Protein function lives on the surface. Shape and chemistry act there as one thing, not two.
Most protein comparison methods split shape from chemistry. Fold-comparison tools align backbone structures and report geometric similarity. Binding-site methods catalog local chemical fingerprints — hydrogen bond donors, hydrophobic patches, charge distributions. Each captures half the picture. The result: proteins with similar folds but different chemistry score as similar. Proteins with conserved functional patches embedded in different folds score as different.
Mallet, Léonard, and Montes (arXiv:2603.09860) introduce IFACE, a framework that refuses to separate the two. It represents each protein surface as an intrinsic manifold decorated with spatially distributed chemical fields — electrostatics, hydrophobicity, hydrogen bonding — and computes correspondences that align geometry and chemistry simultaneously. The alignment uses probabilistic coupling: surface points are matched not by minimizing geometric distance alone, or chemical distance alone, but by minimizing a joint distance that integrates both within a single formulation.
The test case is the cytochrome P450 family. These enzymes share a common fold but vary enormously in substrate specificity, determined by the shape and chemistry of buried catalytic pockets. Fold-based methods see the family as nearly identical. IFACE reveals coherent sub-family organization — clusters of enzymes with similar functional surfaces embedded in the shared fold. The conserved catalytic pockets, invisible to fold comparison because they're local features, emerge as the dominant organizational signal.
The distance also distinguishes conformational flexibility (one protein in different states) from true structural divergence (different proteins). A protein moving through its conformational landscape stays nearby in the joint distance. A different protein, even if momentarily adopting a similar shape, remains far.
When you measure the right thing — shape and chemistry together, at the surface where function happens — the classification problem becomes tractable.
Mallet, Léonard, and Montes, "Joint Geometric-Chemical Distance for Protein Surfaces," arXiv:2603.09860 (2026).