GatorSC: Multi-Scale Cell and Gene Graphs with Mixture-of-Experts Fusion for Single-Cell Transcriptomics

GatorSC is a unified, self-supervised representation learning framework that integrates multi-scale cell and gene graphs via a Mixture-of-Experts architecture to produce robust embeddings, outperforming state-of-the-art methods in cell clustering, gene imputation, and biological discovery across diverse single-cell RNA sequencing datasets.

The Gist

Imagine you are trying to understand a massive, chaotic city. You have a list of every person living there (the cells) and a list of every job they do or tool they use (the genes).

The problem is that this city is noisy. Some people forgot to write down their jobs (missing data), some reports are filled with typos (technical noise), and the city is so huge that looking at just one street or just one profession doesn't tell you the whole story.

This is the challenge of scRNA-seq (single-cell RNA sequencing). Scientists want to group people into neighborhoods (clustering), guess what jobs they should have had (imputation), and label who they are (annotation), but the data is messy.

Enter GatorSC. Think of GatorSC as a super-smart city planner who uses a "Mixture of Experts" to organize this chaos. Here is how it works, broken down into simple concepts:

1. The Three Maps (Hierarchical Graphs)

Instead of looking at the city from just one angle, GatorSC draws three different maps to understand the structure:

• Map A: The Neighborhood Map (Global Cell-Cell Graph)

  • The Analogy: This map connects people who live near each other. It asks, "Who hangs out with whom?" It captures the big picture of the city's layout.
  • The Science: It connects cells that look similar to each other, creating a global view of the population.

• Map B: The Industry Map (Global Gene-Gene Graph)

  • The Analogy: This map ignores who the people are and focuses on what they do. It connects "Carpenters" to "Carpenters" and "Doctors" to "Doctors," even if they live on opposite sides of the city. It shows how tools and jobs are related across the whole city.
  • The Science: It finds relationships between genes that function together, regardless of which specific cell they are in.

• Map C: The Local Block Map (Local Gene-Gene Graph)

  • The Analogy: This is a zoomed-in view of specific neighborhoods. Maybe in the "Downtown" neighborhood, carpenters and electricians work together closely, but in "Suburbia," they don't. This map captures those unique, local partnerships.
  • The Science: It looks at gene relationships within specific groups of cells, capturing context-specific interactions.

2. The "Mixture of Experts" (The Smart Fusion)

Now, you have three different maps. A bad planner might just glue them all together into one giant, confusing mess. A better planner might just pick one map and ignore the others.

GatorSC uses a Mixture-of-Experts (MoE) system. Imagine a team of three specialized detectives:

• Detective A is great at seeing the big neighborhood layout.
• Detective B is great at understanding industry trends.
• Detective C is great at spotting local street-level details.

Instead of forcing them to agree on a single answer, GatorSC has a Gating Network (like a wise manager). When looking at a specific person (cell), the manager asks: "Who is the best detective for this specific case?"

• If the person is part of a complex, large group, the manager listens more to Detective A.
• If the person has a unique local job, the manager listens more to Detective C.

This allows the system to adaptively combine the best parts of all three maps to create a perfect profile for every single cell.

3. The "Self-Teaching" Method (Self-Supervised Learning)

Usually, to teach a computer to sort things, you need a teacher with a stack of answer keys (labeled data). But in biology, we often don't know the answers yet!

GatorSC teaches itself using two tricks:

• The "Fill-in-the-Blanks" Game (Reconstruction): The system takes a map, hides some connections (like erasing a few streets), and tries to redraw them. If it can successfully guess the missing streets, it proves it understands the city's structure.
• The "Spot the Difference" Game (Contrastive Learning): The system takes two slightly different versions of the same map (like a photo and a slightly blurry photo of the same city) and learns to recognize that they are the same city, despite the noise.

By playing these games over and over, GatorSC learns to ignore the noise and find the true, underlying structure of the cells without needing a teacher.

What Did They Find?

The researchers tested GatorSC on 19 different "cities" (datasets) from around the world.

• Better Grouping: It sorted cells into neighborhoods more accurately than any other method.
• Better Guessing: It could fill in missing data (like guessing a job title for someone who forgot to write it down) better than competitors.
• Real-World Application: They used it on data from Alzheimer's disease patients. It successfully identified different brain cell types and found specific "pathways" (like traffic jams in the brain's signaling system) that were broken in the disease, revealing new biological insights.

The Bottom Line

GatorSC is like a master architect who doesn't just look at a city from the sky or the ground, but uses a team of experts to combine all those views into a single, crystal-clear 3D model. It helps scientists see the hidden patterns in messy biological data, leading to better understanding of diseases and how our bodies work.

Technical Summary

1. Problem Statement

Single-cell RNA sequencing (scRNA-seq) data presents significant analytical challenges due to high dimensionality, extreme sparsity (caused by dropout events), and technical noise. While existing methods have made progress, they suffer from three critical limitations regarding information fusion:

• Lack of Multi-Scale Integration: Most methods model either global cell-cell relationships or gene-gene dependencies in isolation. They fail to jointly integrate global population structures, global gene networks, and local, context-specific gene dependencies within a single representation.
• Rigid Fusion Strategies: When multiple graph views are used, they are typically combined via fixed schemes (e.g., concatenation or uniform averaging). These methods cannot adaptively weigh the importance of different structural signals across different datasets or even individual cells.
• Incomplete Self-Supervision: Existing self-supervised objectives often prioritize either graph reconstruction (denoising) or contrastive learning (semantic consistency) but rarely both, limiting the model's ability to simultaneously preserve topology and learn robust, noise-invariant embeddings without labels.

