Majid Masso

Accurate prediction of stability changes in protein mutants by combining machine learning with structure based computational mutagenesis

MOTIVATION Accurate predictive models for the impact of single amino acid substitutions on protein stability provide insight into protein structure and function. Such models are also valuable for the design and engineering of new proteins. Previously described methods have utilized properties of protein sequence or structure to predict the free energy change of mutants due to thermal (DeltaDeltaG) and denaturant (DeltaDeltaG(H2O)) denaturations, as well as mutant thermal stability (DeltaT(m)), through the application of either computational energy-based approaches or machine learning techniques. However, accuracy associated with applying these methods separately is frequently far from optimal. RESULTS We detail a computational mutagenesis technique based on a four-body, knowledge-based, statistical contact potential. For any mutation due to a single amino acid replacement in a protein, the method provides an empirical normalized measure of the ensuing environmental perturbation occurring at every residue position. A feature vector is generated for the mutant by considering perturbations at the mutated position and its ordered six nearest neighbors in the 3-dimensional (3D) protein structure. These predictors of stability change are evaluated by applying machine learning tools to large training sets of mutants derived from diverse proteins that have been experimentally studied and described. Predictive models based on our combined approach are either comparable to, or in many cases significantly outperform, previously published results. AVAILABILITY A web server with supporting documentation is available at http://proteins.gmu.edu/automute.

Accurate prediction of enzyme mutant activity based on a multibody statistical potential

MOTIVATION An important area of research in biochemistry and molecular biology focuses on characterization of enzyme mutants. However, synthesis and analysis of experimental mutants is time consuming and expensive. We describe a machine-learning approach for inferring the activity levels of all unexplored single point mutants of an enzyme, based on a training set of such mutants with experimentally measured activity. RESULTS Based on a Delaunay tessellation-derived four-body statistical potential function, a perturbation vector measuring environmental changes relative to wild type (wt) at every residue position uniquely characterizes each enzyme mutant for model development and prediction. First, a measure of model performance utilizing area (AUC) under the receiver operating characteristic (ROC) curve surpasses 0.83 and 0.77 for data sets of experimental HIV-1 protease and T4 lysozyme mutants, respectively. Additionally, a novel method is introduced for evaluating statistical significance associated with the number of correct test set predictions obtained from a trained model. Third, 100 stratified random splits of the protease and T4 lysozyme mutant data sets into training and test sets achieve 77.0% and 80.8% mean accuracy, respectively. Next, protease and T4 lysozyme models trained with experimental mutants are used to predict activity levels for all remaining mutants; a subsequent search for publications reporting on dozens of these test mutants reveals that experimental results are matched by 79% and 86% of predictions, respectively. Finally, learning curves for each mutant enzyme system indicate the influence of training set size on model performance. AVAILABILITY Prediction databases at http://proteins.gmu.edu/automute/

AUTO-MUTE: web-based tools for predicting stability changes in proteins due to single amino acid replacements

Utilizing cutting-edge supervised classification and regression algorithms, three web-based tools have been developed for predicting stability changes upon single residue substitutions in proteins with known native structures. Trained models classify independent mutant test sets with accuracies ranging from 87 to 94%. Attributes representing each mutant protein are based on a computational mutagenesis methodology relying on a four-body statistical potential, illustrating a novel integration of both energy-based and machine learning approaches. The servers are written in PHP and hosted on a Linux platform, and they can be freely accessed online along with detailed data sets, documentation and performance results at http://proteins.gmu.edu/automute.

Knowledge-based computational mutagenesis for predicting the disease potential of human non-synonymous single nucleotide polymorphisms

Certain genetic variations in the human population are associated with heritable diseases, and single nucleotide polymorphisms (SNPs) represent the most common form of such differences in DNA sequence. In particular, substantial interest exists in determining whether a non-synonymous SNP (nsSNP), leading to a single residue replacement in the translated protein product, is neutral or disease-related. The nature of protein structure-function relationships suggests that nsSNP effects, either benign or leading to aberrant protein function possibly associated with disease, are dependent on relative structural changes introduced upon mutation. In this study, we characterize a representative sampling of 1790 documented neutral and disease-related human nsSNPs mapped to 243 diverse human protein structures, by quantifying environmental perturbations in the associated proteins with the use of a computational mutagenesis methodology that relies on a four-body, knowledge-based, statistical contact potential. These structural change data are used as attributes to generate a vector representation for each nsSNP, in combination with additional features reflecting sequence and structure of the corresponding protein. A trained model based on the random forest supervised classification algorithm achieves 76% cross-validation accuracy. Our classifier performs at least as well as other methods that use significantly larger datasets of nsSNPs for model training, and the novelty of our attributes differentiates the model as an orthogonal approach that can be utilized in conjunction with other techniques. A dedicated server for obtaining predictions, as well as supporting datasets and documentation, is available at http://proteins.gmu.edu/automute.

