Computational proteomics: management and analysis of proteomics data

Proteomics is about the study of the proteins expressed in a cell, organism, or tissue. This includes protein identification and quantification (or quantitation), protein–protein interactions, protein complexes prediction, protein modifications and protein localization in the cell. Mass Spectrometry (MS) is one of the main technologies in proteomics and is more and more used for its increasing precision and for the possibility to automate the proteomics analysis pipeline, yielding to large-scale high-throughput experiments. Since proteins play a central role in the life of an organism, proteomics is instrumental in many biomedical applications, such as biomarker discovery and drug treatment evaluation, as well as for investigating the dynamics of cells in Systems Biology. Computational Proteomics is about the computational methods, algorithms, databases and methodologies used to process, manage, analyze and interpret the data produced in proteomics experiments. The broad application of proteomics in different biological and medical fields, as well as the diffusion of high-throughput platforms, leads to increasing volumes of available proteomics data requiring efficient algorithms, new data management capabilities and novel analysis, inference and visualization techniques. Moreover, high-throughput production and collection of data pose new challenges in data handling and reusability as well as in tools interoperability and interconnection. On the other hand, the increasing availability of data and tools opens new research directions and opportunities (e.g. annotated spectral libraries) that can be exploited only through the rigorous application of computer science, machine learning, knowledge discovery, statistics and signal processing techniques. As in many scientific disciplines, the final goal of Computational Proteomics is to infer knowledge models (e.g. verify a hypothesis or identify proteins involved in a disease) from the inspection of biological samples. Such an activity involves different steps, happening both in wet and dry lab, and is the result of combination of many instruments, methods, tools, algorithms, databases, according to established or emerging working methodologies and standards. This special issue explores the current state-of-the-art research taking place in different areas of Computational Proteomics, with special emphasis on methodologies and tools for data handling and analysis, machine learning, knowledge discovery, biomarker discovery, data standardization and information quality, in MS-based proteomics. The discussion of all experimental techniques and computational methods taking place in proteomics, would be too large to be hosted in one journal issue, thus, this special issue is complemented by the March 2008 issue of Briefings in Functional Genomics and Proteomics, that discusses, among others, MS-based techniques for improving the study of protein structure and protein folding, as well as some applications and tools in Interactomics and Systems Biology. MS permits, with high accuracy, the determination of molecular weight of chemical compounds, ranging from small molecules to large, polar biopolymers [1]. The mass spectrometer separates gas phase ions according to their mass to charge ratio values. The output of the spectrometer (the spectrum) is a large sequence of value pairs. Each pair contains a measured intensity and a mass to charge ratio (m/z), which depend, respectively, on the quantity and molecular mass of the detected molecule. Although MS is only able to identify molecular masses, the adoption of advanced sample preparation techniques, such as chromatographic separation or labeling techniques, the use of protein database information, e.g. the theoretical spectrum associated to a protein sequence, and the application of powerful machine learning algorithms, make this technique a basic cornerstone for proteomics applications. In a proteomics experiment different steps can be identified, each one having one or more parameters of choice: (i) sample preparation, including separation and labeling; (ii) MS experiment, including mass spectrometer choice and configuration; (iii) spectra preprocessing, including spectra signal deconvolution, often happening inside the spectrometer software, baseline subtraction, noise removal, dimension reduction, peaks extraction, outlier detection; (iv) peptide/protein identification, including database searching, eventually coupled with de novo sequencing or sequence tagging; (v) peptide/protein quantitation, either performed through stable isotope labeling or through intensity profiling and (vi) knowledge discovery, including biomarker discovery in clinical applications, systems biology modeling and biological behavior explanation. In this step, a major problem to be faced by data mining algorithms is the extremely high dimensionality of proteomics data (i.e. the number of features or variables) that largely exceeds the size of sample set. While recent papers concentrate especially on protein/peptide identification and quantitation [2, 3], this special issue focuses on the overall knowledge discovery process behind computational proteomics, with special emphasis on machine learning methods, spectra data handling, biomarker discovery, standard-based and quality-aware management of proteomics experiments. The first group of papers expounds the biomarker discovery problem, presenting general methodologies for dimensionality reduction (DR) in proteomics data, machine learning methods and way to interconnect them for conducting biomarker discovery studies and two novel classification algorithms with special features for proteomics studies. The second group of papers reviews basic algorithms for spectra analysis and management and describes issues and opportunities when managing large datasets. Peptide/protein identification and quantitation are the basic building blocks for