Feature Selection Methods for Protein Biomarker Discovery from Proteomics or Multiomics Data

Abstract

•New algorithms enable protein biomarker discovery from proteomics or multiomics data.•Superior performance is demonstrated in two clinically important classification problems.•Feature clusters facilitate functional interpretation of the identified protein biomarkers.•Alternative choices are provided for each identified protein biomarker. Untargeted mass spectrometry (MS)-based proteomics provides a powerful platform for protein biomarker discovery, but clinical translation depends on the selection of a small number of proteins for downstream verification and validation. Due to the small sample size of typical discovery studies, protein markers identified from discovery data may not be generalizable to independent datasets. In addition, a good protein marker identified using a discovery platform may be difficult to implement in verification and validation platforms. Moreover, although multiomics characterization is being increasingly used in discovery cohort studies, there is no existing method for multiomics-facilitated protein biomarker selection. Here, we present ProMS, a computational algorithm for protein marker selection. The algorithm is based on the hypothesis that a phenotype is characterized by a few underlying biological functions, each manifested by a group of coexpressed proteins. A weighted k-medoids clustering algorithm is applied to all univariately informative proteins to identify both coexpressed protein clusters and a representative protein for each cluster as markers. In two clinically important classification problems, ProMS shows superior performance compared with existing feature selection methods. ProMS can be extended to the multiomics setting (ProMS_mo) through a constrained weighted k-medoids clustering algorithm, and the protein panels selected by ProMS_mo show improved performance on independent test data compared with ProMS. In addition to superior performance, ProMS and ProMS_mo also have two unique strengths. First, the feature clusters enable functional interpretation of the selected protein markers. Second, the feature clusters provide an opportunity to select replacement protein markers, facilitating a robust transition to the verification and validation platforms. In summary, this study provides a unified and effective computational framework for selecting protein biomarkers using proteomics or multiomics data. The software implementation is publicly available at https://github.com/bzhanglab/proms. Untargeted mass spectrometry (MS)-based proteomics provides a powerful platform for protein biomarker discovery, but clinical translation depends on the selection of a small number of proteins for downstream verification and validation. Due to the small sample size of typical discovery studies, protein markers identified from discovery data may not be generalizable to independent datasets. In addition, a good protein marker identified using a discovery platform may be difficult to implement in verification and validation platforms. Moreover, although multiomics characterization is being increasingly used in discovery cohort studies, there is no existing method for multiomics-facilitated protein biomarker selection. Here, we present ProMS, a computational algorithm for protein marker selection. The algorithm is based on the hypothesis that a phenotype is characterized by a few underlying biological functions, each manifested by a group of coexpressed proteins. A weighted k-medoids clustering algorithm is applied to all univariately informative proteins to identify both coexpressed protein clusters and a representative protein for each cluster as markers. In two clinically important classification problems, ProMS shows superior performance compared with existing feature selection methods. ProMS can be extended to the multiomics setting (ProMS_mo) through a constrained weighted k-medoids clustering algorithm, and the protein panels selected by ProMS_mo show improved performance on independent test data compared with ProMS. In addition to superior performance, ProMS and ProMS_mo also have two unique strengths. First, the feature clusters enable functional interpretation of the selected protein markers. Second, the feature clusters provide an opportunity to select replacement protein markers, facilitating a robust transition to the verification and validation platforms. In summary, this study provides a unified and effective computational framework for selecting protein biomarkers using proteomics or multiomics data. The software implementation is publicly available at https://github.com/bzhanglab/proms. According to the definition from the Food and Drug Administration–National Institutes of Health (FDA-NIH) Biomarker Working Group, a biomarker is “a defined characteristic that is measured as an indicator of normal biological processes, pathogenic processes or responses to an exposure or intervention” (1FDA-NIH Biomarker Working Group, BEST (Biomarkers, EndpointS, and Other Tools) Resource. Maryland: Silver Spring, MD: 2016.Google Scholar). Being the functional molecules of the cell, proteins have long been recognized as an important source of putative biomarkers for disease diagnosis, prognosis, and response to therapeutic intervention. Since the approval of human hemoglobin as a fecal test for the detection of colorectal cancer in 1976, more than 20 tumor protein markers have been approved by the FDA and are currently used in clinical practice (2Füzéry A.K. Levin J. Chan M.M. Chan D.W. Translation of proteomic biomarkers into FDA approved cancer diagnostics: Issues and challenges.Clin. Proteomics. 