# DeepLMI

**Repository Path**: Tomhappy/DeepLMI

## Basic Information

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

## Statistics

- **Stars**: 0
- **Forks**: 0
- **Created**: 2026-10-09
- **Last Updated**: 2026-10-09

## Categories & Tags

**Categories**: Uncategorized

**Tags**: None

## README

# DeepLMI

DeepLMI: Deep Feature Mining with a Globally Enhanced Graph Convolutional Network for Robust lncRNA–miRNA Interaction Prediction

---

## Directory Structure

```
DeepLMI/
├─ cold_data_splits/          # Cold-start fold splits (lncRNA/miRNA/both)
├─ dataset/                   # Cross-validation dataset
│  ├─ lncrna.fasta, mirna.fasta, mirna_id.fasta
│  ├─ mirna_lncrna_interaction.csv
│  ├─ node_link.csv, index_value.csv
│  ├─ mirna_emb/, mirna_contact/, mirna_f/
│  ├─ merged_mirna*, process_rna.py
├─ dataset_in/             # Processed independent dataset
├─ save/                      # Standard 5-fold CV results
├─ save_cold/                 # Cold-start results
├─ save_embedding/            
├─ main.py                    # Standard cross-validation
├─ main_cold_start.py         # Cold-start experiments
├─ main_independent.py              # Independent test
├─ model.py                   # DeepLMI model architecture
├─ utils.py                   # Data loading and metrics
├─ requirements.txt
└─ README.md
```

---

## Environment

Create a new conda environment:

```bash
conda create -n deeplmi python=3.10
conda activate deeplmi
pip install -r requirements.txt
```

---

## Datasets

* **Cross-validation dataset** → `dataset/`
* **Cold-start splits** → `cold_data_splits/`
* **Independent dataset** → `dataset_in/`

---
## Feature construction
RNA sequence representations were obtained using RNA-FM. A pretrained RNA-FM model was applied to encode RNA sequences and generate nucleotide-level embeddings, where each nucleotide is represented by a 640-dimensional vector, resulting in an embedding matrix of size L × 640 (L denotes the RNA sequence length). Detailed instructions for installation and embedding extraction are available in the official RNA-FM repository: [RNA-FM](https://github.com/ml4bio/RNA-FM).

In parallel, RNA secondary structure information was predicted using SPOT-RNA-2D. This method infers base-pairing relationships between nucleotides and produces a binary contact matrix of size L × L, where each element is either 0 or 1, indicating the presence or absence of a structural contact. The detailed usage and implementation of SPOT-RNA-2D can be found in its official documentation:[SPOT-RNA-2D](https://github.com/jaswindersingh2/SPOT-RNA-2D).

---

## Experiments

Three main experimental setups are implemented:

### 1. Standard Cross-Validation

Run:

```bash
python main.py
```

* Outputs to `save/`

### 2. Blind Tests

Run:

```bash
python main_cold_start.py
```

* Modes: blind lncRNA / blind miRNA / blind both
* Outputs to `save_cold/`

### 3. Independent Test

Run:

```bash
python main_independent.py
```

* Outputs to `save/`

---

## Contact

If you have any issues or questions about this paper or need assistance with reproducing the results, please contact me.

Zhijian Huang

School of Computer Science and Engineering,

Central South University

Email: zhijianhuang@csu.edu.cn