Comparison of Discrimination Methods for the Classification of Tumors Using Gene Expression Data

Characterization of the peripheral blood transcriptome in a repeated measures design using a panel of healthy individuals

A repeated measures microarray design with 22 healthy, non-smoking volunteers (aging 32±5years) was set up to study transcriptome profiles in whole blood samples. The results indicate that repeatable data can be obtained with high within-subject correlation. Probes that could discriminate between individuals are associated with immune and inflammatory functions. When investigating possible time trends in the microarray data, we have found no differential expression within a sampling period (within-season effect). Differential expression was observed between sampling seasons and the data suggest a weak response of genes related to immune system functioning. Finally, a high number of probes showed significant season-specific expression variability within subjects. Expression variability increased in springtime and there was an association of the probe list with immune system functioning. Our study suggests that the blood transcriptome of healthy individuals is reproducible over a time period of several months.

A new statistic for identifying batch effects in high-throughput genomic data that uses guided principal component analysis

MOTIVATION

Batch effects are due to probe-specific systematic variation between groups of samples (batches) resulting from experimental features that are not of biological interest. Principal component analysis (PCA) is commonly used as a visual tool to determine whether batch effects exist after applying a global normalization method. However, PCA yields linear combinations of the variables that contribute maximum variance and thus will not necessarily detect batch effects if they are not the largest source of variability in the data.

RESULTS

We present an extension of PCA to quantify the existence of batch effects, called guided PCA (gPCA). We describe a test statistic that uses gPCA to test whether a batch effect exists. We apply our proposed test statistic derived using gPCA to simulated data and to two copy number variation case studies: the first study consisted of 614 samples from a breast cancer family study using Illumina Human 660 bead-chip arrays, whereas the second case study consisted of 703 samples from a family blood pressure study that used Affymetrix SNP Array 6.0. We demonstrate that our statistic has good statistical properties and is able to identify significant batch effects in two copy number variation case studies.

CONCLUSION

We developed a new statistic that uses gPCA to identify whether batch effects exist in high-throughput genomic data. Although our examples pertain to copy number data, gPCA is general and can be used on other data types as well.

AVAILABILITY AND IMPLEMENTATION

The gPCA R package (Available via CRAN) provides functionality and data to perform the methods in this article.

CONTACT

reesese@vcu.edu

A computational study identifies HIV progression-related genes using mRMR and shortest path tracing

Since statistical relationships between HIV load and CD4+ T cell loss have been demonstrated to be weak, searching for host factors contributing to the pathogenesis of HIV infection becomes a key point for both understanding the disease pathology and developing treatments. We applied Maximum Relevance Minimum Redundancy (mRMR) algorithm to a set of microarray data generated from the CD4+ T cells of viremic non-progressors (VNPs) and rapid progressors (RPs) to identify host factors associated with the different responses to HIV infection. Using mRMR algorithm, 147 gene had been identified. Furthermore, we constructed a weighted molecular interaction network with the existing protein-protein interaction data from STRING database and identified 1331 genes on the shortest-paths among the genes identified with mRMR. Functional analysis shows that the functions relating to apoptosis play important roles during the pathogenesis of HIV infection. These results bring new insights of understanding HIV progression.

Sex-specific genomic biomarkers for individualized treatment of life-threatening diseases

Numerous studies have demonstrated sex differences in drug reactions to the same drug treatment, steering away from the traditional view of one-size-fits-all medicine. A premise of this study is that the sex of a patient influences difference in disease characteristics and risk factors. In this study, we intend to exploit and to obtain better sex-specific biomarkers from gene-expression data. We propose a procedure to isolate a set of important genes as sex-specific genomic biomarkers, which may enable more effective patient treatment. A set of sex-specific genes is obtained by a variable importance ranking using a combination of cross-validation methods. The proposed procedure is applied to three gene-expression datasets.

Criteria for the use of omics-based predictors in clinical trials: explanation and elaboration

High-throughput 'omics' technologies that generate molecular profiles for biospecimens have been extensively used in preclinical studies to reveal molecular subtypes and elucidate the biological mechanisms of disease, and in retrospective studies on clinical specimens to develop mathematical models to predict clinical endpoints. Nevertheless, the translation of these technologies into clinical tests that are useful for guiding management decisions for patients has been relatively slow. It can be difficult to determine when the body of evidence for an omics-based test is sufficiently comprehensive and reliable to support claims that it is ready for clinical use, or even that it is ready for definitive evaluation in a clinical trial in which it may be used to direct patient therapy. Reasons for this difficulty include the exploratory and retrospective nature of many of these studies, the complexity of these assays and their application to clinical specimens, and the many potential pitfalls inherent in the development of mathematical predictor models from the very high-dimensional data generated by these omics technologies. Here we present a checklist of criteria to consider when evaluating the body of evidence supporting the clinical use of a predictor to guide patient therapy. Included are issues pertaining to specimen and assay requirements, the soundness of the process for developing predictor models, expectations regarding clinical study design and conduct, and attention to regulatory, ethical, and legal issues. The proposed checklist should serve as a useful guide to investigators preparing proposals for studies involving the use of omics-based tests. The US National Cancer Institute plans to refer to these guidelines for review of proposals for studies involving omics tests, and it is hoped that other sponsors will adopt the checklist as well.

