Structure-based prediction of protein-nucleic acid binding using graph neural networks

Area of Science:

• Structural biology
• Bioinformatics
• Genomics

Background:

• Protein-nucleic acid (PNA) binding is crucial for genome regulation.
• Structural models of bound proteins are scarce, limiting understanding of binding mechanisms.
• Predicting PNA binding from unbound structures is a significant challenge.

Purpose of the Study:

• To develop a deep learning approach (PNAbind) for predicting PNA binding from unbound protein structures.
• To identify both the overall binding function and specific binding residues of proteins.
• To differentiate between DNA and RNA binding specificity.

Main Methods:

• Utilized graph neural networks to encode spatial distributions of physicochemical and geometric properties of protein structures.
• Employed global physicochemical encodings for predicting overall protein binding function.
• Used local encodings for predicting individual nucleic acid binding residues.

Main Results:

• PNAbind achieved high accuracy in predicting binding sites, with AUROC scores of 0.92-0.95.
• The model successfully discriminated between DNA and RNA binding specificity.
• Predictions on computationally derived structures provided mechanistic insights into NA recognition.
• Applied to APOBEC3G, PNAbind's predictions aligned with experimental RNA binding data.

Conclusions:

• PNAbind offers a novel computational approach to predict PNA binding sites and functions.
• The method enhances understanding of the structural and chemical determinants of nucleic acid recognition.
• PNAbind can elucidate mechanisms of action for proteins involved in genome regulation, such as APOBEC3G.