Machine learning integration for predicting the effect of single amino acid substitutions on protein stability

Protein Science Meets Artificial Intelligence: A Systematic Review and a Biochemical Meta-Analysis of an Inter-Field

Proteins are some of the most fascinating and challenging molecules in the universe, and they pose a big challenge for artificial intelligence. The implementation of machine learning/AI in protein science gives rise to a world of knowledge adventures in the workhorse of the cell and proteome homeostasis, which are essential for making life possible. This opens up epistemic horizons thanks to a coupling of human tacit-explicit knowledge with machine learning power, the benefits of which are already tangible, such as important advances in protein structure prediction. Moreover, the driving force behind the protein processes of self-organization, adjustment, and fitness requires a space corresponding to gigabytes of life data in its order of magnitude. There are many tasks such as novel protein design, protein folding pathways, and synthetic metabolic routes, as well as protein-aggregation mechanisms, pathogenesis of protein misfolding and disease, and proteostasis networks that are currently unexplored or unrevealed. In this systematic review and biochemical meta-analysis, we aim to contribute to bridging the gap between what we call binomial artificial intelligence (AI) and protein science (PS), a growing research enterprise with exciting and promising biotechnological and biomedical applications. We undertake our task by exploring "the state of the art" in AI and machine learning (ML) applications to protein science in the scientific literature to address some critical research questions in this domain, including What kind of tasks are already explored by ML approaches to protein sciences? What are the most common ML algorithms and databases used? What is the situational diagnostic of the AI-PS inter-field? What do ML processing steps have in common? We also formulate novel questions such as Is it possible to discover what the rules of protein evolution are with the binomial AI-PS? How do protein folding pathways evolve? What are the rules that dictate the folds? What are the minimal nuclear protein structures? How do protein aggregates form and why do they exhibit different toxicities? What are the structural properties of amyloid proteins? How can we design an effective proteostasis network to deal with misfolded proteins? We are a cross-functional group of scientists from several academic disciplines, and we have conducted the systematic review using a variant of the PICO and PRISMA approaches. The search was carried out in four databases (PubMed, Bireme, OVID, and EBSCO Web of Science), resulting in 144 research articles. After three rounds of quality screening, 93 articles were finally selected for further analysis. A summary of our findings is as follows: regarding AI applications, there are mainly four types: 1) genomics, 2) protein structure and function, 3) protein design and evolution, and 4) drug design. In terms of the ML algorithms and databases used, supervised learning was the most common approach (85%). As for the databases used for the ML models, PDB and UniprotKB/Swissprot were the most common ones (21 and 8%, respectively). Moreover, we identified that approximately 63% of the articles organized their results into three steps, which we labeled pre-process, process, and post-process. A few studies combined data from several databases or created their own databases after the pre-process. Our main finding is that, as of today, there are no research road maps serving as guides to address gaps in our knowledge of the AI-PS binomial. All research efforts to collect, integrate multidimensional data features, and then analyze and validate them are, so far, uncoordinated and scattered throughout the scientific literature without a clear epistemic goal or connection between the studies. Therefore, our main contribution to the scientific literature is to offer a road map to help solve problems in drug design, protein structures, design, and function prediction while also presenting the "state of the art" on research in the AI-PS binomial until February 2021. Thus, we pave the way toward future advances in the synthetic redesign of novel proteins and protein networks and artificial metabolic pathways, learning lessons from nature for the welfare of humankind. Many of the novel proteins and metabolic pathways are currently non-existent in nature, nor are they used in the chemical industry or biomedical field.

Evaluating Pointwise Reliability of Machine Learning prediction

Interest in Machine Learning applications to tackle clinical and biological problems is increasing. This is driven by promising results reported in many research papers, the increasing number of AI-based software products, and by the general interest in Artificial Intelligence to solve complex problems. It is therefore of importance to improve the quality of machine learning output and add safeguards to support their adoption. In addition to regulatory and logistical strategies, a crucial aspect is to detect when a Machine Learning model is not able to generalize to new unseen instances, which may originate from a population distant to that of the training population or from an under-represented subpopulation. As a result, the prediction of the machine learning model for these instances may be often wrong, given that the model is applied outside its "reliable" space of work, leading to a decreasing trust of the final users, such as clinicians. For this reason, when a model is deployed in practice, it would be important to advise users when the model's predictions may be unreliable, especially in high-stakes applications, including those in healthcare. Yet, reliability assessment of each machine learning prediction is still poorly addressed. Here, we review approaches that can support the identification of unreliable predictions, we harmonize the notation and terminology of relevant concepts, and we highlight and extend possible interrelationships and overlap among concepts. We then demonstrate, on simulated and real data for ICU in-hospital death prediction, a possible integrative framework for the identification of reliable and unreliable predictions. To do so, our proposed approach implements two complementary principles, namely the density principle and the local fit principle. The density principle verifies that the instance we want to evaluate is similar to the training set. The local fit principle verifies that the trained model performs well on training subsets that are more similar to the instance under evaluation. Our work can contribute to consolidating work in machine learning especially in medicine.

