D’BiasRNA: A Disentanglement Variational Autoencoder for Predicting RNA-Ligand Binding Sites

Abstract

RNA has emerged as an attractive therapeutic target, driving growing interest in computational RNAligand binding site prediction. Despite recent progress enabled by deep learning, most existing methods rely on structural features extracted from post-binding RNA-ligand complexes, are largely ligand-agnostic, and ignore the ionic environment. This reliance introduces information leakage and limits applicability to realistic settings where holo-form structures are unavailable and binding must be inferred from ligand-free RNA structures.

We propose D’BiasRNA, a generative framework inspired by disentangled variational autoencoders that models RNA-ligand binding as a transition from apo-form RNA structures to post-binding outcomes. D’BiasRNA separates intrinsic RNA geometry separately from ligand- and ion-induced effects using disentangled latent representations, enabling binding site prediction conditioned on arbitrary ligands and ionic environments without direct reliance on post-binding structural features at inference time. We further construct a comprehensive dataset comprising annotated RNA-ligand binding complexes and high quality apo-form RNA structures. Experimental results demonstrate that D’BiasRNA achieves competitive performance while exhibiting improved robustness in blind prediction scenarios involving unseen RNAs, ligands, and ionic conditions. Qualitative analyses confirm that the model captures biologically meaningful RNA-ligand interaction patterns.

Install environment (Linux)

conda env create --file environment.yml
conda activate rbs

---

Download datasets

Please download the rbs_data.tar.gz in the main directory and uncompress it.

pip install gdown
gdown https://drive.google.com/file/d/1SxvFC3qqfZnhStRz3t2DCMFHbcNJh9Fv/view?usp=drive_link
tar -zxvf rbs_data.tar.gz

---

Training models

python train.py --gpu_idx 0,1,2,3

---

Download checkpoints

Please download the ckpt.tar.gz in the main directory and uncompress it.

pip install gdown
gdown https://drive.google.com/file/d/1fasbfngRaLxknmWfOoKKPGdLcgy965gP/view?usp=sharing
tar -zxvf ckpt.tar.gz

---

Inference using the checkpoints

python aggregate_results.py --exp_name best --gpu_idx 0,1,2,3

---

Acknowledgement

Contributors

Name	Affiliation	Email
Junseok Choe†	Data Mining and Information Systems Lab, Korea University, Seoul, South Korea	juns94@korea.ac.kr
Hajung Kim	Data Mining and Information Systems Lab, Korea University, Seoul, South Korea	hajungk@korea.ac.kr
Sohyun Chung	Data Mining and Information Systems Lab, Korea University, Seoul, South Korea	sohyunjung@korea.ac.kr
Mogan Gim*	Digital Health Intelligence Systems Lab, Hankuk University of Foreign Studies, Yongin, South Korea	gimmogan@hufs.ac.kr
Jaewoo Kang*	Data Mining and Information Systems Lab, Korea University, Seoul, South Korea	kangj@korea.ac.kr

• †: First Author
• *: Corresponding Author