 Research
 Open Access
 Published:
AucPR: An AUCbased approach using penalized regression for disease prediction with highdimensional omics data
BMC Genomics volume 15, Article number: S1 (2014)
Abstract
Motivation
It is common to get an optimal combination of markers for disease classification and prediction when multiple markers are available. Many approaches based on the area under the receiver operating characteristic curve (AUC) have been proposed. Existing works based on AUC in a highdimensional context depend mainly on a nonparametric, smooth approximation of AUC, with no work using a parametric AUCbased approach, for highdimensional data.
Results
We propose an AUCbased approach using penalized regression (AucPR), which is a parametric method used for obtaining a linear combination for maximizing the AUC. To obtain the AUC maximizer in a highdimensional context, we transform a classical parametric AUC maximizer, which is used in a lowdimensional context, into a regression framework and thus, apply the penalization regression approach directly. Two kinds of penalization, lasso and elastic net, are considered. The parametric approach can avoid some of the difficulties of a conventional nonparametric AUCbased approach, such as the lack of an appropriate concave objective function and a prudent choice of the smoothing parameter. We apply the proposed AucPR for gene selection and classification using four real microarray and synthetic data. Through numerical studies, AucPR is shown to perform better than the penalized logistic regression and the nonparametric AUCbased method, in the sense of AUC and sensitivity for a given specificity, particularly when there are many correlated genes.
Conclusion
We propose a powerful parametric and easilyimplementable linear classifier AucPR, for gene selection and disease prediction for highdimensional data. AucPR is recommended for its good prediction performance. Beside gene expression microarray data, AucPR can be applied to other types of highdimensional omics data, such as miRNA and protein data.
Background
Nowadays, it is easy and common to measure thousands of markers simultaneously through highthroughput technologies, for example, the microarray study. A disease is usually related to several markers and the combination of multiple makers for classifying a subject into different statuses of a specific disease is widely studied. The performance of a combination of markers is frequently measured by indices related to the Receiver Operating Characteristic (ROC) curve: sensitivity, specificity, or the area under the ROC curve (AUC). Sensitivity (specificity) is defined as the probability of success in classifying a diseased (nondiseased) individual accurately. By varying the decision rules (thresholds), different sensitivities and specificities are obtained. The ROC curve plots all possible sensitivities against 1specificities and expresses the tradeoff between sensitivity and specificity visually. AUC is the most popular summary index for the curve; it has been shown to be the probability that the score of a randomly chosen diseased individual exceeds that of a randomly chosen nondiseased subject [1].
Therefore, it is natural to construct a combination of markers in order to maximize the ROCbased metrics. A number of combinations based on ROC indices have been suggested by [2–9]. Among these, [3] and [5] developed distributionfree methods to achieve the best linear combination for maximizing the smoothed AUC for highdimensional situations. They developed algorithms based on optimizing a sigmoid approximation of AUC. The sigmoid approximation of AUC relies on a smoothing parameter, which should be carefully chosen, though there are no theoretical guidelines for choosing this parameter. The rule of thumb for the choice of the smoothing parameter may reduce the power of the method. Moreover, the sigmoid approximation of AUC is not a concave function and multiple local maxima may exist. The global maximum is not guaranteed to be attained through commonly used numeric algorithms. For example, the performance of the linear combination decided by [5] is very poor for microarray data [9]. To avoid the difficulties of maximizing a nonparametric approximation of AUC, we can use a parametric method. To our knowledge, there is no published parametric method for maximizing the AUC under a highdimensional context. This paper tries to fill this gap.