Comprehensive mutagenesis of HIV-1 protease: a computational geometry approach.

A computational geometry technique based on Delaunay tessellation of protein structure, represented by C(alpha) atoms, is used to study effects of single residue mutations on sequence-structure compatibility in HIV-1 protease. Profiles of residue scores derived from the four-body statistical potential are constructed for all 1881 mutants of the HIV-1 protease monomer and compared with the profile of the wild-type protein. The profiles for an isolated monomer of HIV-1 protease and the identical monomer in a dimeric state with an inhibitor are analyzed to elucidate changes to structural stability. Protease residues shown to undergo the greatest impact are those forming the dimer interface and flap region, as well as those known to be involved in inhibitor binding.

Statistical geometry based prediction of nonsynonymous SNP functional effects using random forest and neuro-fuzzy classifiers.

There is substantial interest in methods designed to predict the effect of nonsynonymous single nucleotide polymorphisms (nsSNPs) on protein function, given their potential relationship to heritable diseases. Current state‐of‐the‐art supervised machine learning algorithms, such as random forest (RF), train models that classify single amino acid mutations in proteins as either neutral or deleterious to function. However, it is frequently the case that the functional effect of a polymorphism on a protein resides between these two extremes. The utilization of classifiers that incorporate fuzzy logic provides a natural extension in order to account for the spectrum of possible functional consequences. We generated a dataset of single amino acid substitutions in human proteins having known three‐dimensional structures. Each variant was uniquely represented as a feature vector that included computational geometry and knowledge‐based statistical potential predictors obtained though application of Delaunay tessellation of protein structures. Additional attributes consisted of physicochemical properties of the native and replacement amino acids as well as topological location of the mutated residue position in the solved structure. Classification performance of the RF algorithm was evaluated on a training set consisting of the disease‐associated and neutral nsSNPs taken from our dataset, and attributes were ranked according to their relative importance. Similarly, we evaluated the performance of adaptive neuro‐fuzzy inference system (ANFIS). The utility of statistical geometry predictors was compared with that of traditional structural and evolutionary attributes employed by other researchers, revealing an equally effective yet complementary methodology. Among all attributes in our feature set, the statistical geometry predictors were found to be the most highly ranked. On the basis of the AUC (area under the ROC curve) measure of performance, the ANFIS and RF models were equally effective when only statistical geometry features were utilized. Tenfold cross‐validation studies evaluating AUC, balanced error rate (BER), and Matthews correlation coefficient (MCC) showed that our RF model was at least comparable with the well‐established methods of SIFT and PolyPhen. The trained RF and ANFIS models were each subsequently used to predict the disease potential of human nsSNPs in our dataset that are currently unclassified (http://rna.gmu.edu/FuzzySnps/). Proteins 2008.

Computational Mutagenesis Studies of Protein Structure- Function Correlations

Topological scores, measures of sequence‐structure compatibility, are calculated for all 1,881 single point mutants of the human immunodeficiency virus (HIV)‐1 protease using a four‐body statistical potential function based on Delaunay tessellation of protein structure. Comparison of the mutant topological score data with experimental data from alanine scan studies specifically on the dimer interface residues supports previous findings that 1) L97 and F99 contribute greatly to the Gibbs energy of HIV‐1 protease dimerization, 2) Q2 and T4 contribute the least toward the Gibbs energy, and 3) C‐terminal residues are more sensitive to mutations than those at the N‐terminus. For a more comprehensive treatment of the relationship between protease structure and function, mutant topological scores are compared with the activity levels for a set of 536 experimentally synthesized protease mutants, and a significant correlation is observed. Finally, this structure‐function correlation is similarly identified by examining model systems consisting of 2,015 single point mutants of bacteriophage T4 lysozyme as well as 366 single point mutants of HIV‐1 reverse transcriptase and is hypothesized to be a property generally applicable to all proteins. Proteins 2006.

