Single Protein Structure Research Articles

EuDockScore: Euclidean graph neural networks for scoring protein-protein interfaces.

Protein-protein interactions are essential for a variety of biological phenomena including mediating biochemical reactions, cell signaling, and the immune response. Proteins seek to form interfaces which reduce overall system energy. Although determination of single polypeptide chain protein structures has been revolutionized by deep learning techniques, complex prediction has still not been perfected. Additionally, experimentally determining structures is incredibly resource and time expensive. An alternative is the technique of computational docking, which takes the solved individual structures of proteins to produce candidate interfaces (decoys). Decoys are then scored using a mathematical function that assess the quality of the system, known as scoring functions. Beyond docking, scoring functions are a critical component of assessing structures produced by many protein generative models. Scoring models are also used as a final filtering in many generative deep learning models including those that generate antibody binders, and those which perform docking. In this work, we present improved scoring functions for protein-protein interactions which utilizes cutting-edge Euclidean graph neural network architectures, to assess protein-protein interfaces. These Euclidean docking score models are known as EuDockScore, and EuDockScore-Ab with the latter being antibody-antigen dock specific. Finally, we provided EuDockScore-AFM a model trained on antibody-antigen outputs from AlphaFold-Multimer (AFM) which proves useful in reranking large numbers of AFM outputs. The code for these models is available at https://gitlab.com/mcfeemat/eudockscore.

Leveraging neural networks to correct FoldX free energy estimates.

Proteins play a pivotal role in many biological processes, and changes in their amino acid sequences can lead to dysfunction and disease. These changes can affect protein folding or interaction with other biomolecules, such as preventing antibodies from inhibiting a viral infection or causing proteins to misfold. The ability to predict the effects of mutations in proteins is crucial. Although experimental techniques can accurately quantify the effect of mutations on protein folding free energies and protein-protein binding free energies, they are often time-consuming and costly. By contrast, computational techniques offer fast and cost-effective alternatives for estimating free energies, but they typically suffer from lower accuracy. Enhancing the accuracy of computational predictions is therefore of high importance, with the potential to greatly impact fields ranging from drug design to understanding disease mechanisms. One such widely used computational method, FoldX, is capable of rapidly predicting the relative folding stability ( ) for a protein as well as the relative binding affinity ( ) between proteins using a single protein structure as input. However, it can suffer from low accuracy, especially for antibody-antigen systems. In this work, we trained a neural network on FoldX output to enhance its prediction accuracy. We first performed FoldX calculations on the largest datasets available for mutations that affect binding (SKEMPIv2) and folding (ProTherm4) with experimentally measured . Features were then extracted from the FoldX output files including its prediction for . We then developed and optimized a neural network framework to predict the difference between FoldX's estimated and the experimental data, creating a model capable of producing a correction factor. Our approach showed significant improvements in Pearson correlation performance. For single mutations affecting folding, the correlation improved from a baseline of 0.3 to 0.66. In terms of binding, performance increased from 0.37 to 0.61 for single mutations and from 0.52 to 0.81 for double mutations. For epistasis, the correlation for binding affinity (both singles and doubles) improved from 0.19 to 0.59. Our results also indicated that models trained on double mutations enhanced accuracy when predicting higher-order mutations (such as triple or quadruple mutations), whereas models trained on singles did not. This suggests that interaction energy and epistasis effects present in the FoldX output are not fully utilized by FoldX itself. Once trained, these models add minimal computational time but provide a substantial increase in performance, especially for higher-order mutations and epistasis. This makes them a valuable addition to any free energy prediction pipeline using FoldX. Furthermore, we believe this technique can be further optimized and tested for predicting antibody escape, aiding in the efficient development of watch lists.

Predicting Mutation-Induced Allosteric Changes in Structures and Conformational Ensembles of the ABL Kinase Using AlphaFold2 Adaptations with Alanine Sequence Scanning.

