Machine learning in biological physics: From biomolecular prediction to design.

Machine learning has been proposed as an alternative to theoretical modeling when dealing with complex problems in biological physics. However, in this perspective, we argue that a more successful approach is a proper combination of these two methodologies. We discuss how ideas coming from physical modeling neuronal processing led to early formulations of computational neural networks, e.g., Hopfield networks. We then show how modern learning approaches like Potts models, Boltzmann machines, and the transformer architecture are related to each other, specifically, through a shared energy representation. We summarize recent efforts to establish these connections and provide examples on how each of these formulations integrating physical modeling and machine learning have been successful in tackling recent problems in biomolecular structure, dynamics, function, evolution, and design. Instances include protein structure prediction; improvement in computational complexity and accuracy of molecular dynamics simulations; better inference of the effects of mutations in proteins leading to improved evolutionary modeling and finally how machine learning is revolutionizing protein engineering and design. Going beyond naturally existing protein sequences, a connection to protein design is discussed where synthetic sequences are able to fold to naturally occurring motifs driven by a model rooted in physical principles. We show that this model is "learnable" and propose its future use in the generation of unique sequences that can fold into a target structure.

Reversible switching between two common protein folds in a designed system using only temperature.

Naturally occurring metamorphic proteins have the ability to interconvert from one folded state to another through either a limited set of mutations or by way of a change in the local environment. Here, we show in a designed system that it is possible to switch reversibly between two of the most common monomeric folds employing only temperature changes. We demonstrate that a latent 3α state can be unmasked from an α/ß-plait topology with a single V90T amino acid substitution, populating both forms simultaneously. The equilibrium between these two states exhibits temperature dependence, such that the 3α state is predominant (>90%) at 5 °C, while the α/ß-plait fold is the major species (>90%) at 30 °C. We describe the structure and dynamics of these topologies, how mutational changes affect the temperature dependence, and the energetics and kinetics of interconversion. Additionally, we demonstrate how ligand-binding function can be tightly regulated by large amplitude changes in protein structure over a relatively narrow temperature range that is relevant to biology. The 3α/αß switch thus represents a potentially useful approach for designing proteins that alter their fold topologies in response to environmental triggers. It may also serve as a model for computational studies of temperature-dependent protein stability and fold switching.

The Statistical Trends of Protein Evolution: A Lesson from AlphaFold Database.

The recent development of artificial intelligence provides us with new and powerful tools for studying the mysterious relationship between organism evolution and protein evolution. In this work, based on the AlphaFold Protein Structure Database (AlphaFold DB), we perform comparative analyses of the proteins of different organisms. The statistics of AlphaFold-predicted structures show that, for organisms with higher complexity, their constituent proteins will have larger radii of gyration, higher coil fractions, and slower vibrations, statistically. By conducting normal mode analysis and scaling analyses, we demonstrate that higher organismal complexity correlates with lower fractal dimensions in both the structure and dynamics of the constituent proteins, suggesting that higher functional specialization is associated with higher organismal complexity. We also uncover the topology and sequence bases of these correlations. As the organismal complexity increases, the residue contact networks of the constituent proteins will be more assortative, and these proteins will have a higher degree of hydrophilic-hydrophobic segregation in the sequences. Furthermore, by comparing the statistical structural proximity across the proteomes with the phylogenetic tree of homologous proteins, we show that, statistical structural proximity across the proteomes may indirectly reflect the phylogenetic proximity, indicating a statistical trend of protein evolution in parallel with organism evolution. This study provides new insights into how the diversity in the functionality of proteins increases and how the dimensionality of the manifold of protein dynamics reduces during evolution, contributing to the understanding of the origin and evolution of lives.

A drug discovery toolbox for Nuclear Magnetic Resonance (NMR) characterization of ligands and their targets.

Information about the structure, dynamics, and ligand-binding properties of biomolecules can be derived from Nuclear Magnetic Resonance (NMR) spectroscopy and provides valuable information for drug discovery. A multitude of experimental approaches provides a wealth of information that can be tailored to the system of interest. Methods to study the behavior of ligands upon target binding enable the identification of weak binders in a robust manner that is critical for the identification of truly novel binding interactions. This is particularly important for challenging targets. Observing the solution behavior of biomolecules yields information about their structure, dynamics, and interactions. This review describes the breadth of approaches that are available, many of which are under-utilized in a drug-discovery environment, and focuses on recent advances that continue to emerge.

DNP NMR of biomolecular assemblies.

