MoE-DTI

KO | EN

gitlite — search

#javascript #python #react #ai #typescript #go #golang #deep-learning #docker #java #nodejs #kubernetes #mcp #ios

(★ 81)

No description available.

MoE-DTI Latest Version Download

Download Latest Version (.zip)

// repository documentation

Was this content helpful?

★ 0 (0 ratings)

# DTI-MoE
### 蛋白质-药物分子靶标相互作用预测（多模态 MoE）

[![English](https://img.shields.io/badge/README-English-2563EB?style=for-the-badge)](https://raw.githubusercontent.com/ZBK211/MoE-DTI/main/README_EN.md)
[![中文](https://img.shields.io/badge/README-中文-16A34A?style=for-the-badge)](https://raw.githubusercontent.com/ZBK211/MoE-DTI/main/README.md)

</div>

---

## 项目简介

DTI-MoE 是我们大创项目的开源实现，围绕 DTI（Drug-Target Interaction）任务构建可复现、可演示、可扩展的工程化方案。

融合三类核心模态：

- 分子图结构表征（SMILES -> GNN embedding）
- 蛋白 3D 结构表征（PDB -> residue graph embedding）
- 蛋白序列表征（ESM2 embedding）

联合优化两个任务：

- Affinity regression
- Interaction classification

---

## 功能矩阵

| 模块 | 能力 | 说明 |
|---|---|---|
| Molecular Encoder | SMILES -> 图表征 | 基于 GNN 的分子表示学习 |
| Protein Encoder | PDB -> 残基图表征 | 引入蛋白 3D 拓扑信息 |
| Sequence Branch | ESM2 序列表征融合 | 与 3D 表征进行多模态融合 |
| DTI-MoE Core | 稀疏专家路由 | 增强跨样本表达能力 |
| Multi-task Head | 回归 + 分类联合学习 | 同时输出亲和力与相互作用概率 |
| Evaluation Suite | 指标与可视化 | 散点图、ROC/PR、路由分布、t-SNE |
| Streamlit App | 交互式预测 | 输入 SMILES + PDB + Protein ID 即可推理 |

---

## 架构流程

```mermaid
flowchart LR
    A[SMILES] --> B[Drug Graph Encoder]
    C[PDB Structure] --> D[Protein 3D Encoder]
    E[ESM2 Embedding] --> F[Sequence Feature]
    D --> G[Protein Fusion]
    F --> G
    B --> H[DTI-MoE Backbone]
    G --> H
    H --> I[Affinity Regression Head]
    H --> J[Interaction Classification Head]
```

---

## 快速开始

```bash
pip install -r requirements.txt
python run_moe.py --mode quick_test
streamlit run app.py
```

---

## 训练与评估

```bash
# Train
python run_moe.py --mode train --dataset KIBA
python run_moe.py --mode train --dataset DAVIS

# Test
python run_moe.py --mode test --dataset KIBA --checkpoint <checkpoint_path>
python run_moe.py --mode test --dataset DAVIS --checkpoint <checkpoint_path>

# Visualization
python eval_visualize.py
```

---

## Benchmark 结果

| Dataset | RMSE | CI | AUC | AUPR |
|---|---:|---:|---:|---:|
| KIBA | 0.4117 | 0.8866 | 0.9341 | 0.8332 |
| DAVIS | 0.4775 | 0.8830 | 0.9329 | 0.8898 |

---

## 可视化画廊

| KIBA Training | Joint Training |
|---|---|
| ![KIBA training](https://raw.githubusercontent.com/ZBK211/MoE-DTI/main/figures/1_kiba_training_curves.png) | ![Joint training](https://raw.githubusercontent.com/ZBK211/MoE-DTI/main/figures/2_joint_training_curves.png) |

| Expert Routing | Feature t-SNE |
|---|---|
| ![Expert routing](https://raw.githubusercontent.com/ZBK211/MoE-DTI/main/figures/5_expert_routing.png) | ![TSNE](https://raw.githubusercontent.com/ZBK211/MoE-DTI/main/figures/6_tsne_features.png) |

---

## 开源协作

- 贡献指南：[CONTRIBUTING.md](CONTRIBUTING.md)
- 行为准则：[CODE_OF_CONDUCT.md](CODE_OF_CONDUCT.md)
- 版本记录：[CHANGELOG.md](CHANGELOG.md)
- 引用信息：[CITATION.cff](CITATION.cff)

## License

MIT License. See [LICENSE](LICENSE).

MoE-DTI

MoE-DTI Latest Version Download

Today's picks

Was this content helpful?

Recent Feedback

MoE-DTI

MoE-DTI Latest Version Download

Today's picks

Was this content helpful?

Recent Feedback

Download README