Despite the success of AlphaFold2 approaches in predicting single protein structures, these methods showed intrinsic limitations in predicting multiple functional conformations of allosteric proteins and have been challenged to accurately capture the effects of single point mutations that induced significant structural changes. We examined several implementations of AlphaFold2 methods to predict conformational ensembles for state-switching mutants of the ABL kinase. The results revealed that a combination of randomized alanine sequence masking with shallow multiple sequence alignment subsampling can significantly expand the conformational diversity of the predicted structural ensembles and capture shifts in populations of the active and inactive ABL states. Consistent with the NMR experiments, the predicted conformational ensembles for M309L/L320I and M309L/H415P ABL mutants that perturb the regulatory spine networks featured the increased population of the fully closed inactive state. The proposed adaptation of AlphaFold can reproduce the experimentally observed mutation-induced redistributions in the relative populations of the active and inactive ABL states and capture the effects of regulatory mutations on allosteric structural rearrangements of the kinase domain. The ensemble-based network analysis complemented AlphaFold predictions by revealing allosteric hotspots that correspond to state-switching mutational sites which may explain the global effect of regulatory mutations on structural changes between the ABL states. This study suggested that attention-based learning of long-range dependencies between sequence positions in homologous folds and deciphering patterns of allosteric interactions may further augment the predictive abilities of AlphaFold methods for modeling of alternative protein sates, conformational ensembles and mutation-induced structural transformations.

Integration of a Randomized Sequence Scanning Approach in AlphaFold2 and Local Frustration Profiling of Conformational States Enable Interpretable Atomistic Characterization of Conformational Ensembles and Detection of Hidden Allosteric States in the ABL1 Protein Kinase.

Despite the success of AlphaFold methods in predicting single protein structures, these methods showed intrinsic limitations in the characterization of multiple functional conformations of allosteric proteins. The recent NMR-based structural determination of the unbound ABL kinase in the active state and discovery of the inactive low-populated functional conformations that are unique for ABL kinase present an ideal challenge for the AlphaFold2 approaches. In the current study, we employ several adaptations of the AlphaFold2 methodology to predict protein conformational ensembles and allosteric states of the ABL kinase including randomized alanine sequence scanning combined with the multiple sequence alignment subsampling proposed in this study. We show that the proposed new AlphaFold2 adaptation combined with local frustration profiling of conformational states enables accurate prediction of the protein kinase structures and conformational ensembles, also offering a robust approach for interpretable characterization of the AlphaFold2 predictions and detection of hidden allosteric states. We found that the large high frustration residue clusters are uniquely characteristic of the low-populated, fully inactive ABL form and can define energetically frustrated cracking sites of conformational transitions, presenting difficult targets for AlphaFold2. The results of this study uncovered previously unappreciated fundamental connections between local frustration profiles of the functional allosteric states and the ability of AlphaFold2 methods to predict protein structural ensembles of the active and inactive states. This study showed that integration of the randomized sequence scanning adaptation of AlphaFold2 with a robust landscape-based analysis allows for interpretable atomistic predictions and characterization of protein conformational ensembles, providing a physical basis for the successes and limitations of current AlphaFold2 methods in detecting functional allosteric states that play a significant role in protein kinase regulation.

AHoJ-DB: A PDB-wide Assignment of apo & holo Relationships Based on Individual Protein–Ligand Interactions

A single protein structure is rarely sufficient to capture the conformational variability of a protein. Both bound and unbound (holo and apo) forms of a protein are essential for understanding its geometry and making meaningful comparisons. Nevertheless, docking or drug design studies often still consider only single protein structures in their holo form, which are for the most part rigid. With the recent explosion in the field of structural biology, large, curated datasets are urgently needed. Here, we use a previously developed application (AHoJ) to perform a comprehensive search for apo-holo pairs for 468,293 biologically relevant protein–ligand interactions across 27,983 proteins. In each search, the binding pocket is captured and mapped across existing structures within the same UniProt, and the mapped pockets are annotated as apo or holo, based on the presence or absence of ligands. We assemble the results into a database, AHoJ-DB (www.apoholo.cz/db), that captures the variability of proteins with identical sequences, thereby exposing the agents responsible for the observed differences in geometry. We report several metrics for each annotated pocket, and we also include binding pockets that form at the interface of multiple chains. Analysis of the database shows that about 24% of the binding sites occur at the interface of two or more chains and that less than 50% of the total binding sites processed have an apo form in the PDB. These results can be used to train and evaluate predictors, discover potentially druggable proteins, and reveal protein- and ligand-specific relationships that were previously obscured by intermittent or partial data.Availability: www.apoholo.cz/db

Toward physics-based precision medicine: Exploiting protein dynamics to design new therapeutics and interpret variants.

