Deep Learning-based Prediction of Chromatin-Associated RNA-Binding Proteins

This repository contains a deep learning framework designed to identify Chromatin-associated RNA-binding Proteins (caRBPs) using Protein Language Models (pLM) and advanced hybrid neural architectures.

Key Features

• pLM Integration: Leverages ProtT5-XL-UniRef50 embeddings to capture high-dimensional evolutionary and biophysical information.
• Hybrid Loss Function: Implements a Hybrid Dice-Focal Loss to specifically address class imbalance in biological datasets.

Dependencies

• Python 3.8
• PyTorch 2.4.1
• TensorFlow 2.13.1
• Transformers 4.46.3
• Scikit-learn 1.3.2
• Pandas 2.0.3
• Numpy 1.24.3
• Matplotlib 3.7.5

Contents

• data/: Directory containing the datasets in .csv format.
• inference/: Directory containing datasets in .csv format for inference and scripts to predict potential caRBPs.
• model_training/: Directory containing scripts with or without pLM for peptide/protein training.
• model/: Directory containing the model generated by pLM-CNN-BiLSTM for inference.
• README.md: Project documentation.

Code Details

• inference.py: A script for high-throughput prediction on new FASTA sequences.
• gen_T5_feature.py: Extracts features from protein sequences using the T5EncoderModel.
• plm_protein_model.py: The final training script for full-length proteins, incorporating pLM features and a dual-pathway CNN-BiLSTM architecture.
• peptide_model.py: A specialized model for short peptide fragments without pLM, used for classifying chromatin-binding peptides.
• protein_model.py: A specialized model for full-length protein without pLM, used for classifying caRBPs.
• plm_peptide_model.py: A specialized model for short peptide fragments with pLM, used for classifying chromatin-binding peptides.

Installation

git clone https://github.com/callmeracle/caRBP-Pred.git    
cd caRBP-Pred  
conda create -n carbp_env python=3.8  
conda activate carbp_env

The ProtT5-XL model can be downloaded from here (https://huggingface.co/Rostlab/prot_t5_xl_uniref50/tree/main) Embedding files can be downloaded from here (https://drive.google.com/drive/folders/1TrboNK1oMGsd03l1G9CyPhQWVOuyt2Gt?usp=sharing)

Usage

1. Feature Extraction

Use gen_T5_feature.py to extract pre-trained language model features from protein sequences:

python gen_T5_feature.py

• Input Format

protein_id,sequence
P12345,MKTLLILGL...  
P67890,GSHMASSNA...  

• Output:

protein_plm_residue.npy

2. Model Training Run plm_protein_model.py for 5-fold cross-validation training:

python plm_protein_model.py

3. Model Inference Use the previously trained model to predict new sequences:

python inference/inference.py

• Configuration (modify in main):

task = {  
    "input_csv": "inference/all_fasta.csv",           # Input sequences  
    "ptm_npy_path": "./inference_protein_plm_residue.npy",  # Pre-computed features  
    "model_path": "../model/model.h5",                # Model path  
    "output_csv": "Final_BiLSTM_pLM_Inference.csv"      # Output results  
}  

• Output results:

Gene_Name,Prediction_Prob,Decision,Threshold  
Protein_A,0.8923,Positive,0.7  
Protein_B,0.2341,Negative,0.7  
...