RGTBind: RBF-gate graph transformer with spatially biased attention for protein-DNA binding-site prediction.

Area of Science:

• Computational Biology
• Structural Bioinformatics
• Genomics

Background:

• Protein-DNA binding is crucial for gene regulation and protein function.
• Accurate prediction is challenging due to the complex interplay of sequence and 3D structural information.
• Existing methods struggle to effectively integrate these diverse data types.

Purpose of the Study:

• To develop an advanced computational model for predicting protein-DNA binding sites.
• To improve the accuracy of binding-site prediction by incorporating both sequence and structural features.
• To enhance our understanding of gene regulation mechanisms.

Main Methods:

• Developed RGTBind, a graph transformer model utilizing multi-scale radial basis function distance encoding.
• Integrated AlphaFold2-predicted protein structures, DSSP descriptors, PSSM, and HMM profiles.
• Employed spatially biased multi-head self-attention and a learnable threshold gate for residue interactions.

Main Results:

• RGTBind achieved superior performance on independent benchmarks (Test_129, Test_181) compared to existing methods.
• Demonstrated state-of-the-art results in F1, AUC, and MCC scores for binding-site prediction.
• Validated the effectiveness of distance-aware attention and structure-guided neighbor selection.

Conclusions:

• RGTBind represents a significant advancement in residue-level protein-DNA binding-site prediction.
• The model's ability to integrate structural and sequence data offers new insights into protein-DNA interactions.
• This approach holds promise for future research in gene regulation and protein function.