DeepRNA-DTI: A Deep Learning Approach for RNA-Compound Interaction Prediction with Binding Site Interpretability

Abstract

RNA-targeted therapeutics represent a promising frontier for expanding the druggable genome beyond conventional protein targets. However, computational prediction of RNA-compound interactions remains challenging due to limited experimental data and the inherent complexity of RNA structures. Here, we present DeepRNA-DTI, a novel sequence-based deep learning approach for RNA-compound interaction prediction with binding site interpretability. Our model leverages transfer learning from pretrained embeddings, RNA-FM for RNA sequences and Mole-BERT for compounds, and employs a multitask learning framework that simultaneously predicts both presence of interactions and nucleotide-level binding sites. This dual prediction strategy provides mechanistic insights into RNA-compound recognition patterns. Trained on a comprehensive dataset integrating resources from the Protein Data Bank and literature sources, DeepRNA-DTI demonstrates superior performance compared to existing methods. The model shows consistent effectiveness across diverse RNA subtypes, highlighting its robust generalization capabilities. Application to high-throughput virtual screening of over 48 million compounds against oncogenic pre-miR-21 successfully identified known binders and novel chemical scaffolds with RNA-specific physicochemical properties. By combining sequence-based predictions with binding site interpretability, DeepRNA-DTI advances our ability to identify promising RNA-targeting compounds and offers new opportunities for RNA-directed drug discovery.

Overview

!Overview.png

Setup

1. Environment Setup

Create a conda environment using the provided environment.yml:

cd DeepRNA-DTI
conda env create --file environment.yml -n DeepRNA_DTI
conda activate DeepRNA_DTI

2. Pretrained Model Setup

DeepRNA-DTI uses pretrained weights from RNA-FM and Mole-BERT. Download the model code and weights:

mkdir -p Model/pretrained_model Model/trained_weight
cd Model/pretrained_model

# Clone RNA-FM
git clone https://github.com/ml4bio/RNA-FM.git

# Clone Mole-BERT
git clone https://github.com/junxia97/Mole-BERT.git

Model Sources:

• RNA-FM: https://github.com/ml4bio/RNA-FM
• Mole-BERT: https://github.com/junxia97/Mole-BERT

3. Download Trained Weights

We provide pretrained weights for all four generalization scenarios. Download them using gdown:

pip install gdown
cd Model/trained_weight
gdown "https:// ... --folder"

Available Pretrained Weights

Scenario	Description	Download
unseen_pair	Pair unseen at test time	gdown "https://drive.google.com/drive/folders/109jdGX0yKuC7AgD50y4ZPs1rAa2h0Q9f" --folder
unseen_rna	RNA unseen, compounds from training set	gdown "https://drive.google.com/drive/folders/1UxpOBIyFzw90H4Gy5wyffYSgSVDUvMLi" --folder
unseen_compound	Compounds unseen, RNA from training set	gdown "https://drive.google.com/drive/folders/1uWNXatSLZ1qxhCVkhAV1Y4yiPt3NCW2V" --folder
unseen_both	Both RNA and compound unseen at test time	gdown "https://drive.google.com/drive/folders/1d0UIwn9sYiAeBgHkqXco4NQDOPnrk28S" --folder

After downloading, organize the weights as follows:

Model/trained_weight/
├── unseen_pair/
│   ├── model_fold0.pt
│   ├── model_fold1.pt
│   ├── model_fold2.pt
│   ├── model_fold3.pt
│   └── model_fold4.pt
├── unseen_rna/
│   └── ...
├── unseen_compound/
│   └── ...
└── unseen_both/
    └── ...

Usage

Testing (Inference)

Evaluate the model on test data:

python test.py --test_type unseen_pair

Arguments:

Argument	Default	Description
--test_type	unseen_pair	Dataset split: unseen_pair, unseen_rna, unseen_compound, unseen_both
--model_folder_path	./Model/trained_weight	Path to trained model weights
--data_folder_path	./Dataset/	Path to dataset
--batch_size	32	Batch size for evaluation

Training

Train the model from scratch:

python train.py --train_type unseen_pair --num_epochs 100

Arguments:

Argument	Default	Description
--train_type	unseen_pair	Training split type: unseen_pair, unseen_rna, unseen_compound, unseen_both
--batch_size	32	Batch size
--num_epochs	100	Number of training epochs
--learning_rate	0.001	Learning rate
--weight_decay	1e-4	L2 regularization

Dataset Structure

Dataset/
├── unseen_pair/          # Pair unseen at test time
├── unseen_rna/           # RNA unseen, compounds from training
├── unseen_compound/      # Compounds unseen, RNA from training
└── unseen_both/          # Both RNA and compound unseen at test time
    ├── dti_data/         # Drug-target interaction data
    │   ├── train_fold0/
    │   │   └── raw/
    │   │       └── interactions.csv
    │   ├── train_fold1/
    │   ├── ...
    │   ├── val_fold0/
    │   └── test_fold/
    └── bs_data/          # Binding site data
        ├── train_fold0/
        ├── ...
        └── test_fold/

Data Format (interactions.csv)

Column	Description
sequence	RNA sequence
smiles	Compound SMILES string
interactions	Binary label (1 = interacts, 0 = no interaction)
binding_site_index	List of nucleotide indices that bind to the compound

Evaluation Metrics

Drug-Target Interaction (DTI)

• AUC-ROC: Area under the receiver operating characteristic curve
• AUPR: Area under the precision-recall curve

Binding Site (BS)

• Micro-averaged AUC/AUPR: Per-RNA sample average (only for positive interactions)

Project Structure

DeepRNA_DTI_github/
├── train.py              # Training script
├── test.py               # Evaluation script
├── environment.yml       # Conda environment
├── README.md             # This file
├── src/
│   ├── model.py          # DeepRNA_DTI model architecture
│   ├── data_utils.py     # Dataset classes and data loaders
│   └── utils.py          # Loss functions and evaluation metrics
├── Model/
│   ├── pretrained_model/
│   │   ├── RNA-FM/       # RNA-FM pretrained model
│   │   └── Mole-BERT/    # Mole-BERT pretrained model
│   └── trained_weight/   # Trained DeepRNA-DTI weights
└── Dataset/
    ├── unseen_pair/
    ├── unseen_rna/
    ├── unseen_compound/
    └── unseen_both/

Contact

• Haelee Bae: haeleeeeleah@gm.gist.ac.kr
• Hojung Nam (Corresponding Author): hjnam@gist.ac.kr