Machine Learning in Complex Organic Mixtures: Applying Domain Knowledge Allows for Meaningful Performance with Small Data Sets.

The ability to quantify individual components of complex mixtures is a challenge found throughout the life and physical sciences. An improved capacity to generate large data sets along with the uptake of machine-learning (ML)-based analysis tools has allowed for various "omics" disciplines to realize exceptional advances. Other areas of chemistry that deal with complex mixtures often do not leverage these advances. Environmental samples, for example, can be more difficult to access, and the resulting small data sets are less appropriate for unconstrained ML approaches. Herein, we present an approach to address this latter issue. Using a very small environmental data set─35 high-resolution mass spectra gathered from various solvent extractions of Canadian petroleum fractions─we show that the application of specific domain knowledge can lead to ML models with notable performance.

Understanding FABP7 binding to fatty acid micelles and membranes.

Fatty acid-binding proteins (FABPs) are chaperones that facilitate the transport of long-chain fatty acids within the cell and can provide cargo-dependent localization to specific cellular compartments. Understanding the nature of this transport is important because lipid signaling functions are associated with metabolic pathways impacting disease pathologies including cancer, autism, and schizophrenia. FABPs often associate with cell membranes to acquire and deliver their bound cargo as part of transport. We focus on brain FABP (FABP7), which demonstrates localization to the cytoplasm and nucleus, influencing transcription and fatty acid metabolism. We use a combined biophysical-computational approach to elucidate the interaction between FABP7 and model membranes. Specifically, we use multiple experiments to demonstrate that FABP7 can bind oleic acid and docosahexaenoic acid micelles. Data from NMR and multiscale molecular dynamics simulations reveal that the interaction with micelles is through FABP7's portal region residues. Simulations suggest that binding to membranes occurs through the same residues as micelles. Simulations also capture binding events where fatty acids dissociate from the membrane and enter FABP7's binding pocket. Overall, our data shed light on the interactions between FABP7 and OA or DHA micelles and provide insight into the transport of long-chain fatty acids.

A combined computational-biophysical approach to understanding fatty acid binding to FABP7.

Members of the fatty acid binding protein (FABP) family function as intracellular transporters of long-chain fatty acids and other hydrophobic molecules to different cellular compartments. Brain FABP (FABP7) exhibits ligand-directed differences in cellular transport. For example, when FABP7 binds to docosahexaenoic acid (DHA), the complex relocates to the nucleus and influences transcriptional activity, whereas FABP7 bound with monosaturated fatty acids remains in the cytosol. Preferential binding of FABP7 to polyunsaturated fatty acids like DHA has been previously observed and is thought to play a role in differential localization. However, we find that at 37°C, FABP7 does not display strong selectivity, suggesting that the conformational ensemble of FABP7 and its perturbation upon binding may be important. We use molecular dynamics simulations, NMR, and a variety of biophysical techniques to better understand the conformational ensemble of FABP7, how it is perturbed by fatty acid binding, and how this may be related to ligand-directed transport. We find that FABP7 has high degree of conformational heterogeneity that is substantially reduced upon ligand binding. We also observe substantial heterogeneity in ligand binding poses, which is consistent with our finding that ligand binding is resistant to mutations in key polar residues in the binding pocket. Our NMR experiments show that DHA binding leads to chemical shift perturbations in residues near the nuclear localization signal, which may point toward a mechanism of differential transport.

An implementation of the Martini coarse-grained force field in OpenMM.

We describe a complete implementation of Martini 2 and Martini 3 in the OpenMM molecular dynamics software package. Martini is a widely used coarse-grained force field with applications in biomolecular simulation, materials, and broader areas of chemistry. It is implemented as a force field but makes extensive use of facilities unique to the GROMACS software, including virtual sites and bonded terms that are not commonly used in standard atomistic force fields. OpenMM is a flexible molecular dynamics package widely used for methods development and is competitive in speed on GPUs with other commonly used packages. OpenMM has facilities to easily implement new force field terms, external forces and fields, and other nonstandard features, which we use to implement all force field terms used in Martini 2 and Martini 3. This allows Martini simulations, starting with GROMACS topology files that are processed by custom scripts, with all the added flexibility of OpenMM. We provide a GitHub repository with test cases, compare accuracy and performance between GROMACS and OpenMM, and discuss the limitations of our implementation in terms of direct comparison with GROMACS. We describe a use case that implements the Modeling Employing Limited Data method to apply experimental constraints in a Martini simulation to efficiently determine the structure of a protein complex. We also discuss issues and a potential solution with the Martini 2 topology for cholesterol.

CryoFold 2.0: Cryo-EM Structure Determination with MELD.

