ProtHGT Knowledge Graph Data & Pretrained Checkpoints

This repository provides the knowledge graph (KG) .pt files and pretrained model checkpoints used in ProtHGT: Heterogeneous Graph Transformers for Automated Protein Function Prediction Using Knowledge Graphs and Language Models.

• Code (training & prediction): https://github.com/HUBioDataLab/ProtHGT

---

What’s Inside

Files are organized by task: go/ for GO term prediction and ec/ for EC number prediction, under both knowledge_graphs/ and models/. In each task, the top-level files use the default TAPE protein embeddings, and alternative_protein_embeddings/ holds versions whose protein node features use ESM-2 or ProtT5 instead.

knowledge_graphs/

PyTorch Geometric-compatible KG files: a full KG plus train/validation/test splits per task.

GO term prediction — knowledge_graphs/go/

├── prothgt-kg.pt                      # The default full knowledge graph (TAPE embeddings).
├── prothgt-train-graph.pt             # Training set (80% of the full KG).
├── prothgt-val-graph.pt               # Validation set (10%).
├── prothgt-test-graph.pt              # Test set (10%).
└── alternative_protein_embeddings/    # ESM-2 / ProtT5 variants.
    ├── esm2/   prothgt-esm2-{kg,train-graph,val-graph,test-graph}.pt
    └── prott5/ prothgt-prott5-{kg,train-graph,val-graph,test-graph}.pt

EC number prediction — knowledge_graphs/ec/

├── prothgt-ec-kg.pt                   # The default full EC knowledge graph (TAPE embeddings).
├── prothgt-ec-train-graph.pt          # Training set.
├── prothgt-ec-val-graph.pt            # Validation set.
├── prothgt-ec-test-graph.pt           # Test set (transductive: training proteins' EC edges injected).
└── alternative_protein_embeddings/    # ESM-2 / ProtT5 variants.
    ├── esm2/   prothgt-ec-esm2-{kg,train-graph,val-graph,test-graph}.pt
    └── prott5/ prothgt-ec-prott5-{kg,train-graph,val-graph,test-graph}.pt

models/

Pretrained ProtHGT checkpoints (.pt). Use a checkpoint that matches the KG embedding variant you are using.

GO term prediction — models/go/ (one model per GO sub-ontology × embedding)

├── prothgt-model-molecular-function.pt      # default TAPE-based KG
├── prothgt-model-biological-process.pt
├── prothgt-model-cellular-component.pt
└── alternative_protein_embeddings/          # ESM-2 / ProtT5 variants.
    ├── esm2/   prothgt-esm2-model-{molecular-function,biological-process,cellular-component}.pt
    └── prott5/ prothgt-prott5-model-{molecular-function,biological-process,cellular-component}.pt

EC number prediction — models/ec/ (one joint hierarchical model per embedding, predicting all four EC levels)

├── prothgt-ec-model.pt                      # default TAPE-based KG
└── alternative_protein_embeddings/          # ESM-2 / ProtT5 variants.
    ├── esm2/   prothgt-ec-esm2-model.pt
    └── prott5/ prothgt-ec-prott5-model.pt

---

How to Use (Training & Prediction)

To train or run inference, follow the instructions in the GitHub repository: https://github.com/HUBioDataLab/ProtHGT

Key scripts:

• train.py — trains ProtHGT using the provided KG splits
• predict.py — runs inference using pretrained checkpoints

GO term prediction lives at the top level of src/; EC number prediction is in src/ec/.

---

Citation

Please refer to our preprint for more information. If you use the ProtHGT method or the datasets provided in this repository, please cite this paper: Ulusoy, E., & Dogan, T. (2025). ProtHGT: Heterogeneous Graph Transformers for Automated Protein Function Prediction Using Biological Knowledge Graphs and Language Models (p. 2025.04.19.649272). bioRxiv. Link

---

Licensing

Copyright (C) 2025 HUBioDataLab

This dataset is released under GPL-3.0.