Uncovering substrate specificity determinants of class IIb aminoacyl-tRNA synthetases with machine learning.

Specific amino acid (AA) binding by aminoacyl-tRNA synthetases (aaRSs) is necessary for correct translation of the genetic code. Sequence and structure analyses have revealed the main specificity determinants and allowed a partitioning of aaRSs into two classes and several subclasses. However, the information contributed by each determinant has not been precisely quantified, and other, minor determinants may still be unidentified. Growth of genomic data and development of machine learning classification methods allow us to revisit these questions. This work considered the subclass IIb, formed by the three enzymes aspartyl-, asparaginyl-, and lysyl-tRNA synthetase (LysRS). Over 35,000 sequences from the Pfam database were considered, and used to train a machine-learning model based on ensembles of decision trees. The model was trained to reproduce the existing classification of each sequence as AspRS, AsnRS, or LysRS, and to identify which sequence positions were most important for the classification. A few positions (5-8 depending on the AA substrate) sufficed for accurate classification. Most but not all of them were well-known specificity determinants. The machine learning models thus identified sets of mutations that distinguish the three subclass members, which might be targeted in engineering efforts to alter or swap the AA specificities for biotechnology applications.

Classifying protein kinase conformations with machine learning.

Protein kinases are key actors of signaling networks and important drug targets. They cycle between active and inactive conformations, distinguished by a few elements within the catalytic domain. One is the activation loop, whose conserved DFG motif can occupy DFG-in, DFG-out, and some rarer conformations. Annotation and classification of the structural kinome are important, as different conformations can be targeted by different inhibitors and activators. Valuable resources exist; however, large-scale applications will benefit from increased automation and interpretability of structural annotation. Interpretable machine learning models are described for this purpose, based on ensembles of decision trees. To train them, a set of catalytic domain sequences and structures was collected, somewhat larger and more diverse than existing resources. The structures were clustered based on the DFG conformation and manually annotated. They were then used as training input. Two main models were constructed, which distinguished active/inactive and in/out/other DFG conformations. They considered initially 1692 structural variables, spanning the whole catalytic domain, then identified ("learned") a small subset that sufficed for accurate classification. The first model correctly labeled all but 3 of 3289 structures as active or inactive, while the second assigned the correct DFG label to all but 17 of 8826 structures. The most potent classifying variables were all related to well-known structural elements in or near the activation loop and their ranking gives insights into the conformational preferences. The models were used to automatically annotate 3850 kinase structures predicted recently with the Alphafold2 tool, showing that Alphafold2 reproduced the active/inactive but not the DFG-in proportions seen in the Protein Data Bank. We expect the models will be useful for understanding and engineering kinases.

Enzyme redesign and genetic code expansion.

Enzyme design is an important application of computational protein design (CPD). It can benefit enormously from the additional chemistries provided by noncanonical amino acids (ncAAs). These can be incorporated into an 'expanded' genetic code, and introduced in vivo into target proteins. The key step for genetic code expansion is to engineer an aminoacyl-transfer RNA (tRNA) synthetase (aaRS) and an associated tRNA that handles the ncAA. Experimental directed evolution has been successfully used to engineer aaRSs and incorporate over 200 ncAAs into expanded codes. But directed evolution has severe limits, and is not yet applicable to noncanonical AA backbones. CPD can help address several of its limitations, and has begun to be applied to this problem. We review efforts to redesign aaRSs, studies that designed new proteins and functionalities with the help of ncAAs, and some of the method developments that have been used, such as adaptive landscape flattening Monte Carlo, which allows an enzyme to be redesigned with substrate or transition state binding as the design target.

PLP-Dependent Enzyme Methionine γ-Lyase: Insights into the Michaelis Complex from Molecular Dynamics and Free Energy Simulations.

Methionine γ-lyase (MGL) breaks down methionine, with the help of its cofactor pyridoxal-5'-phosphate (PLP), or vitamin B6. Methionine depletion is damaging for cancer cells but not normal cells, so MGL is of interest as a therapeutic protein. To increase our understanding and help engineer improved activity, we focused on the reactive, Michaelis complex M between MGL, covalently bound PLP, and substrate Met. M is not amenable to crystallography, as it proceeds to products. Experimental activity measurements helped exclude a mechanism that would bypass M. We then used molecular dynamics and alchemical free energy simulations to elucidate its structure and dynamics. We showed that the PLP phosphate has a pKa strongly downshifted by the protein, whether Met is present or not. Met binding affects the structure surrounding the reactive atoms. With Met, the Schiff base linkage between PLP and a nearby lysine shifts from a zwitterionic, keto form to a neutral, enol form that makes it easier for Met to approach its labile, target atom. The Met ligand also stabilizes the correct orientation of the Schiff base, more strongly than in simulations without Met, and in agreement with structures in the Protein Data Bank, where the Schiff base orientation correlates with the presence or absence of a co-bound anion or substrate analogue in the active site. Overall, the Met ligand helps organize the active site for the enzyme reaction by reducing fluctuations and shifting protonation states and conformational populations.