The goal of precision medicine is to utilize our knowledge of the molecular causes of disease to better diagnose and treat patients. However, there is a substantial mismatch between the small number of food and drug administration (FDA)-approved drugs and annotated coding variants compared to the needs of precision medicine. This review introduces the concept of physics-based precision medicine, a scalable framework that promises to improve our understanding of sequence-function relationships and accelerate drug discovery. We show that accounting for the ensemble of structures a protein adopts in solution with computer simulations overcomes many of the limitations imposed by assuming a single protein structure. We highlight studies of protein dynamics and recent methods for the analysis of structural ensembles. These studies demonstrate that differences in conformational distributions predict functional differences within protein families and between variants. Thanks to new computational tools that are providing unprecedented access to protein structural ensembles, this insight may enable accurate predictions of variant pathogenicity for entire libraries of variants. We further show that explicitly accounting for protein ensembles, with methods like alchemical free energy calculations or docking to Markov state models, can uncover novel lead compounds. To conclude, we demonstrate that cryptic pockets, or cavities absent in experimental structures, provide an avenue to target proteins that are currently considered undruggable. Taken together, our review provides a roadmap for the field of protein science to accelerate precision medicine.

Critical assessment of methods of protein structure prediction (CASP)-Round XV.

Computing protein structure from amino acid sequence information has been a long-standing grand challenge. Critical assessment of structure prediction (CASP) conducts community experiments aimed at advancing solutions to this and related problems. Experiments are conducted every 2 years. The 2020 experiment (CASP14) saw major progress, with the second generation of deep learning methods delivering accuracy comparable with experiment for many single proteins. There is an expectation that these methods will have much wider application in computational structural biology. Here we summarize results from the most recent experiment, CASP15, in 2022, with an emphasis on new deep learning-driven progress. Other papers in this special issue of proteins provide more detailed analysis. For single protein structures, the AlphaFold2 deep learning method is still superior to other approaches, but there are two points of note. First, although AlphaFold2 was the core of all the most successful methods, there was a wide variety of implementation and combination with other methods. Second, using the standard AlphaFold2 protocol and default parameters only produces the highest quality result for about two thirds of the targets, and more extensive sampling is required for the others. The major advance in this CASP is the enormous increase in the accuracy of computed protein complexes, achieved by the use of deep learning methods, although overall these do not fully match the performance for single proteins. Here too, AlphaFold2 based method perform best, and again more extensive sampling than the defaults is often required. Also of note are the encouraging early results on the use of deep learning to compute ensembles of macromolecular structures. Critically for the usability of computed structures, for both single proteins and protein complexes, deep learning derived estimates of both local and global accuracy are of high quality, however the estimates in interface regions are slightly less reliable. CASP15 also included computation of RNA structures for the first time. Here, the classical approaches produced better agreement with experiment than the new deep learning ones, and accuracy is limited. Also, for the first time, CASP included the computation of protein-ligand complexes, an area of special interest for drug design. Here too, classical methods were still superior to deep learning ones. Many new approaches were discussed at the CASP conference, and it is clear methods will continue to advance.

Insilico molecular modelling to identify PDK-1 targeting agents based on its protein–protein docking interaction

PDK1, an attractive cancer target that downstreams 23 other kinases towards cell growth, survival and metabolism has gaining attention due to allosteric effect of ligands bound to it. Generally, the drug design strategy using pharmacophores is either a single protein structure or ensemble or ligand-based. Apart from these methods, yet another new approach of protein–protein docking with state of art computational tool like Schrodinger Suite to generate pharmacophores based on the interacting partners of the protein is proposed in this work. The structure-based pharmacophoric features were picked up from docking the ten interacting partners of PDK1 and screened against the Enamine libraries containing protein–protein interacting compound collection, advanced, protein mimetic and allosteric compounds. High throughput virtual screening against the PIF pocket of PDK1 yields an indole scaffold. The identified indole derivative is proposed to be a strong activator that binds in the protein–protein interaction site of PDK1 which was further confirmed by molecular metadynamics simulations, free energy surface analysis and MM-GBSA calculations. Thus, the pharmacophores generated by the interacting proteins for PPI can facilitate the virtual screening in structure-based drug discovery of similar therapeutic targets. Communicated by Ramaswamy H. Sarma

Applications of bioinformatics in the post-genome era

Bioinformatics uses various informatics technical tools to study biological data that are difficult to process by general methods, to predict and compare biological information, etc. In the post-genomic era, the application of bioinformatics is highlighted and the methods of information processing have been expanded through innovation based on the original ones. In the post-genomic era, bioinformatics has attracted tremendous interest from medical and industrial communities worldwide, and it studies various biological genome sequences, single nucleotide polymorphism analysis, comparative genomics and protein structures, playing a profound role in biology, medicine, agriculture and information technology.

