Predicting Proteome-Scale Protein Structure with Artificial Intelligence

Future for cardiogenic shock research

PURPOSE OF REVIEW

To discuss future research themes and study design in cardiogenic shock.

RECENT FINDINGS

Cardiogenic shock research faces multiple challenges, hindering progress in understanding and treating this life-threatening condition. Cardiogenic shock's heterogeneous nature poses challenges in patient selection for clinical trials, potentially leading to variability in treatment responses and outcomes. Ethical considerations arise due to the acuity and severity of the condition, posing challenges in obtaining informed consent and conducting randomized controlled trials where time to treatment is pivotal.

SUMMARY

This review discusses research in this area focusing on the importance of phenotyping patients with cardiogenic shock, based on artificial intelligence, machine learning, and unravel new molecular mechanisms using proteomics and metabolomics. Further, the future research focus in mechanical circulatory support and targeting inflammation is reviewed. Finally, newer trial designs including adaptive platform trials are discussed.

Generative artificial intelligence performs rudimentary structural biology modeling

Natural language-based generative artificial intelligence (AI) has become increasingly prevalent in scientific research. Intriguingly, capabilities of generative pre-trained transformer (GPT) language models beyond the scope of natural language tasks have recently been identified. Here we explored how GPT-4 might be able to perform rudimentary structural biology modeling. We prompted GPT-4 to model 3D structures for the 20 standard amino acids and an α-helical polypeptide chain, with the latter incorporating Wolfram mathematical computation. We also used GPT-4 to perform structural interaction analysis between nirmatrelvir and its target, the SARS-CoV-2 main protease. Geometric parameters of the generated structures typically approximated close to experimental references. However, modeling was sporadically error-prone and molecular complexity was not well tolerated. Interaction analysis further revealed the ability of GPT-4 to identify specific amino acid residues involved in ligand binding along with corresponding bond distances. Despite current limitations, we show the capacity of natural language generative AI to perform basic structural biology modeling and interaction analysis with atomic-scale accuracy.

Lysine lactylation in the regulation of tumor biology

Lysine lactylation (Kla), a newly discovered post-translational modification (PTM) of lysine residues, is progressively revealing its crucial role in tumor biology. A growing body of evidence supports its capacity of transcriptional regulation through histone modification and modulation of non-histone protein function. It intricately participates in a myriad of events in the tumor microenvironment (TME) by orchestrating the transitions of immune states and augmenting tumor malignancy. Its preferential modification of metabolic proteins underscores its specific regulatory influence on metabolism. This review focuses on the effect and the probable mechanisms of Kla-mediated regulation of tumor metabolism, the upstream factors that determine Kla intensity, and its potential implications for the clinical diagnosis and treatment of tumors.

SIGMA leverages protein structural information to predict the pathogenicity of missense variants

Leveraging protein structural information to evaluate pathogenicity has been hindered by the scarcity of experimentally determined 3D protein. With the aid of AlphaFold2 predictions, we developed the structure-informed genetic missense mutation assessor (SIGMA) to predict missense variant pathogenicity. In comparison with existing predictors across labeled variant datasets and experimental datasets, SIGMA demonstrates superior performance in predicting missense variant pathogenicity (AUC = 0.933). We found that the relative solvent accessibility of the mutated residue contributed greatly to the predictive ability of SIGMA. We further explored combining SIGMA with other top-tier predictors to create SIGMA+, proving highly effective for variant pathogenicity prediction (AUC = 0.966). To facilitate the application of SIGMA, we pre-computed SIGMA scores for over 48 million possible missense variants across 3,454 disease-associated genes and developed an interactive online platform (https://www.sigma-pred.org/). Overall, by leveraging protein structure information, SIGMA offers an accurate structure-based approach to evaluating the pathogenicity of missense variants.