Redesigning methionyl-tRNA synthetase for ß-methionine activity with adaptive landscape flattening and experiments.

Amino acids (AAs) with a noncanonical backbone would be a valuable tool for protein engineering, enabling new structural motifs and building blocks. To incorporate them into an expanded genetic code, the first, key step is to obtain an appropriate aminoacyl-tRNA synthetase. Currently, directed evolution is not available to optimize AAs with noncanonical backbones, since an appropriate selective pressure has not been discovered. Computational protein design (CPD) is an alternative. We used a new CPD method to redesign MetRS and increase its activity towards ß-Met, which has an extra backbone methylene. The new method considered a few active site positions for design and used a Monte Carlo exploration of the corresponding sequence space. During the exploration, a bias energy was adaptively learned, such that the free energy landscape of the apo enzyme was flattened. Enzyme variants could then be sampled, in the presence of the ligand and the bias energy, according to their ß-Met binding affinities. Eighteen predicted variants were chosen for experimental testing; 10 exhibited detectable activity for ß-Met adenylation. Top predicted hits were characterized experimentally in detail. Dissociation constants, catalytic rates, and Michaelis constants for both α-Met and ß-Met were measured. The best mutant retained a preference for α-Met over ß-Met; however, the preference was reduced, compared to the wildtype, by a factor of 29. For this mutant, high resolution crystal structures were obtained in complex with both α-Met and ß-Met, indicating that the predicted, active conformation of ß-Met in the active site was retained.

A Computational Model for the PLP-Dependent Enzyme Methionine γ-Lyase.

Pyridoxal-5'-phosphate (PLP) is a cofactor in the reactions of over 160 enzymes, several of which are implicated in diseases. Methionine γ-lyase (MGL) is of interest as a therapeutic protein for cancer treatment. It binds PLP covalently through a Schiff base linkage and digests methionine, whose depletion is damaging for cancer cells but not normal cells. To improve MGL activity, it is important to understand and engineer its PLP binding. We develop a simulation model for MGL, starting with force field parameters for PLP in four main states: two phosphate protonation states and two tautomeric states, keto or enol for the Schiff base moiety. We used the force field to simulate MGL complexes with each form, and showed that those with a fully-deprotonated PLP phosphate, especially keto, led to the best agreement with MGL structures in the PDB. We then confirmed this result through alchemical free energy simulations that compared the keto and enol forms, confirming a moderate keto preference, and the fully-deprotonated and singly-protonated phosphate forms. Extensive simulations were needed to adequately sample conformational space, and care was needed to extrapolate the protonation free energy to the thermodynamic limit of a macroscopic, dilute protein solution. The computed phosphate pK a was 5.7, confirming that the deprotonated, -2 form is predominant. The PLP force field and the simulation methods can be applied to all PLP enzymes and used, as here, to reveal fine details of structure and dynamics in the active site.

Knowledge-Based Unfolded State Model for Protein Design.

The design of proteins and miniproteins is an important challenge. Designed variants should be stable, meaning the folded/unfolded free energy difference should be large enough. Thus, the unfolded state plays a central role. An extended peptide model is often used, where side chains interact with solvent and nearby backbone, but not each other. The unfolded energy is then a function of sequence composition only and can be empirically parametrized. If the space of sequences is explored with a Monte Carlo procedure, protein variants will be sampled according to a well-defined Boltzmann probability distribution. We can then choose unfolded model parameters to maximize the probability of sampling native-like sequences. This leads to a well-defined maximum likelihood framework. We present an iterative algorithm that follows the likelihood gradient. The method is presented in the context of our Proteus software, as a detailed downloadable tutorial. The unfolded model is combined with a folded model that uses molecular mechanics and a Generalized Born solvent. It was optimized for three PDZ domains and then used to redesign them. The sequences sampled are native-like and similar to a recent PDZ design study that was experimentally validated.

Computational protein design repurposed to explore enzyme vitality and help predict antibiotic resistance.