Dynamic Nuclear Polarization (DNP) is an effective approach to alleviate the inherently low sensitivity of solid-state NMR (ssNMR) under magic angle spinning (MAS) towards large-sized multi-domain complexes and assemblies. DNP relies on a polarization transfer at cryogenic temperatures from unpaired electrons to adjacent nuclei upon continuous microwave irradiation. This is usually made possible via the addition in the sample of a polarizing agent. The first pioneering experiments on biomolecular assemblies were reported in the early 2000s on bacteriophages and membrane proteins. Since then, DNP has experienced tremendous advances, with the development of extremely efficient polarizing agents or with the introduction of new microwaves sources, suitable for NMR experiments at very high magnetic fields (currently up to 900 MHz). After a brief introduction, several experimental aspects of DNP enhanced NMR spectroscopy applied to biomolecular assemblies are discussed. Recent demonstration experiments of the method on viral capsids, the type III and IV bacterial secretion systems, ribosome and membrane proteins are then described.

Molecular differences between human liver fatty acid binding protein and its T94A variant in their unbound and lipid-bound states.

Liver fatty acid binding protein (L-FABP) is an abundant cytosolic protein playing a central role in intracellular lipid trafficking. The L-FABP T94A variant, originating from one of the most common polymorphisms in the FABP family, is associated with several lipid-related disorders. However, the molecular factors that determine the observed functional differences are currently unknown. In our work, we performed a high resolution comparative molecular analysis of L-FABP T94T and L-FABP T94A in their unbound states and in the presence of representative ligands of the fatty acid and bile acid classes. We collected residue-resolved NMR spectral fingerprints of the two variants, and compared secondary structures, backbone dynamics, side chain arrangements, binding site occupation, and intermolecular contacts. We found that threonine to alanine replacement did not result in strongly perturbed structural and dynamic features, although differences in oleic acid binding by the two variants were detected. Based on chemical shift perturbations at sites distant from position 94 and on differences in intermolecular contacts, we suggest that long-range communication networks in L-FABP propagate the effect of amino acid substitution at sites relevant for ligand binding or biomolecular recognition.

Channelrhodopsins: a bioinformatics perspective.

Channelrhodopsins are microbial-type rhodopsins that function as light-gated cation channels. Understanding how the detailed architecture of the protein governs its dynamics and specificity for ions is important, because it has the potential to assist in designing site-directed channelrhodopsin mutants for specific neurobiology applications. Here we use bioinformatics methods to derive accurate alignments of channelrhodopsin sequences, assess the sequence conservation patterns and find conserved motifs in channelrhodopsins, and use homology modeling to construct three-dimensional structural models of channelrhodopsins. The analyses reveal that helices C and D of channelrhodopsins contain Cys, Ser, and Thr groups that can engage in both intra- and inter-helical hydrogen bonds. We propose that these polar groups participate in inter-helical hydrogen-bonding clusters important for the protein conformational dynamics and for the local water interactions. This article is part of a Special Issue entitled: Retinal Proteins - You can teach an old dog new tricks.

Probing Antibody Structures by Hydrogen/Deuterium Exchange Mass Spectrometry.

Hydrogen/deuterium exchange (HDX) followed by mass spectrometry detection (MS) provides a fast, reliable, and detailed solution for the assessment of a protein structure. It has been widely recognized as an indispensable tool and already approved by several regulatory agencies as a structural technique for the validation of protein biopharmaceuticals, including antibody-based drugs. Antibodies are of a key importance in life and medical sciences but considered to be challenging analytical targets because of their compact structure stabilized by disulfide bonds and due to the presence of glycosylation. Despite these difficulties, there are already numerous excellent studies describing MS-based antibody structure characterization. In this chapter, we describe a universal HDX-MS workflow. Deeper attention is paid to sample handling, optimization procedures, and feasibility stages, as these elements of the HDX experiment are crucial for obtaining reliable detailed and spatially well-resolved information.

Structural characterization of human tryptophan hydroxylase 2 reveals that L-Phe is superior to L-Trp as the regulatory domain ligand.