Cryo-electron microscopy data are becoming more prevalent and accessible at higher resolution levels, leading to the development of new computational tools to determine the atomic structure of macromolecules. However, while existing tools adapted from X-ray crystallography are suitable for the highest-resolution maps, new tools are needed for lower-resolution levels and to account for map heterogeneity. In this article, we introduce CryoFold 2.0, an integrative physics-based approach that combines Bayesian inference and the ability to handle multiple data sources with the molecular dynamics flexible fitting (MDFF) approach to determine the structures of macromolecules by using cryo-EM data. CryoFold 2.0 is incorporated into the MELD (modeling employing limited data) plugin, resulting in a pipeline that is more computationally efficient and accurate than running MELD or MDFF alone. The approach requires fewer computational resources and shorter simulation times than the original CryoFold, and it minimizes manual intervention. We demonstrate the effectiveness of the approach on eight different systems, highlighting its various benefits.

High Accuracy Protein Structures from Minimal Sparse Paramagnetic Solid-State NMR Restraints.

There is a pressing need for new computational tools to integrate data from diverse experimental approaches in structural biology. We present a strategy that combines sparse paramagnetic solid-state NMR restraints with physics-based atomistic simulations. Our approach explicitly accounts for uncertainty in the interpretation of experimental data through the use of a semi-quantitative mapping between the data and the restraint energy that is calibrated by extensive simulations. We apply our approach to solid-state NMR data for the model protein GB1 labeled with Cu2+ -EDTA at six different sites. We are able to determine the structure to 0.9 Šaccuracy within a single day of computation on a GPU cluster. We further show that in some cases, the data from only a single paramagnetic tag are sufficient for accurate folding.

Accelerating molecular simulations of proteins using Bayesian inference on weak information.

Atomistic molecular dynamics (MD) simulations of protein molecules are too computationally expensive to predict most native structures from amino acid sequences. Here, we integrate "weak" external knowledge into folding simulations to predict protein structures, given their sequence. For example, we instruct the computer "to form a hydrophobic core," "to form good secondary structures," or "to seek a compact state." This kind of information has been too combinatoric, nonspecific, and vague to help guide MD simulations before. Within atomistic replica-exchange molecular dynamics (REMD), we develop a statistical mechanical framework, modeling using limited data with coarse physical insight(s) (MELD + CPI), for harnessing weak information. As a test, we apply MELD + CPI to predict the native structures of 20 small proteins. MELD + CPI samples to within less than 3.2 Å from native for all 20 and correctly chooses the native structures (<4 Å) for 15 of them, including ubiquitin, a millisecond folder. MELD + CPI is up to five orders of magnitude faster than brute-force MD, satisfies detailed balance, and should scale well to larger proteins. MELD + CPI may be useful where physics-based simulations are needed to study protein mechanisms and populations and where we have some heuristic or coarse physical knowledge about states of interest.

Determining protein structures by combining semireliable data with atomistic physical models by Bayesian inference.

More than 100,000 protein structures are now known at atomic detail. However, far more are not yet known, particularly among large or complex proteins. Often, experimental information is only semireliable because it is uncertain, limited, or confusing in important ways. Some experiments give sparse information, some give ambiguous or nonspecific information, and others give uncertain information-where some is right, some is wrong, but we don't know which. We describe a method called Modeling Employing Limited Data (MELD) that can harness such problematic information in a physics-based, Bayesian framework for improved structure determination. We apply MELD to eight proteins of known structure for which such problematic structural data are available, including a sparse NMR dataset, two ambiguous EPR datasets, and four uncertain datasets taken from sequence evolution data. MELD gives excellent structures, indicating its promise for experimental biomolecule structure determination where only semireliable data are available.

Molecular modeling of biomolecules by paramagnetic NMR and computational hybrid methods.

The 3D atomic structures of biomolecules and their complexes are key to our understanding of biomolecular function, recognition, and mechanism. However, it is often difficult to obtain structures, particularly for systems that are complex, dynamic, disordered, or exist in environments like cell membranes. In such cases sparse data from a variety of paramagnetic NMR experiments offers one possible source of structural information. These restraints can be incorporated in computer modeling algorithms that can accurately translate the sparse experimental data into full 3D atomic structures. In this review, we discuss various types of paramagnetic NMR/computational hybrid modeling techniques that can be applied to successful modeling of not only the atomic structure of proteins but also their interacting partners. This article is part of a Special Issue entitled: Biophysics in Canada, edited by Lewis Kay, John Baenziger, Albert Berghuis and Peter Tieleman.

Hydrophobicity scales: a thermodynamic looking glass into lipid-protein interactions.