In response to antibiotics that inhibit a bacterial enzyme, resistance mutations inevitably arise. Predicting them ahead of time would aid target selection and drug design. The simplest resistance mechanism would be to reduce antibiotic binding without sacrificing too much substrate binding. The property that reflects this is the enzyme "vitality", defined here as the difference between the inhibitor and substrate binding free energies. To predict such mutations, we borrow methodology from computational protein design. We use a Monte Carlo exploration of mutation space and vitality changes, allowing us to rank thousands of mutations and identify ones that might provide resistance through the simple mechanism considered. As an illustration, we chose dihydrofolate reductase, an essential enzyme targeted by several antibiotics. We simulated its complexes with the inhibitor trimethoprim and the substrate dihydrofolate. 20 active site positions were mutated, or "redesigned" individually, then in pairs or quartets. We computed the resulting binding free energy and vitality changes. Out of seven known resistance mutations involving active site positions, five were correctly recovered. Ten positions exhibited mutations with significant predicted vitality gains. Direct couplings between designed positions were predicted to be small, which reduces the combinatorial complexity of the mutation space to be explored. It also suggests that over the course of evolution, resistance mutations involving several positions do not need the underlying point mutations to arise all at once: they can appear and become fixed one after the other.

How much can physics do for protein design?

Physics and physical chemistry are an important thread in computational protein design, complementary to knowledge-based tools. They provide molecular mechanics scoring functions that need little or no ad hoc parameter readjustment, methods to thoroughly sample equilibrium ensembles, and different levels of approximation for conformational flexibility. They led recently to the successful redesign of a small protein using a physics-based folded state energy. Adaptive Monte Carlo or molecular dynamics schemes were discovered where protein variants are populated as per their ligand-binding free energy or catalytic efficiency. Molecular dynamics have been used for backbone flexibility. Implicit solvent models have been refined, polarizable force fields applied, and many physical insights obtained.

Computational Design of PDZ-Peptide Binding.

This chapter describes two computational methods for PDZ-peptide binding: high-throughput computational protein design (CPD) and a medium-throughput approach combining molecular dynamics for conformational sampling with a Poisson-Boltzmann (PB) Linear Interaction Energy for scoring. A new CPD method is outlined, which uses adaptive Monte Carlo simulations to efficiently sample peptide variants that tightly bind a PDZ domain, and provides at the same time precise estimates of their relative binding free energies. A detailed protocol is described based on the Proteus CPD software. The medium-throughput approach can be performed with standard MD and PB software, such as NAMD and Charmm. For 40 complexes between Tiam1 and peptide ligands, it gave high a2ccuracy, with mean errors of around 0.5 kcal/mol for relative binding free energies and no large errors. It requires a moderate amount of parameter fitting before it can be applied, and its transferability to other protein families is still untested.

Physics-Based Computational Protein Design: An Update.

We describe methods for physics-based protein design and some recent applications from our work. We present the physical interpretation of a MC simulation in sequence space and show that sequences and conformations form a well-defined statistical ensemble, explored with Monte Carlo and Boltzmann sampling. The folded state energy combines molecular mechanics for solutes with continuum electrostatics for solvent. We usually assume one or a few fixed protein backbone structures and discrete side chain rotamers. Methods based on molecular dynamics, which introduce additional backbone and side chain flexibility, are under development. The redesign of a PDZ domain and an aminoacyl-tRNA synthetase enzyme were successful. We describe a versatile, adaptive, Wang-Landau MC method that can be used to design for substrate affinity, catalytic rate, catalytic efficiency, or the specificity of these properties. The methods are transferable to all biomolecules, can be systematically improved, and give physical insights.

Application of Various Molecular Modelling Methods in the Study of Estrogens and Xenoestrogens.

In this review, applications of various molecular modelling methods in the study of estrogens and xenoestrogens are summarized. Selected biomolecules that are the most commonly chosen as molecular modelling objects in this field are presented. In most of the reviewed works, ligand docking using solely force field methods was performed, employing various molecular targets involved in metabolism and action of estrogens. Other molecular modelling methods such as molecular dynamics and combined quantum mechanics with molecular mechanics have also been successfully used to predict the properties of estrogens and xenoestrogens. Among published works, a great number also focused on the application of different types of quantitative structure-activity relationship (QSAR) analyses to examine estrogen's structures and activities. Although the interactions between estrogens and xenoestrogens with various proteins are the most commonly studied, other aspects such as penetration of estrogens through lipid bilayers or their ability to adsorb on different materials are also explored using theoretical calculations. Apart from molecular mechanics and statistical methods, quantum mechanics calculations are also employed in the studies of estrogens and xenoestrogens. Their applications include computation of spectroscopic properties, both vibrational and Nuclear Magnetic Resonance (NMR), and also in quantum molecular dynamics simulations and crystal structure prediction. The main aim of this review is to present the great potential and versatility of various molecular modelling methods in the studies on estrogens and xenoestrogens.