We suggest an AUCbased approach using penalized regression (AucPR), based on a classical parametric linear combination derived by [2] in a lowdimensional context. The problem is then transformed into a linear regression framework, and the existing software for solving linear regression with penalization can be used directly, which facilitates the implementation of the proposed method. There are many penalty functions available, for example, the elastic net criterion [10], which is a mixture of penalties of L_{1} and L_{2} norms of the linear coefficients. The lasso penalty [11] is a special form of elastic net. Both the lasso and the elastic net have been widely used for marker selection and disease classification for highdimensional data [3, 5, 9, 10, 12, 13]. In this work, we maximize AUC through elastic net or lasso penalty. We compare the proposed AucPR to a logistic regression with elastic net or lasso penalty and the AUCbased nonparametric method proposed by [3], through four microarray data sets and synthetic data. The performance is gauged on the AUC and sensitivity given specificity equals to 0.95 on testing samples. AucPR achieves better prediction performance.
Methods
AucPR: An AUCbased approach using penalized regression
Suppose nondiseased samples {X_{ i }; 1 ≤ i ≤ m} and diseased samples {Y_{ j }; 1 ≤ j ≤ n} are independent and identically distributed (i.i.d.) from multivariate normal distributions N (µ_{ x }, Σ_{ x }) and N (µ_{ y } , Σ_{ y } ), respectively, where µ_{ x } and µ_{ y } are pdimension mean vectors, and Σ_{ x } and Σ_{ y } arep × p covariance matrices.
Under the multivariate normal distribution assumption, [2] showed that, among all possible linear combination of markers, without a positive constant multiplier, the combination with the coefficient vector
is optimum for maximizing the AUC. Furthermore, they also proved that if Σ_{ x } is proportional to Σ_{ y }, βis uniformly optimum, that is, it achieves the highest ROC curve among all linear combinations for all possible values of specificity.
Although this approach has been widely used in disease classification [14–17], it cannot be applied directly to highdimensional problems, where the number of markers (p) are larger than the number of observations in the sample. Penalized regression methods such as lasso [11] and elastic net [10], are effective tools for variable selection in highdimensional problems. We thus try to restate our problem in a regression framework.
Note that from Equation (1), µ_{ y } − µ_{ x } = (Σ_{ y }+ Σ_{ x })βholds. Instead of solving this equation, we suggest approximating β by solving the following linear regression problem:
where Iis a p × p identity matrix. By this transformation, we can avoid calculating the inverse of a large covariance matrix in (1), which is intractable due to lack of samples.
We then propose using a regularized linear regression method to obtain β. Let Σ = Σ_{ y } + Σ_{ x } = ((σ_{ ij })), 1 ≤ i, j ≤ p, and µ= µ_{ y } − µ_{ x } = (µ_{1}, . . . , µ_{ p })ʹ. Then, using the elastic net, we have
where λ is a parameter controlling the strength of the penalty and α is a mixing parameter that determines the relative strength of the L_{1} norm to the L_{2} norm, with 0 ≤ α ≤ 1. When α = 1, the elastic net reduces to lasso. The elastic net encourages a group of highly correlated markers to enter the model together, while lasso is quite parsimonious in selecting correlated markers. Under some conditions, both penalties were shown to have consistency in model selection [18, 19], or in other words, the selected model includes the true model with a high probability.
In practice, we replace the covariance matrices with the sample covariance matrices, and the mean vectors with the sample mean vectors. Formally,
and
The idea of the proposed AucPR is similar to a procedure proposed by [20] for the sparse linear discriminant analysis (LDA), where they restrict L_{∞} error and obtain the combination by linear programming. When Σ_{ x } and Σ_{ y } are proportional to each other, Σ ^{− 1}µis proportional to the coefficient vector of LDA. In this sense, AucPR also provides a solution for sparse LDA.
There are several computationally efficient algorithms to implement penalized linear regression for highdimensional data, for example, program lars by [21] and glmnet by [22]. In this paper, we use glmnet to solve Equation (3), since it is more efficient than lars [22].
Remark 1: We use sample mean vectors and sample covariance matrices, which are quite sensitive to the outlier observations. Therefore, intuitively, it may lead to the proposed method being inefficient under a general mean and a covariance structure without any restriction, especially when the sample sizes are small. However, AucPR can be powerful for some structures of Σ and µ, for example, when Σ or µare sparse, which is common in highdimensional data. We illustrate this with numerical studies in the Result and discussion Section.