2013; 10: 13Crossref PubMed Scopus (237) Google Scholar). Protein biomarker development typically includes three phases: discovery, verification, and validation (3Parker C.E. Borchers C.H. Mass spectrometry based biomarker discovery, verification, and validation--quality assurance and control of protein biomarker assays.Mol. Oncol. 2014; 8: 840-858Crossref PubMed Scopus (137) Google Scholar, 4Rifai N. Gillette M.A. Carr S.A. Protein biomarker discovery and validation: The long and uncertain path to clinical utility.Nat. Biotechnol. 2006; 24: 971-983Crossref PubMed Scopus (1263) Google Scholar). The discovery phase is now powered by the mass spectrometry (MS)-based untargeted proteomics technology, which enables the identification and quantification of more than 10,000 proteins in clinical specimens (5Mertins P. Tang L.C. Krug K. Clark D.J. Gritsenko M.A. Chen L. Clauser K.R. Clauss T.R. Shah P. Gillette M.A. Petyuk V.A. Thomas S.N. Mani D.R. Mundt F. Moore R.J. et al.Reproducible workflow for multiplexed deep-scale proteome and phosphoproteome analysis of tumor tissues by liquid chromatography-mass spectrometry.Nat. Protoc. 2018; 13: 1632-1661Crossref PubMed Scopus (117) Google Scholar). This provides an excellent opportunity to identify new protein biomarker candidates in an unbiased manner. Moreover, it is well recognized that a combination of biomarkers, rather than an individual protein, is needed to distinguish biological states. By quantifying all proteins simultaneously, MS proteomics provides an ideal platform for identifying biomarker combinations. Despite the immense promise of MS proteomics in protein biomarker discovery, few new biomarkers have been introduced into clinical practice during the past decade. One of the rate-limiting steps in protein biomarker development is the identification of a small number of promising candidates from thousands of proteins quantified by untargeted MS proteomics for downstream verification and validation using targeted assays. Although MS-based discovery platforms provide measurements for a large number of proteins (i.e., features), they are often carried out using a limited number of samples, leading to the “large p, small n” problem (6T. Hastie, R. Tibshirani and J. Friedman, The Elements of Statistical Learning: Data Mining, Inference, and Prediction. Berlin, Germany: Springer Science & Business Media, 2009.Google Scholar). This challenge is typical in all omics-based association studies and is commonly addressed by dimension reduction techniques such as principal component analysis (PCA) and its supervised alternatives (7Bair E. Hastie T. Paul D. Tibshirani R. Prediction by supervised principal components.J. Am. Stat. Assoc. 2006; 101: 119-137Crossref Scopus (401) Google Scholar). The goal of PCA is to rotate the data into a new axis system where the greatest amount of variance is captured in a few dimensions. Transformed data are represented by a set of principal components (PCs) ordered by the amount of variance they capture. Usually, a small number of PCs capture most of the variance of a dataset, leading to dimension reduction. However, because each PC is a linear combination of all original features, predictive models constructed based on the PCs require genome-wide measurements as inputs and cannot be implemented as targeted clinical assays. Feature selection algorithms can be used to select biomarker combinations from high-dimensional data for predictive model construction. Throughout the paper, we use the terms “feature” and “marker” interchangeably. Feature selection algorithms can be categorized into filter methods, wrapper methods, and embedded methods. Filter methods evaluate the importance of features according to some univariate or multivariate evaluation criteria (8N. Sánchez-Maroño, A. Alonso-Betanzos and M. Tombilla-Sanromán, Filter methods for feature selection – a comparative study, In: P. E. and Data and in Science Springer Berlin, Scholar, A of feature selection and feature methods applied on Scopus Google Scholar). A method proteins according to univariate association with the phenotype of and the proteins. Due to the functional and proteins are leading to performance of this The and algorithm this by selecting features that have the with the phenotype and are also they are to each as as and feature selection from of the Scholar). methods the of feature based on the performance of a algorithm Chen J. A wrapper method for feature selection and its Scopus Google Scholar, Feature selection wrapper based on and Scopus Google Scholar, R. A wrapper method for feature selection using Scopus Google Scholar). The feature with the performance is as the selected Due to the number of methods are used in biomarker selection where the number of original features typically are in the methods an to select features during the model F. A on feature selection 2014; Scopus Google Scholar, F. D. feature selection with PubMed Scopus Google Scholar). One commonly used is through the of the and selection method of a linear model by some to the selection of features with