# ASHL-DTI

**Repository Path**: Tomhappy/ASHL-DTI

## Basic Information

- **Project Name**: ASHL-DTI
- **Description**: No description available
- **Primary Language**: Unknown
- **License**: Not specified
- **Default Branch**: main
- **Homepage**: None
- **GVP Project**: No

## Statistics

- **Stars**: 0
- **Forks**: 0
- **Created**: 2026-08-08
- **Last Updated**: 2026-08-08

## Categories & Tags

**Categories**: Uncategorized

**Tags**: None

## README

# ASHL-DTI
**Adaptive Feature Selection with Hierarchical Learning for Drug-Target Interaction Prediction**

[![Python 3.8](https://img.shields.io/badge/python-3.8-blue.svg)](https://www.python.org/downloads/release/python-380/)
[![PyTorch 2.2](https://img.shields.io/badge/PyTorch-2.2.1-EE4C2C.svg?logo=pytorch)](https://pytorch.org/)
[![DGL 1.0.2](https://img.shields.io/badge/DGL-1.0.2-12B522.svg)](https://www.dgl.ai/)
[![License](https://img.shields.io/badge/License-MIT-green.svg)](https://opensource.org/licenses/MIT)

---

## 📌 Table of Contents
- [📖 Introduction](#-introduction)
- [🛠️ Setup](#%EF%B8%8F-setup)
- [📂 Data Preparation](#-data-preparation)
- [🧬 Data Preprocessing](#-data-preprocessing)
- [⚙️ Configuration](#%EF%B8%8F-configuration)
- [🚀 Run](#-run)
- [📝 Citation](#-citation)
- [📚 References](#-references)

---

## 📖 Introduction
**ASHL-DTI** is a novel framework that integrates **hierarchical learning** with **adaptive feature selection** to significantly boost feature quality and model generalizability. Specifically, it captures multi-level intramolecular associations to learn discriminative representations while employing an **adaptive Top-k selection mechanism** to retain the most predictive features. Extensive experiments demonstrate that ASHL-DTI outperforms state-of-the-art approaches on public benchmarks and exhibits strong generalization in predicting novel drug-target pairs.

## 🛠️ Setup

### 1. Environment Requirements
The project is implemented using Python 3.8. We recommend using a virtual environment (Conda) to manage dependencies.

```bash
conda create -n ashldti python=3.8
conda activate ashldti
```

### 2. Install Dependencies

Install the required packages using pip:

```bash
# Core DL Frameworks
pip install torch==2.2.1 torchvision==0.17.1 torchaudio==2.2.1
pip install dgl==1.0.2 dgllife==0.3.2

# Data & Utils
pip install numpy==1.24.4 pandas==2.0.3 scipy==1.10.1 scikit-learn==1.3.2
pip install rdkit==2024.3.5 einops==0.8.0 tqdm==4.67.1 yacs==0.1.8 zipp==3.20.2
```

## 📂 Data Preparation

The datasets folder contains all experimental data utilized in **ASHL-DTI**, including **BindingDB** [1], **BioSNAP** [2], and **Human** [3].

#### 1. Directory Structure

Please organize the datasets within the `datasets/` directory as follows.

```bash
datasets/
├── bindingdb/
│   ├── train.csv
│   ├── val.csv
│   └── test.csv
├── biosnap/
│   ├── train.csv
│   ├── val.csv
│   └── test.csv
└── human/
    ├── train.csv
    ├── val.csv
    └── test.csv
```

#### 2. Data Splitting

For all three datasets, the samples are randomly split into **Training**, **Validation**, and **Testing** sets with a fixed ratio.

| **Dataset** | **Train Split** | Val Split | Test Split |
| :---------: | :-------------: | :-------: | :--------: |
|  **Ratio**  |     **70%**     |  **10%**  |  **20%**   |

#### 3. File Format

Each `.csv` file contains the following three columns:

| **Column**  | **Description**                                         |
| :---------: | ------------------------------------------------------- |
| **SMILES**  | Molecular structure representation of the drug (string) |
| **Protein** | Amino acid sequence of the target protein (string)      |
|    **Y**    | Interaction Label (1 for positive, 0 for negative)      |

**Example:**

```bash
SMILES,Protein,Y
CC1=C(C=C(C=C1)NC(=O)...,MALIPDLAMET...,1
```

## 🧬 Data Preprocessing

The raw data is processed into graph and sequence formats suitable for the model. The preprocessing pipeline includes **Molecular Graph Construction** and **Protein Sequence Encoding**.

