TL;DR
This paper introduces GNNLink, a novel framework that uses graph neural networks (GNNs) to predict gene regulatory networks (GRNs) from single-cell RNA sequencing (scRNA-seq) data. By framing GRN inference as a link prediction problem and employing a graph convolutional network (GCN) encoder, GNNLink effectively captures gene dependencies and achieves superior accuracy and efficiency compared to existing methods, even with noisy and sparse scRNA-seq data.
What Problem Does This Paper Address?
Understanding how genes interact and regulate each other within a cell is fundamental to deciphering complex biological processes and diseases. Gene Regulatory Networks (GRNs) are models that represent these interactions. Single-cell RNA sequencing (scRNA-seq) technology has revolutionized our ability to study gene expression at an unprecedented resolution, offering insights into cellular heterogeneity. However, scRNA-seq data is notoriously challenging to work with due to its high sparsity, noise, and dropout events (where a gene's expression is detected as zero even if it's present). These issues make it difficult to accurately infer GRNs. While some methods leverage existing knowledge of transcription factor (TF)-DNA binding, this data is often incomplete. This paper aims to bridge this gap by developing a robust method for GRN inference that can effectively handle the inherent challenges of scRNA-seq data and leverage both expression data and existing network information.
Methods
The core of GNNLink is its innovative approach to GRN inference, framed as a link prediction task. This means the model learns to predict whether a regulatory link (an edge) exists between two genes (nodes) in a network.
Here's a breakdown of the key components:
- Data Preprocessing: Standard scRNA-seq preprocessing steps are applied, including filtering low-expressed genes and focusing on those with significant variability.
- GCN-based Interaction Graph Encoder: This is the novel part. GNNLink uses a Graph Convolutional Network (GCN) to learn rich "gene features." The GCN takes an initial gene interaction graph (derived from prior knowledge or preliminary analysis) and the preprocessed gene expression data as input. It iteratively aggregates information from a gene's neighbors in the graph to refine its feature representation. This allows the model to capture complex interdependencies between genes.
- Link Prediction: After the GCN encoder generates refined gene features, GNNLink predicts the regulatory link score between any two genes by performing a simple dot product operation on their feature vectors. This score indicates the strength of the predicted regulatory relationship.
- Model Optimization: The model is trained using a supervised approach with both positive (known regulatory links) and negative (non-existent links) samples. A key aspect here is the use of a negative sampling strategy to help the model distinguish between true and false regulatory connections. The model is optimized using the mean squared error (MSE) loss and the Adam optimizer.
Datasets
The study utilizes a comprehensive set of datasets to evaluate GNNLink's performance:
- scRNA-seq Datasets: Seven scRNA-seq datasets from different cell types are used, including human embryonic stem cells (hESC), human mature hepatocytes (hHEP), mouse dendritic cells (mDC), mouse embryonic stem cells (mESC), and various mouse hematopoietic stem cells (mHSC-E, mHSC-GM, mHSC-L).
- Ground Truth Networks: To validate the inferred GRNs, four types of ground truth networks are employed:
- STRING database: Functional interaction networks.
- Non-specific ChIP-seq data: Provides information on TF binding across the genome.
- Cell-type-specific ChIP-seq data: More targeted TF binding information.
- Loss of Function/Gain of Function (LOF/GOF) data: Experimental data indicating the effect of gene perturbations.
The authors also mention using datasets with varying numbers of highly variable genes (top 500 and 1000) and transcription factors (TFs) to assess the model's scalability.
Key Insights and Ideas
- GRN Inference as Link Prediction: Reframing the complex problem of GRN inference as a link prediction task simplifies the problem and allows for the effective application of graph-based machine learning models.
- GCNs for Feature Enrichment: The GCN-based encoder is crucial for extracting meaningful gene features by effectively capturing the network structure and gene expression patterns simultaneously. This is a key differentiator from methods that only rely on expression data.
- Robustness to Sparsity: GNNLink demonstrates strong performance even with the inherent sparsity and noise in scRNA-seq data, a significant challenge for many existing methods.
- Efficiency Gains: The model not only achieves high accuracy but also significantly reduces training time compared to several baseline methods.
- Imputation Benefits: The study highlights that imputing missing values in scRNA-seq data (using tools like DeepImpute) can further improve GNNLink's GRN inference accuracy.
Key Results
GNNLink consistently outperformed six state-of-the-art baseline methods across various scRNA-seq datasets and ground truth networks.
- Accuracy: GNNLink achieved superior performance in terms of AUROC (Area Under the Receiver Operating Characteristic Curve) and AUPRC (Area Under the Precision-Recall Curve). For instance, AUROC values were approximately 9.2% and 18.3% higher on average than the next best methods (GENELink and GNE, respectively). GNNLink achieved the best prediction performance on 93.18% of the benchmark datasets for AUPRC.
- Efficiency: GNNLink demonstrated significantly faster running times compared to many baseline methods, making it more practical for large-scale analyses. For example, it was considerably faster than CNNC and DeepDRIM.
- Robustness: The model showed consistent performance across datasets of varying scales and was robust to changes in training data size and noise levels. It reached a performance plateau with as little as 40% of the training data.
Limitations and Future Work
The authors acknowledge several limitations and propose future directions:
- Suboptimal Node Features: The current node features might not be optimal, especially given the sparsity of known ground truth networks.
- Negative Sampling Strategy: The uniform negative sampling strategy might not fully capture the nuances of negative samples, potentially introducing bias.
- Future Work:
- Integrating Self-Supervised Learning: Incorporating self-supervised learning techniques could further enhance gene feature learning.
- Multi-modal Data Integration: Combining scRNA-seq data with other single-cell modalities like scATAC-seq and single-cell methylation data could provide a more comprehensive understanding of regulatory mechanisms.
- Advanced Negative Sampling: Exploring alternative negative sampling strategies to better capture data intricacies and improve model performance.
- Generalizability: The authors suggest GNNLink can be applied to other data types (e.g., plant cells) if the input data aligns with its characteristics. They also propose using inferred GRNs from other methods as prior information when specific prior knowledge is scarce.
Summary updated