2. Methodology: The GatorSC Framework

GatorSC is a unified representation learning framework designed to address these gaps through hierarchical graph modeling and Mixture-of-Experts (MoE) fusion.

A. Hierarchical Graph Modeling

The framework constructs three distinct graphs from the gene expression matrix ($X$) to capture biological structure at different scales:

1. Global Cell–Cell Graph ($G^{(1)}$): Captures population-level organization. It uses a multi-head attention mechanism to compute pairwise cell similarities, delineating the global cellular architecture.
2. Global Gene–Gene Graph ($G^{(2)}$): Encodes cross-cell functional dependencies. By applying attention to the transposed expression matrix, it models robust gene-gene associations that remain stable across diverse cellular contexts.
3. Local Gene–Gene Graph ($G^{(3)}$): Captures context-specific interactions. Based on the global cell graph, it extracts $k$-hop subgraphs (local neighborhoods) for each cell and recomputes gene-gene similarities within these neighborhoods. This emphasizes fine-grained, context-dependent gene programs.

B. Adaptive Fusion via Mixture-of-Experts (MoE)

Instead of simple concatenation, GatorSC employs an MoE architecture to fuse the embeddings ($Z^{(1)}, Z^{(2)}, Z^{(3)}$) learned from the three graphs:

• Experts: Each graph view is processed by a dedicated "expert" network (a combination of Graph Convolutional Networks and MLPs).
• Gating Network: A learnable gating network dynamically assigns weights to each expert based on the input cell's characteristics. This allows the model to adaptively emphasize the most informative structural signal for specific datasets or cell populations.
• Output: The final representation ($\tilde{Z}$) is a weighted sum of the expert outputs, creating a unified, low-dimensional embedding.

C. Unified Self-Supervised Learning

The model is trained without cell-type labels using a dual objective that combines Graph Reconstruction and Graph Contrastive Learning:

• Graph Augmentation: The original graphs are perturbed via random walk-based edge deletion and addition to create augmented views ($G_{aug}$).
• Reconstruction Loss: The model learns to reconstruct the original adjacency matrices and edge sets from the augmented graphs (using Binary Cross-Entropy), ensuring structural fidelity and denoising capabilities.
• Contrastive Loss: The model maximizes similarity between positive pairs (original node, its augmented counterpart, and its neighbors) and minimizes similarity with negative pairs. This enforces semantic consistency and noise invariance.
• Total Loss: A weighted sum of reconstruction and contrastive losses across all hierarchical graphs, balancing global structural preservation with semantic alignment.

3. Key Contributions

• Multi-Scale Graph Construction: Introduced a novel framework that simultaneously models global cell-cell topology, global gene-gene networks, and local context-specific gene dependencies.
• Adaptive MoE Fusion: Developed a gating mechanism that dynamically fuses heterogeneous graph embeddings, outperforming static fusion strategies.
• Dual Self-Supervision: Integrated generative (reconstruction) and discriminative (contrastive) objectives into a unified training loop, enabling robust learning from noisy, sparse data without supervision.
• Comprehensive Benchmarking: Validated the framework on 19 diverse public scRNA-seq datasets covering various tissues, species, and sequencing platforms.

4. Experimental Results

GatorSC was evaluated on three primary tasks and compared against state-of-the-art deep generative models (e.g., scVI, DeepBID), graph-based methods (e.g., scGNN), and contrastive learning frameworks (e.g., scGPCL).

• Cell Clustering: On 14 benchmark datasets, GatorSC consistently achieved the best or second-best scores in Adjusted Rand Index (ARI), Normalized Mutual Information (NMI), Accuracy (ACC), and Homogeneity. It showed particular superiority on heterogeneous datasets with many cell types (e.g., Global, Pancreas).
• Gene Expression Imputation: Evaluated under simulated dropout rates (10%–50%), GatorSC demonstrated higher Pearson Correlation Coefficients (PCC) and lower L1 errors than competitors. It remained stable even under severe dropout (40–50%), where other methods degraded significantly.
• Cell-Type Annotation: In supervised annotation tasks, GatorSC matched or outperformed specialized classifiers (e.g., scDeepSort, CellTypist) in F1 score, Accuracy, and Matthews Correlation Coefficient (MCC).
• Downstream Applications:
  • Trajectory Inference: Successfully reconstructed the branching structure of the "binary tree 8" dataset, accurately ordering cells along developmental pseudotime.
  • Marker Gene Analysis: Correctly identified canonical marker genes for distinct cell types in the Quake diaphragm dataset.
  • Disease Case Study (Alzheimer's): Applied to human snRNA-seq data, GatorSC segregated brain cell types and revealed cell-type-specific pathway alterations (e.g., oxidative phosphorylation, MAPK signaling suppression in oligodendrocytes) associated with Alzheimer's disease.

5. Significance

GatorSC represents a significant advancement in single-cell analysis by treating the integration of multi-scale biological structures as an adaptive information fusion problem.

• Biological Fidelity: By preserving both global topology and local context, the learned embeddings are more biologically meaningful, facilitating accurate trajectory inference and pathway analysis.
• Robustness: The unified self-supervised objective makes the model highly resilient to the sparsity and noise inherent in scRNA-seq data.
• Flexibility: The framework is modular and can be readily extended to multi-omic and spatial transcriptomics modalities, providing a versatile foundation for future single-cell research.

The source code is publicly available, and the method sets a new state-of-the-art baseline for unsupervised and semi-supervised single-cell transcriptomic analysis.