Context <p>Protein–DNA binding-site prediction is essential for understanding gene regulation and protein function, but remains difficult because DNA recognition depends on both sequence context and three-dimensional structure. We developed RGTBind, a graph transformer that combines multi-scale radial basis function distance encoding with a learnable threshold-gating mechanism to model spatially informative residue interactions. On the independent Test_129 and Test_181 benchmarks, RGTBind achieved the best F1, AUC, and MCC among the compared methods, supporting the value of distance-aware attention with structure-guided neighbor selection for residue-level protein–DNA binding-site prediction.</p> Methods <p>Each protein was represented as a residue-level graph derived from AlphaFold2-predicted structures. Residue features included AlphaFold2 single representations, DSSP-derived structural descriptors, PSI-BLAST position-specific scoring matrices (PSSM), and HHblits hidden Markov model (HMM) profiles. Pairwise C<InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(\alpha \)</EquationSource> <EquationSource Format="MATHML"><math> <mi>α</mi> </math></EquationSource> </InlineEquation>–C<InlineEquation ID="IEq2"> <EquationSource Format="TEX">\(\alpha \)</EquationSource> <EquationSource Format="MATHML"><math> <mi>α</mi> </math></EquationSource> </InlineEquation> distances were encoded using a multi-scale radial basis function scheme and incorporated into a graph transformer through spatially biased multi-head self-attention and a learnable threshold gate. Sequence redundancy was reduced with CD-HIT. The model was trained with AdamW using five-fold cross-validation on Train_573 and evaluated on the Test_129 and Test_181 benchmark datasets.</p>