Granular clustering of de novo protein models

TL;DR

This paper introduces a novel granular clustering algorithm for de novo protein models that improves the accuracy of partial models and enhances their effectiveness in molecular replacement applications.

Contribution

The paper presents a new granular clustering method that operates on partial protein models, offering more accurate structural details than traditional global alignment approaches.

Findings

Clusters of partial models are more accurate than global alignments.

Partial models from the new algorithm improve molecular replacement success.

The method outperforms existing clustering software in structure detail accuracy.

Abstract

Modern algorithms for de novo prediction of protein structures typically output multiple full-length models (decoys) rather than a single solution. Subsequent clustering of such decoys is used both to gauge the success of the modelling and to decide on the most native-like conformation. At the same time, partial protein models are sufficient for some applications such as crystallographic phasing by molecular replacement (MR) in particular, provided these models represent a certain part of the target structure with reasonable accuracy. Here we propose a novel clustering algorithm that natively operates in the space of partial models through an approach known as granular clustering (GC). The algorithm is based on growing local similarities found in a pool of initial decoys. We demonstrate that the resulting clusters of partial models provide a substantially more accurate structural detail on the target protein than those obtained upon a global alignment of decoys. As the result, the partial models output by our GC algorithm are also much more effective towards the MR procedure, compared to the models produced by existing software. The source code is freely available at https://github.com/biocryst/gc