Abstract
Single-cell RNA sequencing (scRNA-seq) technologies allow researchers to uncover the biological states of a single cell at high resolution. For computational efficiency and easy visualization, dimensionality reduction is necessary to capture gene expression patterns in low-dimensional space. Here we propose an ensemble method for simultaneous dimensionality reduction and feature gene extraction (EDGE) of scRNA-seq data. Different from existing dimensionality reduction techniques, the proposed method implements an ensemble learning scheme that utilizes massive weak learners for an accurate similarity search. Based on the similarity matrix constructed by those weak learners, the low-dimensional embedding of the data is estimated and optimized through spectral embedding and stochastic gradient descent. Comprehensive simulation and empirical studies show that EDGE is well suited for searching for meaningful organization of cells, detecting rare cell types, and identifying essential feature genes associated with certain cell types.
Introduction
Single-cell RNA sequencing (scRNA-seq) provides a high-resolution look at individual cells, but the resulting data presents a significant "curse of dimensionality": thousands of genes across thousands of cells, often plagued by "dropout" events where gene expression is missed due to technical noise. To interpret this data, researchers use dimensionality reduction (DR) to visualize patterns. However, current tools have trade-offs: t-SNE captures local clusters well but often destroys global relationships, while UMAP can struggle to separate rare cell types from dominant ones. Furthermore, most DR methods are "black boxes" regarding gene importance; they show that cells are different, but not which genes make them so.
We introduce EDGE (Ensemble Dimensionality reduction and feature Gene Extraction), a new framework that uses massive "weak learners" to simultaneously map high-dimensional single-cell data into 2D/3D space and identify the specific genes driving those patterns.
Methodology: The EDGE Framework
EDGE borrows the concept of "ensemble learning" (similar to Random Forests) and applies it to similarity searches. The algorithm proceeds through four key steps:
- Ensemble Sketching: The algorithm generates thousands of "weak learners." Each learner randomly selects a small subset of genes and creates a "sketch" (a bit vector).
- Hashing for Similarity: These sketches are used to group cells into "hash codes." If two cells consistently end up in the same box across thousands of learners, they are highly similar. This ensemble approach makes the similarity matrix incredibly robust against the "dropout" noise typical of scRNA-seq.
- Optimization: Once the similarity matrix is built, EDGE uses spectral embedding and stochastic gradient descent (SGD) to minimize cross-entropy, placing similar cells close together in a low-dimensional plot.
- Feature Extraction: Because the model knows which genes were used in which "boxes," it calculates an importance score based on information entropy. Genes that consistently help separate different cell types receive higher scores, identifying them as "feature genes."
Results and Performance
The authors validated EDGE using synthetic data and real-world datasets, including Jurkat/293T mixtures, Peripheral Blood Mononuclear Cells (PBMC), and Mouse Brain cortex/hippocampus cells.
Rare Cell Sensitivity and Hierarchical Preservation
EDGE is particularly effective at identifying rare cell populations. While UMAP might pull a rare cell into a nearby dominant cluster, EDGE’s hashing mechanism preserves the distinctiveness of rare entities. In the Jurkat dataset, EDGE perfectly isolated a 2.5% rare cell population that t-SNE and UMAP partially contaminated. Furthermore, in mouse brain data, EDGE successfully maintained hierarchical structures, keeping subtypes like different astrocyte groups together in a single "neighborhood" while showing clear boundaries.
Biological Relevance of Feature Genes
For the PBMC dataset, EDGE identified 35 top feature genes, including well-known markers like PPBP (platelets) and NKG7 (natural killer cells). This demonstrates that the algorithm identifies biologically valid targets without requiring separate differential expression analysis.
Discussion
EDGE outperforms industry standards like t-SNE and UMAP in preserving hierarchical cell structures and detecting rare cell populations. Despite using "massive" learners, the sketching method is computationally linear and efficient. Future work may extend this ensemble sketching framework to integrate multi-omics data, combining RNA-seq with protein or epigenetic data for even more robust cell state mapping.
Summary updated