Choice of tuning parameter
The tuning parameter λ controls the tradeoff between data fitting and model complexity. Given a larger λ, fewer markers are selected and the data may not be well fitted, while for a smaller λ, a larger number of markers are chosen and overfitting may occur. We tune λ in our numeric studies by a threefold crossvalidation (CV) method. Note that when the sample sizes are large, we can use a Kfold CV with K >3.
For the Kfold CV, we randomly divide the samples into K subsets of equal size. We select λ that maximizes the following CV score:
where ${\widehat{\beta}}_{\lambda}^{\left(i\right)}$ is the coefficient vector estimated without the samples in the ith fold, and ${\hat{AUC}}_{\lambda}^{\left(i\right)}$ is the empirical AUC estimator with the data in the i th fold, for a given λ, i = 1, ... , K. The empirical AUC estimator for a given βis defined as $\sum \sum I\left({\beta}^{\prime}\left({Y}_{i}{X}_{j}\right)>0\right)/nm$, with I (·) being the indicator function.
For the elastic net, α is fixed at 0.5 in this investigation. We note that although α can be tuned in the same fashion as λ, a simple, fixed α still captures the characteristics of the elastic net and is widely used in the literature as well [13, 23].
Another practical issue about tuning the parameter λ is how to provide the candidates of λ for CV, as it has not been specified clearly in the literature. We propose finding the range of λ using the whole data, and then generating a fixed number of candidates within that range such that they are evenly distributed in the log scale. Denoting the range of candidates for λ as [λ_{ l }, λ_{ u }], where λ_{ l } corresponds to the most complex model (for example, 100 markers are selected) while λ_{ u } corresponds to the least complex model (for example, 1 marker is chosen). It is easy to use the bisection method [24] to fix λ = λ_{ k }, such that there are exactly k nonzero coefficients (k = 1,..., p). To do this, we first have an initial guess at the value of λ. Let r(λ) be the number of nonzero coefficients of the tuning parameter λ. If r(λ) = k, we are done. If r(λ) < k, we let λ = λ/2, continuing this until r(λ) ≥ k. Once we have an interval [λ_{1}, λ_{2}], we employ the bisection method. We test the middle point λ_{ m } = (λ_{1} + λ_{2})/2, and if r(λ_{ m }) = k, we are done. If r(λ_{ m }) < k, set λ_{2} = λ_{ m }; otherwise, set λ_{1} = λ_{ m }. Repeat until r(λ_{ m }) = k.
Results and discussion
Application to gene selection and cancer classification
In this section, we apply our proposed AucPR, the penalized logistic regressions, and the AUC based nonparametric method proposed by [3], which maximizes a sigmoid approximation of AUC, to four microarray datasets for gene selection and cancer classification. We refer to our approaches to AucPR with elastic net and lasso penalty as AucEN and AucL, respectively, the logistic regression approaches with elastic net and lasso penalties as LogEN and LogL, respectively, and maximizing the sigmoid approximation of AUC as MSauc in the following content. The four microarray data sets are:
· Brain cancer data: The original data have five different types of tumors, and 42 samples with 5597 expressions. This data set was also studied by [25], and we use their preprocessed data and denote the first two types as the control group and the other three as the case group. It can be downloaded from http://stat.ethz.ch/~dettling/bagboost.html.

Colon cancer data: Expression levels of 40 tumors and 22 normal colon tissues for 2000 human genes, with the highest minimal intensity from 62 subjects measured [26]. The data can be downloaded from colonCA package on the Bioconductor website (http://www.bioconductor.org).

Leukemia data: We consider two types of leukemia cancer: acute myeloid leukemia (AML) and acute lymphoblast leukemia (ALL). Samples used by [27] were derived from 47 patients with ALL and 25 patients with AML, with 7129 genes. The data set is available in the golubEsets package on the Bioconductor website (http://www.bioconductor.org).