AUTO-MUTE 2.0: A Portable Framework with Enhanced Capabilities for Predicting Protein Functional Consequences upon Mutation

The AUTO-MUTE 2.0 stand-alone software package includes a collection of programs for predicting functional changes to proteins upon single residue substitutions, developed by combining structure-based features with trained statistical learning models. Three of the predictors evaluate changes to protein stability upon mutation, each complementing a distinct experimental approach. Two additional classifiers are available, one for predicting activity changes due to residue replacements and the other for determining the disease potential of mutations associated with nonsynonymous single nucleotide polymorphisms (nsSNPs) in human proteins. These five command-line driven tools, as well as all the supporting programs, complement those that run our AUTO-MUTE web-based server. Nevertheless, all the codes have been rewritten and substantially altered for the new portable software, and they incorporate several new features based on user feedback. Included among these upgrades is the ability to perform three highly requested tasks: to run “big data” batch jobs; to generate predictions using modified protein data bank (PDB) structures, and unpublished personal models prepared using standard PDB file formatting; and to utilize NMR structure files that contain multiple models.

Accurate and efficient gp120 V3 loop structure based models for the determination of HIV-1 co-receptor usage

BackgroundHIV-1 targets human cells expressing both the CD4 receptor, which binds the viral envelope glycoprotein gp120, as well as either the CCR5 (R5) or CXCR4 (X4) co-receptors, which interact primarily with the third hypervariable loop (V3 loop) of gp120. Determination of HIV-1 affinity for either the R5 or X4 co-receptor on host cells facilitates the inclusion of co-receptor antagonists as a part of patient treatment strategies. A dataset of 1193 distinct gp120 V3 loop peptide sequences (989 R5-utilizing, 204 X4-capable) is utilized to train predictive classifiers based on implementations of random forest, support vector machine, boosted decision tree, and neural network machine learning algorithms. An in silico mutagenesis procedure employing multibody statistical potentials, computational geometry, and threading of variant V3 sequences onto an experimental structure, is used to generate a feature vector representation for each variant whose components measure environmental perturbations at corresponding structural positions.ResultsClassifier performance is evaluated based on stratified 10-fold cross-validation, stratified dataset splits (2/3 training, 1/3 validation), and leave-one-out cross-validation. Best reported values of sensitivity (85%), specificity (100%), and precision (98%) for predicting X4-capable HIV-1 virus, overall accuracy (97%), Matthews correlation coefficient (89%), balanced error rate (0.08), and ROC area (0.97) all reach critical thresholds, suggesting that the models outperform six other state-of-the-art methods and come closer to competing with phenotype assays.ConclusionsThe trained classifiers provide instantaneous and reliable predictions regarding HIV-1 co-receptor usage, requiring only translated V3 loop genotypes as input. Furthermore, the novelty of these computational mutagenesis based predictor attributes distinguishes the models as orthogonal and complementary to previous methods that utilize sequence, structure, and/or evolutionary information. The classifiers are available online at http://proteins.gmu.edu/automute.

Sequence and structure based models of HIV-1 protease and reverse transcriptase drug resistance

BackgroundSuccessful management of chronic human immunodeficiency virus type 1 (HIV-1) infection with a cocktail of antiretroviral medications can be negatively affected by the presence of drug resistant mutations in the viral targets. These targets include the HIV-1 protease (PR) and reverse transcriptase (RT) proteins, for which a number of inhibitors are available on the market and routinely prescribed. Protein mutational patterns are associated with varying degrees of resistance to their respective inhibitors, with extremes that can range from continued susceptibility to cross-resistance across all drugs.ResultsHere we implement statistical learning algorithms to develop structure- and sequence-based models for systematically predicting the effects of mutations in the PR and RT proteins on resistance to each of eight and eleven inhibitors, respectively. Employing a four-body statistical potential, mutant proteins are represented as feature vectors whose components quantify relative environmental perturbations at amino acid residue positions in the respective target structures upon mutation. Two approaches are implemented in developing sequence-based models, based on use of either relative frequencies or counts of n-grams, to generate vectors for representing mutant proteins. To the best of our knowledge, this is the first reported study on structure- and sequence-based predictive models of HIV-1 PR and RT drug resistance developed by implementing a four-body statistical potential and n-grams, respectively, to generate mutant attribute vectors. Performance of the learning methods is evaluated on the basis of tenfold cross-validation, using previously assayed and publicly available in vitro data relating mutational patterns in the targets to quantified inhibitor susceptibility changes.ConclusionOverall performance results are competitive with those of a previously published study utilizing a sequence-based strategy, while our structure- and sequence-based models provide orthogonal and complementary prediction methodologies, respectively. In a novel application, we describe a technique for identifying every possible pair of RT inhibitors as either potentially effective together as part of a cocktail, or a combination that is to be avoided.