Hybrid MC/MD for protein design.

Computational protein design relies on simulations of a protein structure, where selected amino acids can mutate randomly, and mutations are selected to enhance a target property, such as stability. Often, the protein backbone is held fixed and its degrees of freedom are modeled implicitly to reduce the complexity of the conformational space. We present a hybrid method where short molecular dynamics (MD) segments are used to explore conformations and alternate with Monte Carlo (MC) moves that apply mutations to side chains. The backbone is fully flexible during MD. As a test, we computed side chain acid/base constants or pKa's in five proteins. This problem can be considered a special case of protein design, with protonation/deprotonation playing the role of mutations. The solvent was modeled as a dielectric continuum. Due to cost, in each protein we allowed just one side chain position to change its protonation state and the other position to change its type or mutate. The pKa's were computed with a standard method that scans a range of pH values and with a new method that uses adaptive landscape flattening (ALF) to sample all protonation states in a single simulation. The hybrid method gave notably better accuracy than standard, fixed-backbone MC. ALF decreased the computational cost a factor of 13.

A physics-based energy function allows the computational redesign of a PDZ domain.

Computational protein design (CPD) can address the inverse folding problem, exploring a large space of sequences and selecting ones predicted to fold. CPD was used previously to redesign several proteins, employing a knowledge-based energy function for both the folded and unfolded states. We show that a PDZ domain can be entirely redesigned using a "physics-based" energy for the folded state and a knowledge-based energy for the unfolded state. Thousands of sequences were generated by Monte Carlo simulation. Three were chosen for experimental testing, based on their low energies and several empirical criteria. All three could be overexpressed and had native-like circular dichroism spectra and 1D-NMR spectra typical of folded structures. Two had upshifted thermal denaturation curves when a peptide ligand was present, indicating binding and suggesting folding to a correct, PDZ structure. Evidently, the physical principles that govern folded proteins, with a dash of empirical post-filtering, can allow successful whole-protein redesign.

Adaptive landscape flattening allows the design of both enzyme: Substrate binding and catalytic power.

Designed enzymes are of fundamental and technological interest. Experimental directed evolution still has significant limitations, and computational approaches are a complementary route. A designed enzyme should satisfy multiple criteria: stability, substrate binding, transition state binding. Such multi-objective design is computationally challenging. Two recent studies used adaptive importance sampling Monte Carlo to redesign proteins for ligand binding. By first flattening the energy landscape of the apo protein, they obtained positive design for the bound state and negative design for the unbound. We have now extended the method to design an enzyme for specific transition state binding, i.e., for its catalytic power. We considered methionyl-tRNA synthetase (MetRS), which attaches methionine (Met) to its cognate tRNA, establishing codon identity. Previously, MetRS and other synthetases have been redesigned by experimental directed evolution to accept noncanonical amino acids as substrates, leading to genetic code expansion. Here, we have redesigned MetRS computationally to bind several ligands: the Met analog azidonorleucine, methionyl-adenylate (MetAMP), and the activated ligands that form the transition state for MetAMP production. Enzyme mutants known to have azidonorleucine activity were recovered by the design calculations, and 17 mutants predicted to bind MetAMP were characterized experimentally and all found to be active. Mutants predicted to have low activation free energies for MetAMP production were found to be active and the predicted reaction rates agreed well with the experimental values. We suggest the present method should become the paradigm for computational enzyme design.

Variable Neighborhood Search with Cost Function Networks To Solve Large Computational Protein Design Problems.

Computational protein design (CPD) aims to predict amino acid sequences that fold to specific structures and perform desired functions. CPD depends on a rotamer library, an energy function, and an algorithm to search the sequence/conformation space. Variable neighborhood search (VNS) with cost function networks is a powerful framework that can provide tight upper bounds on the global minimum energy. We propose a new CPD heuristic based on VNS in which a subset of the solution space (a "neighborhood") is explored, whose size is gradually increased with a dedicated probabilistic heuristic. The algorithm was tested on 99 protein designs with fixed backbones involving nine proteins from the SH2, SH3, and PDZ families. The number of mutating positions was 20, 30, or all of the amino acids, while the rest of the protein explored side-chain rotamers. VNS was more successful than Monte Carlo (MC), replica-exchange MC, and a heuristic steepest-descent energy minimization, providing solutions with equal or lower best energies in most cases. For complete protein redesign, it gave solutions that were 2.5 to 11.2 kcal/mol lower in energy than those obtained with the other approaches. VNS is implemented in the toulbar2 software. It could be very helpful for large and/or complex design problems.