Tryptophan hydroxylase 2 (TPH2) catalyzes the rate-limiting step in serotonin biosynthesis in the brain. Consequently, regulation of TPH2 is relevant for serotonin-related diseases, yet the regulatory mechanism of TPH2 is poorly understood and structural and dynamical insights are missing. We use NMR spectroscopy to determine the structure of a 47 N-terminally truncated variant of the regulatory domain (RD) dimer of human TPH2 in complex with L-Phe, and show that L-Phe is the superior RD ligand compared with the natural substrate, L-Trp. Using cryo-EM, we obtain a low-resolution structure of a similarly truncated variant of the complete tetrameric enzyme with dimerized RDs. The cryo-EM two-dimensional (2D) class averages additionally indicate that the RDs are dynamic in the tetramer and likely exist in a monomer-dimer equilibrium. Our results provide structural information on the RD as an isolated domain and in the TPH2 tetramer, which will facilitate future elucidation of TPH2's regulatory mechanism.

Triple-action inhibitory mechanism of allosteric TYK2-specific inhibitors.

Deucravacitinib, 6-(cyclopropanecarbonylamido)-4-[2-methoxy-3-(1-methyl-1,2,4-triazol-3-yl)anilino]-N-(trideuteriomethyl)pyridazine-3-carboxamide, is a highly selective inhibitor of protein tyrosine kinase 2 (TYK2) that targets the Janus kinase (JAK)-signal transducer and activator of transcription (STAT) pathway. The structural basis for its selectivity and allosteric inhibition remains poorly understood. Here, we investigate the inhibition mechanism through analysis of available structures relevant to the STAT pathway, including crystal structures of the truncated TYK2 FERM-SH2 domain bound to the IFNα type I receptor (IFNαR1) and the truncated TYK2 JH2-JH1 domain. Our computational analysis provides a mechanistic hypothesis for the relatively rapid interferon-induced gene expression mediated by TYK2 relative to other cytokines. We find that deucravacitinib inhibits TYK2 kinase in three distinct states: the autoinhibited state and two activated states for autophosphorylation and phosphorylation of downstream protein substrates. Its binding to the TYK2 pseudokinase domain in the autoinhibited state restricts the essential dynamics of the TYK2 kinase domain required for kinase activity. Furthermore, it binds competitively with ATP in the pseudokinase domain, and also directly prevents formation of the active state of TYK2 through steric clashes.

Molecular rationale for the impairment of the MexAB-OprM efflux pump by a single mutation in MexA.

Efflux pumps of the Resistance-Nodulation-cell Division (RND) superfamily contribute to intrinsic and acquired resistance in Gram-negative pathogens by expelling chemically unrelated antibiotics with high efficiency. They are tripartite systems constituted by an inner-membrane-anchored transporter, an outer membrane factor protein, and a membrane fusion protein. Multimerization of the membrane fusion protein is an essential prerequisite for full functionality of these efflux pumps. In this work, we employed complementary computational techniques to investigate the stability of a dimeric unit of MexA (the membrane fusion protein of the MexAB-OprM RND efflux pump of Pseudomonas aeruginosa), and to provide a molecular rationale for the effect of the G72S substitution, which affects MexAB-OprM functionality by impairing the assembly of MexA. Our findings indicate that: i) dimers of this protein are stable in multiple µs-long molecular dynamics simulations; ii) the mutation drastically alters the conformational equilibrium of MexA, favouring a collapsed conformation that is unlikely to form dimers or higher order assemblies. Unveiling the mechanistic aspects underlying large conformational distortions induced by minor sequence changes is informative to efforts at interfering with the activity of this elusive bacterial weapon. In this respect, our work further confirms how molecular simulations can give important contribution and useful insights to characterize the mechanism of highly complex biological systems.

Solution NMR backbone assignments of the N-terminal Zα-linker-Zß segment from Homo sapiens ADAR1p150.

Adenosine-to-inosine (A-to-I) editing of a subset of RNAs in a eukaryotic cell is required in order to avoid triggering the innate immune system. Editing is carried out by ADAR1, which exists as short (p110) and long (p150) isoforms. ADAR1p150 is mostly cytoplasmic, possesses a Z-RNA binding domain (Zα), and is only expressed during the innate immune response. A structurally homologous domain to Zα, the Zß domain, is separated by a long linker from Zα on the N-terminus of ADAR1 but its function remains unknown. Zß does not bind to RNA in isolation, yet the binding kinetics of the segment encompassing Zα, Zß and the 95-residue linker between the two domains (Zα-Zß) are markedly different compared to Zα alone. Here we present the solution NMR backbone assignment of Zα-Zß from H. Sapiens ADAR1. The predicted secondary structure of Zα-Zß based on chemical shifts is in agreement with previously determined structures of Zα and Zß in isolation, and indicates that the linker is intrinsically disordered. Comparison of the chemical shifts between the individual Zα and Zß domains to the full Zα-Zß construct suggests that Zß may interact with the linker, the function of which is currently unknown.

