Reproduce “From Surfaces to Volumes: Registered Geometry for Protein Representation Learning” on OpenResearch

About this paper

Siyuan Chen, Cai Zhou, Jinrui Zhang, Zhaokang Liang, Taku Komura, Wojciech Matusik, Stephen Bates, Tommi Jaakkola and 3 more

Protein-TetSphere represents each residue-associated volume by fitting it to a shared tetrahedral template, then combines the resulting deformation features with surface and chemical information. The authors evaluate it on ligand-pocket classification, protein–protein interface prediction, and binder-design screening. On clustered PINDER holo contact prediction, it reaches 0.932 ± 0.001 AUROC versus 0.919 ± 0.002 without structure-specific TetSphere input. The results do not establish experimental binding.

Key insights

• A common geometric parameterization lets models compare residue-wise volumes despite their irregular shapes; 64 spectral modes reconstruct the fitted volumes with about 2% error relative to each residue’s longest extent.
• TetSphere’s contribution is measurable beyond adding a trainable pathway: the full model outperforms a control that retains the pathway but omits structure-specific TetSphere inputs.
• Structural representations can inform a binder generator through training-time alignment even when the representation’s structural inputs are unavailable during inference.

Summary by alphaXiv.