Protein-nucleic acid binding site prediction using interpretable Kolmogorov-Arnold networks with hypergraph...

Area of Science:

• Computational biology
• Bioinformatics
• Structural biology

Background:

• Protein language models (pLMs) and graph neural networks (GNNs) excel at modeling protein-RNA/DNA interactions.
• Simple graphs in existing methods fail to capture complex, high-order residue interactions crucial for nucleic acid binding.
• Spatially continuous yet sequence-discontinuous residues often cooperatively determine nucleic acid binding.

Purpose of the Study:

• To develop an advanced computational approach for identifying nucleic acid binding residues (NBRs) in proteins.
• To overcome the limitations of existing methods in modeling complex residue interactions.
• To enhance the accuracy and interpretability of NBR prediction.

Main Methods:

• Introduced IKANbind, a method combining hypergraph representation learning and interpretable Kolmogorov-Arnold Networks (KANs).
• Leveraged protein language models (pLMs) to implicitly learn physicochemical properties of binding residues.
• Employed symbolic KANs with a weighted mechanism for identifying key predictive features.

Main Results:

• IKANbind significantly outperforms existing methods on multiple NBR benchmark datasets.
• The pLM component effectively learned residue properties like charge and hydrophobicity.
• Symbolic KAN identified polarity and charge as the most significant features for NBR prediction.
• IKANbind demonstrated strong performance in predicting other ligand-binding residues.

Conclusions:

• IKANbind offers a superior approach for identifying nucleic acid binding residues.
• The method provides insights into the key physicochemical properties driving protein-nucleic acid interactions.
• IKANbind shows potential for broader applications in ligand-binding residue prediction.