Constructing sequence-dependent protein models using coevolutionary information

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.

Expanding Direct Coupling Analysis to Identify Heterodimeric Interfaces from Limited Protein Sequence Data

Direct coupling analysis (DCA) is a global statistical approach that uses information encoded in protein sequence data to predict spatial contacts in a three-dimensional structure of a folded protein. DCA has been widely used to predict the monomeric fold at amino acid resolution and to identify biologically relevant interaction sites within a folded protein. Going beyond single proteins, DCA has also been used to identify spatial contacts that stabilize the interaction in protein complex formation. However, extracting this higher order information necessary to predict dimer contacts presents a significant challenge. A DCA evolutionary signal is much stronger at the single protein level (intraprotein contacts) than at the protein-protein interface (interprotein contacts). Therefore, if DCA-derived information is to be used to predict the structure of these complexes, there is a need to identify statistically significant DCA predictions. We propose a simple Z-score measure that can filter good predictions despite noisy, limited data. This new methodology not only improves our prediction ability but also provides a quantitative measure for the validity of the prediction.

Disease-relevant mutations alter amino acid co-evolution networks in the second nucleotide binding domain of CFTR

Cystic Fibrosis (CF) is an inherited disease caused by mutations in the cystic fibrosis transmembrane conductance regulator (CFTR) ion channel. Mutations in CFTR cause impaired chloride ion transport in the epithelial tissues of patients leading to cardiopulmonary decline and pancreatic insufficiency in the most severely affected patients. CFTR is composed of twelve membrane-spanning domains, two nucleotide-binding domains (NBDs), and a regulatory domain. The most common mutation in CFTR is a deletion of phenylalanine at position 508 (ΔF508) in NBD1. Previous research has primarily concentrated on the structure and dynamics of the NBD1 domain; However numerous pathological mutations have also been found in the lesser-studied NBD2 domain. We have investigated the amino acid co-evolved network of interactions in NBD2, and the changes that occur in that network upon the introduction of CF and CF-related mutations (S1251N(T), S1235R, D1270N, N1303K(T)). Extensive coupling between the α- and β-subdomains were identified with residues in, or near Walker A, Walker B, H-loop and C-loop motifs. Alterations in the predicted residue network varied from moderate for the S1251T perturbation to more severe for N1303T. The S1235R and D1270N networks varied greatly compared to the wildtype, but these CF mutations only affect ion transport preference and do not severely disrupt CFTR function, suggesting dynamic flexibility in the network of interactions in NBD2. Our results also suggest that inappropriate interactions between the β-subdomain and Q-loop could be detrimental. We also identified mutations predicted to stabilize the NBD2 residue network upon introduction of the CF and CF-related mutations, and these predicted mutations are scored as benign by the MUTPRED2 algorithm. Our results suggest the level of disruption of the co-evolution predictions of the amino acid networks in NBD2 does not have a straightforward correlation with the severity of the CF phenotypes observed.

Structural complementarity of distance constraints obtained from chemical cross-linking and amino acid coevolution

The analysis of amino acid coevolution has emerged as a practical method for protein structural modeling by providing structural contact information from alignments of amino acid sequences. In parallel, chemical cross-linking/mass spectrometry (XLMS) has gained attention as a universally applicable method for obtaining low-resolution distance constraints to model the quaternary arrangements of proteins, and more recently even protein tertiary structures. Here, we show that the structural information obtained by XLMS and coevolutionary analysis are effectively complementary: the distance constraints obtained by each method are almost exclusively associated with non-coincident pairs of residues, and modeling results obtained by the combination of both sets are improved relative to considering the same total number of constraints of a single type. The structural rationale behind the complementarity of the distance constraints is discussed and illustrated for a representative set of proteins with different sizes and folds.

