Introduction
Single-cell RNA sequencing (scRNA-seq) has transformed biological research by providing high-resolution insights into gene expression at the individual cell level. However, the resulting data presents significant computational challenges: it is high-dimensional, extremely sparse, and inherently noisy. Effective clustering is essential for identifying distinct cell populations, yet existing methods often struggle to balance structural information with computational efficiency.
Traditional clustering approaches generally fall into two categories. Feature-based methods focus on individual cell expression but often neglect the complex structural relationships between cells. Conversely, Graph Neural Network (GNN)-based methods capture these relationships but are prone to "over-smoothing"—where node representations become indistinguishable—and suffer from high computational costs on large datasets. To address these limitations, we present scCDCG, a novel deep learning framework that leverages deep cut-informed graph embeddings and optimal transport to achieve state-of-the-art clustering performance.
Methodology
The scCDCG framework integrates three synergistic modules designed to handle the unique characteristics of scRNA-seq data.
1. Autoencoder-based Feature Learning
The first module employs an unsupervised autoencoder to perform initial dimensionality reduction. By mapping the high-dimensional gene expression matrix into a lower-dimensional latent space, the model filters out noise and captures the most salient biological features, providing a robust foundation for downstream clustering.
2. Deep Cut-informed Graph Embedding
The core innovation of scCDCG lies in its approach to structural information. Rather than relying on computationally intensive graph convolutions, the model utilizes a "deep cut-informed" embedding strategy. It constructs two distinct graphs:
- Probability Metrics Graph: Based on covariance to capture statistical dependencies.
- Spatial Metrics Graph: Based on cosine similarity to capture geometric relationships.
The model optimizes a Joint Normalized Cut (NCut) loss across these graphs. This allows scCDCG to learn embeddings that respect the natural boundaries between cell populations without the iterative overhead or over-smoothing issues associated with standard GNNs.
3. Optimal Transport-based Self-Supervision
To refine cluster assignments, scCDCG incorporates a self-supervised module guided by Optimal Transport (Sinkhorn distance). Standard self-supervision can sometimes lead to "degenerate" solutions where the model collapses all data points into a single cluster. By using Sinkhorn distance, scCDCG enforces constraints that ensure the predicted cluster distribution aligns with the expected proportions of the data, maintaining diversity and accuracy in the final output.
Experimental Evaluation
The performance of scCDCG was evaluated on six diverse real-world scRNA-seq datasets, including Human Pancreas, Mouse Bladder, Worm Neuron, CITE-CMBC, and Human Liver cells. These datasets range from approximately 1,700 to over 8,600 cells and cover various levels of sparsity and dimensionality.
Performance and Efficiency
Benchmarked against seven established models (such as DEC, scGNN, and scDSC), scCDCG consistently achieved superior results across three key metrics: Accuracy (ACC), Normalized Mutual Information (NMI), and Adjusted Rand Index (ARI). Notably, scCDCG improved average accuracy by 5.58% over the previous state-of-the-art models.
Beyond accuracy, scCDCG demonstrated significant gains in computational efficiency. In runtime comparisons, it was substantially faster than GNN-based alternatives, particularly as dataset sizes scaled. Visualizations using t-SNE further confirmed that scCDCG produces tighter, more distinct clusters with well-defined boundaries compared to competing deep learning approaches.
Conclusion and Future Directions
scCDCG offers a faster and more accurate approach to mapping cellular landscapes by combining the strengths of autoencoders with efficient graph-cutting techniques. By avoiding the pitfalls of over-smoothing and high computational latency, it provides a scalable solution for modern transcriptomics.
Future research will focus on extending the framework to support multi-omics integration (e.g., combining RNA-seq with protein or epigenetic data) and improving the interpretability of the deep learning components to provide biologists with deeper insights into the drivers of specific cluster assignments.
Summary updated