Expression of Cell-Adhesion Molecules in E. coli: A High Throughput Screening to Identify Paracellular Modulators.

Cell-adhesion molecules (CAMs) are responsible for cell-cell, cell-extracellular matrix, and cell-pathogen interactions. Claudins (CLDNs), occludin (OCLN), and junctional adhesion molecules (JAMs) are CAMs' components of the tight junction (TJ), the single protein structure tasked with safeguarding the paracellular space. The TJ is responsible for controlling paracellular permeability according to size and charge. Currently, there are no therapeutic solutions to modulate the TJ. Here, we describe the expression of CLDN proteins in the outer membrane of E. coli and report its consequences. When the expression is induced, the unicellular behavior of E. coli is replaced with multicellular aggregations that can be quantified using Flow Cytometry (FC). Our method, called iCLASP (inspection of cell-adhesion molecules aggregation through FC protocols), allows high-throughput screening (HTS) of small-molecules for interactions with CAMs. Here, we focused on using iCLASP to identify paracellular modulators for CLDN2. Furthermore, we validated those compounds in the mammalian cell line A549 as a proof-of-concept for the iCLASP method.

Abstract PL05-03: What if all your favorite proteins are druggable

Abstract Many proteins that play key roles in cancer are considered difficult drug targets, if not entirely undruggable, due to an apparent lack of pockets where small molecules can bind with the affinity and specificity required for a drug. ‘Cryptic’ pockets that are absent in known structures of proteins but form due to protein dynamics could alleviate this problem. However, it has been hard to assess or exploit the opportunities that cryptic pockets present due to the inherent difficulty in identifying such pockets. Here, I will discuss progress from my lab on combining biophysical experiments, computer simulations, and machine learning to identify and target cryptic pockets. In one case study, we have used computer simulations deployed on the Folding@home distributed computing environment to predict a cryptic pocket that can allosterically control a protein-RNA interaction, and then experimentally confirmed our predictions. Based on data from this study and others in our lab, we have trained a machine learning algorithm to predict where cryptic pockets are likely to form from single protein structures, enabling us to assess how prevalent cryptic pockets are on a large scale. Citation Format: Gregory R. Bowman. What if all your favorite proteins are druggable. [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2023; Part 1 (Regular and Invited Abstracts); 2023 Apr 14-19; Orlando, FL. Philadelphia (PA): AACR; Cancer Res 2023;83(7_Suppl):Abstract nr PL05-03.

Predicting locations of cryptic pockets from single protein structures using the PocketMiner graph neural network

Cryptic pockets expand the scope of drug discovery by enabling targeting of proteins currently considered undruggable because they lack pockets in their ground state structures. However, identifying cryptic pockets is labor-intensive and slow. The ability to accurately and rapidly predict if and where cryptic pockets are likely to form from a structure would greatly accelerate the search for druggable pockets. Here, we present PocketMiner, a graph neural network trained to predict where pockets are likely to open in molecular dynamics simulations. Applying PocketMiner to single structures from a newly curated dataset of 39 experimentally confirmed cryptic pockets demonstrates that it accurately identifies cryptic pockets (ROC-AUC: 0.87) >1,000-fold faster than existing methods. We apply PocketMiner across the human proteome and show that predicted pockets open in simulations, suggesting that over half of proteins thought to lack pockets based on available structures likely contain cryptic pockets, vastly expanding the potentially druggable proteome.

Predicting the locations of cryptic pockets from single protein structures using the PocketMiner graph neural network.

Cryptic pockets expand the scope of drug discovery by enabling targeting of proteins currently considered undruggable because they lack pockets in their ground state structures. However, identifying cryptic pockets is labor-intensive and slow. The ability to accurately and rapidly predict if and where cryptic pockets are likely to form from a protein structure would greatly accelerate the search for druggable pockets. Here, we present PocketMiner, a graph neural network trained to predict where pockets are likely to open in molecular dynamics simulations. Applying PocketMiner to single structures from a newly-curated dataset of 39 experimentally confirmed cryptic pockets demonstrates that it accurately identifies cryptic pockets (ROC-AUC: 0.87) >1,000-fold faster than existing methods. We apply PocketMiner across the human proteome and show that predicted pockets open in simulations, suggesting that over half of proteins thought to lack pockets based on available structures are likely to contain cryptic pockets, vastly expanding the druggable proteome.

