Abstract
Applied to human bone and mouse embryonic data, Cell Decoder reveals the multi-scale heterogeneity of cell identities, providing a powerful framework for advancing our understanding of cellular diversity.
Introduction
Single-cell RNA sequencing (scRNA-seq) has transformed our ability to observe biological diversity at the cellular level. However, translating raw gene expression data into accurate cell-type annotations remains a significant bottleneck. Traditional methods are often manual and labor-intensive, relying on marker genes that may be biased by researcher expertise. Conversely, modern deep learning models, while highly accurate, often function as "black boxes," lacking the biological interpretability required for scientific discovery. Furthermore, these models frequently struggle with batch effects and data shifts across different experimental conditions.
To address these challenges, we present Cell Decoder, a novel hierarchical graph neural network (GNN) that identifies cell types by integrating multi-scale biological knowledge—from protein-protein interactions to complex metabolic pathways. By embedding known biological relationships directly into the model architecture, Cell Decoder achieves state-of-the-art accuracy while providing human-readable explanations for its predictions.
The Cell Decoder Framework
Cell Decoder utilizes a hierarchical GNN to simulate the functional organization of biological systems. The model incorporates three distinct layers of prior knowledge:
- Gene Level: Protein-protein interaction (PPI) networks sourced from the STRING database.
- Pathway Level: Gene-to-pathway mappings from MSigDB and Reactome.
- Biological Process (BP) Level: Hierarchical relationships between pathways and broader biological functions.
The architecture employs intra-scale layers to share information within a specific level and inter-scale layers to aggregate information upward through the hierarchy. To optimize the model for specific datasets, Automated Machine Learning (AutoML) is used to select the most effective graph operations, such as Graph Attention Networks (GAT). Explainability is a core feature, achieved through a hierarchical version of Gradient-weighted Class Activation Mapping (Grad-CAM), which allows researchers to identify the specific biological pathways driving a classification.
Experimental Evaluation and Results
Cell Decoder was benchmarked across seven diverse human and mouse datasets, including complex developmental and multi-batch immune cell data.
Superior Accuracy and Robustness
Cell Decoder achieved an average accuracy of 0.87, outperforming nine popular baselines, including SingleR (0.84) and Seurat v5 (0.79). On the Macro F1 score, which accounts for performance on rare cell types, it scored 0.81, significantly higher than its closest competitors.
A key advantage of this approach is its robustness to noise. In feature perturbation tests, Cell Decoder maintained high performance even as data quality was intentionally degraded, whereas traditional deep learning models experienced sharp declines. This resilience is attributed to the use of biological prior knowledge as a regularizer, forcing the model to adhere to established biological structures.
Biological Discovery and Integration
Beyond classification, Cell Decoder demonstrated the ability to refine existing annotations. In a human bone dataset, it correctly identified "Erythrocytes" that had been previously mislabeled as "Progenitor Cells." Additionally, the model can identify "novel" cell types not present in the training data with a 0.94 recall rate. By mapping data to a common biological space (pathways and processes), Cell Decoder also facilitates effective batch correction, aligning datasets from different laboratories without requiring explicit batch labels.
Conclusion
Cell Decoder represents a significant advancement in explainable AI for genomics. While its performance is currently dependent on the quality of available biological maps, it offers a robust and interpretable framework for cell identity decoding. Future work will focus on multi-omics integration and predicting cellular responses to genetic perturbations, moving the field closer to predictive biological engineering.