CABA-Bind: Confounder-aligned backdoor adjustment for debiased RNA-ligand binding prediction

Area of Science:

• Computational Biology
• Drug Discovery
• Bioinformatics

Background:

• RNA-ligand interactions are crucial for drug discovery.
• Existing models often suffer from sequence bias, limiting their reliability.
• This bias hinders predictions for novel RNA targets and ligands.

Purpose of the Study:

• To introduce CABA-Bind, a causal debiasing framework for RNA-ligand binding prediction.
• To mitigate the impact of sequence-driven bias in predictive models.
• To enhance the reliability and interpretability of RNA-ligand recognition models.

Main Methods:

• Utilized RNA-FM and ChemBERTa for encoding RNA sequences and ligand SMILES.
• Employed K-means clustering for confounder center construction.
• Applied RNA-Confounder Alignment and backdoor adjustment to reduce sequence bias.

Main Results:

• CABA-Bind significantly reduced score-prior bias by 40.8%.
• Achieved a 0.65 MRR in hard decoy evaluation, indicating improved performance.
• Identified potential RNA regions (G17, A53/A54) involved in ligand recognition.

Conclusions:

• CABA-Bind effectively reduces sequence bias in RNA-ligand binding prediction.
• The framework enhances model reliability and provides ligand-specific interpretability.
• This approach advances RNA-targeted drug discovery and molecular recognition modeling.