Abstract
Spatial context is central to understanding health and disease. Yet reference protein interaction networks lack such contextualization, thereby limiting the study of where protein interactions likely occur in the human body and how they may be altered in disease. Contextualized protein interactions could better characterize genes with disease-specific interactions and elucidate diseases' manifestation in specific cell types. Here, we introduce AWARE, a graph neural message passing approach to inject cellular and tissue context into protein embeddings. AWARE optimizes for a multi-scale embedding space, whose structure reflects network topology at a single-cell resolution. We construct a multi-scale network of the Human Cell Atlas and apply AWARE to learn protein, cell type, and tissue embeddings that uphold cell type and tissue hierarchies. We demonstrate AWARE's utility on the novel task of predicting whether a protein is altered in disease and where that association most likely manifests in the human body. To this end, AWARE outperforms generic embeddings without contextual information by at least 12.5%, showing AWARE's potential to reveal context-dependent roles of proteins in disease.
Introduction: The Contextual Challenge
Proteins do not function in isolation; their interactions are highly dependent on the specific cell type, tissue environment, and disease state. However, standard protein-protein interaction (PPI) networks are typically presented as static, "one-size-fits-all" maps. While previous research has attempted to build cell-type-specific subnetworks, these efforts are often descriptive rather than predictive. They fail to generate the mathematical representations (embeddings) required for advanced machine learning tasks that can distinguish, for instance, a protein's critical role in a lung T-cell versus its irrelevance in a brain neuron.
The AWARE Framework
To address this, we developed AWARE, a multi-scale graph neural network (GNN) designed to capture the hierarchical nature of biology (Proteins $\rightarrow$ Cells $\rightarrow$ Tissues). The framework operates through several key stages:
- Subnetwork Construction: Utilizing single-cell RNA sequencing (scRNA-seq) data, AWARE identifies essential genes for specific cell types to build contextualized subnetworks.
- Meta-Graph Integration: A high-level graph is constructed where nodes represent cell types and tissues, connected by biological relationships (e.g., a specific cell type residing within a specific tissue).
- Hierarchical Pooling: The core innovation lies in up-pooling, which passes information from individual proteins to the cell-type level, and down-pooling, which uses tissue context to refine protein embeddings. This ensures that a protein's representation reflects the identity of its host cell.
- Semantic Attention: AWARE employs attention mechanisms to weigh different relationship types, such as protein-to-protein versus cell-to-tissue interactions.
Datasets
The model was trained and evaluated using a global PPI network of ~207k edges, scRNA-seq data from the Human Cell Atlas (HCA) defining 64 cell-type subnetworks, and tissue hierarchies from the BRENDA Tissue Ontology. Evaluation relied on the SC2disease dataset for single-cell disease-gene associations.
Results: Predictive Power and Biological Organization
AWARE demonstrates that biological hierarchy is a feature, not a limitation. By explicitly modeling cells within tissues, the framework captures constraints that "flat" models overlook.
- Superior Performance: AWARE outperformed generic, non-contextual embeddings by at least 12.5% in predicting disease-gene associations.
- Disease-Specific Gains: Significant improvements were noted in specific contexts, such as a 200% increase in accuracy for predicting Rheumatoid Arthritis-associated genes in monocytes and a 50% increase for Breast Cancer in T-cells.
- Meaningful Clustering: Unlike standard models that produce disorganized "hairballs," UMAP visualizations show that AWARE successfully organizes proteins in the embedding space according to their tissue and cell-type hierarchy.
Conclusion and Future Directions
While AWARE shows immense promise, it is currently limited by the sparsity of high-quality single-cell disease-gene association data and the computational complexity of training across dozens of subnetworks. Future work will focus on using AWARE to identify novel cell-type biomarkers—proteins the model identifies as critical to a cell's identity that have yet to be experimentally verified. By treating context as a regularizer, AWARE forces machine learning models to respect biological boundaries, paving the way for more precise preventative medicine and therapeutic discovery.
Summary updated