📖 English | 简体中文

LLM-GNN Collaborative Framework for Synthetic Lethality Prediction

Complementary Information Channels for Synthetic Lethality Prediction: Augmenting Graph Topology with Protein Language Models and LLM Weak Supervision

Tao Zhu — ¹ Department of Obstetrics and Gynecology, National Clinical Research Center for Obstetrics and Gynecology, Tongji Hospital, Tongji Medical College, Huazhong University of Science and Technology, Wuhan, China; ² Key Laboratory of Cancer Invasion and Metastasis (Ministry of Education), Hubei Key Laboratory of Tumor Invasion and Metastasis, Tongji Hospital, Tongji Medical College, Huazhong University of Science and Technology, Wuhan, China

---

Overview

This repository contains the complete code, data, and trained models for our LLM-GNN collaborative framework for synthetic lethality (SL) prediction. The framework integrates three complementary information channels:

1. LLM Weak Supervision — DeepSeek-V4-Pro NER on 1,844 PubMed SL abstracts
2. GATv2 Graph Neural Network — Message passing on STRING v12.0 PPI network (15,956 nodes, 932K edges)
3. ESM-2 Protein Language Model — Sequence-level protein priors via esm2_t12_35M

Key Results:

• 5-fold CV AUC: 0.862 (ESM-2) / 0.693 (base)
• Cold-start AUPRC: 0.84 (CV3 gene-level split)
• DepMap CRISPR validation: 1.73× enrichment (Fisher's exact p=0.041)
• Two methodological contributions: SL adjacency leakage quantification (13.7%) + zero-vector ablation bias correction (14×)

---

Repository Structure

SL_LLM_GNN/
├── README.md                    ← This file
├── requirements.txt             ← Python dependencies
├── .gitignore
│
├── code/                        ← All analysis scripts
│   ├── config.py                ← Shared paths and constants
│   ├── 01_pubmed_search.py      ← PubMed API search
│   ├── 02_deepseek_ner.py       ← LLM gene entity extraction
│   ├── 03_ppi_network.py        ← STRING PPI download & processing
│   ├── 04_synlethdb.py          ← SynLethDB label processing
│   ├── 05_build_graph.py        ← Graph construction (PPI + NER + labels)
│   ├── 06_train_gnn.py          ← Main GNN training (5-fold CV)
│   ├── 07_ablation.py           ← Ablation experiments (shuffled-feature)
│   ├── 08_cv3_benchmark.py      ← Feng-benchmark CV1/CV3 evaluation
│   └── 09_esm2_train.py         ← ESM-2 enhanced model training
│
├── data/processed/              ← Processed data (included in repo)
│   ├── graph_data.pkl           ← Constructed PyG graph (23 MB)
│   ├── ner_results.jsonl        ← LLM NER output (2,846 genes)
│   ├── predicted_sl_final.json  ← Top-500 SL predictions
│   ├── depmap_full_validation.json ← DepMap validation results
│   ├── cv3_results.json         ← CV1/CV3 benchmark results
│   ├── slmgae_head_to_head.json ← SLMGAE head-to-head comparison
│   ├── ablation_extended.json   ← Complete ablation data
│   └── ppm1d_tp53_lineage.json  ← PPM1D-TP53 lineage stratification
│
├── results/models/              ← Trained model weights
│   ├── fold1_best.pt ~ fold5_best.pt
│   └── esm_best.pt
│
├── figures/                     ← Publication-quality figures
│   ├── Fig1-5.pdf               ← Main text figures
│   └── FigS1-S3.pdf             ← Supplementary figures
│
├── paper/                       ← Compiled submission PDFs
│   ├── main.pdf
│   ├── supplementary.pdf
│   └── cover_letter.pdf
│
└── latex/                       ← LaTeX source files
    ├── main.tex
    ├── supplementary.tex
    └── cover_letter.tex

---

Reproducibility

Environment Setup

conda create -n sl_gnn python=3.10
conda activate sl_gnn
pip install -r requirements.txt

Data Preparation

External data must be downloaded from public sources:

Source	URL	File
STRING v12.0	https://string-db.org	string_human_v12.0_links.txt.gz, string_human_v12.0_aliases.txt.gz
SynLethDB 3.0	https://synlethdb.sist.shanghaitech.edu.cn	Human.SL.detailed.tsv, Human.non.SL.detailed.tsv
DepMap 26Q1	https://depmap.org	CRISPRGeneEffect.csv
ESM-2	HuggingFace facebook/esm2_t12_35M_UR50D	Auto-downloaded by transformers

Place external files in data/external/. Then run scripts in numerical order:

# Phase 1: Data Collection
python code/01_pubmed_search.py        # → data/raw/
python code/02_deepseek_ner.py         # → data/processed/ner_results.jsonl
python code/03_ppi_network.py          # → data/processed/ppi_string_edges.tsv
python code/04_synlethdb.py            # → data/processed/sl_labels.json

# Phase 2: Graph Construction
python code/05_build_graph.py          # → data/processed/graph_data.pkl

# Phase 3: Training & Evaluation
python code/06_train_gnn.py            # → results/models/fold*_best.pt
python code/07_ablation.py             # → ablation results
python code/08_cv3_benchmark.py        # → CV1/CV3 results
python code/09_esm2_train.py           # → ESM-2 enhanced model

Hardware Requirements

• Consumer GPU (RTX 4090 or Apple M-series MPS) — 5-fold CV completes in <10 minutes
• ~8 GB GPU memory for ESM-2 model
• ~30 GB disk for all external data

---

Citation

If you use this code or data, please cite:

Zhu T. Complementary Information Channels for Synthetic Lethality Prediction:
Augmenting Graph Topology with Protein Language Models and LLM Weak Supervision.
Briefings in Bioinformatics, 2026.

---

License

MIT License. See repository for details.

Contact

Tao Zhu — zhutao@tjh.tjmu.edu.cn Department of Obstetrics and Gynecology, Tongji Hospital, HUST, Wuhan, China