Abstract
GPCR proteins belong to diverse families defined at multiple hierarchical levels. Inspecting relationships between GPCR proteins on this hierarchical structure is critical, as protein characteristics can be inferred from similar hierarchical information. Traditionally, modeling of GPCR families has been performed separately for each level (family, subfamily, and sub-subfamily), ignoring inter-level relationships. We propose DeepHier, a deep learning model that simultaneously learns representations of the GPCR family hierarchy from protein sequences within a unified model. By introducing a novel loss term based on metric learning, we incorporate hierarchical relations directly into the embedding space. Testing on a public GPCR dataset demonstrates that metric distances in the deep feature space correspond to hierarchical relations. Furthermore, we show that downstream tasks like phylogenetic reconstruction and motif discovery are feasible in the constructed embedding space, capturing both technical and biological aspects of the hierarchy.
Introduction
G protein-coupled receptors (GPCRs) are the largest family of transmembrane proteins and serve as primary targets for drug discovery. They are organized in a strict hierarchy based on sequence similarity and evolutionary history. Traditionally, researchers have modeled these levels using separate, disconnected models or computationally expensive pairwise sequence alignments. These approaches often fail to capture the continuous relationships across the entire hierarchy. Furthermore, as protein databases grow, traditional alignment-based tools like BLAST struggle with computational overhead. DeepHier addresses these challenges by learning a single, unified representation of GPCRs that respects this hierarchy while remaining computationally efficient.
Methodology
DeepHier builds upon a 1D-CNN architecture, introducing two critical innovations to handle hierarchical data:
- Multi-branch Classifier: Instead of a single output, the model features three distinct Multilayer Perceptron (MLP) branches, each dedicated to one level of the GPCR hierarchy (family, subfamily, and sub-subfamily). This architecture forces the shared embedding layer to learn features relevant to all levels simultaneously.
- Metric Learning with Center Loss: Beyond standard cross-entropy loss, which focuses on class separation, DeepHier implements center loss. This minimizes the distance between a protein's embedding and the mean vector (center) of its class, creating compact clusters where Euclidean distance directly correlates to biological similarity.
The model is trained using a three-phase curriculum, starting with broad family-level features and progressively focusing on fine-grained sub-subfamily details.
Datasets
The study utilized the BIAS-PROFS GPCR dataset, containing sequence labels across three hierarchical levels. After filtering for classes with at least 10 sequences, the final dataset consisted of 8,222 sequences across 86 sub-subfamilies. Sequences were one-hot encoded and padded to a fixed length of 1,000 amino acids.
Hierarchical Metric Embedding
By integrating center loss, DeepHier transforms high-dimensional sequence data into a unified metric space. In this space, the distance between two points represents evolutionary and functional proximity, allowing for alignment-free biological analysis such as phylogenetic reconstruction and motif discovery.
Results
DeepHier achieved significantly higher Silhouette and Adjusted Mutual Information (AMI) scores than baselines, proving that its embedding space accurately reflects the biological hierarchy. Key findings include:
- Phylogenetic Accuracy: Trees reconstructed from embeddings accurately grouped proteins by family and subfamily, matching established evolutionary models.
- Motif Discovery: The model successfully identified biological markers, such as the DRY and NSY motifs characteristic of Rhodopsin-like GPCRs.
- Search Performance: DeepHier performed similarity searches in ~19ms per query, compared to 266ms for BLAST. Crucially, DeepHier's execution time remained nearly constant as the database size increased, whereas BLAST's time increased linearly.
- Classification Accuracy: At the sub-subfamily level, DeepHier achieved 79.7% accuracy, outperforming BLAST (77.7%).
Conclusion
DeepHier demonstrates that hierarchical relations between sequences can be successfully captured in a unified embedding space. While currently focused on sequence data, future work could incorporate 3D structural data and expand the training set to include non-GPCR sequences. Although initial database embedding is a one-time computational hurdle, the resulting constant-time scalability makes it a powerful alternative for large-scale biological databases.
Summary updated