the comparative analysis of quantitative information (i.e. protein expression level) about identified proteins. Methods for protein quantitation and in particular for label-free comparative quantitation are reviewed and compared. The third group of papers presents two important and timely issues: standardization of proteomics data and methodologies and tools to ensure information quality in proteomics experiments. The main goal of biomarker discovery is to find the most discriminating features in a classification of samples (e.g. peaks or peptides/proteins discriminating healthy versus diseased subjects). An important problem is the huge number of features (e.g. peaks or peptides/proteins) that may represent potential biomarkers in a mass spectrum. The low number of cases/controls with respect to the number of features is another important problem. Moreover, collected data are sparse. The combination of such characteristics yields the so-called high-dimensional small-sample problem that may lead to selection bias and information leakage when analyzing data. Hilario and Kalousis describe in a comprehensive way the computational methods involved in biomarker discovery studies. They first focus on the high-dimensional small-sample problem, underlying significant data-analytical issues raised by biomarker discovery. Then, they describe the phases of the knowledge discovery process—see step (vi) explained earlier—in biomarker discovery, focusing on DR methods, learning methods and the way they can be coupled. The article provides taxonomy of DR methods that includes feature transformation and feature selection methods. The former transform or combine old features generating a smaller set of new features, while the latter eliminate directly irrelevant or redundant features. Considering how DR is coupled to the learning process, the authors identify filter methods that perform DR as a preprocessing step before the learning method, and wrapper methods that wrap DR around a specific learning process. The article then discusses the so-called DR method selection problem, i.e. the criteria and an objective evaluation methodology, which may be used to select a DR method. Barla et al. propose a Design Analysis Protocol (DAP) for organizing and controlling the different steps involved in biomarker discovery. The aim of the proposed methodology, inspired by similar initiatives in the microarray community, is to provide a set of guidelines for the development and validation of predictive models or classifiers. In order to ensure reproducibility of experiments and reliability of produced results, the authors focus on two critical problems in MS-based proteomics: information leakage and selection bias, also considering the above cited high-dimensional small-sample problem. After defining main blocks of a general predictive proteomics pipeline and illustrating key characteristics of elementary functions (e.g. spectra preprocessing, feature selection and learning models) they argue that reproducibility strongly depends on the methods used for selecting training and test datasets from the original dataset. Experimental studies based on publicly available datasets and on synthetic data are provided, including an example showing the effects of varying the preprocessing module in the pipeline. Villmann et al. present two recently developed classification algorithms for analysis of spectra that face the high-dimensional small-sample problem: the supervised neural gas and the fuzzy labeled self-organizing map algorithms. From a mathematical point of view, a consequence of the high-dimensional small-sample problem is that the data space to be explored is sparsely filled. Data separation during classification requires detecting the underlying data regularities. The presented algorithms are inherently self-regularizing and thus are able to deal with high-dimensional, sparse and noisy data. Moreover, the fuzzy labeled self-organizing map algorithm allows the processing of uncertain class information and returns a fuzzy classification scheme that allows deep class analysis. Moreover, encoding similar class information with similar colors allows class dependent data visualization. The way in which the proteomics steps (i)−(vi) are applied makes computational proteomics a multidimensional activity, leading to possibly diverse intermediate results on the basis of the choices made at each step. The second and third groups of papers deal with these aspects. For instance, the different combination of separation techniques [e.g. two-dimensional gel electrophoresis (2-DE) or liquid chromatography (LC)], ionization techniques [e.g. surface enhanced laser desorption ionization (SELDI) or matrix-assisted laser desorption ionization (MALDI)] and mass spectrometer type and configuration (e.g. MS, tandem MS/MS, etc.), as well as any eventual labeling techniques [e.g. isotope-coded affinity tags (ICAT) or stable isotope labeling with amino acids in cell culture (SILAC)], leads to different kind of result spectra, that need to be interpreted accordingly. Thus, steps (i)–(ii) determine the kind of produced spectra and the kind of proteomics analysis that can be performed. The preprocessing step (iii) contains other elements of choice and represents a tradeoff between the need to produce compact data to be analyzed and the need to avoid information leakage and selection bias in further knowledge discovery processes. Veltri, in his paper, discusses characteristics of spectra data, describing the data management issues posed when large datasets have to be managed, and underlying how the results obtained by the database, XML and information retrieval communities, can be beneficial to Computational Proteomics. As an example, an improvement of the peptide/protein identification process based on a library of peptides and advanced database technologies is presented. Moreover, an application of time series analysis to biomarker discovery is also discussed. Stable isotope