inputs to the Subgraph Module. where m is the trainable model designed to extract features and
predict yij , and L denotes the corresponding loss function.Where the paper says it
Where the paper says it
Where the Gains Come From
The superior performance of scGPL stems from its integrative design that fuses global representations with subgraph-enhanced local features The Multi-Scale Fusion Layer hierarchically integrates these global and local biological perspectives to yield comprehensive and robust link predictions. Key contributions include:
We propose scGPL, an end-toend unified framework that enhances pairwise modeling with subgraph-derived topological context
This design allows the model to effectively capture both local topological structures and global network information while retaining the advantage of large-scale training samples
Robustness and Interpretability
scGPL demonstrates remarkable robustness against noise and sparsity inherent in single-cell data For instance, under 30% Edge Noise, the AUPRC of scGPL decreases by merely 0.159%, significantly outperforming the baseline GENELink, which suffers a 13% decline. p. 7 For instance, under 30% Edge Noise,
the AUPRC of scGPL decreases by merely 0. 159%, significantly
outperforming the baseline GENELink, which suffers a 13%
decline. Embeddings derived directly from raw expression
profiles exhibited substantial overlap between target and nontarget gene pairs, indicating poor separability in the original
expression space. In contrast, the latent representations
extracted from the penultimate layer of scGPL—which
seamlessly integrate gene expression, global topological
properties, and local subgraph contexts—demonstrated distinct
clustering. systematically examined them as potential missing regulatory
relationships to determine whether our model could recover
biologically plausible associations beyond current annotations. As illustrated in Figure 7, we selected four representative
genes from the hESC dataset under the Specific network
and constructed localized GRN subnetworks using their top13 predicted TF–target pairs ranked by model confidence
scores.Where the paper says it
Where the paper says it
Where the paper says it
Conclusion
scGPL is an end-to-end unified framework that synergizes macro-level global topology with fine-grained local interactions via a dynamic subgraph sampling mechanism By bridging graph-based and pairwise-based paradigms, scGPL leverages data abundance to effectively alleviate sample scarcity and overfitting, thereby consistently outperforming state-of-the-art methods across diverse datasets. Future work aims to extend this framework by integrating multi-omics modalities like paired scRNA-seq and scATAC-seq. p. 8 Future work aims to
extend this framework by integrating multi-omics modalities
(e. , paired scRNA-seq and scATAC-seq) and leveragingWhere the paper says it
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Dual-stream architecture for scGPL allows it to extract multi-scale topological embeddings from a Global Branch and fine-grained local interactions from a Pairwise Branch, enabling it to
effectively bridges the gap between the two paradigms.
-
Dynamic subgraph sampling mechanism in scGPL captures
fine-grained local dependencies
by iteratively expanding subgraphs based on adynamic priority scoring
that balances topological importance and functional co-expression relevance. -
The framework can perform state-of-the-art GRN inference across diverse datasets, achieving up to 13.7% improvement over the strongest baseline, such as scGREAT on TFs+1000 data.
-
scGPL can generalize learned regulatory rules to
completely unseen cell types
by treating existing interactions asincomplete structural priors,
shifting naturally toward expression-based pairwise representations when topological information is limited. -
The model exhibits superior robustness against noise and sparsity, showing that under 30% Edge Noise, the AUPRC decreases by only 0.159%, significantly outperforming the baseline GENELink which suffers a 13% decline.
-
The representation learning capability demonstrates that scGPL
effectively transforms high-dimensional, sparse singlecell expression data into a robust feature space enriched with regulatory semantics,
as shown by t-SNE visualizations showing distinct clustering of true TF-target gene pairs. -
The framework can generate biologically plausible hypotheses beyond existing databases by analyzing
topranked, high-confidence predictions generated by scGPL that were absent from the benchmark ground-truth networks.
Abstract
Motivation: Gene regulatory networks (GRNs) capture the underlying interactions through which transcription factors (TFs) regulate genes. Based on gene expression data, existing GRN inference approaches generally fall into two categories: graph-based methods, which model the GRN as a whole graph, and pairwise-based methods, which decompose the GRN into individual TF-target gene pairs for modeling. However, each approach exhibits limitations that are precisely the strengths of its counterpart. Graph-based methods tend to overfit due to their reliance on a single training graph, whereas pairwise-based methods benefit from abundant TF-target training pairs. In contrast, pairwise-based methods overlook the global topological structure, which is essential for comprehensively capturing regulatory mechanisms. Results: To address these limitations, we propose scGPL, an end-to-end unified framework that jointly models global regulatory topology and local interactions. scGPL extracts macro-level features from the global regulatory graph and introduces a dynamic, importance-guided subgraph sampling mechanism to capture fine-grained local dependencies. By fusing global representations with subgraph-enhanced local features, our framework bridges the gap between the two paradigms, incorporating essential whole-graph topological dependencies while retaining the data abundance advantage of pairwise methods. Extensive experiments on seven benchmark datasets demonstrate that scGPL consistently achieves state-of-the-art performance, with up to 13.7% improvement over the strongest baseline, validating the effectiveness of our integrative design. Availability and implementation: The source code is maintained on GitHub (https://github.com/talent-hang/scGPL), and the datasets and versioned releases are archived at https://doi.org/10.5281/zenodo.21608485 for reproducibility.
Related papers
- SeekRBP: Leveraging Sequence-Structure Integration with Reinforcement Learning for Receptor-Binding Protein Identification
- Genome-Factory: A Library for Tuning, Deploying, and Interpreting Genomic Foundation Models
- AVA: Agentic Variant Adjudication for Automated, Auditable Read-Level Review in Cancer Diagnostic Sequencing
- C3P: Contrastive promoter-protein pretraining yields representations capturing bacterial gene regulation
- Leveraging genomic deep learning models for the prediction of non-coding variant effects
- Adaptive Allele-Haplotype Incidence Indexing for Pangenome Queries: Representation, Reconstruction, and Benchmarking