Tora-PPI

Tora-PPI is a topology-aware framework for robust multilabel PPI prediction. By combining PPR-enhanced residue topology, ego-subgraph interaction context, and representation alignment across sequence, structure, and network views, Tora-PPI improves prediction stability under unseen proteins, dataset shifts, species transfer, structural noise, and mutation-induced perturbations.

Repository Structure

Tora_PPI/
├── train.py                    # Training entry point
├── configs/                    # YAML training configurations
├── data/SHS27k/                # SHS27k input data and precomputed features
├── split_data/                 # Train/test split JSON files
├── model/
│   ├── data/                   # PPI data loading and graph construction
│   ├── features/               # Sequence and molecular/structure preprocessing
│   ├── graph/                  # PPR and subgraph utilities
│   ├── layers/                 # Reusable neural network layers
│   ├── models/                 # Tora-PPI model architecture
│   ├── training/               # Losses, trainer, profiling, helper functions
│   └── utils/                  # Metrics, logging, seed control, split utilities
├── environment.yml             # Conda environment file
└── requirements.txt            # Pip package snapshot

Model Overview

The main model is implemented in:

model/models/gnn_models.py

The training pipeline uses:

• residue-level graph features from FSP_residue_Embedding.pt
• protein-level sequence features from FSP_sequence_Embedding.pt
• PPR-enhanced residue graph edges from edge_index_ppr.npy and edge_attr_ppr.npy
• PPI train/test split files from split_data/
• a multi-label BCE loss with auxiliary InfoNCE and VICReg losses

The default prediction target contains seven interaction types:

reaction, binding, ptmod, activation, inhibition, catalysis, expression

Data

The SHS27k data used by the provided configs is stored under:

data/SHS27k/

Expected files:

protein.actions.SHS27k.txt
protein.SHS27k.sequences.dictionary.tsv
vec5_CTC.txt
FSP_residue_Embedding.pt
FSP_sequence_Embedding.pt
edge_index_ppr.npy
edge_attr_ppr.npy

Large .pt and .npy files are tracked with Git LFS. After cloning the repository, run:

git lfs install
git lfs pull

Environment

Create the conda environment:

conda env create -f environment.yml
conda activate Tora-PPI

If your environment name is different, activate the corresponding environment before running the training script.

Training

Run training from the repository root:

cd /path/to/Tora_PPI
python train.py --config configs/shs27k_random_seed9.yaml

Command-line arguments after --config override YAML values. For example:

python train.py \
  --config configs/shs27k_random_seed9.yaml \
  --cuda 0 \
  --epoch_num 500

Outputs

Training outputs are written to:

results/SHS27k/<split>/seed_<seed>_run_id_<run_id>_<timestamp>/

Typical files:

config.txt
valid_results.txt
gnn_model_train.ckpt
gnn_model_valid_best.ckpt
gnn_model_final.ckpt

Notes

• Run commands from the repository root so that relative paths in YAML configs resolve correctly.
• If a config uses a split file whose seed does not match seed_num, train.py raises an error to prevent accidental mixed-seed experiments.
• Large data files require Git LFS.