Coevolutionary Landscape of Kinase Family Proteins: Sequence Probabilities and Functional Motifs

The protein kinase catalytic domain is one of the most abundant domains across all branches of life. Although kinases share a common core function of phosphoryl-transfer, they also have wide functional diversity and play varied roles in cell signaling networks, and for this reason are implicated in a number of human diseases. This functional diversity is primarily achieved through sequence variation, and uncovering the sequence-function relationships for the kinase family is a major challenge. In this study we use a statistical inference technique inspired by statistical physics, which builds a coevolutionary "Potts" Hamiltonian model of sequence variation in a protein family. We show how this model has sufficient power to predict the probability of specific subsequences in the highly diverged kinase family, which we verify by comparing the model's predictions with experimental observations in the Uniprot database. We show that the pairwise (residue-residue) interaction terms of the statistical model are necessary and sufficient to capture higher-than-pairwise mutation patterns of natural kinase sequences. We observe that previously identified functional sets of residues have much stronger correlated interaction scores than are typical.

Non-equilibrium coupling of protein structure and function to translation-elongation kinetics

Protein folding research has been dominated by the assumption that thermodynamics determines protein structure and function. And that when the folding process is compromised in vivo the proteostasis machinery-chaperones, deaggregases, the proteasome-work to restore proteins to their soluble, functional form or degrade them to maintain the cellular pool of proteins in a quasi-equilibrium state. During the past decade, however, more and more proteins have been identified for which altering only their speed of synthesis alters their structure and function, the efficiency of the down-stream processes they take part in, and cellular phenotype. Indeed, evidence has emerged that evolutionary selection pressures have encoded translation-rate information into mRNA molecules to coordinate diverse co-translational processes. Thus, non-equilibrium physics can play a fundamental role in influencing nascent protein behavior, mRNA sequence evolution, and disease. Here, we discuss how our understanding of this phenomenon is being advanced by the application of theoretical tools from the physical sciences.

Inferring repeat-protein energetics from evolutionary information

Natural protein sequences contain a record of their history. A common constraint in a given protein family is the ability to fold to specific structures, and it has been shown possible to infer the main native ensemble by analyzing covariations in extant sequences. Still, many natural proteins that fold into the same structural topology show different stabilization energies, and these are often related to their physiological behavior. We propose a description for the energetic variation given by sequence modifications in repeat proteins, systems for which the overall problem is simplified by their inherent symmetry. We explicitly account for single amino acid and pair-wise interactions and treat higher order correlations with a single term. We show that the resulting evolutionary field can be interpreted with structural detail. We trace the variations in the energetic scores of natural proteins and relate them to their experimental characterization. The resulting energetic evolutionary field allows the prediction of the folding free energy change for several mutants, and can be used to generate synthetic sequences that are statistically indistinguishable from the natural counterparts.

Potts Hamiltonian models of protein co-variation, free energy landscapes, and evolutionary fitness

Potts Hamiltonian models of protein sequence co-variation are statistical models constructed from the pair correlations observed in a multiple sequence alignment (MSA) of a protein family. These models are powerful because they capture higher order correlations induced by mutations evolving under constraints and help quantify the connections between protein sequence, structure, and function maintained through evolution. We review recent work with Potts models to predict protein structure and sequence-dependent conformational free energy landscapes, to survey protein fitness landscapes and to explore the effects of epistasis on fitness. We also comment on the numerical methods used to infer these models for each application.

Connecting the Sequence-Space of Bacterial Signaling Proteins to Phenotypes Using Coevolutionary Landscapes

Two-component signaling (TCS) is the primary means by which bacteria sense and respond to the environment. TCS involves two partner proteins working in tandem, which interact to perform cellular functions whereas limiting interactions with non-partners (i.e., cross-talk). We construct a Potts model for TCS that can quantitatively predict how mutating amino acid identities affect the interaction between TCS partners and non-partners. The parameters of this model are inferred directly from protein sequence data. This approach drastically reduces the computational complexity of exploring the sequence-space of TCS proteins. As a stringent test, we compare its predictions to a recent comprehensive mutational study, which characterized the functionality of 20⁴ mutational variants of the PhoQ kinase in Escherichia coli We find that our best predictions accurately reproduce the amino acid combinations found in experiment, which enable functional signaling with its partner PhoP. These predictions demonstrate the evolutionary pressure to preserve the interaction between TCS partners as well as prevent unwanted cross-talk. Further, we calculate the mutational change in the binding affinity between PhoQ and PhoP, providing an estimate to the amount of destabilization needed to disrupt TCS.