#### 1. Drug Molecule Representation

We utilize `dgllife` to convert SMILES strings into molecular graphs with the following enhancements:

- **Featurization:**
  - **Atoms:** `CanonicalAtomFeaturizer` (74 dimensions).
  - **Bonds:** `CanonicalBondFeaturizer` (with self-loops).
- **Virtual Node Padding:**
  - To handle variable molecular sizes efficiently, we align all molecular graphs to a fixed size of **290 nodes** (`max_drug_nodes`).
  - **Virtual Nodes** are added to graphs with fewer atoms.
  - **Indicator Bit:** A binary channel is appended to the node features to distinguish between real atoms (`0`) and virtual nodes (`1`).
  - **Final Dimension:** The resulting node feature dimension is **75** .

#### 2. Target Protein Representation

- **Integer Labeling:** Protein sequences are mapped to integer sequences based on a comprehensive amino acid vocabulary.
- **Truncation/Padding:** All sequences are normalized to a fixed length of **1200** (`max_protein_length`) to facilitate batch processing.

## ⚙️ Configuration

Model hyperparameters and training settings are defined in YAML files located in the `configs/` directory.

#### 1. Config Files

We provide default configurations for each dataset:

```bash
configs/
├── bindingdb.yaml
├── biosnap.yaml
└── human.yaml
```

#### 2. Parameter Details

You can modify the hyperparameters in the `.yaml` files. Below is an explanation of the key parameters:

```xml
Drug:
  Nodes: 290          # Maximum number of nodes in the drug graph (padding size)

Protein:
  Length: 1200        # Maximum length of the input protein sequence
  CNN_Length: 1185    # Effective sequence length after CNN processing

fusion:
  Hidden_Dim: 256     # Dimension of the hidden layers
  Num_Heads: 8        # Number of attention heads in the fusion module
  Dropout_Rate: 0.3   # Dropout rate to prevent overfitting

Global:
  Batch_Size: 32      # Batch size for training
  Epoch: 100          # Total number of training epochs
  LR: 5e-5            # Initial learning rate
```

## 🚀 Run

To train and evaluate the model, simply run the `main.py` script with the corresponding configuration file.

#### For Human Dataset

```bash
python main.py --cfg ./configs/human.yaml --outname human_model --data human --num_worker 0
```

#### For BioSNAP Dataset

```bash
python main.py --cfg ./configs/biosnap.yaml --outname biosnap_model --data biosnap --num_worker 0
```

#### For BindingDB Dataset

```bash
python main.py --cfg ./configs/bindingdb.yaml --outname bindingdb_model --data bindingdb --num_worker 0
```

## 📝 Citation
```bash
@ARTICLE{11457244,
  author={Tian, Zhen and Jiang, Miao and Li, Jin and Zhang, Wenjie and Xu, Mingliang},
  journal={IEEE Journal of Biomedical and Health Informatics}, 
  title={Adaptive Feature Selection With Hierarchical Learning for Drug-Target Interaction Prediction}, 
  year={2026},
  volume={},
  number={},
  pages={1-12},
  keywords={Feeds;Filtering;Contacts;Filters;Circuits and systems;Electronic mail;LoRa;Protocols;Communication systems;Wide area networks;Drug-target interaction;deep learning;hierarchical learning;adaptive selection},
  doi={10.1109/JBHI.2026.3678266}}
```

## 📚 References

- [1] Bai, P.; Miljković, F.; Ge, Y.; Greene, N.; John, B.; Lu, H. Hierarchical Clustering Split for Low-Bias Evaluation of Drug-Target Interaction Prediction. *2021 IEEE International Conference on Bioinformatics and Biomedicine (BIBM)*. 2021; pp 641–644.
- [2] Zitnik, M.; Sosic, R.; Leskovec, J. BioSNAP Datasets: Stanford Biomedical Network Dataset Collection. http://snap.stanford.edu/biodata 2018, 5.

- [3] Liu, H.; Sun, J.; Guan, J.; Zheng, J.; Zhou, S. Improving Compound–Protein Interaction Prediction by Building up Highly Credible Negative Samples. *Bioinformatics* 2015, 31, i221–i229.