Multiscale Graph Neural Networks on the SES Surface for Protein Binding Site Prediction
Received: 10 Oct 2026 Revised: 21 Oct 2026 Accepted: 05 Nov 2026
Published: 2026, vol. 30, issue 4, pp. 539–542
Abstract
We present SES-GNN, a multiscale graph neural network on the solvent-excluded surface (SES) for protein binding site prediction. Instead of a uniform graph with fixed edge radius, a three-scale representation (0–2, 2–4, 4–8 Å) with Farthest Point Sampling centroid selection is built. The architecture uses 12 TransformerConv layers. The multiscale graph gives a +0.015–0.017 ROC-AUC gain over the uniform one. The method achieves ROC-AUC = 0.838 on our own PPI dataset. Ablation shows that vertex features carry the main information but realize it only through the graph structure: a pointwise classifier without message passing falls substantially below the full model, while topology alone is useless.
Keywords: protein binding site prediction, SES surface, graph neural networks.
BibTeX
@article{IS-Polovnikov2026,
author = {Polovnikov, Vladimir Sergeevich},
title = {{Multiscale Graph Neural Networks on the SES Surface for Protein Binding Site Prediction}},
journal = {Intelligent Systems. Theory and Applications},
year = {2026},
volume = {30},
number = {4},
pages = {539--542},
}
AMSBIB
\Bibitem{IS-Polovnikov2026}
\by V.\,S.~Polovnikov
\paper Multiscale Graph Neural Networks on the SES Surface for Protein Binding Site Prediction
\jour Intelligent Systems. Theory and Applications
\yr 2026
\vol 30
\issue 4
\pages 539--542
\lang In Russian
RU