Effects of cryo-EM cooling on structural ensembles

Structure determination by cryo electron microscopy (cryo-EM) provides information on structural heterogeneity and ensembles at atomic resolution. To obtain cryo-EM images of macromolecules, the samples are first rapidly cooled down to cryogenic temperatures. To what extent the structural ensemble is perturbed by the cooling is currently unknown. Here, to quantify the effects of cooling, we combined continuum model calculations of the temperature drop, molecular dynamics simulations of a ribosome complex before and during cooling with kinetic models. Our results suggest that three effects markedly contribute to the narrowing of the structural ensembles: thermal contraction, reduced thermal motion within local potential wells, and the equilibration into lower free-energy conformations by overcoming separating free-energy barriers. During cooling, barrier heights below 10 kJ/mol were found to be overcome resulting in reduction of B-factors in the ensemble imaged by cryo-EM. Our approach now enables the quantification of the heterogeneity of room-temperature ensembles from cryo-EM structures.

pyFoldX: enabling biomolecular analysis and engineering along structural ensembles

Recent years have seen an increase in the number of structures available, not only for new proteins but also for the same protein crystallized with different molecules and proteins. While protein design software have proven to be successful in designing and modifying proteins, they can also be overly sensitive to small conformational differences between structures of the same protein. To cope with this, we introduce here pyFoldX, a python library that allows the integrative analysis of structures of the same protein using FoldX, an established forcefield and modeling software. The library offers new functionalities for handling different structures of the same protein, an improved molecular parametrization module, and an easy integration with the data analysis ecosystem of the python programming language.Availability and implementationpyFoldX is an open-source library that uses the FoldX software for energy calculations and modelling. The latter can be downloaded upon registration in http://foldxsuite.crg.eu/ and is free of charge for academics. Full details on installation, tutorials covering the library functionality, and the scripts used to generate the data and figures presented in this paper are available at https://github.com/leandroradusky/pyFoldX.

Improving the analysis of biological ensembles through extended similarity measures

We present new algorithms to classify structural ensembles of macromolecules, based on the recently proposed extended similarity measures. Molecular Dynamics provides a wealth of structural information on systems of biologically interest. As computer power increases we capture larger ensembles and larger conformational transitions between states. Typically, structural clustering provides the statistical mechanics treatment of the system to identify relevant biological states. The key advantage of our approach is that the newly introduced extended similiarity indices reduce the computational complexity of assessing the similarity of a set of structures from O(N2) to O(N). Here we take advantage of this favorable cost to develop several highly efficient techniques, including a linear-scaling algorithm to determine the medoid of a set (which we effectively use to select the most representative structure of a cluster). Moreover, we use our extended similarity indices as a linkage criterion in a novel hierarchical agglomerative clustering algorithm. We apply these new metrics to analyze the ensembles of several systems of biological interest such as folding and binding of macromolecules (peptide, protein, DNA-protein). In particular, we design a new workflow that is capable of identifying the most important conformations contributing to the protein folding process. We show excellent performance in the resulting clusters (surpassing traditional linkage criteria), along with faster performance and an efficient cost-function to identify when to merge clusters.

Deep learning the structural determinants of protein biochemical properties by comparing structural ensembles with DiffNets

Understanding the structural determinants of a protein’s biochemical properties, such as activity and stability, is a major challenge in biology and medicine. Comparing computer simulations of protein variants with different biochemical properties is an increasingly powerful means to drive progress. However, success often hinges on dimensionality reduction algorithms for simplifying the complex ensemble of structures each variant adopts. Unfortunately, common algorithms rely on potentially misleading assumptions about what structural features are important, such as emphasizing larger geometric changes over smaller ones. Here we present DiffNets, self-supervised autoencoders that avoid such assumptions, and automatically identify the relevant features, by requiring that the low-dimensional representations they learn are sufficient to predict the biochemical differences between protein variants. For example, DiffNets automatically identify subtle structural signatures that predict the relative stabilities of β-lactamase variants and duty ratios of myosin isoforms. DiffNets should also be applicable to understanding other perturbations, such as ligand binding.

Patterns in Protein Flexibility: A Comparison of NMR “Ensembles”, MD Trajectories, and Crystallographic B-Factors

Proteins are molecular machines requiring flexibility to function. Crystallographic B-factors and Molecular Dynamics (MD) simulations both provide insights into protein flexibility on an atomic scale. Nuclear Magnetic Resonance (NMR) lacks a universally accepted analog of the B-factor. However, a lack of convergence in atomic coordinates in an NMR-based structure calculation also suggests atomic mobility. This paper describes a pattern in the coordinate uncertainties of backbone heavy atoms in NMR-derived structural “ensembles” first noted in the development of FindCore2 (previously called Expanded FindCore: DA Snyder, J Grullon, YJ Huang, R Tejero, GT Montelione, Proteins: Structure, Function, and Bioinformatics 82 (S2), 219–230) and demonstrates that this pattern exists in coordinate variances across MD trajectories but not in crystallographic B-factors. This either suggests that MD trajectories and NMR “ensembles” capture motional behavior of peptide bond units not captured by B-factors or indicates a deficiency common to force fields used in both NMR and MD calculations.

Perturbation of ACE2 structural ensembles by SARS-CoV-2 spike protein binding

The human ACE2 enzyme serves as a critical first recognition point of coronaviruses, including SARS-CoV-2. In particular, the extracellular domain of ACE2 interacts directly with the S1 tailspike protein of the SARS-CoV-2 virion through a broad protein-protein interface. Although this interaction has been characterized by X-ray crystallography and Cryo-EM, these structures do not reveal significant differences in ACE2 structure upon S1 protein binding. In this work, using several all-atom molecular dynamics simulations, we show persistent differences in ACE2 structure upon binding. These differences are determined with the Linear Discriminant Analysis (LDA) machine learning method and validated using independent training and testing datasets, including long trajectories generated by D. E. Shaw Research on the Anton 2 supercomputer. In addition, long trajectories for 78 potent ACE2-binding compounds, also generated by D. E. Shaw Research, were projected onto the LDA classification vector in order to determine whether the ligand-bound ACE2 structures were compatible with S1 protein binding. This allows us to predict which compounds are “apo-like” vs “complex-like”, as well as to pinpoint long-range ligand-induced allosteric changes of ACE2 structure.