Abstract
High-throughput measurements that profile the transcriptome or the epigenome of single-cells are becoming a common way to study cell identity. These data are high dimensional, sparse and non linear. Here we present SEAGALL (Single-cell Explainable Geometry-Aware Graph Attention Learning pipeLine), a hypothesis free method to extract biologically relevant features from single-cell experiments based on geometry regularised autoencoders (GRAE) and explainable graph attention networks (GAT). We use a GRAE to embed the data into a latent space preserving the data geometry and we construct a cell-to-cell graph computing distances in the GRAE bottleneck. Exploiting the attention mechanism to dynamically learn the relevant edges, we use GATs to classify the cells and we explain the predictions of the model with XAI methods to unravel the features which are driving cell identity beyond marker genes. We apply our method to data sets from scRNA-seq, scATAC-seq and scChIP-seq experiments. SEAGALL can extract cell type specific and stable signatures which not only differ from the ones found in classical linear approaches but are less biassed by coverage and high expression.
Introduction
Single-cell sequencing technologies, such as scRNA-seq and scATAC-seq, provide unprecedented resolution into cellular heterogeneity. However, the resulting data matrices are characterized by high dimensionality, extreme sparsity, and inherent non-linearity. Traditional analysis pipelines often rely on linear dimensionality reduction techniques like Principal Component Analysis (PCA). While computationally efficient, PCA assumes a linear structure that often fails to capture the complex "manifold" of biological states. Consequently, subtle biological signals can be lost, and downstream interpretations may be biased toward highly expressed marker genes or technical artifacts like sequencing depth (coverage).
Methods: The SEAGALL Pipeline
To overcome these limitations, we developed SEAGALL (Single-cell ExplAinable Geometry-Aware Graph Attention Learning pipeLine). The framework integrates geometry-preserving deep learning with explainable graph neural networks to identify cell-type-specific signatures in a hypothesis-free manner.
The pipeline consists of four primary stages:
- Geometry-Regularized Autoencoder (GRAE): Instead of standard Variational Autoencoders (VAEs), GRAE incorporates a geometric regularization term to ensure the latent space preserves both local and global data structures.
- Graph Construction: A k-Nearest Neighbor (k-NN) graph is built using Euclidean distances calculated within the GRAE bottleneck, representing cells as nodes and their similarities as edges.
- Graph Attention Network (GAT): A GAT is trained to classify cell types. The attention mechanism allows the model to prioritize biologically relevant neighbors while filtering out noise.
- Explainability (XAI): Utilizing GNNExplainer, the model identifies the specific features (genes or genomic peaks) that most significantly influence the GAT's classification decisions.
Results and Benchmarking
We evaluated SEAGALL across ten diverse datasets, including scRNA-seq, scATAC-seq, and scChIP-seq. A critical component of our evaluation was the reconstruction of "corrupted" data—input matrices where artificial dropouts were introduced to simulate technical noise.
Reconstruction and Classification Performance
GRAE demonstrated superior performance in data reconstruction, achieving the lowest Mean Squared Error (MSE) compared to state-of-the-art methods like scVI, PeakVI, and PCA across multiple modalities. Furthermore, the GAT classifier achieved high accuracy, with F1 scores frequently exceeding 0.90, demonstrating robustness to biological noise.
Biological Insights: Chromatin Priming and Stability
SEAGALL’s non-linear approach revealed biological features often missed by standard differential analysis. Notably, the model identified "primed" chromatin states—genomic regions that become accessible in progenitor cells before the onset of gene expression in mature lineages. For instance, in human brain data, SEAGALL highlighted the importance of the SOX9 motif in the chromatin state of astrocyte progenitors. Additionally, the features identified by SEAGALL were found to be more stable and less influenced by technical biases such as high expression levels or read coverage compared to linear methods.
Discussion
SEAGALL offers a robust, geometry-aware framework for single-cell analysis that prioritizes biological interpretability. By respecting the non-linear structure of single-cell data, it uncovers stable signatures that define cell identity. While the computational cost of graph-based methods is higher than linear alternatives, the gain in biological resolution—particularly for epigenetic "priming"—is significant. Future developments will focus on multi-omic integration to provide a unified view of the regulatory landscape across different cellular modalities.
Summary updated