R. and selection the R. Stat. Scholar). feature selection algorithms are to to small at protein markers selected based on a discovery cohort may not be generalizable to new test This a computational challenge in the protein biomarker development In addition, because platforms are used in the discovery and validation a good protein marker identified in the discovery platform may be difficult to implement in the validation In although MS-based targeted proteomics been increasingly used in biomarker verification and validation of mass Oncol. PubMed Scopus Google the most used in the proteins not have in clinical assays. In this paper, we present a new protein marker selection algorithm to to algorithm also to identify a number of features with the with the phenotype and is also However, algorithm is based on the that there are typically a number of informative features for a that features are often with a number of biological underlying the and that coexpressed proteins to biological J. Carr S.A. P. Chan D.W. Chen M. et for based Proteomics. PubMed Scopus Google Scholar). algorithm all informative features through univariate association into clusters based on and most representative feature from each cluster to a set of markers. that protein markers on biological defined by coexpressed proteins of the markers. Moreover, there is a selected protein marker on the verification and validation algorithm provides by the selected marker with coexpressed protein in the In addition, the methods algorithm is to the multiomics a unique to multiomics data to protein marker selection. algorithm and its multiomics as and algorithms using proteomics and multiomics data from two and cancer studies J. J. P. et characterization of human and 2014; PubMed Scopus Google Scholar, Petyuk V.A. Gritsenko M.A. Clauss T.R. Moore R.J. et analysis of human cancer new therapeutic PubMed Scopus Google and two studies A. L. L. M. Chen N. et new therapeutic of PubMed Scopus Google Scholar, L. Chen R. J. L. J. et characterization of PubMed Scopus Google Scholar). we select protein markers to the which both and therapeutic The of in colorectal cancer for 2018; PubMed Scopus Google Scholar, F. D. R. et response of to PubMed Scopus Google Scholar). we select protein markers to in proteomic data and data from a study J. J. P. et characterization of human and 2014; PubMed Scopus Google for protein marker selection and model proteomic data from study Petyuk V.A. Gritsenko M.A. Clauss T.R. Moore R.J. et analysis of human cancer new therapeutic PubMed Scopus Google used for independent Protein quantification for both based on Petyuk V.A. Gritsenko M.A. Clauss T.R. Moore R.J. et analysis of human cancer new therapeutic PubMed Scopus Google Scholar). quantification based on of both proteomic and with Data to by each with or as and with as The data and samples, the test data and in mass proteomic and data and data from a study L. Chen R. J. L. J. et characterization of PubMed Scopus Google for marker selection and model and data using the method that are at and more than data applied to the data. quantification using by by and proteomic data from study A. L. L. M. Chen N. et new therapeutic of PubMed Scopus Google used for independent Protein quantification using the quantification method implemented in J. M. enables identification mass and protein Biotechnol. PubMed Scopus Google Scholar). The quantification data from A. J. M. D.J. A. J. M. E. J. J. T. et and and in for quantification PubMed Scopus Google with number The and with and all features as a classification into two according to with than as good and than as The data and samples, the test data and Data of size is to the protein where to and to proteins to identify protein markers that can be used to the phenotype The algorithm ProMS as a to features, ProMS each feature to the of the the feature and the phenotype A the characteristic is defined to evaluate such The is that or than typically some predictive of the feature although in the the feature to the of the This not be a problem in a can that and the of that feature Since the feature is at a we can use the data and phenotype to the the of is the as that of the original ProMS the features with the is a that to be with of the the data is to of size where the the features, ProMS features into clusters with k-medoids clustering A and algorithm for Scopus Google in sample we the data by in an sample The goal is to a and that the is In the of the of the data to to be k-medoids clustering is to clustering for a into The is that the are selected from the data (i.e., features in this to the most data in the cluster because the the and all of the cluster is provide a to the biomarker selection a unique that not with where the of a cluster is the the data in the cluster of an data The two data and is where the and ProMS a weighted of k-medoids algorithm where each feature is a in the the new is This is based on the that features with univariate predictive be of being selected as a The are selected as the markers for a have data of measurements that the set of is a of size where to and to features in data the of we use to the proteomic data from which we to select a set of informative markers that can be used to the hypothesis is that by from a set of more informative protein markers can be test this we ProMS to ProMS_mo as to ProMS, the of ProMS_mo out features from each data source we use as the ProMS_mo the univariate to data source and the features with the the features as data ProMS_mo features with than data are into a new of size and is the number of features in the data source weighted k-medoids clustering is to the features into clusters in sample that protein markers are selected as ProMS_mo the to protein markers. the steps of a can be by protein marker such the This can be as a constrained the are selected as the protein markers for a compared ProMS with feature selection or dimension reduction methods, a method and feature selection from of the a model based method and a supervised PCA method (7Bair E. Hastie T. Paul D. Tibshirani R. Prediction by supervised principal components.J. Am. Stat. Assoc. 