Structural Evolution of Primate Glutamate Dehydrogenase 2 as Revealed by In Silico Predictions and Experimentally Determined Structures

Glutamate dehydrogenase (GDH) interconverts glutamate to a-ketoglutarate and ammonia, interconnecting amino acid and carbohydrate metabolism. In humans, two functional GDH genes, GLUD1 and GLUD2, encode for hGDH1 and hGDH2, respectively. GLUD2 evolved from retrotransposition of the GLUD1 gene in the common ancestor of modern apes. These two isoenzymes are involved in the pathophysiology of human metabolic, neoplastic, and neurodegenerative disorders. The 3D structures of hGDH1 and hGDH2 have been experimentally determined; however, no information is available about the path of GDH2 structure changes during primate evolution. Here, we compare the structures predicted by the AlphaFold Colab method for the GDH2 enzyme of modern apes and their extinct primate ancestors. Also, we analyze the individual effect of amino acid substitutions emerging during primate evolution. Our most important finding is that the predicted structure of GDH2 in the common ancestor of apes was the steppingstone for the structural evolution of primate GDH2s. Two changes with a strong functional impact occurring at the first evolutionary step, Arg443Ser and Gly456Ala, had a destabilizing and stabilizing effect, respectively, making this step the most important one. Subsequently, GDH2 underwent additional modifications that fine-tuned its enzymatic properties to adapt to the functional needs of modern-day primate tissues.

Modeling of ACE2 and antibodies bound to SARS-CoV-2 provides insights into infectivity and immune evasion

Given the COVID-19 pandemic, there is interest in understanding ligand-receptor features and targeted antibody-binding attributes against emerging SARS-CoV-2 variants. Here, we developed a large-scale structure-based pipeline for analysis of protein-protein interactions regulating SARS-CoV-2 immune evasion. First, we generated computed structural models of the Spike protein of 3 SARS-CoV-2 variants (B.1.1.529, BA.2.12.1, and BA.5) bound either to a native receptor (ACE2) or to a large panel of targeted ligands (n = 282), which included neutralizing or therapeutic monoclonal antibodies. Moreover, by using the Barnes classification, we noted an overall loss of interfacial interactions (with gain of new interactions in certain cases) at the receptor-binding domain (RBD) mediated by substituted residues for neutralizing complexes in classes 1 and 2, whereas less destabilization was observed for classes 3 and 4. Finally, an experimental validation of predicted weakened therapeutic antibody binding was performed in a cell-based assay. Compared with the original Omicron variant (B.1.1.529), derivative variants featured progressive destabilization of antibody-RBD interfaces mediated by a larger set of substituted residues, thereby providing a molecular basis for immune evasion. This approach and findings provide a framework for rapidly and efficiently generating structural models for SARS-CoV-2 variants bound to ligands of mechanistic and therapeutic value.

One half-century of advances in the evaluation and management of disorders of bone and mineral metabolism in children and adolescents

The past 50 years of research in pediatric bone and mineral metabolism have led to remarkable progress in the identification and characterization of disorders that affect the developing skeleton. Progress has been facilitated through advances in both technology and biology and this paper provides a brief description of some but not all of the key findings, including identification of the calcium sensing receptor and the polypeptides parathyroid hormone and parathyroid hormone-related protein as well as their shared receptor and signal generating pathways; the elucidation of vitamin D metabolism and actions; discovery of fibroblast growth factor 23 (FGF23), the sodium-phosphate co-transporters and the other components that regulate phosphate metabolism. Moreover, the past half-century of research has led to the delineation of the molecular bases for genetic forms of hypoparathyroidism, pseudohypoparathyroidism, and primary hyperparathyroidism as well as the determination of the genetic causes of osteogenesis imperfecta, osteopetrosis, hypophosphatasia, and other disorders of mineral/bone homeostasis. During the next decade we expect that many of these fundamental discoveries will lead to the development of innovative treatments that will improve the lives of children with these disorders.