The partitioning of amino acid sidechains into the membrane is a key aspect of membrane protein folding. However, lipid bilayers exhibit rapidly changing physicochemical properties over their nanometer-scale thickness, which complicates understanding the thermodynamics and microscopic details of membrane partitioning. Recent data from diverse approaches, including protein insertion by the Sec translocon, folding of a small beta-barrel membrane protein and computer simulations of the exact distribution of a variety of small molecules and peptides, have joined older hydrophobicity scales for membrane protein prediction. We examine the correlations among the scales and find that they are remarkably correlated even though there are large differences in magnitude. We discuss the implications of these scales for understanding membrane protein structure and function.

Constraint methods that accelerate free-energy simulations of biomolecules.

Atomistic molecular dynamics simulations of biomolecules are critical for generating narratives about biological mechanisms. The power of atomistic simulations is that these are physics-based methods that satisfy Boltzmann's law, so they can be used to compute populations, dynamics, and mechanisms. But physical simulations are computationally intensive and do not scale well to the sizes of many important biomolecules. One way to speed up physical simulations is by coarse-graining the potential function. Another way is to harness structural knowledge, often by imposing spring-like restraints. But harnessing external knowledge in physical simulations is problematic because knowledge, data, or hunches have errors, noise, and combinatoric uncertainties. Here, we review recent principled methods for imposing restraints to speed up physics-based molecular simulations that promise to scale to larger biomolecules and motions.

Extracting representative structures from protein conformational ensembles.

A large number of methods generate conformational ensembles of biomolecules. Often one structure is selected to be representative of the whole ensemble, usually by clustering and selecting the structure closest to the center of the most populated cluster. We find that this structure is not necessarily the best representation of the cluster and present here two computationally inexpensive averaging protocols that can systematically provide better representations of the system, which can be more directly compared with structures from X-ray crystallography. In practice, systematic errors in the generated conformational ensembles appear to limit the maximum improvement of averaging methods.

Hybrid computational methods combining experimental information with molecular dynamics.

A goal of structural biology is to understand how macromolecules carry out their biological roles by identifying their metastable states, mechanisms of action, pathways leading to conformational changes, and the thermodynamic and kinetic relationships between those states. Integrative modeling brings structural insights into systems where traditional structure determination approaches cannot help. We focus on the synergies and challenges of integrative modeling combining experimental data with molecular dynamics simulations.

A Synthetically Accessible Small-Molecule Inhibitor of USP5-Cav3.2 Calcium Channel Interactions with Analgesic Properties.

Cav3.2 calcium channels are important mediators of nociceptive signaling in the primary afferent pain pathway, and their expression is increased in various rodent models of chronic pain. Previous work from our laboratory has shown that this is in part mediated by an aberrant expression of deubiquitinase USP5, which associates with these channels and increases their stability. Here, we report on a novel bioactive rhodanine compound (II-1), which was identified in compound library screens. II-1 inhibits biochemical interactions between USP5 and the Cav3.2 domain III-IV linker in a dose-dependent manner, without affecting the enzymatic activity of USP5. Molecular docking analysis reveals two potential binding pockets at the USP5-Cav3.2 interface that are distinct from the binding site of the deubiquitinase inhibitor WP1130 (a.k.a. degrasyn). With an understanding of the ability of some rhodanines to produce false positives in high-throughput screening, we have conducted several orthogonal assays to confirm the validity of this hit, including in vivo experiments. Intrathecal delivery of II-1 inhibited both phases of formalin-induced nocifensive behaviors in mice, as well as abolished thermal hyperalgesia induced by the delivery of complete Freund's adjuvant (CFA) to the hind paw. The latter effects were abolished in Cav3.2 null mice, thus confirming that Cav3.2 is required for the action of II-1. II-1 also mediated a robust inhibition of mechanical allodynia induced by injury to the sciatic nerve. Altogether, our data uncover a novel class of analgesics─well suited to rapid structure-activity relationship studies─that target the Cav3.2/USP5 interface.

Transfer of arginine into lipid bilayers is nonadditive.

Computer simulations suggest that the translocation of arginine through the hydrocarbon core of a lipid membrane proceeds by the formation of a water-filled defect that keeps the arginine molecule hydrated even at the center of the bilayer. We show here that adding additional arginine molecules into one of these water defects causes only a small change in free energy. The barrier for transferring multiple arginines through the membrane is approximately the same as for a single arginine and may even be lower depending on the exact geometry of the system. We discuss these results in the context of arginine-rich peptides such as antimicrobial and cell-penetrating peptides.

Assessment of protein structure refinement in CASP9.