2006; 101: 119-137Crossref Scopus (401) Google Scholar). is implemented as a of an J. K. F. Tang J. Feature A data 2018; Scopus Google Scholar). in the original we used as the to the of and is implemented in F. A. M. P. R. J. A. D. M. M. et in and we that there be features with in the The amount of to the model to this The implementation with a as in ProMS. by the PCA method on the data selected the components as the new This is not a feature selection method because each PC is a linear combination of original ProMS and ProMS_mo with the and datasets. cancer a set and an independent test set to select a few of markers with each we the is into two for feature selection and for validation feature selection methods ProMS, and using the selected and A number of using with the set T. from univariate and to the individual The with the validation set for each of the as well as the independent test set for all we used all data in to the feature selection and and a model to be with the independent test set analysis using J. set analysis with and PubMed Scopus Google Scholar, J. system for in biological PubMed Scopus Google through used as the used for the ProMS and ProMS_mo implemented in the The source and are available at https://github.com/bzhanglab/proms. and and using of features selected by ProMS, and PCs to predictive models where each PC is a linear combination of all original it cannot be used for marker selection. as a for on both data from the cohort and test data from an independent to the small of selected features, or performance compared with more algorithms in all we present from in the on both validation data and independent test data for all feature This because the PC features capture more than the small number of protein features selected by methods. performance in all Although both ProMS and are methods, ProMS and or performance compared with the method A and performance with feature but the three methods performance all that there a small set of protein markers that can performance on the and independent test PCA the performance on the test the performance on the test data compared with methods, a of for than that for ProMS, but both performance the validation and test data. in the in compared with in and However, most of the in the analysis in the One new is that ProMS not but also in this more clinical performance with feature and it also performance in the independent test data. the performance ProMS and more in the test data and a number of features The performance of ProMS or that of the in all although the combinations of features in the original Moreover, both ProMS and performance the validation and test no during performance of the ProMS algorithm in selecting markers from proteomics data. A of ProMS_mo is the to multiomics data but select protein Feature selection with and cannot be to multiomics data to facilitate protein marker selection. we on the performance of models using features selected by ProMS_mo with using features selected by ProMS. as a for the from models on in and in feature In we not performance models features selected by ProMS_mo and ProMS. However, the models on the independent test ProMS_mo ProMS in all with for in the performance of ProMS_mo or performance compared with that using from multiomics data may the of protein marker leading to improved performance on independent test data. performance of the ProMS_mo using we applied the to the data to identify the marker panels for model in all models with protein marker from to 20 excellent performance on the independent test data In the model with protein markers an of The markers selected by ProMS_mo two proteins with in and and three proteins with and proteins and in the data The but in the test data in more The model with protein markers an of and the number of protein markers not to performance The markers selected by ProMS_mo two proteins with in with and and three proteins with and proteins with and good in the data Although the of individual markers in the test data the good performance of the selected markers have in the phenotype of and of is a response and are with Petyuk V.A. Gritsenko M.A. Clauss T.R. Moore R.J. et analysis of human cancer new therapeutic PubMed Scopus Google Scholar). of by is to and it been as a to T. K. N. T. T. A. K. K. M. M. and 2013; PubMed Scopus Google Scholar). In addition, in with J. T. and in PubMed Scopus Google and F. Chen L. R. J. of by PubMed Scopus Google is with clinical of is a of and is to to and studies show that can tumor in Chen D. J. Chen P. A in human 2014; PubMed Scopus Google and it is for the detection and of T. M. J. Chen P. of the of