The Epigenetic Dimension of Protein Structure Is an Intrinsic Weakness of the AlphaFold Program

One of the most important lessons we have learned from sequencing the human genome is that not all proteins have a 3D structure. In fact, a large part of the human proteome is made up of intrinsically disordered proteins (IDPs) which can adopt multiple structures, and therefore, multiple functions, depending on the ligands with which they interact. Under these conditions, one can wonder about the value of algorithms developed for predicting the structure of proteins, in particular AlphaFold, an AI which claims to have solved the problem of protein structure. In a recent study, we highlighted a particular weakness of AlphaFold for membrane proteins. Based on this observation, we have proposed a paradigm, referred to as “Epigenetic Dimension of Protein Structure” (EDPS), which takes into account all environmental parameters that control the structure of a protein beyond the amino acid sequence (hence “epigenetic”). In this new study, we compare the reliability of the AlphaFold and Robetta algorithms’ predictions for a new set of membrane proteins involved in human pathologies. We found that Robetta was generally more accurate than AlphaFold for ascribing a membrane-compatible topology. Raft lipids (e.g., gangliosides), which control the structural dynamics of membrane protein structure through chaperone effects, were identified as major actors of the EDPS paradigm. We conclude that the epigenetic dimension of a protein structure is an intrinsic weakness of AI-based protein structure prediction, especially AlphaFold, which warrants further development.

Assessing PDB macromolecular crystal structure confidence at the individual amino acid residue level

Approximately 87% of the more than 190,000 atomic-level three-dimensional (3D) biostructures in the PDB were determined using macromolecular crystallography (MX). Agreement between 3D atomic coordinates and experimental data for >100 million individual amino acid residues occurring within ∼150,000 PDB MX structures was analyzed in detail. The real-space correlation coefficient (RSCC) calculated using the 3D atomic coordinates for each residue and experimental-data-derived electron density enables outlier detection of unreliable atomic coordinates (particularly important for poorly resolved side-chain atoms) and ready evaluation of local structure quality by PDB users. For human protein MX structures in PDB, comparisons of the per-residue RSCC metric with AlphaFold2-computed structure model confidence (pLDDT-predicted local distance difference test) document (1) that RSCC values and pLDDT scores are correlated (median correlation coefficient ∼0.41), and (2) that experimentally determined MX structures (3.5 Å resolution or better) are more reliable than AlphaFold2-computed structure models and should be used preferentially whenever possible.

Protein Data Bank: A Comprehensive Review of 3D Structure Holdings and Worldwide Utilization by Researchers, Educators, and Students

The Research Collaboratory for Structural Bioinformatics Protein Data Bank (RCSB PDB), funded by the United States National Science Foundation, National Institutes of Health, and Department of Energy, supports structural biologists and Protein Data Bank (PDB) data users around the world. The RCSB PDB, a founding member of the Worldwide Protein Data Bank (wwPDB) partnership, serves as the US data center for the global PDB archive housing experimentally-determined three-dimensional (3D) structure data for biological macromolecules. As the wwPDB-designated Archive Keeper, RCSB PDB is also responsible for the security of PDB data and weekly update of the archive. RCSB PDB serves tens of thousands of data depositors (using macromolecular crystallography, nuclear magnetic resonance spectroscopy, electron microscopy, and micro-electron diffraction) annually working on all permanently inhabited continents. RCSB PDB makes PDB data available from its research-focused web portal at no charge and without usage restrictions to many millions of PDB data consumers around the globe. It also provides educators, students, and the general public with an introduction to the PDB and related training materials through its outreach and education-focused web portal. This review article describes growth of the PDB, examines evolution of experimental methods for structure determination viewed through the lens of the PDB archive, and provides a detailed accounting of PDB archival holdings and their utilization by researchers, educators, and students worldwide.

Comparative studies of AlphaFold, RoseTTAFold and Modeller: a case study involving the use of G-protein-coupled receptors

Neural network (NN)-based protein modeling methods have improved significantly in recent years. Although the overall accuracy of the two non-homology-based modeling methods, AlphaFold and RoseTTAFold, is outstanding, their performance for specific protein families has remained unexamined. G-protein-coupled receptor (GPCR) proteins are particularly interesting since they are involved in numerous pathways. This work directly compares the performance of these novel deep learning-based protein modeling methods for GPCRs with the most widely used template-based software—Modeller. We collected the experimentally determined structures of 73 GPCRs from the Protein Data Bank. The official AlphaFold repository and RoseTTAFold web service were used with default settings to predict five structures of each protein sequence. The predicted models were then aligned with the experimentally solved structures and evaluated by the root-mean-square deviation (RMSD) metric. If only looking at each program’s top-scored structure, Modeller had the smallest average modeling RMSD of 2.17 Å, which is better than AlphaFold’s 5.53 Å and RoseTTAFold’s 6.28 Å, probably since Modeller already included many known structures as templates. However, the NN-based methods (AlphaFold and RoseTTAFold) outperformed Modeller in 21 and 15 out of the 73 cases with the top-scored model, respectively, where no good templates were available for Modeller. The larger RMSD values generated by the NN-based methods were primarily due to the differences in loop prediction compared to the crystal structures.