labeling is a common technique for comparing protein levels between samples, but it requires a careful sample preparation that can be costly for large numbers of samples. The comparative quantification of label-free LC-MS/MS data is an emerging alternative to isotope labeling. The main idea is that under well-controlled conditions identical peptides across different LC-MS/MS datasets can be compared directly without the use of isotope labeling. Wong et al. review two main computational methods for comparative quantitative analysis of LC-MS proteomics data: extraction of peptide ion intensities and spectral counting. They survey available algorithms and tools and discuss statistical tests for evaluating the significance of the comparative results. They argue that the computational process behind intensity-based quantification is significantly more complex compared with spectral counting. On the other hand, spectral counting is more likely to be influenced by the acquisition program of the mass spectrometer and high abundance proteins can mask low abundance ones. Thus, although spectral counting represents the simplest method, both methods could be used simultaneously to improve confidence. Continuing along the computational proteomics pipeline, it is possible to note as the binding between the kind of spectra produced and the type of performed analysis is done by manual association or is implicit when using the software tools provided by the mass spectrometer vendor. While portable, open and vendor-neutral representation of structured data like spectra can be provided by the XML and RDF World Wide Web languages, semantic associations between the type, meaning and format of the generated data and the applicable preprocessing and analysis activities could be managed through domain ontologies. The use of ontologies to model the overall analysis pipeline and the relationship and interconnection between the steps of proteomics analysis is in its beginning [4]. Orchard and Hermjakob describe the HUPO (Human Proteome Organization) Proteomics Standards Initiative, which aims to provide the data standards and interchange formats for the storage and sharing of proteomics data. They present the current state of these standards and describe some standards compliant proteomics data repositories. It should be noted that such emerging initiative includes controlled vocabularies for guiding metadata generation and data annotation and domain ontologies (e.g. the HUPO PSI Mass Spectrometry mzOntology) to model relationship between objects of the domain. Currently available proteomics pipelines, although offering a set of tools for preprocessing and analysis in a comprehensive suite, do not offer the possibility of varying the internal preprocessing or analysis tools. While interoperability at the data layer is starting to be implemented through the exchange of data in standard formats (e.g. the HUPO PSI mzML standard discussed in Orchard and Hermjakob), interoperability at the application level is far from being completed. Web services and workflows are key technologies for enabling application-to-application interoperability, overcoming the limits of current web-based bioinformatics tools that are accessed through human intervention [5, 6]. Different bioinformatics tools are beginning to be offered as web services that can be composed through workflows, but agreed semantics and standard result formats are needed. Orchard and Hermjakob also report on the HUPO PSI common syntax for peptide/protein identification and for protein modification description, useful for capturing results from MS search engines and for giving them as input parameters to further analysis tools. Because the proteomics pipeline presents many steps and different options for each of them, the presence of inaccuracies throughout the processing pipeline may prevent reuse of data for comparative analysis and for experiment validation. Information quality of proteomics experiments and proteomics data repositories is thus an important issue that needs to be managed in an explicit way throughout the entire proteomics pipeline. Stead et al. describe factors that impact on the quality of experimental data and review current approaches for information quality management in proteomics. Data quality issues are considered along the entire pipeline of a proteomics experiment, from experiment design and technique selection, through data analysis, to archiving and sharing. To do this they define a set of quality parameters (e.g. completeness, reproducibility, accuracy and uniformity of information), and attempt to identify how they are affected, focusing on 2-DE and MS technologies and peptide/protein identification. They describe minimum information guidelines, good practice guidelines and standard data formats for achieving and managing information quality. Finally, they introduce an information quality management system with the capability to measure, filter and flag proteomics data based on its quality, and able to access quality-checking web services. In summary, high-throughput MS facilities must be coupled to flexible and accurate computational proteomics software platforms to result in an effective data analysis pipeline. Basic components of such platforms are specialized spectra services as well as advanced machine learning tools, surrounded by orthogonal services for data validation and quality management. Interoperability of basic software tools can be obtained through web services and workflows, while proteomics data repositories and ontologies can allow knowledge sharing and integration, opening new research directions. This special issue has been originated by the Special Track on Computational Proteomics held at the IEEE Symposium on Computer-Based Medical Systems since 2006. It would not have been possible without the encouragement of Tim Clark. A special thanks goes to my friends Gianni Cuda and Marco Gaspari for having introduced me in the world of proteomics.