Performance and estimation of the true error rate of classification rules built with additional information. An application to a cancer trial

Classification rules that incorporate additional information usually present in discrimination problems are receiving certain attention during the last years as they perform better than the usual rules. Fernández, M. A., C. Rueda and B. Salvador (2006): "Incorporating additional information to normal linear discriminant rules," J. Am. Stat. Assoc., 101, 569-577, proved that these rules have lower total misclassification probability than the usual Fisher's rule. In this paper we consider two issues; on the one hand, we compare these rules with those based on shrinkage estimators of the mean proposed by Tong, T., L. Chen and H. Zhao (2012): "Improved mean estimation and its application to diagonal discriminant analysis," Bioinformatics, 28(4): 531-537. with regard to four criteria: total misclassification probability, area under ROC curve, well-calibratedness and refinement; on the other hand, we consider the estimation of the true error rate, which is a very interesting parameter in applications. We prove results on the apparent error rate of the rules that expose the need of new estimators of their true error rate. We propose four such new estimators. Two of them are defined incorporating the additional information into the leave-one-out-bootstrap. The other two are the corresponding cross-validation after bootstrap versions. We compare these estimators with the usual ones in a simulation study and in a cancer trial application, showing the good behavior of the rules that incorporate additional information and of the new leave-one-out bootstrap estimators of their true error rate.

Fuzzy support vector machine: an efficient rule-based classification technique for microarrays

BACKGROUND

The abundance of gene expression microarray data has led to the development of machine learning algorithms applicable for tackling disease diagnosis, disease prognosis, and treatment selection problems. However, these algorithms often produce classifiers with weaknesses in terms of accuracy, robustness, and interpretability. This paper introduces fuzzy support vector machine which is a learning algorithm based on combination of fuzzy classifiers and kernel machines for microarray classification.

RESULTS

Experimental results on public leukemia, prostate, and colon cancer datasets show that fuzzy support vector machine applied in combination with filter or wrapper feature selection methods develops a robust model with higher accuracy than the conventional microarray classification models such as support vector machine, artificial neural network, decision trees, k nearest neighbors, and diagonal linear discriminant analysis. Furthermore, the interpretable rule-base inferred from fuzzy support vector machine helps extracting biological knowledge from microarray data.

CONCLUSIONS

Fuzzy support vector machine as a new classification model with high generalization power, robustness, and good interpretability seems to be a promising tool for gene expression microarray classification.

Discovering combinatorial interactions in survival data

Motivation: Although several methods exist to relate high-dimensional gene expression data to various clinical phenotypes, finding combinations of features in such input remains a challenge, particularly when fitting complex statistical models such as those used for survival studies.

Results: Our proposed method builds on existing ‘regularization path-following’ techniques to produce regression models that can extract arbitrarily complex patterns of input features (such as gene combinations) from large-scale data that relate to a known clinical outcome. Through the use of the data’s structure and itemset mining techniques, we are able to avoid combinatorial complexity issues typically encountered with such methods, and our algorithm performs in similar orders of duration as single-variable versions. Applied to data from various clinical studies of cancer patient survival time, our method was able to produce a number of promising gene-interaction candidates whose tumour-related roles appear confirmed by literature.

Availability: An R implementation of the algorithm described in this article can be found at https://github.com/david-duverle/regularisation-path-following

Contact:dave.duverle@aist.go.jp

Supplementary information:Supplementary data are available at Bioinformatics online.

A classification framework applied to cancer gene expression profiles

Classification of cancer based on gene expression has provided insight into possible treatment strategies. Thus, developing machine learning methods that can successfully distinguish among cancer subtypes or normal versus cancer samples is important. This work discusses supervised learning techniques that have been employed to classify cancers. Furthermore, a two-step feature selection method based on an attribute estimation method (e.g., ReliefF) and a genetic algorithm was employed to find a set of genes that can best differentiate between cancer subtypes or normal versus cancer samples. The application of different classification methods (e.g., decision tree, k-nearest neighbor, support vector machine (SVM), bagging, and random forest) on 5 cancer datasets shows that no classification method universally outperforms all the others. However, k-nearest neighbor and linear SVM generally improve the classification performance over other classifiers. Finally, incorporating diverse types of genomic data (e.g., protein-protein interaction data and gene expression) increase the prediction accuracy as compared to using gene expression alone.

Improved Sparse Multi-Class SVM and Its Application for Gene Selection in Cancer Classification

BACKGROUND

Microarray techniques provide promising tools for cancer diagnosis using gene expression profiles. However, molecular diagnosis based on high-throughput platforms presents great challenges due to the overwhelming number of variables versus the small sample size and the complex nature of multi-type tumors. Support vector machines (SVMs) have shown superior performance in cancer classification due to their ability to handle high dimensional low sample size data. The multi-class SVM algorithm of Crammer and Singer provides a natural framework for multi-class learning. Despite its effective performance, the procedure utilizes all variables without selection. In this paper, we propose to improve the procedure by imposing shrinkage penalties in learning to enforce solution sparsity.

RESULTS

The original multi-class SVM of Crammer and Singer is effective for multi-class classification but does not conduct variable selection. We improved the method by introducing soft-thresholding type penalties to incorporate variable selection into multi-class classification for high dimensional data. The new methods were applied to simulated data and two cancer gene expression data sets. The results demonstrate that the new methods can select a small number of genes for building accurate multi-class classification rules. Furthermore, the important genes selected by the methods overlap significantly, suggesting general agreement among different variable selection schemes.

CONCLUSIONS

High accuracy and sparsity make the new methods attractive for cancer diagnostics with gene expression data and defining targets of therapeutic intervention.

AVAILABILITY

The source MATLAB code are available from http://math.arizona.edu/~hzhang/software.html.