Structural propensities of kinase family proteins from a Potts model of residue co-variation

Understanding the conformational propensities of proteins is key to solving many problems in structural biology and biophysics. The co-variation of pairs of mutations contained in multiple sequence alignments of protein families can be used to build a Potts Hamiltonian model of the sequence patterns which accurately predicts structural contacts. This observation paves the way to develop deeper connections between evolutionary fitness landscapes of entire protein families and the corresponding free energy landscapes which determine the conformational propensities of individual proteins. Using statistical energies determined from the Potts model and an alignment of 2896 PDB structures, we predict the propensity for particular kinase family proteins to assume a "DFG-out" conformation implicated in the susceptibility of some kinases to type-II inhibitors, and validate the predictions by comparison with the observed structural propensities of the corresponding proteins and experimental binding affinity data. We decompose the statistical energies to investigate which interactions contribute the most to the conformational preference for particular sequences and the corresponding proteins. We find that interactions involving the activation loop and the C-helix and HRD motif are primarily responsible for stabilizing the DFG-in state. This work illustrates how structural free energy landscapes and fitness landscapes of proteins can be used in an integrated way, and in the context of kinase family proteins, can potentially impact therapeutic design strategies.

SMOG 2: A Versatile Software Package for Generating Structure-Based Models

Molecular dynamics simulations with coarse-grained or simplified Hamiltonians have proven to be an effective means of capturing the functionally important long-time and large-length scale motions of proteins and RNAs. Originally developed in the context of protein folding, structure-based models (SBMs) have since been extended to probe a diverse range of biomolecular processes, spanning from protein and RNA folding to functional transitions in molecular machines. The hallmark feature of a structure-based model is that part, or all, of the potential energy function is defined by a known structure. Within this general class of models, there exist many possible variations in resolution and energetic composition. SMOG 2 is a downloadable software package that reads user-designated structural information and user-defined energy definitions, in order to produce the files necessary to use SBMs with high performance molecular dynamics packages: GROMACS and NAMD. SMOG 2 is bundled with XML-formatted template files that define commonly used SBMs, and it can process template files that are altered according to the needs of each user. This computational infrastructure also allows for experimental or bioinformatics-derived restraints or novel structural features to be included, e.g. novel ligands, prosthetic groups and post-translational/transcriptional modifications. The code and user guide can be downloaded at http://smog-server.org/smog2.

Sequence co-evolutionary information is a natural partner to minimally-frustrated models of biomolecular dynamics

Experimentally derived structural constraints have been crucial to the implementation of computational models of biomolecular dynamics. For example, not only does crystallography provide essential starting points for molecular simulations but also high-resolution structures permit for parameterization of simplified models. Since the energy landscapes for proteins and other biomolecules have been shown to be minimally frustrated and therefore funneled, these structure-based models have played a major role in understanding the mechanisms governing folding and many functions of these systems. Structural information, however, may be limited in many interesting cases. Recently, the statistical analysis of residue co-evolution in families of protein sequences has provided a complementary method of discovering residue-residue contact interactions involved in functional configurations. These functional configurations are often transient and difficult to capture experimentally. Thus, co-evolutionary information can be merged with that available for experimentally characterized low free-energy structures, in order to more fully capture the true underlying biomolecular energy landscape.

Influences of heterogeneous native contact energy and many-body interactions on the prediction of protein folding mechanisms

Combining heterogenous native contact energies and many-body interactions could improve the prediction of Brønsted plots using a structure-based model.