Oncogenic mutations on Rac1 affect global intrinsic dynamics underlying GTP and PAK1 binding

Rac1 is a small member of the Rho GTPase family. One of the most important downstream effectors of Rac1 is a serine/threonine kinase, p21-activated kinase 1 (PAK1). Mutational activation of PAK1 by Rac1 has oncogenic signaling effects. Here, although we focus on Rac1-PAK1 interaction by atomic-force-microscopy-based single-molecule force spectroscopy experiments, we explore the effect of active mutations on the intrinsic dynamics and binding interactions of Rac1 by Gaussian network model analysis and molecular dynamics simulations. We observe that Rac1 oncogenic mutations are at the hinges of three global modes of motion, suggesting the mechanical changes as potential markers of oncogenicity. Indeed, the dissociation of wild-type Rac1-PAK1 complex shows two distinct unbinding dynamic states that are reduced to one with constitutively active Q61L and oncogenic Y72C mutant Rac1, as revealed by single-molecule force spectroscopy experiments. Q61L and Y72C mutations change the mechanics of the Rac1-PAK1 complex by increasing the elasticity of the protein and slowing down the transition to the unbound state. On the other hand, Rac1's intrinsic dynamics reveal more flexible GTP and PAK1-binding residues on switches I and II with Q61L, Y72C, oncogenic P29S and Q61R, and negative T17N mutations. The cooperativity in the fluctuations of GTP-binding sites around the p-loop and switch I decreases in all mutants, mostly in Q61L, whereas some PAK1-binding residues display enhanced coupling with GTP-binding sites in Q61L and Y72C and within each other in P29S. The predicted binding free energies of the modeled Rac1-PAK1 complexes show that the change in the dynamic behavior likely means a more favorable PAK1 interaction. Overall, these findings suggest that the active mutations affect intrinsic functional dynamic events and alter the mechanics underlying the binding of Rac1 to GTP and upstream and downstream partners including PAK1.

Computational design of enzymes for biotechnological applications

Enzymes are the natural catalysts that execute biochemical reactions upholding life. Their natural effectiveness has been fine-tuned as a result of millions of years of natural evolution. Such catalytic effectiveness has prompted the use of biocatalysts from multiple sources on different applications, including the industrial production of goods (food and beverages, detergents, textile, and pharmaceutics), environmental protection, and biomedical applications. Natural enzymes often need to be improved by protein engineering to optimize their function in non-native environments. Recent technological advances have greatly facilitated this process by providing the experimental approaches of directed evolution or by enabling computer-assisted applications. Directed evolution mimics the natural selection process in a highly accelerated fashion at the expense of arduous laboratory work and economic resources. Theoretical methods provide predictions and represent an attractive complement to such experiments by waiving their inherent costs. Computational techniques can be used to engineer enzymatic reactivity, substrate specificity and ligand binding, access pathways and ligand transport, and global properties like protein stability, solubility, and flexibility. Theoretical approaches can also identify hotspots on the protein sequence for mutagenesis and predict suitable alternatives for selected positions with expected outcomes. This review covers the latest advances in computational methods for enzyme engineering and presents many successful case studies.

Convolution Neural Network-Based Prediction of Protein Thermostability

Most natural proteins exhibit poor thermostability, which limits their industrial application. Computer-aided rational design is an efficient purpose-oriented method that can improve protein thermostability. Numerous machine-learning-based methods have been designed to predict the changes in protein thermostability induced by mutations. However, all of these methods have certain limitations due to existing mutation coding methods that overlook protein sequence features. Here we propose a method to predict protein thermostability using convolutional neural networks based on an in-depth study of thermostability-related protein properties. This method comprises a three-dimensional coding algorithm, including protein mutation information and a strategy to extract neighboring features at protein mutation sites based on multiscale convolution. The accuracies on the S1615 and S388 data sets, which are widely used for protein thermostability predictions, reached 86.4 and 87%, respectively. The Matthews correlation coefficient was nearly double those produced using other methods. Furthermore, a model was constructed to predict the thermostability of Rhizomucor miehei lipase mutants based on the S3661 data set, a single amino acid mutation data set screened from the ProTherm protein thermodynamics database. Compared with the RIF strategy, which consists of three algorithms, i.e., Rosetta ddg monomer, I Mutant 3.0, and FoldX, the accuracy of the proposed method was higher (75.0 vs 66.7%), and the negative sample resolution was simultaneously enhanced. These results indicate that our prediction method more effectively assessed the protein thermostability and distinguished its features, making it a powerful tool to devise mutations that enhance the thermostability of proteins, particularly enzymes.

