Paper

AtomSurf : Surface Representation for Learning on Protein Structures

An essential aspect of learning from protein structures is the choice of their representation as a geometric object (be it a grid, graph, or surface), which conditions the associated learning method. The performance of a given approach will then depend on both the representation and its corresponding learning model. In this paper, we investigate representing proteins as $\textit{surfaces embedded in 3D}$ and evaluate this representation within an established benchmark: atom3d. Our first finding is that despite promising results, state-of-the-art surface-based learning approaches alone are not competitive with other modalities on this benchmark. Building on this, we introduce a novel synergistic approach that incorporates graph and surface-based approaches within a single learnable architecture. We show that using this combination, which inherits the strengths of the two representations, we obtain state-of-the-art results across $\textit{all tested tasks}$, on the atom3d benchmark, as well as on binding pocket classification. Our code and data can be found online: https://github.com/Vincentx15/atom2D.

Results in Papers With Code
(↓ scroll down to see all results)