A simulation to analyze feature selection methods utilizing gene ontology for gene expression classification

Gene expression profile classification is a pivotal research domain assisting in the transformation from traditional to personalized medicine. A major challenge associated with gene expression data classification is the small number of samples relative to the large number of genes. To address this problem, researchers have devised various feature selection algorithms to reduce the number of genes. Recent studies have been experimenting with the use of semantic similarity between genes in Gene Ontology (GO) as a method to improve feature selection. While there are few studies that discuss how to use GO for feature selection, there is no simulation study that addresses when to use GO-based feature selection. To investigate this, we developed a novel simulation, which generates binary class datasets, where the differentially expressed genes between two classes have some underlying relationship in GO. This allows us to investigate the effects of various factors such as the relative connectedness of the underlying genes in GO, the mean magnitude of separation between differentially expressed genes denoted by δ, and the number of training samples. Our simulation results suggest that the connectedness in GO of the differentially expressed genes for a biological condition is the primary factor for determining the efficacy of GO-based feature selection. In particular, as the connectedness of differentially expressed genes increases, the classification accuracy improvement increases. To quantify this notion of connectedness, we defined a measure called Biological Condition Annotation Level BCAL(G), where G is a graph of differentially expressed genes. Our main conclusions with respect to GO-based feature selection are the following: (1) it increases classification accuracy when BCAL(G) ≥ 0.696; (2) it decreases classification accuracy when BCAL(G) ≤ 0.389; (3) it provides marginal accuracy improvement when 0.389<BCAL(G)<0.696 and δ<1; (4) as the number of genes in a biological condition increases beyond 50 and δ ≥ 0.7, the improvement from GO-based feature selection decreases; and (5) we recommend not using GO-based feature selection when a biological condition has less than ten genes. Our results are derived from datasets preprocessed using RMA (Robust Multi-array Average), cases where δ is between 0.3 and 2.5, and training sample sizes between 20 and 200, therefore our conclusions are limited to these specifications. Overall, this simulation is innovative and addresses the question of when SoFoCles-style feature selection should be used for classification instead of statistical-based ranking measures.

DepthTools: an R package for a robust analysis of gene expression data

Background

The use of DNA microarrays and oligonucleotide chips of high density in modern biomedical research provides complex, high dimensional data which have been proven to convey crucial information about gene expression levels and to play an important role in disease diagnosis. Therefore, there is a need for developing new, robust statistical techniques to analyze these data.

Results

depthTools is an R package for a robust statistical analysis of gene expression data, based on an efficient implementation of a feasible notion of depth, the Modified Band Depth. This software includes several visualization and inference tools successfully applied to high dimensional gene expression data. A user-friendly interface is also provided via an R-commander plugin.

Conclusion

We illustrate the utility of the depthTools package, that could be used, for instance, to achieve a better understanding of genome-level variation between tumors and to facilitate the development of personalized treatments.

svapls: an R package to correct for hidden factors of variability in gene expression studies

Background

Hidden variability is a fundamentally important issue in the context of gene expression studies. Collected tissue samples may have a wide variety of hidden effects that may alter their transcriptional landscape significantly. As a result their actual differential expression pattern can be potentially distorted, leading to inaccurate results from a genome-wide testing for the important transcripts.

Results

We present an R package svapls that can be used to identify several types of unknown sample-specific sources of heterogeneity in a gene expression study and adjust for them in order to provide a more accurate inference on the original expression pattern of the genes over different varieties of samples. The proposed method implements Partial Least Squares regression to extract the hidden signals of sample-specific heterogeneity in the data and uses them to find the genes that are actually correlated with the phenotype of interest. We also compare our package with three other popular softwares for testing differential gene expression along with a detailed illustration on the widely popular Golub dataset. Results from the sensitivity analyes on simulated data with widely different hidden variation patterns reveal the improved detection power of our R package compared to the other softwares along with reasonably smaller error rates. Application on the real-life dataset exhibits the efficacy of the R package in detecting potential batch effects from the dataset.

Conclusions

Overall, Our R package provides the user with a simplified framework for analyzing gene expression data with a wide range of hidden variation patterns and delivering a differential gene expression analysis with substantially improved power and accuracy.

The R package svapls is freely available at http://cran.r-project.org/web/packages/svapls/index.html.

Sample size requirements for training high-dimensional risk predictors

A common objective of biomarker studies is to develop a predictor of patient survival outcome. Determining the number of samples required to train a predictor from survival data is important for designing such studies. Existing sample size methods for training studies use parametric models for the high-dimensional data and cannot handle a right-censored dependent variable. We present a new training sample size method that is non-parametric with respect to the high-dimensional vectors, and is developed for a right-censored response. The method can be applied to any prediction algorithm that satisfies a set of conditions. The sample size is chosen so that the expected performance of the predictor is within a user-defined tolerance of optimal. The central method is based on a pilot dataset. To quantify uncertainty, a method to construct a confidence interval for the tolerance is developed. Adequacy of the size of the pilot dataset is discussed. An alternative model-based version of our method for estimating the tolerance when no adequate pilot dataset is available is presented. The model-based method requires a covariance matrix be specified, but we show that the identity covariance matrix provides adequate sample size when the user specifies three key quantities. Application of the sample size method to two microarray datasets is discussed.

Exploring correlations in gene expression microarray data for maximum predictive-minimum redundancy biomarker selection and classification

An important issue in the analysis of gene expression microarray data is concerned with the extraction of valuable genetic interactions from high dimensional data sets containing gene expression levels collected for a small sample of assays. Past and ongoing research efforts have been focused on biomarker selection for phenotype classification. Usually, many genes convey useless information for classifying the outcome and should be removed from the analysis; on the other hand, some of them may be highly correlated, which reveals the presence of redundant expressed information. In this paper we propose a method for the selection of highly predictive genes having a low redundancy in their expression levels. The predictive accuracy of the selection is assessed by means of Classification and Regression Trees (CART) models which enable assessment of the performance of the selected genes for classifying the outcome variable and will also uncover complex genetic interactions. The method is illustrated throughout the paper using a public domain colon cancer gene expression data set.