Integrative approaches to reconstruct regulatory networks from multi-omics data: A review of state-of-the-art methods

Data generation using high throughput technologies has led to the accumulation of diverse types of molecular data. These data have different types (discrete, real, string, etc.) and occur in various formats and sizes. Datasets including gene expression, miRNA expression, protein-DNA binding data (ChIP-Seq/ChIP-ChIP), mutation data (copy number variation, single nucleotide polymorphisms), annotations, interactions, and association data are some of the commonly used biological datasets to study various cellular mechanisms of living organisms. Each of them provides a unique, complementary and partly independent view of the genome and hence embed essential information about the regulatory mechanisms of genes and their products. Therefore, integrating these data and inferring regulatory interactions from them offer a system level of biological insight in predicting gene functions and their phenotypic outcomes. To study genome functionality through regulatory networks, different methods have been proposed for collective mining of information from an integrated dataset. We survey here integration methods that reconstruct regulatory networks using state-of-the-art techniques to handle multi-omics (i.e., genomic, transcriptomic, proteomic) and other biological datasets.

Text mining electronic hospital records to automatically classify admissions against disease: Measuring the impact of linking data sources

OBJECTIVE

Text and data mining play an important role in obtaining insights from Health and Hospital Information Systems. This paper presents a text mining system for detecting admissions marked as positive for several diseases: Lung Cancer, Breast Cancer, Colon Cancer, Secondary Malignant Neoplasm of Respiratory and Digestive Organs, Multiple Myeloma and Malignant Plasma Cell Neoplasms, Pneumonia, and Pulmonary Embolism. We specifically examine the effect of linking multiple data sources on text classification performance.

METHODS

Support Vector Machine classifiers are built for eight data source combinations, and evaluated using the metrics of Precision, Recall and F-Score. Sub-sampling techniques are used to address unbalanced datasets of medical records. We use radiology reports as an initial data source and add other sources, such as pathology reports and patient and hospital admission data, in order to assess the research question regarding the impact of the value of multiple data sources. Statistical significance is measured using the Wilcoxon signed-rank test. A second set of experiments explores aspects of the system in greater depth, focusing on Lung Cancer. We explore the impact of feature selection; analyse the learning curve; examine the effect of restricting admissions to only those containing reports from all data sources; and examine the impact of reducing the sub-sampling. These experiments provide better understanding of how to best apply text classification in the context of imbalanced data of variable completeness.

RESULTS

Radiology questions plus patient and hospital admission data contribute valuable information for detecting most of the diseases, significantly improving performance when added to radiology reports alone or to the combination of radiology and pathology reports.

CONCLUSION

Overall, linking data sources significantly improved classification performance for all the diseases examined. However, there is no single approach that suits all scenarios; the choice of the most effective combination of data sources depends on the specific disease to be classified.

Predicting protein thermal stability changes upon point mutations using statistical potentials: Introducing HoTMuSiC

The accurate prediction of the impact of an amino acid substitution on the thermal stability of a protein is a central issue in protein science, and is of key relevance for the rational optimization of various bioprocesses that use enzymes in unusual conditions. Here we present one of the first computational tools to predict the change in melting temperature ΔT_m upon point mutations, given the protein structure and, when available, the melting temperature T_m of the wild-type protein. The key ingredients of our model structure are standard and temperature-dependent statistical potentials, which are combined with the help of an artificial neural network. The model structure was chosen on the basis of a detailed thermodynamic analysis of the system. The parameters of the model were identified on a set of more than 1,600 mutations with experimentally measured ΔT_m. The performance of our method was tested using a strict 5-fold cross-validation procedure, and was found to be significantly superior to that of competing methods. We obtained a root mean square deviation between predicted and experimental ΔT_m values of 4.2 °C that reduces to 2.9 °C when ten percent outliers are removed. A webserver-based tool is freely available for non-commercial use at soft.dezyme.com.

Protein thermostability engineering

Using structure and sequence based analysis we can engineer proteins to increase their thermal stability.