Genetic and Structural Analysis of SARS-CoV-2 Spike Protein for Universal Epitope Selection

Evaluation of immunogenic epitopes for universal vaccine development in the face of ongoing SARS-CoV-2 evolution remains a challenge. Herein, we investigate the genetic and structural conservation of an immunogenically relevant epitope (C662-C671) of spike (S) protein across SARS-CoV-2 variants to determine its potential utility as a broad-spectrum vaccine candidate against coronavirus diseases. Comparative sequence analysis, structural assessment, and molecular dynamics simulations of C662-C671 epitope were performed. Mathematical tools were employed to determine its mutational cost. We found that the amino acid sequence of C662-C671 epitope is entirely conserved across the observed major variants of SARS-CoV-2 in addition to SARS-CoV. Its conformation and accessibility are predicted to be conserved, even in the highly mutated Omicron variant. Costly mutational rate in the context of energy expenditure in genome replication and translation can explain this strict conservation. These observations may herald an approach to developing vaccine candidates for universal protection against emergent variants of coronavirus.

Structural models of SARS-CoV-2 Omicron variant in complex with ACE2 receptor or antibodies suggest altered binding interfaces

There is enormous ongoing interest in characterizing the binding properties of the SARS-CoV-2 Omicron Variant of Concern (VOC) (B.1.1.529), which continues to spread towards potential dominance worldwide. To aid these studies, based on the wealth of available structural information about several SARS-CoV-2 variants in the Protein Data Bank (PDB) and a modeling pipeline we have previously developed for tracking the ongoing global evolution of SARS-CoV-2 proteins, we provide a set of computed structural models (henceforth models) of the Omicron VOC receptor-binding domain (omRBD) bound to its corresponding receptor Angiotensin-Converting Enzyme (ACE2) and a variety of therapeutic entities, including neutralizing and therapeutic antibodies targeting previously-detected viral strains. We generated bound omRBD models using both experimentally-determined structures in the PDB as well as machine learningbased structure predictions as starting points. Examination of ACE2-bound omRBD models reveals an interdigitated multi-residue interaction network formed by omRBD-specific substituted residues (R493, S496, Y501, R498) and ACE2 residues at the interface, which was not present in the original Wuhan-Hu-1 RBD-ACE2 complex. Emergence of this interaction network suggests optimization of a key region of the binding interface, and positive cooperativity among various sites of residue substitutions in omRBD mediating ACE2 binding. Examination of neutralizing antibody complexes for Barnes Class 1 and Class 2 antibodies modeled with omRBD highlights an overall loss of interfacial interactions (with gain of new interactions in rare cases) mediated by substituted residues. Many of these substitutions have previously been found to independently dampen or even ablate antibody binding, and perhaps mediate antibody-mediated neutralization escape ( e.g ., K417N). We observe little compensation of corresponding interaction loss at interfaces when potential escape substitutions occur in combination. A few selected antibodies ( e.g ., Barnes Class 3 S309), however, feature largely unaltered or modestly affected protein-protein interfaces. While we stress that only qualitative insights can be obtained directly from our models at this time, we anticipate that they can provide starting points for more detailed and quantitative computational characterization, and, if needed, redesign of monoclonal antibodies for targeting the Omicron VOC Spike protein. In the broader context, the computational pipeline we developed provides a framework for rapidly and efficiently generating retrospective and prospective models for other novel variants of SARS-CoV-2 bound to entities of virological and therapeutic interest, in the setting of a global pandemic.