One particle per residue is sufficient to describe all-atom protein structures

Abstract

ABSTRACTAtomistic resolution is considered the standard for high-resolution biomolecular structures, but coarse-grained models are often necessary to reflect limited experimental resolution or to achieve feasibility in computational studies. It is generally assumed that reduced representations involve a loss of detail, accuracy, and transferability. This study explores the use of advanced machine-learning networks to learn from known structures of proteins how to reconstruct atomistic models from reduced representations to assess how much information is lost when the vast knowledge about protein structures is taken into account. The main finding is that highly accurate and stereochemically realistic all-atom structures can be recovered with minimal loss of information from just a single bead per amino acid residue, especially when placed at the side chain center of mass. High-accuracy reconstructions with better than 1 Å heavy atom root-mean square deviations are still possible when only Cα coordinates are used as input. This suggests that lower-resolution representations are essentially sufficient to represent protein structures when combined with a machine-learning framework that encodes knowledge from known structures. Practical applications of this high-accuracy reconstruction scheme are illustrated for adding atomistic detail to low-resolution structures from experiment or coarse-grained models generated from computational modeling. Moreover, a rapid, deterministic all-atom reconstruction scheme allows the implementation of an efficient multi-scale framework. As a demonstration, the rapid refinement of accurate models against cryoEM densities is shown where sampling at the coarse-grained level is guided by map correlation functions applied at the atomistic level. With this approach, the accuracy of standard all-atom simulation based refinement schemes can be matched at a fraction of the computational cost.STATEMENT OF SIGNIFICANCEThe fundamental insight of this work is that atomistic detail of proteins can be recovered with minimal loss of information from highly reduced representations with just a single bead per amino acid residue. This is possible by encoding the existing knowledge about protein structures in a machine-learning model. This suggests that it is not strictly necessary to resolve structures in atomistic detail in experiments, computational modeling, or the generation of protein conformations via neural networks since atomistic details can inferred quickly via the neural network. This increases the relevance of experimental structures obtained at lower resolutions and broadens the impact of coarse-grained modeling.