Determination of minimum training sample size for microarray-based cancer outcome prediction-an empirical assessment

The promise of microarray technology in providing prediction classifiers for cancer outcome estimation has been confirmed by a number of demonstrable successes. However, the reliability of prediction results relies heavily on the accuracy of statistical parameters involved in classifiers. It cannot be reliably estimated with only a small number of training samples. Therefore, it is of vital importance to determine the minimum number of training samples and to ensure the clinical value of microarrays in cancer outcome prediction. We evaluated the impact of training sample size on model performance extensively based on 3 large-scale cancer microarray datasets provided by the second phase of MicroArray Quality Control project (MAQC-II). An SSNR-based (scale of signal-to-noise ratio) protocol was proposed in this study for minimum training sample size determination. External validation results based on another 3 cancer datasets confirmed that the SSNR-based approach could not only determine the minimum number of training samples efficiently, but also provide a valuable strategy for estimating the underlying performance of classifiers in advance. Once translated into clinical routine applications, the SSNR-based protocol would provide great convenience in microarray-based cancer outcome prediction in improving classifier reliability.

Mathematical modeling in cancer drug discovery

Mathematical models have the potential to help discover new therapeutic targets and treatment strategies. In this review, we discuss how the latest developments in mathematical modeling can provide useful context for the rational design, validation and prioritization of novel cancer drug targets and their combinations. We give special attention to two modeling approaches: network-based modeling and multiscale modeling, because they have begun to show promise in facilitating the process of effective cancer drug discovery. Both modeling approaches are integrated with a variety of experimental methods to ensure proper parameterization and to maximize their predictive value. We also discuss several challenges faced in modeling-based drug discovery.

Broad spectrum detection and "barcoding" of water pollutants by a genome-wide bacterial sensor array

An approach for the rapid detection and classification of a broad spectrum of water pollutants, based on a genome-wide reporter bacterial live cell array, is proposed and demonstrated. An array of ca. 2000 Escherichia coli fluorescent transcriptional reporters was exposed to 25 toxic compounds as well as to unpolluted water, and its responses were recorded after 3 h. The 25 toxic compounds represented 5 pollutant classes: genotoxicants, metals, detergents, alcohols, and monoaromatic hydrocarbons. Identifying unique gene expression patterns, a nearest neighbour-based model detected pollutant presence and predicted class attribution with an estimated accuracy of 87%. Sensitivity and positive predictive values varied among classes, being higher for pollutant classes that were defined by mode of action than for those defined by structure only. Sensitivity for unpolluted water was 0.90 and the positive predictive value was 0.79. All pollutant classes induced the transcription of a statistically significant proportion of membrane associated genes; in addition, the sets of genes responsive to genotoxicants, detergents and alcohols were enriched with genes involved in DNA repair, iron utilization and the translation machinery, respectively. Following further development, a methodology of the type described herein may be suitable for integration in water monitoring schemes in conjunction with existing analytical and biological detection techniques.

Statistical analysis of big data on pharmacogenomics

This paper discusses statistical methods for estimating complex correlation structure from large pharmacogenomic datasets. We selectively review several prominent statistical methods for estimating large covariance matrix for understanding correlation structure, inverse covariance matrix for network modeling, large-scale simultaneous tests for selecting significantly differently expressed genes and proteins and genetic markers for complex diseases, and high dimensional variable selection for identifying important molecules for understanding molecule mechanisms in pharmacogenomics. Their applications to gene network estimation and biomarker selection are used to illustrate the methodological power. Several new challenges of Big data analysis, including complex data distribution, missing data, measurement error, spurious correlation, endogeneity, and the need for robust statistical methods, are also discussed.

Overfitting in prediction models - is it a problem only in high dimensions?

The growing recognition that human diseases are molecularly heterogeneous has stimulated interest in the development of prognostic and predictive classifiers for patient selection and stratification. In the process of classifier development, it has been repeatedly emphasized that in situations where the number of candidate predictor variables is much larger than the number of observations, the apparent (training set, resubstitution) accuracy of the classifiers can be highly optimistically biased and hence, classification accuracy should be reported based on evaluation of the classifier on a separate test set or using complete cross-validation. Such evaluation methods have however not been the norm in the case of low-dimensional, p<n data that arise, for example, in clinical trials when a classifier is developed on a combination of clinico-pathological variables and a small number of genetic biomarkers selected from an understanding of the biology of the disease. We undertook simulation studies to investigate the existence and extent of the problem of overfitting with low-dimensional data. The results indicate that overfitting can be a serious problem even for low-dimensional data, especially if the relationship of outcome to the set of predictor variables is not strong. We hence encourage the adoption of either a separate test set or complete cross-validation to evaluate classifier accuracy, even when the number of candidate predictor variables is substantially smaller than the number of cases.