We assess performance in the structure refinement category in CASP9. Two years after CASP8, the performance of the best groups has not improved. There are few groups that improve any of our assessment scores with statistical significance. Some predictors, however, are able to consistently improve the physicality of the models. Although we cannot identify any clear bottleneck in improving refinement, several points arise: (1) The refinement portion of CASP has too few targets to make many statistically meaningful conclusions. (2) Predictors are usually very conservative, limiting the possibility of large improvements in models. (3) No group is actually able to correctly rank their five submissions-indicating that potentially better models may be discarded. (4) Different sampling strategies work better for different refinement problems; there is no single strategy that works on all targets. In general, conservative strategies do better, while the greatest improvements come from more adventurous sampling-at the cost of consistency. Comparison with experimental data reveals aspects not captured by comparison to a single structure. In particular, we show that improvement in backbone geometry does not always mean better agreement with experimental data. Finally, we demonstrate that even given the current challenges facing refinement, the refined models are useful for solving the crystallographic phase problem through molecular replacement. Proteins 2011;. © 2011 Wiley-Liss, Inc.

An Integrative Approach to Determine 3D Protein Structures Using Sparse Paramagnetic NMR Data and Physical Modeling.

Paramagnetic nuclear magnetic resonance (NMR) methods have emerged as powerful tools for structure determination of large, sparsely protonated proteins. However traditional applications face several challenges, including a need for large datasets to offset the sparsity of restraints, the difficulty in accounting for the conformational heterogeneity of the spin-label, and noisy experimental data. Here we propose an integrative approach to structure determination combining sparse paramagnetic NMR with physical modelling to infer approximate protein structural ensembles. We use calmodulin in complex with the smooth muscle myosin light chain kinase peptide as a model system. Despite acquiring data from samples labeled only at the backbone amide positions, we are able to produce an ensemble with an average RMSD of ∼2.8 Å from a reference X-ray crystal structure. Our approach requires only backbone chemical shifts and measurements of the paramagnetic relaxation enhancement and residual dipolar couplings that can be obtained from sparsely labeled samples.

Bacterial cyclic diguanylate signaling networks sense temperature.

Many bacteria use the second messenger cyclic diguanylate (c-di-GMP) to control motility, biofilm production and virulence. Here, we identify a thermosensory diguanylate cyclase (TdcA) that modulates temperature-dependent motility, biofilm development and virulence in the opportunistic pathogen Pseudomonas aeruginosa. TdcA synthesizes c-di-GMP with catalytic rates that increase more than a hundred-fold over a ten-degree Celsius change. Analyses using protein chimeras indicate that heat-sensing is mediated by a thermosensitive Per-Arnt-SIM (PAS) domain. TdcA homologs are widespread in sequence databases, and a distantly related, heterologously expressed homolog from the Betaproteobacteria order Gallionellales also displayed thermosensitive diguanylate cyclase activity. We propose, therefore, that thermotransduction is a conserved function of c-di-GMP signaling networks, and that thermosensitive catalysis of a second messenger constitutes a mechanism for thermal sensing in bacteria.

Correlation between Labeling Yield and Surface Accessibility in Covalent Labeling Mass Spectrometry.

The functional properties of a protein are strongly influenced by its topography, or the solvent-facing contour map of its surface. Together with crosslinking, covalent labeling mass spectrometry (CL-MS) has the potential to contribute topographical data through the measurement of surface accessibility. However, recent efforts to correlate measures of surface accessibility with labeling yield have been met with mixed success. Most applications of CL-MS involve differential analysis of protein interactions (i.e., footprinting experiments) where such inconsistencies have limited effect. Extending CL-MS into structural analysis requires an improved evaluation of the relationship between labeling and surface exposure. In this study, we applied recently developed diazirine reagents to obtain deep coverage of the large motor domain of Eg5 (a mitotic kinesin), and together with computational methods we correlated labeling yields with accessibility data in a number of ways. We observe that correlations can indeed be seen at a local structural level, but these correlations do not extend across the structure. The lack of correlation arises from the influence of protein dynamics and chemical composition on reagent partitioning and, thus, also on labeling yield. We conclude that our use of CL-MS data should be considered in light of "chemical accessibility" rather than "solvent accessibility" and suggest that CL-MS data would be a useful tool in the fundamental study of protein-solute interactions.

Assessment of the protein-structure refinement category in CASP8.

Here, we summarize the assessment of protein structure refinement in CASP8. Twenty-four groups refined a total of 12 target proteins. Averaging over all groups and all proteins, there was no net improvement over the original starting models. However, there are now some individual research groups who consistently do improve protein structures relative to a starting starting model. We compare various measures of quality assessment, including (i) standard backbone-based methods, (ii) new methods from the Richardson group, and (iii) ensemble-based methods for comparing experimental structures, such as NMR NOE violations and the suitability of the predicted models to serve as templates for molecular replacement. On the whole, there is a general correlation among various measures. However, there are interesting differences. Sometimes a structure that is in better agreement with the experimental data is judged to be slightly worse by GDT-TS. This suggests that for comparing protein structures that are already quite close to the native, it may be preferable to use ensemble-based experimentally derived measures of quality, in addition to single-structure-based methods such as GDT-TS.