CoBRA: compound binding site prediction using RNA language model.

Area of Science:

• Biochemistry
• Computational Biology
• Drug Discovery

Background:

• Ribonucleic acid (RNA) plays crucial roles in cellular functions and is implicated in numerous human diseases.
• Targeting RNA with small molecules presents a promising therapeutic strategy due to the limited druggability of proteins.

Purpose of the Study:

• To develop a computational tool for accurate prediction of small-molecule binding sites on RNA molecules.
• To evaluate the efficacy of a deep learning approach utilizing RNA language models for binding site prediction.

Main Methods:

• Introduction of compound binding site prediction for RNA (CoBRA), a lightweight deep learning program.
• Utilized residue-level embeddings from a pre-trained RNA language model, bypassing the need for structural information.
• Employed a multi-layer perceptron classifier for binary classification of nucleotide binding sites.

Main Results:

• CoBRA achieved a 22.1% relative improvement in Matthew correlation coefficient and a 45.6% increase in sensitivity compared to existing methods.
• Sequence-based language model embeddings demonstrated performance comparable to or exceeding structure-based prediction methods.
• The model was trained on TR60 and HARIBOSS datasets and validated on four independent benchmark sets.

Conclusions:

• CoBRA provides a flexible and effective tool for predicting RNA-drug binding sites without requiring structural data.
• This sequence-based approach advances the development of RNA-targeted therapeutics.
• The findings highlight the potential of language models in understanding RNA-ligand interactions.