Methods for biological data integration: perspectives and challenges

Rapid technological advances have led to the production of different types of biological data and enabled construction of complex networks with various types of interactions between diverse biological entities. Standard network data analysis methods were shown to be limited in dealing with such heterogeneous networked data and consequently, new methods for integrative data analyses have been proposed. The integrative methods can collectively mine multiple types of biological data and produce more holistic, systems-level biological insights. We survey recent methods for collective mining (integration) of various types of networked biological data. We compare different state-of-the-art methods for data integration and highlight their advantages and disadvantages in addressing important biological problems. We identify the important computational challenges of these methods and provide a general guideline for which methods are suited for specific biological problems, or specific data types. Moreover, we propose that recent non-negative matrix factorization-based approaches may become the integration methodology of choice, as they are well suited and accurate in dealing with heterogeneous data and have many opportunities for further development.

Stability curve prediction of homologous proteins using temperature-dependent statistical potentials

The unraveling and control of protein stability at different temperatures is a fundamental problem in biophysics that is substantially far from being quantitatively and accurately solved, as it requires a precise knowledge of the temperature dependence of amino acid interactions. In this paper we attempt to gain insight into the thermal stability of proteins by designing a tool to predict the full stability curve as a function of the temperature for a set of 45 proteins belonging to 11 homologous families, given their sequence and structure, as well as the melting temperature () and the change in heat capacity () of proteins belonging to the same family. Stability curves constitute a fundamental instrument to analyze in detail the thermal stability and its relation to the thermodynamic stability, and to estimate the enthalpic and entropic contributions to the folding free energy. In summary, our approach for predicting the protein stability curves relies on temperature-dependent statistical potentials derived from three datasets of protein structures with targeted thermal stability properties. Using these potentials, the folding free energies () at three different temperatures were computed for each protein. The Gibbs-Helmholtz equation was then used to predict the protein's stability curve as the curve that best fits these three points. The results are quite encouraging: the standard deviations between the experimental and predicted 's, 's and folding free energies at room temperature () are equal to 13 , 1.3 ) and 4.1 , respectively, in cross-validation. The main sources of error and some further improvements and perspectives are briefly discussed.

The prediction of protein stability remains one of the key goals of protein science. Despite the significant efforts of the last decades, faster and more accurate stability predictors on the proteomic-wide scale are currently demanded. The determination and control of protein stability are indeed fundamental steps on the path towards de novo design. In this paper we develop a method for predicting the stability curve of proteins. This curve encodes the temperature dependence of the folding free energy (). Its knowledge is important in the study of protein stability since all the thermodynamic parameters characterizing the folding transition can be extracted from it. Our prediction method is based on temperature-dependent mean force potentials and uses the tertiary structure of the target protein as well as the melting temperature () and the heat capacity change () of some other proteins belonging to the same family. From the predicted stability curves, the , the and the at room temperature can be inferred. The predictions obtained are compared with experimental data and show reasonable performances.

Computational and experimental approaches to reveal the effects of single nucleotide polymorphisms with respect to disease diagnostics

DNA mutations are the cause of many human diseases and they are the reason for natural differences among individuals by affecting the structure, function, interactions, and other properties of DNA and expressed proteins. The ability to predict whether a given mutation is disease-causing or harmless is of great importance for the early detection of patients with a high risk of developing a particular disease and would pave the way for personalized medicine and diagnostics. Here we review existing methods and techniques to study and predict the effects of DNA mutations from three different perspectives: in silico, in vitro and in vivo. It is emphasized that the problem is complicated and successful detection of a pathogenic mutation frequently requires a combination of several methods and a knowledge of the biological phenomena associated with the corresponding macromolecules.

Feature-based multiple models improve classification of mutation-induced stability changes

Background

Reliable prediction of stability changes in protein variants is an important aspect of computational protein design. A number of machine learning methods that allow a classification of stability changes knowing only the sequence of the protein emerged. However, their performance on amino acid substitutions of previously unseen non-homologous proteins is rather limited. Moreover, the performance varies for different types of mutations based on the secondary structure or accessible surface area of the mutation site.

Results

We proposed feature-based multiple models with each model designed for a specific type of mutations. The new method is composed of five models trained for mutations in exposed, buried, helical, sheet, and coil residues. The classification of a mutation as stabilising or destabilising is made as a consensus of two models, one selected based on the predicted accessible surface area and the other based on the predicted secondary structure of the mutation site. We refer to our new method as Evolutionary, Amino acid, and Structural Encodings with Multiple Models (EASE-MM). Cross-validation results show that EASE-MM provides a notable improvement to our previous work reaching a Matthews correlation coefficient of 0.44. EASE-MM was able to correctly classify 73% and 75% of stabilising and destabilising protein variants, respectively. Using an independent test set of 238 mutations, we confirmed our results in a comparison with related work.