Differences in Outcomes Reported by Patients With Inflammatory Bowel Diseases vs Their Health Care Professionals.

BACKGROUND & AIMS: Inflammatory bowel disease (IBD) scoring systems combine patient-reported data with physicians' observations to determine patient outcomes, but these systems are believed to have limitations. We used real-world data from a large IBD cohort in Switzerland to compare results between patients and healthcare professionals from scoring systems for Crohn's disease (CD) and ulcerative colitis (UC). METHODS: We collected data from the Swiss IBD cohort, beginning in 2006, using 2453 reports for 1385 patients (52% female, 58% with CD). During office visits, physicians asked patients about signs and symptoms and recorded their answers (health care professional-reported outcomes). On a later date, patients received a questionnaire at home (independently of the medical visit), complete it, and sent it back to the data center. Patients also completed the short form 36 and IBD quality of life (QoL) questionnaires. We calculated Cohen's kappa (κ) statistics to assess the level of agreement in scores between patients and health care professionals (Δt between reports collected less than 2 months apart). We used Spearman correlation coefficients (ρ) to compare general well-being (GWB) and QoL scores determined by patients vs health care professionals. Our primary aim was to investigate the overall and individual level of agreement on signs and symptoms reported by health care professionals vs patients. RESULTS: The best level of agreement (although moderate) was observed for number of stools last week in patients with CD (κ = 0.47), and nocturnal diarrhea in patients with UC (κ = 0.52). Agreement was low on level of abdominal pain (κ = 0.31 for patients with CD and κ = 0.37 for patients with UC) and GWB (κ = 0.23 for patients with CD and κ = 0.26 for patients with UC). Patients reported less severe abdominal pain and worse GWB (CD) or better GWB (UC) than that determined by health care professionals. Patient self-rated GWB correlated with IBD quality of life (ρ = 0.68 for patients with CD and ρ = 0.70 for patients with UC) and SF-36 physical scores (ρ = 0.55 for patients with CD and ρ = 0.60 for patients with UC); there was no correlation between health care professional-rated GWB and QoL. CONCLUSIONS: In a comparison of patient vs health care provider-reported outcomes in a Swiss IBD cohort, we found that health care professionals seem to misinterpret patients' complaints. Patients self-rated GWB correlated with QoL scores, indicating that reporting GWB in a single question is possible and relevant, but can vary based on how the data are collected.

Protein p Ka's from Adaptive Landscape Flattening Instead of Constant-pH Simulations.

Protein acid/base constants, or p Ka's are often computed from Monte Carlo or molecular dynamics simulations at a series of constant pH values. Instead, we propose to adaptively flatten the free energy landscape in the space of protonation states. The flattening is achieved by a Wang-Landau Monte Carlo, where a bias potential is constructed adaptively during an initial phase, such that all protonation states achieve comparable probabilities. Biased ensembles of states are then reweighted by subtracting out the bias and adding a pH-dependent free energy term. Titration curves constructed for three test proteins agreed, within the small numerical uncertainty, with those obtained earlier from the constant-pH approach.

Adaptive landscape flattening in amino acid sequence space for the computational design of protein:peptide binding.

For the high throughput design of protein:peptide binding, one must explore a vast space of amino acid sequences in search of low binding free energies. This complex problem is usually addressed with either simple heuristic scoring or expensive sequence enumeration schemes. Far more efficient than enumeration is a recent Monte Carlo approach that adaptively flattens the energy landscape in sequence space of the unbound peptide and provides formally exact binding free energy differences. The method allows the binding free energy to be used directly as the design criterion. We propose several improvements that allow still more efficient sampling and can address larger design problems. They include the use of Replica Exchange Monte Carlo and landscape flattening for both the unbound and bound peptides. We used the method to design peptides that bind to the PDZ domain of the Tiam1 signaling protein and could serve as inhibitors of its activity. Four peptide positions were allowed to mutate freely. Almost 75 000 peptide variants were processed in two simulations of 109 steps each that used 1 CPU hour on a desktop machine. 96% of the theoretical sequence space was sampled. The relative binding free energies agreed qualitatively with values from experiment. The sampled sequences agreed qualitatively with an experimental library of Tiam1-binding peptides. The main assumption limiting accuracy is the fixed backbone approximation, which could be alleviated in future work by using increased computational resources and multi-backbone designs.