Utilization of never-medicated bipolar disorder patients towards development and validation of a peripheral biomarker profile

There are currently no biological tests that differentiate patients with bipolar disorder (BPD) from healthy controls. While there is evidence that peripheral gene expression differences between patients and controls can be utilized as biomarkers for psychiatric illness, it is unclear whether current use or residual effects of antipsychotic and mood stabilizer medication drives much of the differential transcription. We therefore tested whether expression changes in first-episode, never-medicated BPD patients, can contribute to a biological classifier that is less influenced by medication and could potentially form a practicable biomarker assay for BPD. We employed microarray technology to measure global leukocyte gene expression in first-episode (n=3) and currently medicated BPD patients (n=26), and matched healthy controls (n=25). Following an initial feature selection of the microarray data, we developed a cross-validated 10-gene model that was able to correctly predict the diagnostic group of the training sample (26 medicated patients and 12 controls), with 89% sensitivity and 75% specificity (p<0.001). The 10-gene predictor was further explored via testing on an independent cohort consisting of three pairs of monozygotic twins discordant for BPD, plus the original enrichment sample cohort (the three never-medicated BPD patients and 13 matched control subjects), and a sample of experimental replicates (n=34). 83% of the independent test sample was correctly predicted, with a sensitivity of 67% and specificity of 100% (although this result did not reach statistical significance). Additionally, 88% of sample diagnostic classes were classified correctly for both the enrichment (p=0.015) and the replicate samples (p<0.001). We have developed a peripheral gene expression biomarker profile, that can classify healthy controls from patients with BPD receiving antipsychotic or mood stabilizing medication, which has both high sensitivity and specificity. Moreover, assay of three first-episode patients who had never received such medications, to first enrich the expression dataset for disease-related genes independent of medication effects, and then to test the 10-gene predictor, validates the peripheral biomarker approach for BPD.

Functional genomics identifies type I interferon pathway as central for host defense against Candida albicans

Candida albicans is the most common human fungal pathogen causing mucosal and systemic infections. However, human antifungal immunity remains poorly defined. Here, by integrating transcriptional analysis and functional genomics, we identified Candida-specific host defense mechanisms in humans. Candida induced significant expression of genes from the type I interferon (IFN) pathway in human peripheral blood mononuclear cells. This unexpectedly prominent role of type I IFN pathway in anti-Candida host defense was supported by additional evidence. Polymorphisms in type I IFN genes modulated Candida-induced cytokine production and were correlated with susceptibility to systemic candidiasis. In in-vitro experiments, type I IFNs skewed Candida-induced inflammation from a Th17-response toward a Th1-response. Patients with chronic mucocutaneaous candidiasis displayed defective expression of genes in the type I IFN pathway. These findings indicate that the type I IFN pathway is a main signature of Candida-induced inflammation and plays a crucial role in anti-Candida host defense in humans.

Sparse logistic regression with a L1/2 penalty for gene selection in cancer classification

Background

Microarray technology is widely used in cancer diagnosis. Successfully identifying gene biomarkers will significantly help to classify different cancer types and improve the prediction accuracy. The regularization approach is one of the effective methods for gene selection in microarray data, which generally contain a large number of genes and have a small number of samples. In recent years, various approaches have been developed for gene selection of microarray data. Generally, they are divided into three categories: filter, wrapper and embedded methods. Regularization methods are an important embedded technique and perform both continuous shrinkage and automatic gene selection simultaneously. Recently, there is growing interest in applying the regularization techniques in gene selection. The popular regularization technique is Lasso (L₁), and many L₁ type regularization terms have been proposed in the recent years. Theoretically, the Lq type regularization with the lower value of q would lead to better solutions with more sparsity. Moreover, the L_1/2 regularization can be taken as a representative of Lq (0 <q < 1) regularizations and has been demonstrated many attractive properties.

Results

In this work, we investigate a sparse logistic regression with the L_1/2 penalty for gene selection in cancer classification problems, and propose a coordinate descent algorithm with a new univariate half thresholding operator to solve the L_1/2 penalized logistic regression. Experimental results on artificial and microarray data demonstrate the effectiveness of our proposed approach compared with other regularization methods. Especially, for 4 publicly available gene expression datasets, the L_1/2 regularization method achieved its success using only about 2 to 14 predictors (genes), compared to about 6 to 38 genes for ordinary L₁ and elastic net regularization approaches.

Conclusions

From our evaluations, it is clear that the sparse logistic regression with the L_1/2 penalty achieves higher classification accuracy than those of ordinary L₁ and elastic net regularization approaches, while fewer but informative genes are selected. This is an important consideration for screening and diagnostic applications, where the goal is often to develop an accurate test using as few features as possible in order to control cost. Therefore, the sparse logistic regression with the L_1/2 penalty is effective technique for gene selection in real classification problems.

BreastPRS is a gene expression assay that stratifies intermediate-risk Oncotype DX patients into high- or low-risk for disease recurrence

Molecular prognostic assays, such as Oncotype DX, are increasingly incorporated into the management of patients with invasive breast carcinoma. BreastPRS is a new molecular assay developed and validated from a meta-analysis of publically available genomic datasets. We applied the assay to matched fresh-frozen (FF) and formalin-fixed paraffin-embedded (FFPE) tumor samples to translate the assay to FFPE. A linear relationship of the BreastPRS prognostic score was observed between tissue preservation formats. BreastPRS recurrence scores were compared with Oncotype DX recurrence scores from 246 patients with invasive breast carcinoma and known Oncotype DX results. Using this series, a 120-gene Oncotype DX approximation algorithm was trained to predict Oncotype DX risk groups and then applied to series of untreated, node-negative, estrogen receptor (ER)-positive patients from previously published studies with known clinical outcomes. Correlation of recurrence score and risk group between Oncotype DX and BreastPRS was statistically significant (P < 0.0001). 59 of 260 (23 %) patients from four previously published studies were classified as intermediate-risk when the 120-gene Oncotype DX approximation algorithm was applied. BreastPRS reclassified the 59 patients into binary risk groups (high- vs. low-risk). 23 (39 %) patients were classified as low-risk and 36 (61 %) as high-risk (P = 0.029, HR: 3.64, 95 % CI: 1.40-9.50). At 10 years from diagnosis, the low-risk group had a 90 % recurrence-free survival (RFS) rate compared to 60 % for the high-risk group. BreastPRS recurrence score is comparable with Oncotype DX and can reclassify Oncotype DX intermediate-risk patients into two groups with significant differences in RFS. Further studies are needed to validate these findings.