Computational structural biology in the era of deep learning.

By the time of the BPS conference, it will be almost three years since CASP14 results showed the AlphaFold2 method from DeepMind is able to model protein structures with an accuracy that in many cases is competitive with the best experiments. Since then, there has been an explosion of work applying deep learning and related methods to other problems in computational structural biology as well as further development of single protein approaches. In this talk I'll examine where the field has now advanced to, drawing primarily on the results of the most recent CASP (CASP15, results of which will be public at the end of 2022). In particular, what, if any, are the remaining limitations for high accuracy modeling of single protein structures? How close are protein complexes models to the accuracy achieved for single proteins? Have the methods been successively extended to modeling RNA structures? Are protein-organic ligand complexes treatable? How far have we progressed towards generating ensembles of conformations, not just single structures? How seriously should we take the estimated accuracy of models? I’ll also attempt to relate progress to the astonishing diversification of deep learning methods that has occurred and what the results tell us about how the methods work: is it all just dumb pattern recognition, or are principles of protein structure learned and applied?

Investigation of Single and Multiple Mutations Prediction Using Binary Classification Approach

The mutation is a critical element in determining the proteins’ stability, becoming a core element in portraying the effects of a drug in the pharmaceutical industry. Doing wet laboratory tests to provide a better perspective on protein mutations is expensive and time-intensive since there are so many potential mutations, computational approaches that can reliably anticipate the consequences of amino acid mutations are critical. This work presents a robust methodology to analyze and identify the effects of mutation on a single protein structure. Initially, the context in a collection of words is determined using a knowledge graph for feature selection purposes. The proposed prediction is based on an easier and simpler logistic regression inferred binary classification technique. This approach can able to obtain a classification accuracy (AUC) Area Under the Curve of 87% when randomly validated against experimental energy changes. Moreover, for each cross-fold validation, the precision, recall, and F-Score are presented. These results support the validity of our strategy since it performs the vast majority of prior studies in this domain.

Entropies Derived from the Packing Geometries within a Single Protein Structure.

A fast, simple, yet robust method to calculate protein entropy from a single protein structure is presented here. The focus is on the atomic packing details, which are calculated by combining Voronoi diagrams and Delaunay tessellations. Even though the method is simple, the entropies computed exhibit an extremely high correlation with the entropies previously derived by other methods based on quasi-harmonic motions, quantum mechanics, and molecular dynamics simulations. These packing-based entropies account directly for the local freedom and provide entropy for any individual protein structure that could be used to compute free energies directly during simulations for the generation of more reliable trajectories and also for better evaluations of modeled protein structures. Physico-chemical properties of amino acids are compared with these packing entropies to uncover the relationships with the entropies of different residue types. A public packing entropy web server is provided at packing-entropy.bb.iastate.edu, and the application programing interface is available within the PACKMAN (https://github.com/Pranavkhade/PACKMAN) package.

Improved prediction of protein-protein interactions using AlphaFold2

Predicting the structure of interacting protein chains is a fundamental step towards understanding protein function. Unfortunately, no computational method can produce accurate structures of protein complexes. AlphaFold2, has shown unprecedented levels of accuracy in modelling single chain protein structures. Here, we apply AlphaFold2 for the prediction of heterodimeric protein complexes. We find that the AlphaFold2 protocol together with optimised multiple sequence alignments, generate models with acceptable quality (DockQ ≥ 0.23) for 63% of the dimers. From the predicted interfaces we create a simple function to predict the DockQ score which distinguishes acceptable from incorrect models as well as interacting from non-interacting proteins with state-of-art accuracy. We find that, using the predicted DockQ scores, we can identify 51% of all interacting pairs at 1% FPR.

Extended-ensemble docking to probe dynamic variation of ligand binding sites during large-scale structural changes of proteins.