DLBCL data: The diffused large Bcell lymphoma (DLBCL) data set contains 58 DLBCL patients and 19 follicular lymphoma patients from a related germinal center Bcell lymphoma [28]. The data are available from the Broad Institute website (http://www.genome.wi.mit.edu/MPR/lymphoma).
All data sets are further processed using quantile normalization and logarithm transformation (except the Brain cancer data, since it has been preprocessed). To save computation time, we screen the genes such that the 1000 genes with the largest absolute moderated tstatistics [29] are kept. Filtering genes by ttyped statistics has been widely used in the literature, for example, [3, 5, 20] among others. Our empirical study shows that including more than 1000 genes does not significantly change the patterns found. LogEN and LogL are also implemented by R package glmnet and the tuning parameter is chosen by a threefold CV, using the CV score defined in Equation (4).
Then, we randomly split the data into training and testing sets, comprising 2/3 and 1/3 of the sample, respectively. The AUC value and sensitivity when specificity = 0.95 are evaluated based on the testing set. This procedure is repeated 100 times and the boxplots of the two comparison metrics are plotted in Figures 1  4.
We can see that the proposed AucPR outperforms the other approaches for all the four datasets,. The AucEN has the best prediction performance. The AucL is slightly less powerful than the AucEN, but better than the other three methods. The penalized logistic regression and MSauc perform poorly for the Brain and Colon cancer data. Even though the differences of AUC between these approaches are small for Leukemia and DLBCL data, the superiority of the proposed AUCbased methods becomes larger in sensitivity when specificity is as high as 0.95. This finding is very meaningful, since high sensitivity and high specificity are greatly appreciated for real cancer studies.
For the four data sets, Table 1 shows the median number of nonzero coefficients for each method. The AucPR selects more markers than the others. The approaches with the elastic net provide more genes than the lasso penalized approaches, which is consistent with the literature [10]. MSauc generally selects more genes than the penalized logistic regressions, but does not always give a better prediction performance than the penalized logistic regressions (Figures 1  4). The averaged ROC curves are plotted in Figure 5, showing that the ROC curve of the proposed AucPR lies above the curves of others, especially in Brain and Colon cancer data.
In summary, the proposed AucPR selects more genes than the other three competing approaches and also have better prediction performance. Although a sparse model is good for interpretation, a better prediction performance is the primary objective and more appealing in many real world applications. As pointed out by [20], sparse models commonly ignore the correlations between the variables, which are generally inefficient even when the zero markers (or "unimportant" markers) are known in advance and all the important markers are selected correctly. It was demonstrated that those unimportant markers are in fact useful and even potentially important for classification because of the correlations. In addition, including a sufficient number of genes to the model has a practical implication; more potential important genes may be incorporated and these genes would have a higher chance for further investigation. In Table 2 the top 10 frequently selected genes by AucEN are listed for Colon, Leukaemia and DLBCL data sets (The gene information is not included in the preprocessed Brain cancer data, so we omit those results). The genes which are commonly selected by other approaches are marked. Gene description and related studies in the literature are shown too. The top frequently selected genes by AucEN were also reported in the literature.
There are several other popular approaches available for classification in highdimensional situation. For example, the "SIS" function in package SIS (http://cran.rproject.org/web/packages/SIS/index.html), which first implements the Iterative Sure Independence Screening [30], and then fits the final model by penalized regression; treebased method "randomForest" in randomForest package [31]; LogEN with alternative CV score ("type.measure = deviance" in glmnet package). As a demonstration, we implemented the third approach on the Brain cancer data. The result is improved and has been updated. However, Our method still outperforms others (see Figure 1).
Remark 2: Prediction accuracy and interpretation are two major concerns for microarray cancer classification study. A sparse model is generally easier to interpret but may not reflect the true biological phenomena or have poor prediction. For example, many genes are highly correlated in microarray data, and these genes may work together. Therefore, it is worthwhile to identify these genes jointly to increase prediction performance and to provide a sufficient number of potential risks for a further validation study. Note that the lasso penalized logistic regression is too parsimonious, as it cannot select a sufficient number of genes in a highly correlated group and thus, has poor prediction performance, while our AucL method, although with the lasso penalty, seems to be able to alleviate this problem by selecting more genes.
