The determination and analysis of three-dimensional structures of biological macromolecules, using techniques like crystallography, NMR, and cryo-EM to understand structure-function relationships and drug design
12 papersThis study introduces a new method for analyzing protein structures within living cells using a fast-acting chemical probe called vinyl thianthrenium tetrafluoroborate (VTT). VTT creates "snapshots" of protein interactions by quickly linking nearby amino acids, including cysteine and others with nucleophilic side chains, offering insights into protein structure and function in various cell types.
This paper introduces RosettaFold-3 (RF3), a new open-source deep learning model for predicting biomolecular structures, and AtomWorks, a framework for developing such models. RF3 shows improved handling of chirality and user-defined constraints, narrowing the gap between open-source and closed-source models like AlphaFold3. RF3 and AtomWorks together facilitate the creation and training of biomolecular modeling tools, emphasizing data quality and reproducibility.
This paper describes a new method for determining the 3D structures of proteins using cryo-electron microscopy directly from the crude product of a cell-free protein expression system. By skipping the purification step, the researchers reduced the time and resources required to go from gene to a fully refined structure to under 24 hours. They successfully determined the structures of two different proteins using this approach at resolutions of 2.8Å and 2.5Å, demonstrating the method’s potential.
This paper introduces RoseTTAFold, a three-track neural network that predicts protein structures and interactions with high accuracy, approaching that of DeepMind's AlphaFold2. This network integrates information at the sequence, distance map, and 3D coordinate levels, enabling rapid solutions for X-ray crystallography, cryo-EM modeling, and prediction of protein-protein complex structures directly from sequence information.
AlphaFold is a deep learning-based method that predicts protein structures with atomic accuracy, even in cases where no similar structures are known. It achieved high accuracy in the CASP14 assessment, outperforming other methods, and its performance generalizes to a large set of recently released PDB structures.
AlphaFold 3 demonstrates significant improvements in predicting the structure of biomolecular interactions across various categories, including protein-ligand, protein-nucleic acid, and antibody-antigen complexes. The model's updated diffusion-based architecture allows for the prediction of complexes with diverse molecular types, achieving higher accuracy than previous specialized tools in most cases.
The WeNMR project has provided web-based structural biology tools and high-throughput computing resources to researchers for over a decade, transitioning into a Thematic Service within the European Open Science Cloud (EOSC). The ecosystem has served a large international user base, with increased usage during the COVID-19 pandemic, and continues to facilitate research and education in structural biology.
This paper presents Topaz-Denoise, a general deep-learning method and trained models for removing noise from cryoEM micrographs and cryoET tomograms. The authors demonstrate that denoising with Topaz allows for the identification of particles previously unseen due to low SNR, thus allowing for solving protein structures with more complete particle orientations and identifying new conformations, as well as substantially reducing electron dose and microscope exposure time without sacrificing data quality.
This paper introduces new methods in RELION-3.1 to detect and correct for symmetrical and antisymmetrical optical aberrations and anisotropic magnification in cryo-EM data. Applying these corrections to several publicly available datasets improved the resolution of the 3D reconstructions, especially when these optical effects were prominent.
DeepEMhancer, a deep learning-based algorithm, enhances cryo-EM maps by performing masking and sharpening-like operations, leading to improved map quality and easier atomic model building. Tested on 20 different maps, DeepEMhancer demonstrated a median resolution improvement of ~0.6 Å compared to the original maps, outperforming traditional B-factor-based approaches.
IsoNet, a deep learning software package, was developed to enhance the resolution and interpretability of cryo-electron tomography (cryoET) data by iteratively filling in missing information and improving the signal-to-noise ratio. Applied to various cryoET datasets, IsoNet reduced resolution anisotropy and allowed for visualization of previously obscured structural details in viruses, cellular organelles, and neuronal synapses, without needing sub-tomogram averaging.
This paper introduces a new Bayesian approach for subtomogram averaging that operates directly on 2D tilt series images within RELION 4.0. This method eliminates the need for missing wedge correction and offers improved accuracy for subtomogram alignment and classification, demonstrated by achieving resolutions sufficient for de novo atomic modeling in two test cases: the HIV immature capsid and the C. crescentus S-layer.