Proteins can sample a broad landscape as they undergo conformational transition between different functional states. At the same time, as key players in almost all cellular processes, proteins are important drug targets. Considering the different conformational states of a protein is therefore central for a successful drug-design strategy. Here we introduce a novel docking protocol, termed extended-ensemble docking, pertaining to proteins that undergo large-scale (global) conformational changes during their function. In its application to multidrug ABC-transporter P-glycoprotein (Pgp), extensive non-equilibrium molecular dynamics simulations employing system-specific collective variables are first used to describe the transition cycle of the transporter. An extended set of conformations (extended ensemble) representing the full transition cycle between the inward- and the outward-facing states is then used to seed high-throughput docking calculations of known substrates, non-substrates, and modulators of the transporter. Large differences are predicted in the binding affinities to different conformations, with compounds showing stronger binding affinities to intermediate conformations compared to the starting crystal structure. Hierarchical clustering of the binding modes shows all ligands preferably bind to the large central cavity of the protein, formed at the apex of the transmembrane domain (TMD), whereas only small binding populations are observed in the previously described R and H sites present within the individual TMD leaflets. Based on the results, the central cavity is further divided into two major subsites, first preferably binding smaller substrates and high-affinity inhibitors, whereas the second one shows preference for larger substrates and low-affinity modulators. These central subsites along with the low-affinity interaction sites present within the individual TMD leaflets may respectively correspond to the proposed high- and low-affinity binding sites in Pgp. We propose further an optimization strategy for developing more potent inhibitors of Pgp, based on increasing its specificity to the extended ensemble of the protein, instead of using a single protein structure, as well as its selectivity for the high-affinity binding site. In contrast to earlier in silico studies using single static structures of Pgp, our results show better agreement with experimental studies, pointing to the importance of incorporating the global conformational flexibility of proteins in future drug-discovery endeavors.

Benchmarking Docking Protocols for Virtual Screenings of Novel Acetylcholinesterase Inhibitors

Molecular docking is emerging as a frequently applied structure-based virtual technique in the drug design processes. The method could significantly reduce the time required for the development of novel and effective molecules compared to high-throughput screening. However, a major drawback of the docking simulations is the high number of false-positive ligands in the top-ranked solutions. Thus, this work focuses on the optimization of genetic optimization for GOLD, and Glide docking protocols in the active sites of crystallographic acetylcholinesterase proteins, which could be used in the processes of design and optimization of novel acetylcholinesterase inhibitors. The performance of GOLD and Glide was assessed by their ability to reproduce known inhibitor conformations of co-crystallized ligands, and to efficiently detect known active compounds seeded into a decoy set. In addition, ensemble docking and molecular mechanics with generalised Born and surface area solvation (MM/GBSA) recalculations were introduced to observe the alterations in the enrichment factors. The variances in the enrichment values between both docking softwares were significant, considering the weak performance of Glide. In all of the employed crystallographic structures, GOLD 5.3 showed drastically better results. Interestingly, the enrichment factors were not increased after utilizing the ensemble docking simulations and the free binding energy recalculations with MM/GBSA. Overall, it was noted that the application of ChemPLP (GOLD 5.3) scoring function in a single acetylcholinesterase protein structure (PDB: 1Q84) displays the most reliable docking results. The obtained data will be beneficial for future virtual screenings of novel acetylcholinesterase inhibitors.

Highly accurate protein structure prediction with AlphaFold

Proteins are essential to life, and understanding their structure can facilitate a mechanistic understanding of their function. Through an enormous experimental effort1–4, the structures of around 100,000 unique proteins have been determined5, but this represents a small fraction of the billions of known protein sequences6,7. Structural coverage is bottlenecked by the months to years of painstaking effort required to determine a single protein structure. Accurate computational approaches are needed to address this gap and to enable large-scale structural bioinformatics. Predicting the three-dimensional structure that a protein will adopt based solely on its amino acid sequence—the structure prediction component of the ‘protein folding problem’8—has been an important open research problem for more than 50 years9. Despite recent progress10–14, existing methods fall far short of atomic accuracy, especially when no homologous structure is available. Here we provide the first computational method that can regularly predict protein structures with atomic accuracy even in cases in which no similar structure is known. We validated an entirely redesigned version of our neural network-based model, AlphaFold, in the challenging 14th Critical Assessment of protein Structure Prediction (CASP14)15, demonstrating accuracy competitive with experimental structures in a majority of cases and greatly outperforming other methods. Underpinning the latest version of AlphaFold is a novel machine learning approach that incorporates physical and biological knowledge about protein structure, leveraging multi-sequence alignments, into the design of the deep learning algorithm.