Cancer subtype discovery and biomarker identification via a new robust network clustering algorithm

In cancer biology, it is very important to understand the phenotypic changes of the patients and discover new cancer subtypes. Recently, microarray-based technologies have shed light on this problem based on gene expression profiles which may contain outliers due to either chemical or electrical reasons. These undiscovered subtypes may be heterogeneous with respect to underlying networks or pathways, and are related with only a few of interdependent biomarkers. This motivates a need for the robust gene expression-based methods capable of discovering such subtypes, elucidating the corresponding network structures and identifying cancer related biomarkers. This study proposes a penalized model-based Student’s t clustering with unconstrained covariance (PMT-UC) to discover cancer subtypes with cluster-specific networks, taking gene dependencies into account and having robustness against outliers. Meanwhile, biomarker identification and network reconstruction are achieved by imposing an adaptive penalty on the means and the inverse scale matrices. The model is fitted via the expectation maximization algorithm utilizing the graphical lasso. Here, a network-based gene selection criterion that identifies biomarkers not as individual genes but as subnetworks is applied. This allows us to implicate low discriminative biomarkers which play a central role in the subnetwork by interconnecting many differentially expressed genes, or have cluster-specific underlying network structures. Experiment results on simulated datasets and one available cancer dataset attest to the effectiveness, robustness of PMT-UC in cancer subtype discovering. Moveover, PMT-UC has the ability to select cancer related biomarkers which have been verified in biochemical or biomedical research and learn the biological significant correlation among genes.

Smoothed Linear Modeling for Smooth Spectral Data

Classification and prediction problems using spectral data lead to high-dimensional data sets. Spectral data are, however, different from most other high-dimensional data sets in that information usually varies smoothly with wavelength, suggesting that fitted models should also vary smoothly with wavelength. Functional data analysis, widely used in the analysis of spectral data, meets this objective by changing perspective from the raw spectra to approximations using smooth basis functions. This paper explores linear regression and linear discriminant analysis fitted directly to the spectral data, imposing penalties on the values and roughness of the fitted coefficients, and shows by example that this can lead to better fits than existing standard methodologies.

The Marker State Space (MSS) method for classifying clinical samples

The development of accurate clinical biomarkers has been challenging in part due to the diversity between patients and diseases. One approach to account for the diversity is to use multiple markers to classify patients, based on the concept that each individual marker contributes information from its respective subclass of patients. Here we present a new strategy for developing biomarker panels that accounts for completely distinct patient subclasses. Marker State Space (MSS) defines “marker states” based on all possible patterns of high and low values among a panel of markers. Each marker state is defined as either a case state or a control state, and a sample is classified as case or control based on the state it occupies. MSS was used to define multi-marker panels that were robust in cross validation and training-set/test-set analyses and that yielded similar classification accuracy to several other classification algorithms. A three-marker panel for discriminating pancreatic cancer patients from control subjects revealed subclasses of patients based on distinct marker states. MSS provides a straightforward approach for modeling highly divergent subclasses of patients, which may be adaptable for diverse applications.

Prediction of lung tumor types based on protein attributes by machine learning algorithms

Early diagnosis of lung cancers and distinction between the tumor types (Small Cell Lung Cancer (SCLC) and Non-Small Cell Lung Cancer (NSCLC) are very important to increase the survival rate of patients. Herein, we propose a diagnostic system based on sequence-derived structural and physicochemical attributes of proteins that involved in both types of tumors via feature extraction, feature selection and prediction models. 1497 proteins attributes computed and important features selected by 12 attribute weighting models and finally machine learning models consist of seven SVM models, three ANN models and two NB models applied on original database and newly created ones from attribute weighting models; models accuracies calculated through 10-fold cross and wrapper validation (just for SVM algorithms). In line with our previous findings, dipeptide composition, autocorrelation and distribution descriptor were the most important protein features selected by bioinformatics tools. The algorithms performances in lung cancer tumor type prediction increased when they applied on datasets created by attribute weighting models rather than original dataset. Wrapper-Validation performed better than X-Validation; the best cancer type prediction resulted from SVM and SVM Linear models (82%). The best accuracy of ANN gained when Neural Net model applied on SVM dataset (88%). This is the first report suggesting that the combination of protein features and attribute weighting models with machine learning algorithms can be effectively used to predict the type of lung cancer tumors (SCLC and NSCLC).

An industry statistician's perspective on PHC drug development

In the past decade, the cost of drug development has increased significantly. The estimates vary widely but frequently quoted numbers are staggering-it takes 10-15 years and billions of dollars to bring a drug to patients. To a large extent this is due to many long, expensive and ultimately unsuccessful drug trials. While one approach to combat the low yield on investment could be to continue searching for new blockbusters, an alternative method would lead us to focus on testing new targeted treatments that have a strong underlying scientific rationale and are more likely to provide enhanced clinical benefit in population subsets defined by molecular diagnostics. Development of these new treatments, however, cannot follow the usual established path; new strategies and approaches are required for the co-development of novel therapeutics and the diagnostic. In this paper we will review, from the point of view of industry, the approaches to, and challenges of drug development strategies incorporating predictive biomarkers into clinical programs. We will outline the basic concepts behind co-development with predictive biomarkers and summarize the current regulatory paradigm. We will present guiding principles of personalized health care (PHC) development and review the statistical, strategic, regulatory and operational challenges that statisticians regularly encounter on development programs with a PHC component. Some practical recommendations for team statisticians involved in PHC drug development are included. The majority of the examples and recommendations are drawn from oncology but broader concepts apply across all therapeutic areas.