Simulation
In this section, we demonstrate our approaches using synthetic data under two scenarios; genes are generated from a normal distribution or a mixture of normal distributions.
We first simulate gene expressions following a setting similar to [32], where they mimicked the real microarray data. We generate data under a different number of independent blocks (block = 1, 2, 3), and the number of genes per block (size=5, 20, 40). The data are simulated from multivariate normal distributions N (µ_{ x }, Σ_{ x }(ρ)) and N (µ_{ y }, Σ_{ y }(ρ)) for diseased and nondiseased classes, respectively. All genes have a variance of 1, and the correlation between genes within a block is ρ (ρ = 0.3, 0.6, 0.9), whereas the correlation between genes among blocks is 0. In other words, the covariance matrix is a blockdiagonal matrix
where
The mean vectors are set as µ_{ y } = (0.6, 0.6,... , 0.6) and µ_{ x } = (−0.6, −0.6, ..., −0.6). Here the mean vectors are selected such that the AUC of each single gene is 0.8.
After the informative genes described above are generated, we evenly add a type of noninformative "genes" from N (0, 1) and another type of noninformative "genes" from U [−1, 1], for both diseased and nondiseased observations, and make 1000 markers in total.
We generate n = m = 40 i.i.d. individuals as a training set for diseased and nondiseased samples, respectively, from the above distributions. Under the same structure as the training set, another n = m = 20 samples are simulated independently as a testing set. Each method is applied to the training set and the prediction performance is measured on the testing set. We repeat this procedure 100 times, as we have done in the examples with real data.
For the synthetic data, the AucPR shows better prediction accuracy than the other three approaches in most scenarios. The median values of AUC, sensitivity when specificity = 0.95, the number of true informative markers being selected (nIMS ), and the number of total markers being selected (nTMS ), are summarized in Tables 3, 4, 5. There are some facts we can state, based on the simulation results:
1 Given ρ and the number of blocks (block), as the block size increases, our AucPR dominates the other approaches. We summarize the results when size = 5 and size = 40 in Table 3.
2 Given block size and the number of blocks, as ρ becomes larger, the performance of our methods do not vary much, while those of the other three methods become worse. Specifically, the sensitivities of our methods are getting larger than others when ρ is getting larger. The results for ρ = 0.3 and ρ = 0.9 are given in Table 4.
3 As the number of independent blocks increases, all methods have improved performances. When the number of blocks is 3, except LogL and MSauc, the other three methods seem to be similar in each case with AucEN performing slightly better (Table 5).
4 Penalized logistic regression performs better only when ρ is small (for example, 0.3) and the number of the informative genes is small. Approaches with elastic net penalty always lead to better results than the approaches with lasso penalty (Tables 3, 4, 5).
5 Generally, our AucPR approaches select more informative genes, and the approaches with elastic net penalty incorporate more informative genes than the approaches with lasso penalty (Tables 3, 4, 5). Note that as block and/or size increase (or equivalently, as the number of informative genes increases), the number of selected informative genes for our AUCbased methods increase faster, but logistics regression based approaches and MSauc do not. This fact may be interpreted as that our approaches show better prediction accuracy.
Next, we also study the scenario where the genes are generated from a nonGaussian setting. We simulate 50 informative genes from 0.8N (µ_{ y }, Σ_{ y }(0.8)) + 0.2N (0 , I) and 0.8N (µ_{ x }, Σ_{ x }(0.8))+0.2N (0 , I) for diseased and nondiseased groups, respectively. The noninformative genes are generated in the same way as in the first scenario. Similar patterns can be found as in the normal distribution scenario (data are not shown).
In summary, through selecting more genes, the proposed AucPR performs better when there are a lot of informative genes or the correlations between them are high (larger than 0.6 for example).
