Graph Representation Learning on 3D Mesh-Based Protein Structures

Exploring 3D protein structure is one of the most persistent and complex challenges in computational biology. While geometric deep learning has driven breakthroughs in modeling non-Euclidean data, the direct application of graph-based methods to protein structures, inherently representable as 3D graphs, remains underexplored. Addressing this gap, we introduce a pipeline that converts protein structures into 3D surface meshes, constructs graph representations, and applies a Graph Auto-Encoder (GAE) to learn meaningful embeddings. These embeddings are used for unsupervised node clustering, revealing spatially conserved functional regions and structural motifs within proteins. Extensive experiments in diverse protein datasets demonstrate the efficacy of the framework, achieving state-of-the-art performance in structural feature reconstruction and region clustering. In particular, our approach identifies subtle conformational variations related to functional dynamics, outperforming traditional methods by 15–20% in cross-validation tasks.