VoroMarmotte: predicting ensemble-derived protein contact stability from a single structure

Abstract

AbstractProtein structures are dynamic, and contacts observed in a single structure may not remain stable across different conformations. However, most structure-based analyses rely on individual structures and do not explicitly account for this variation. Here, we quantify contact-area persistence in experimentally determined ensembles of monomeric proteins using the minimum and maximum Voronoi contact areas observed across conformations. A contact is classified as stable when its area-based persistence meets a defined threshold, reflecting how consistently its contact area is retained rather than how frequently the contact occurs. We also subdivide Voronoi contact areas into layers and sectors to capture finer geometric differences. We first introduce VoroChipmunk, which extends VoroMQA's occurrence-based statistical potential with persistence statistics and more detailed contact geometry. Its weighted layer- and sector-specific terms improve the selection of inter-chain interface models compared with a VoroMQA-like baseline. Building on VoroChipmunk, we develop VoroMarmotte, which combines its energy components with local geometric features in fixed-length Voronoi Contacts Blocks (VCBlocks). Using a standard multilayer perceptron rather than a graph neural network, VoroMarmotte predicts, from a single protein structure, which contacts meet this ensemble-defined stability criterion. On held-out monomeric proteins, VoroMarmotte achieves approximately 0.78 balanced accuracy and 0.55 Matthews correlation coefficient. These results demonstrate that ensemble-derived contact information can support predictions of contact stability from a single structure. The same contact representation can also be applied to inter-chain interfaces. Software availability VoroMarmotte: https://github.com/kliment-olechnovic/voromarmotte-appVoroChipmunk: https://github.com/kliment-olechnovic/vorochipmunk-app