Conclusions

EASE-MM not only outperformed other related methods but achieved more balanced results for different types of mutations based on the accessible surface area, secondary structure, or magnitude of stability changes. This can be attributed to using multiple models with the most relevant features selected for the given type of mutations. Therefore, our results support the presumption that different interactions govern stability changes in the exposed and buried residues or in residues with a different secondary structure.

Towards sequence-based prediction of mutation-induced stability changes in unseen non-homologous proteins

Background

Reliable prediction of stability changes induced by a single amino acid substitution is an important aspect of computational protein design. Several machine learning methods capable of predicting stability changes from the protein sequence alone have been introduced. Prediction performance of these methods is evaluated on mutations unseen during training. Nevertheless, different mutations of the same protein, and even the same residue, as encountered during training are commonly used for evaluation. We argue that a faithful evaluation can be achieved only when a method is tested on previously unseen proteins with low sequence similarity to the training set.

Results

We provided experimental evidence of the limitations of the evaluation commonly used for assessing the prediction performance. Furthermore, we demonstrated that the prediction of stability changes in previously unseen non-homologous proteins is a challenging task for currently available methods. To improve the prediction performance of our previously proposed method, we identified features which led to over-fitting and further extended the model with new features. The new method employs Evolutionary And Structural Encodings with Amino Acid parameters (EASE-AA). Evaluated with an independent test set of more than 600 mutations, EASE-AA yielded a Matthews correlation coefficient of 0.36 and was able to classify correctly 66% of the stabilising and 74% of the destabilising mutations. For real-value prediction, EASE-AA achieved the correlation of predicted and experimentally measured stability changes of 0.51.

Conclusions

Commonly adopted evaluation with mutations in the same protein, and even the same residue, randomly divided between the training and test sets lead to an overestimation of prediction performance. Therefore, stability changes prediction methods should be evaluated only on mutations in previously unseen non-homologous proteins. Under such an evaluation, EASE-AA predicts stability changes more reliably than currently available methods.

Electronic supplementary material

The online version of this article (doi:10.1186/1471-2164-15-S1-S4) contains supplementary material, which is available to authorized users.

Sequence-only evolutionary and predicted structural features for the prediction of stability changes in protein mutants

Background

Even a single amino acid substitution in a protein sequence may result in significant changes in protein stability, structure, and therefore in protein function as well. In the post-genomic era, computational methods for predicting stability changes from only the sequence of a protein are of importance. While evolutionary relationships of protein mutations can be extracted from large protein databases holding millions of protein sequences, relevant evolutionary features for the prediction of stability changes have not been proposed. Also, the use of predicted structural features in situations when a protein structure is not available has not been explored.

Results

We proposed a number of evolutionary and predicted structural features for the prediction of stability changes and analysed which of them capture the determinants of protein stability the best. We trained and evaluated our machine learning method on a non-redundant data set of experimentally measured stability changes. When only the direction of the stability change was predicted, we found that the best performance improvement can be achieved by the combination of the evolutionary features mutation likelihood and SIFTscore in conjunction with the predicted structural feature secondary structure. The same two evolutionary features in the combination with the predicted structural feature accessible surface area achieved the lowest error when the prediction of actual values of stability changes was assessed. Compared to similar studies, our method achieved improvements in prediction performance.

Conclusion

Although the strongest feature for the prediction of stability changes appears to be the vector of amino acid identities in the sequential neighbourhood of the mutation, the most relevant combination of evolutionary and predicted structural features further improves prediction performance. Even the predicted structural features, which did not perform well on their own, turn out to be beneficial when appropriately combined with evolutionary features. We conclude that a high prediction accuracy can be achieved knowing only the sequence of a protein when the right combination of both structural and evolutionary features is used.

Evaluation of data integration strategies based on kernel method of clinical and microarray data

The cancer classification problem is one of the most challenging problems in bioinformatics. The data provided by Netherland Cancer Institute consists of 295 breast cancer patient; 101 patients are with distant metastases and 194 patients are without distant metastases. Combination of features sets based on kernel method to classify the patient who are with or without distant metastases will be investigated. The single data set will be compared with three data integration strategies and also weighted data integration strategies based on kernel method. Least Square Support Vector Machine (LS-SVM) is chosen as the classifier because it can handle very high dimensional features, for instance, microarray data. The experiment result shows that the performance of weighted late integration and the using of only microarray data are almost similar. The data integration strategy is not always better than using single data set in this case. The performance of classification absolutely depends on the features that are used to represent the object.