protein is with tumor and of PubMed Scopus Google Scholar). The identification of markers the of biomarker selection A unique of framework for protein biomarker selection is that each selected marker is with a cluster of features, biological for functional interpretation of the selected markers. in as an the clusters and features, and features in clusters protein a with all protein features in the clusters from the D. A. K. D. J. M. A. A. M. P. the of PubMed Google Scholar). The of the with to the cluster than the from with clusters of which a of and of This that proteins in the clusters are more to be to proteins in the cluster as compared with proteins in the functional of the A of for features in each cluster and selected protein markers, and cluster a of features from all three platforms. Although a small of by all three most by platform features from all platforms the underlying biological of the which be through each the terms in biological component and are in such as have biological in analysis not existing but also for using feature clusters to enable functional interpretation of selected biomarkers with limited functional of framework is that it provides for markers in the that there are in the selected proteins in a clinical to for the development of assays. cluster analysis to proteins with and biological in the we that the predictive the selected proteins are with proteins in the cluster test this we used the data to models where of the markers by the protein in its with the performance of models on the independent test data to the models with the original set of markers framework provides a robust to facilitate clinical Untargeted MS proteomics provides a powerful platform for protein biomarker discovery, but an effective transition from discovery to verification and validation on the selection of a small of protein biomarkers from discovery proteomics data. ProMS, a new computational algorithm to facilitate protein biomarker selection. ProMS superior performance and for feature selection in the of in and performance more with the more in In addition to good performance, ProMS also a few unique that are in feature selection First, the feature clusters enable functional interpretation of the selected protein markers. Second, the feature clusters provide an opportunity to select replacement protein markers, facilitating a transition to the verification and validation platforms. the algorithm is to the multiomics and ProMS_mo multiomics data to protein biomarker selection. ProMS is to the used algorithms to select features that have association with the but association with each this by an that the and the by the the algorithm features at a in an In to the by ProMS a which is based on the hypothesis that a phenotype is characterized by a few underlying biological functions, each manifested by a group of and coexpressed proteins. A weighted k-medoids clustering algorithm is used to identify both protein and a representative protein for each group as markers. performance in to its to the ProMS is by biological rather than a it not but also performance than the feature selection method which typically shows performance than methods. in ProMS the performance of the which from a more proteins. of the number of may be by clinical In addition, in both the ProMS and ProMS_mo can provide as and the with the performance in be selected for the of the In to the selected is to the software also includes as an the which is not an feature selection method but can as a for performance as in have that multiomics characterization enables a more of biological M. A. to PubMed Scopus Google Scholar, R. S.A. D. of data to PubMed Scopus Google Scholar, is not an is the to biological Proteomics. PubMed Scopus Google Scholar). However, feature selection from multiomics to data and data important of new and to in data in to models with predictive L. and L. Springer Scholar). been to the where the importance of each is into to feature selection M. D. and feature selection for cancer using 2018; PubMed Scopus Google Scholar). A method also been to feature selection with and in the of L. A feature selection method PubMed Scopus Google Scholar). However, methods select features and features from can be in the feature clinical translation of markers is to the of platforms. ProMS_mo is from methods because it proteins although all multiomics data are used to facilitate protein marker selection. features with (i.e., and proteins from the or by a or by a are more to have functional to protein biomarker selection. ProMS performance than the feature selection methods in both the and the studies, the in the This that is more difficult than Although the phenotype is by a more the phenotype may be by more a sample size is the phenotype of is to be with Moreover, for in the study as a phenotype in One development is to enable the analysis of phenotype such as or in ProMS. ProMS_mo ProMS on the independent the are in Moreover, the two algorithms in One is that the data used in this study, and are to proteomics which may the amount of ProMS_mo is not limited by and of such as data or can also be in the a is to test ProMS_mo with of which may proteomics data than and The source and of the ProMS are available at https://github.com/bzhanglab/proms. This data. The no This by from the by from the and of and by from the at The and is a in and a

Topics & Concepts