Remark 3: Note that the penalized logistic regressions are very powerful for marker selection in the sense that all the selected genes are the true informative genes, that is, nIMS = nTMS. For AucPR, the nTMS is larger than the nIMS, that is, there are some noisy genes selected. If the sample size increases, this phenomenon can be avoided or become negligible. Figure 6 shows that when the sample size is larger than 100, the number of noisy genes selected by AucL becomes very small.
Discussion
Note that, in our comparison study, the tuning parameters for all methods are tuned with an empirical (nonparametric) AUC estimator as the CV score. When sample size is very small, some difficulties may occur for calculating such AUC estimators as we did in the brain cancer study. Alternatively, parametric AUC estimators or the deviance from a distribution model can be used as the CV score. Different CV scores may lead to different results, especially when the sample sizes are small. It is worthy of investigating this issue as a future research topic.
Although we only use gene expression microarray data, AucPR can also be applied to other types of highthroughput omics data, such as miRNA and protein data.
AucPR methods rely on sample mean vectors and sample covariance matrices, which may not be stable enough, specifically when only a small number of samples are available. An improvement may exist in practice by replacing them with, for example, sample median and the positivedefinite estimator of a large covariance matrix proposed by [33]. This can be a topic of future research.
Note that after the transformation, we try to solve a regression problem with p "samples" and p "predictors." Thus, the computation cost would grow quickly as p increases. Although screening the original p genes to a smaller number (1000 in our numerical studies) of genes is widely used and does not affect the prediction performance, as seen from our empirical study and the relevant literature [3, 5, 13, 20], it is still worthwhile to develop fast algorithms for large scale and highdimensional regression problem. This, too, needs further investigation.
Conclusions
We propose a powerful parametric and easilyimplementable linear classifier AucPR, for gene selection and disease prediction for highdimensional data. We transform a classical parametric AUC estimator into a linear regression and thus, the existing packages for regularized linear regression can be used directly. This novelty makes the implementation of the proposed methods very easy and efficient, since the regularized regression has been well studied. The proposed parametric method also avoids maximizing a nonconcave objective function and elaborately choosing the smoothing parameter in a conventional nonparametric method. Comparisons among the AucPR, the penalized logistic regression, and a nonparametric AUCbased approach shows that our methods lead to better classifiers in the sense of predictive performance, through application to real microarray and synthetic data. In addition, the proposed AucPR selects more markers than the others and thus, could include more potential important markers for further investigation.
In addition, [34] demonstrated that the linear combination of multiple markers based on maximizing AUC generally performs better than logistic regression when the logistic model does not hold, and the two methods are comparable when the logistic model is satisfied, but their analysis was done under the condition that a very limited number of markers would be considered. This paper states that the AUCbased approach could also be advocated in highdimensional setting, since it achieves better prediction ability than the penalized logistic regression.
Availability and supporting data
This work was implemented in R software. The R source codes are freely available at http://bibs.snu.ac.kr/software/aucpr.
References
 1.
Bamber D: The area above the ordinal dominance graph and the area below the receiver operating characteristic graph. Journal of mathematical psychology. 1975, 12 (4): 387415. 10.1016/00222496(75)900012.
 2.
Su JQ, Liu JS: Linear combinations of multiple diagnostic markers. Journal of the American Statistical Association. 1993, 88 (424): 13501355. 10.1080/01621459.1993.10476417.
 3.
Ma S, Huang J: Regularized ROC method for disease classification and biomarker selection with microarray data. Bioinformatics. 2005, 21 (24): 43564362. 10.1093/bioinformatics/bti724.
 4.
Ma S, Song X, Huang J: Supervised group lasso with applications to microarray data analysis. BMC bioinformatics. 2007, 8 (1): 6010.1186/14712105860.
 5.
Wang Z, Yuanchin IC, Ying Z, Zhu L, Yang Y: A parsimonious thresholdindependent protein feature selection method through the area under receiver operating characteristic curve. Bioinformatics. 2007, 23 (20): 27882794. 10.1093/bioinformatics/btm442.
 6.
Osamu K, Shinto E: A boosting method for maximizing the partial area under the ROC curve. BMC Bioinformatics. 2010, 11:
 7.