Chronic carvedilol treatment partially reverses the right ventricular failure transcriptional profile in experimental pulmonary hypertension

Right ventricular failure (RVF) is the most frequent cause of death in patients with pulmonary arterial hypertension (PAH); however, specific therapies targeted to treat RVF have not been developed. Chronic treatment with carvedilol has been shown to reduce established maladaptive right ventricle (RV) hypertrophy and to improve RV function in experimental PAH. However, the mechanisms by which carvedilol improves RVF are unknown. We have previously demonstrated by microarray analysis that RVF is characterized by a distinct gene expression profile when compared with functional, compensatory hypertrophy. We next sought to identify the effects of carvedilol on gene expression on a genome-wide basis. PAH and RVF were induced in male Sprague-Dawley rats by the combination of VEGF-receptor blockade and chronic hypoxia. After RVF was established, rats were treated with carvedilol or vehicle for 4 wk. RNA was isolated from RV tissue and hybridized for microarray analysis. An initial prediction analysis of carvedilol-treated RVs showed that the gene expression profile resembled the RVF prediction set. However, a more extensive analysis revealed a small group of genes differentially expressed after carvedilol treatment. Further analysis categorized these genes in pathways involved in cardiac hypertrophy, mitochondrial dysfunction, and protein ubiquitination. Genes encoding proteins in the cardiac hypertrophy and protein ubiquitination pathways were downregulated in the RV by carvedilol, while genes encoding proteins in the mitochondrial dysfunction pathway were upregulated by carvedilol. These gene expression changes may explain some of the mechanisms that underlie the functional improvement of the RV after carvedilol treatment.

Multinomial logistic regression ensembles

This article proposes a method for multiclass classification problems using ensembles of multinomial logistic regression models. A multinomial logit model is used as a base classifier in ensembles from random partitions of predictors. The multinomial logit model can be applied to each mutually exclusive subset of the feature space without variable selection. By combining multiple models the proposed method can handle a huge database without a constraint needed for analyzing high-dimensional data, and the random partition can improve the prediction accuracy by reducing the correlation among base classifiers. The proposed method is implemented using R, and the performance including overall prediction accuracy, sensitivity, and specificity for each category is evaluated on two real data sets and simulation data sets. To investigate the quality of prediction in terms of sensitivity and specificity, the area under the receiver operating characteristic (ROC) curve (AUC) is also examined. The performance of the proposed model is compared to a single multinomial logit model and it shows a substantial improvement in overall prediction accuracy. The proposed method is also compared with other classification methods such as the random forest, support vector machines, and random multinomial logit model.

Genomic biomarkers for personalized medicine: development and validation in clinical studies

The establishment of high-throughput technologies has brought substantial advances to our understanding of the biology of many diseases at the molecular level and increasing expectations on the development of innovative molecularly targeted treatments and molecular biomarkers or diagnostic tests in the context of clinical studies. In this review article, we position the two critical statistical analyses of high-dimensional genomic data, gene screening and prediction, in the framework of development and validation of genomic biomarkers or signatures, through taking into consideration the possible different strategies for developing genomic signatures. A wide variety of biomarker-based clinical trial designs to assess clinical utility of a biomarker or a new treatment with a companion biomarker are also discussed.

Computational medicine: translating models to clinical care

Because of the inherent complexity of coupled nonlinear biological systems, the development of computational models is necessary for achieving a quantitative understanding of their structure and function in health and disease. Statistical learning is applied to high-dimensional biomolecular data to create models that describe relationships between molecules and networks. Multiscale modeling links networks to cells, organs, and organ systems. Computational approaches are used to characterize anatomic shape and its variations in health and disease. In each case, the purposes of modeling are to capture all that we know about disease and to develop improved therapies tailored to the needs of individuals. We discuss advances in computational medicine, with specific examples in the fields of cancer, diabetes, cardiology, and neurology. Advances in translating these computational methods to the clinic are described, as well as challenges in applying models for improving patient health.

Current Omics Technologies in Biomarker Discovery

Biomarkers are playing an increasingly important role in drug discovery and development and can be applied for many purposes, including disease mechanism study, diagnosis, prognosis, staging, and treatment selection. Advances in high-throughput “omics” technologies, including genomics, transcriptomics, proteomics and metabolomics, significantly accelerate the pace of biomarker discovery. Comprehensive molecular profiling using these “omics” technology has become a field of intensive research aiming at identifying biomarkers relevant for improved diagnostics and therapeutics. Although each “omics” technology plays important roles in biomarker research, different “omics” platforms have different strengths and limitations. This chapter aims to give an overview of these “omics” technologies and their current application in the biomarker discovery.

Genotype-phenotype matching analysis of 38 Lactococcus lactis strains using random forest methods

Background

Lactococcus lactis is used in dairy food fermentation and for the efficient production of industrially relevant enzymes. The genome content and different phenotypes have been determined for multiple L. lactis strains in order to understand intra-species genotype and phenotype diversity and annotate gene functions. In this study, we identified relations between gene presence and a collection of 207 phenotypes across 38 L. lactis strains of dairy and plant origin. Gene occurrence and phenotype data were used in an iterative gene selection procedure, based on the Random Forest algorithm, to identify genotype-phenotype relations.

