GraphSeqLM: A Unified Graph–Language Framework for Omic Graph Learning

---

Overview

GraphSeqLM is a unified framework that fuses graph neural networks (GNNs) with biological sequence embeddings to learn from multi‑omic graphs at scale. It augments topological signals with LLM‑derived embeddings of DNA, RNA, and proteins, enabling richer node/edge semantics for sample‑specific analyses of signaling pathways and protein–protein interaction networks.

---

Contents

• Overview
• Environment Setup
• Data Preprocessing
• Training
• License

---

Environment Setup

Tested with Python 3.10 and PyTorch CUDA 12.1 wheels.

# Create environment
conda create --name mkg python=3.10
conda activate mkg

# PyTorch (CUDA 12.1)
pip3 install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu121

# PyTorch Geometric core
pip install torch_geometric

# Optional optimizations (match your torch/cu version)
pip install pyg_lib torch_scatter torch_sparse torch_cluster torch_spline_conv   -f https://data.pyg.org/whl/torch-2.2.0+cu121.html

# Transformer backbones for sequence embeddings
pip install transformers

---

Data Preprocessing

Note: This repository uses(https://git-lfs.com/) for large data files. After cloning, please run git lfs pull to download the full data files.

cd data
python processed_data_gen.py

This script prepares graph structures and attaches sequence‑derived features used by the GraphSeqLM encoder.

---

Training

Note: This framework requires a pre-trained DNAGPT model. Please see the instructions in models/finetune_llm/checkpoints/README.md to manually download dna_gpt0.1b_h.pth and place it in the checkpoints directory before running.

python main-graphseqlm-gpt.py

Key flags (see the script for full options):

• --dataset: dataset identifier
• --task: task name (e.g., classification/regression)
• --epochs, --lr, --batch_size: training hyperparameters

---

Citation

If this repository is useful in your research, please consider citing the following related work:

@inproceedings{zhang2025graphseqlm,
  title={GraphSeqLM: A Unified Graph Language Framework for Omic Graph Learning},
  author={Zhang, Heming and Huang, Di and Chen, Yixin and Li, Fuhai},
  booktitle={Companion Proceedings of the ACM on Web Conference 2025},
  pages={1510--1513},
  year={2025}
}

You may also wish to cite the GraphSeqLM paper (see the PDF linked above).

License

MIT License. See LICENSE for details.