The Biophysical Basis for Phosphorylation-Enhanced DNA-Binding Autoinhibition of the ETS1 Transcription Factor.

The eukaryotic transcription factor ETS1 is regulated by an intrinsically disordered serine-rich region (SRR) that transiently associates with the adjacent ETS domain to inhibit DNA binding. In this study, we further elucidated the physicochemical basis for ETS1 autoinhibition by characterizing the interaction of its ETS domain with a series of synthetic peptides corresponding to the SRR. Binding is driven by the hydrophobic effect and enhanced electrostatically by phosphorylation of serines adjacent to aromatic residues in the amphipathic SRR. Structural characterization of the dynamic peptide/protein complex by NMR spectroscopy and X-ray crystallography revealed multiple modes of binding that lead to autoinhibition by synergistically blocking the DNA-binding interface of the ETS domain and stabilizing an appended helical inhibitory module against allosterically induced unfolding. Consistent with these conclusions, the SRR peptide does not interact with DNA-bound ETS1. In addition, we found that the ETS1 SRR phosphopeptide binds to distantly related PU.1 in vitro, indicating that autoinhibition exploits features of the ETS domain that are conserved across this family of transcription factors.

Two decades of progress in structural and dynamic studies of amyloids by solid-state NMR.

In this perspective article I briefly highlight the rapid progress made over the past two decades in atomic level structural and dynamic studies of amyloids, which are representative of non-crystalline biomacromolecular assemblies, by magic-angle spinning solid-state NMR spectroscopy. Given new and continuing developments in solid-state NMR instrumentation and methodology, ongoing research in this area promises to contribute to an improved understanding of amyloid structure, polymorphism, interactions, assembly mechanisms, and biological function and toxicity.

Surface-Binding to Cardiolipin Nanodomains Triggers Cytochrome c Pro-apoptotic Peroxidase Activity via Localized Dynamics.

The peroxidation of cardiolipins by reactive oxygen species, which is regulated and enhanced by cytochrome c (cyt c), is a critical signaling event in mitochondrial apoptosis. We probe the molecular underpinnings of this mitochondrial death signal through structural and functional studies of horse heart cyt c binding to mixed-lipid membranes containing cardiolipin with mono- and polyunsaturated acyl chains. Lipidomics reveal the selective oxidation of polyunsaturated fatty acid (PUFA) cardiolipin (CL), while multidimensional solid-state NMR probes the structure and dynamics of the membrane and the peripherally bound protein. The hydrophilic milieu at the membrane interface stabilizes a native-like fold, but also leads to localized flexibility at the membrane-interacting protein face. PUFA CL acts as both a preferred substrate and a dynamic regulator by affecting the dynamics of the cyt c N70-I85 Ω loop, which covers the heme cavity.

Ensemble MD simulations restrained via crystallographic data: accurate structure leads to accurate dynamics.

Currently, the best existing molecular dynamics (MD) force fields cannot accurately reproduce the global free-energy minimum which realizes the experimental protein structure. As a result, long MD trajectories tend to drift away from the starting coordinates (e.g., crystallographic structures). To address this problem, we have devised a new simulation strategy aimed at protein crystals. An MD simulation of protein crystal is essentially an ensemble simulation involving multiple protein molecules in a crystal unit cell (or a block of unit cells). To ensure that average protein coordinates remain correct during the simulation, we introduced crystallography-based restraints into the MD protocol. Because these restraints are aimed at the ensemble-average structure, they have only minimal impact on conformational dynamics of the individual protein molecules. So long as the average structure remains reasonable, the proteins move in a native-like fashion as dictated by the original force field. To validate this approach, we have used the data from solid-state NMR spectroscopy, which is the orthogonal experimental technique uniquely sensitive to protein local dynamics. The new method has been tested on the well-established model protein, ubiquitin. The ensemble-restrained MD simulations produced lower crystallographic R factors than conventional simulations; they also led to more accurate predictions for crystallographic temperature factors, solid-state chemical shifts, and backbone order parameters. The predictions for (15) N R1 relaxation rates are at least as accurate as those obtained from conventional simulations. Taken together, these results suggest that the presented trajectories may be among the most realistic protein MD simulations ever reported. In this context, the ensemble restraints based on high-resolution crystallographic data can be viewed as protein-specific empirical corrections to the standard force fields.