Wang Z, Chang YCI: Marker selection via maximizing the partial area under the ROC curve of linear risk scores. Biostatistics. 2011, 12 (2): 369385. 10.1093/biostatistics/kxq052.
 8.
Hsu MJ, Hsueh HM: The linear combinations of biomarkers which maximize the partial area under the ROC curves. Computational Statistics. 2013, 120.
 9.
Yu W, Chang YcI, Park E: A modified area under the roc curve and its application to marker selection and classification. Journal of the Korean Statistical Society. 2014, 43 (2): 161175. 10.1016/j.jkss.2013.05.003.
 10.
Zou H, Hastie T: Regularization and variable selection via the elastic net. Journal of the Royal Statistical Society: Series B (Statistical Methodology). 2005, 67 (2): 301320. 10.1111/j.14679868.2005.00503.x.
 11.
Tibshirani R: Regression shrinkage and selection via the lasso. Journal of the Royal Statistical Society Series B (Methodological). 1996, 267288.
 12.
Ghosh D, Chinnaiyan AM: Classification and selection of biomarkers in genomic data using lasso. BioMed Research International. 2005, 2005 (2): 147154.
 13.
Liu Z, Jiang F, Tian G, Wang S, Sato F, Meltzer SJ, Tan M: Sparse logistic regression with lp penalty for biomarker identification. Statistical Applications in Genetics and Molecular Biology. 2007, 6 (1):
 14.
Schisterman E, Faraggi D, Browne R, Freudenheim J, Dorn J, Muti P, Armstrong D, Reiser B, Trevisan M: Minimal and best linear combination of oxidative stress and antioxidant biomarkers to discriminate cardiovascular disease. Nutrition, metabolism, and cardiovascular diseases: NMCD. 2002, 12 (5): 259266.
 15.
Weber F, Shen L, Aldred MA, Morrison CD, Frilling A, Saji M, Schuppert F, Broelsch CE, Ringel MD, Eng C: Genetic classification of benign and malignant thyroid follicular neoplasia based on a threegene combination. Journal of Clinical Endocrinology & Metabolism. 2005, 90 (5): 25122521. 10.1210/jc.20042028.
 16.
Lu LJ, Xia Y, Paccanaro A, Yu H, Gerstein M: Assessing the limits of genomic data integration for predicting protein networks. Genome research. 2005, 15 (7): 945953. 10.1101/gr.3610305.
 17.
Attallah AM, Mosa TE, Omran MM, AboZeid MM, ElDosoky I, Shaker YM: Immunodetection of collagen types i, ii, iii, and iv for differentiation of liver fibrosis stages in patients with chronic hcv. Journal of Immunoassay & Immunochemistry. 2007, 28 (2): 155168. 10.1080/15321810701212088.
 18.
Zhao P, Yu B: On model selection consistency of lasso. The Journal of Machine Learning Research. 2006, 7: 25412563.
 19.
Jia J, Yu B: On model selection consistency of the elastic net when p¿¿ n. Technical report, DTIC Document. 2008
 20.
Cai T, Liu W: A direct estimation approach to sparse linear discriminant analysis. Journal of the American Statistical Association. 2011, 106 (496):
 21.
Efron B, Hastie T, Johnstone I, Tibshirani R: Least angle regression. The Annals of statistics. 2004, 32 (2): 407499. 10.1214/009053604000000067.
 22.
Friedman J, Hastie T, Tibshirani R: Regularization paths for generalized linear models via coordinate descent. Journal of statistical software. 2010, 33 (1): 1
 23.
Ayers KL, Cordell HJ: Snp selection in genomewide and candidate gene studies via penalized logistic regression. Genetic epidemiology. 2010, 34 (8): 879891. 10.1002/gepi.20543.
 24.
Wu TT, Chen YF, Hastie T, Sobel E, Lange K: Genomewide association analysis by lasso penalized logistic regression. Bioinformatics. 2009, 25 (6): 714721. 10.1093/bioinformatics/btp041.