Results

A total of 1388 gene-phenotype relations were found, of which some confirmed known gene-phenotype relations, such as the importance of arabinose utilization genes only for strains of plant origin. We also identified a gene cluster related to growth on melibiose, a plant disaccharide; this cluster is present only in melibiose-positive strains and can be used as a genetic marker in trait improvement. Additionally, several novel gene-phenotype relations were uncovered, for instance, genes related to arsenite resistance or arginine metabolism.

Conclusions

Our results indicate that genotype-phenotype matching by integrating large data sets provides the possibility to identify gene-phenotype relations, possibly improve gene function annotation and identified relations can be used for screening bacterial culture collections for desired phenotypes. In addition to all gene-phenotype relations, we also provide coherent phenotype data for 38 Lactococcus strains assessed in 207 different phenotyping experiments, which to our knowledge is the largest to date for the Lactococcus lactis species.

SMOTE for high-dimensional class-imbalanced data

BACKGROUND

Classification using class-imbalanced data is biased in favor of the majority class. The bias is even larger for high-dimensional data, where the number of variables greatly exceeds the number of samples. The problem can be attenuated by undersampling or oversampling, which produce class-balanced data. Generally undersampling is helpful, while random oversampling is not. Synthetic Minority Oversampling TEchnique (SMOTE) is a very popular oversampling method that was proposed to improve random oversampling but its behavior on high-dimensional data has not been thoroughly investigated. In this paper we investigate the properties of SMOTE from a theoretical and empirical point of view, using simulated and real high-dimensional data.

RESULTS

While in most cases SMOTE seems beneficial with low-dimensional data, it does not attenuate the bias towards the classification in the majority class for most classifiers when data are high-dimensional, and it is less effective than random undersampling. SMOTE is beneficial for k-NN classifiers for high-dimensional data if the number of variables is reduced performing some type of variable selection; we explain why, otherwise, the k-NN classification is biased towards the minority class. Furthermore, we show that on high-dimensional data SMOTE does not change the class-specific mean values while it decreases the data variability and it introduces correlation between samples. We explain how our findings impact the class-prediction for high-dimensional data.

CONCLUSIONS

In practice, in the high-dimensional setting only k-NN classifiers based on the Euclidean distance seem to benefit substantially from the use of SMOTE, provided that variable selection is performed before using SMOTE; the benefit is larger if more neighbors are used. SMOTE for k-NN without variable selection should not be used, because it strongly biases the classification towards the minority class.

Applications of functional data analysis: A systematic review

BACKGROUND

Functional data analysis (FDA) is increasingly being used to better analyze, model and predict time series data. Key aspects of FDA include the choice of smoothing technique, data reduction, adjustment for clustering, functional linear modeling and forecasting methods.

METHODS

A systematic review using 11 electronic databases was conducted to identify FDA application studies published in the peer-review literature during 1995-2010. Papers reporting methodological considerations only were excluded, as were non-English articles.

RESULTS

In total, 84 FDA application articles were identified; 75.0% of the reviewed articles have been published since 2005. Application of FDA has appeared in a large number of publications across various fields of sciences; the majority is related to biomedicine applications (21.4%). Overall, 72 studies (85.7%) provided information about the type of smoothing techniques used, with B-spline smoothing (29.8%) being the most popular. Functional principal component analysis (FPCA) for extracting information from functional data was reported in 51 (60.7%) studies. One-quarter (25.0%) of the published studies used functional linear models to describe relationships between explanatory and outcome variables and only 8.3% used FDA for forecasting time series data.

CONCLUSIONS

Despite its clear benefits for analyzing time series data, full appreciation of the key features and value of FDA have been limited to date, though the applications show its relevance to many public health and biomedical problems. Wider application of FDA to all studies involving correlated measurements should allow better modeling of, and predictions from, such data in the future especially as FDA makes no a priori age and time effects assumptions.

Stem cell-like gene expression in ovarian cancer predicts type II subtype and prognosis

Although ovarian cancer is often initially chemotherapy-sensitive, the vast majority of tumors eventually relapse and patients die of increasingly aggressive disease. Cancer stem cells are believed to have properties that allow them to survive therapy and may drive recurrent tumor growth. Cancer stem cells or cancer-initiating cells are a rare cell population and difficult to isolate experimentally. Genes that are expressed by stem cells may characterize a subset of less differentiated tumors and aid in prognostic classification of ovarian cancer. The purpose of this study was the genomic identification and characterization of a subtype of ovarian cancer that has stem cell-like gene expression. Using human and mouse gene signatures of embryonic, adult, or cancer stem cells, we performed an unsupervised bipartition class discovery on expression profiles from 145 serous ovarian tumors to identify a stem-like and more differentiated subgroup. Subtypes were reproducible and were further characterized in four independent, heterogeneous ovarian cancer datasets. We identified a stem-like subtype characterized by a 51-gene signature, which is significantly enriched in tumors with properties of Type II ovarian cancer; high grade, serous tumors, and poor survival. Conversely, the differentiated tumors share properties with Type I, including lower grade and mixed histological subtypes. The stem cell-like signature was prognostic within high-stage serous ovarian cancer, classifying a small subset of high-stage tumors with better prognosis, in the differentiated subtype. In multivariate models that adjusted for common clinical factors (including grade, stage, age), the subtype classification was still a significant predictor of relapse. The prognostic stem-like gene signature yields new insights into prognostic differences in ovarian cancer, provides a genomic context for defining Type I/II subtypes, and potential gene targets which following further validation may be valuable in the clinical management or treatment of ovarian cancer.