 25.
Dettling M: Bagboosting for tumor classification with gene expression data. Bioinformatics. 2004, 20 (18): 35833593. 10.1093/bioinformatics/bth447.
 26.
Alon U, Barkai N, Notterman DA, Gish K, Ybarra S, Mack D, Levine AJ: Broad patterns of gene expression revealed by clustering analysis of tumor and normal colon tissues probed by oligonucleotide arrays. Proceedings of the National Academy of Sciences. 1999, 96 (12): 67456750. 10.1073/pnas.96.12.6745.
 27.
Golub TR, Slonim DK, Tamayo P, Huard C, Gaasenbeek M, Mesirov JP, Coller H, Loh ML, Downing JR, Caligiuri MA, et al: Molecular classification of cancer: class discovery and class prediction by gene expression monitoring. science. 1999, 286 (5439): 531537. 10.1126/science.286.5439.531.
 28.
Shipp MA, Ross KN, Tamayo P, Weng AP, Kutok JL, Aguiar RC, Gaasenbeek M, Angelo M, Reich M, Pinkus GS, et al: Diffuse large bcell lymphoma outcome prediction by geneexpression profiling and supervised machine learning. Nature medicine. 2002, 8 (1): 6874. 10.1038/nm010268.
 29.
Smyth GK, et al: Linear models and empirical bayes methods for assessing differential expression in microarray experiments. Statistical applications in genetics and molecular biology. 2004, 3 (1): 3
 30.
Fan J, Lv J: Sure independence screening for ultrahigh dimensional feature space. Journal of the Royal Statistical Society: Series B (Statistical Methodology). 2008, 70 (5): 849911. 10.1111/j.14679868.2008.00674.x.
 31.
Liaw A, Wiener M: Classification and regression by randomforest. R news. 2002, 2 (3): 1822.
 32.
DíazUriarte R, De Andres SA: Gene selection and classification of microarray data using random forest. BMC bioinformatics. 2006, 7 (1): 310.1186/1471210573.
 33.
Xue L, Ma S, Zou H: Positivedefinite 1penalized estimation of large covariance matrices. Journal of the American Statistical Association. 2012, 107 (500): 14801491. 10.1080/01621459.2012.725386.
 34.
Pepe MS, Cai T, Longton G: Combining predictors for classification using the area under the receiver operating characteristic curve. Biometrics. 2006, 62 (1): 221229. 10.1111/j.15410420.2005.00420.x.
 35.
Jabari S, da Silveira AB, de Oliveira EC, Quint K, Wirries A, Neuhuber W, Brehmer A: Mucosal layers and related nerve fibres in nonchagasic and chagasic human colona quantitative immunohistochemical study. Cell and tissue research. 2014, 19.
 36.
ÁlvarezChaver P, RodríguezPiñeiro AM, RodríguezBerrocal FJ, GarcíaLorenzo A, Páez de la Cadena M, MartínezZorzano VS: Selection of putative colorectal cancer markers by applying pca on the soluble proteome of tumors: Ndk a as a promising candidate. Journal of proteomics. 2011, 74 (6): 874886. 10.1016/j.jprot.2011.02.031.
 37.
Nambiar PR, Gupta RR, Misra V: An omics based survey of human colon cancer. Mutation Research/Fundamental and Molecular Mechanisms of Mutagenesis. 2010, 693 (1): 318.
 38.
Xq Z, Zhang F, Tao Y, Cm W, Sz L, Fl H, et al: Expression profiling based on graphclustering approach to determine colon cancer pathway. Journal of cancer research and therapeutics. 2013, 9 (3): 46710.4103/09731482.119351.
 39.
Jiang W, Li X, Rao S, Wang L, Du L, Li C, Wu C, Wang H, Wang Y, Yang B: Constructing diseasespecific gene networks using pairwise relevance metric: application to colon cancer identifies interleukin 8, desmin and enolase 1 as the central elements. BMC systems biology. 2008, 2 (1): 7210.1186/17520509272.
 40.
Tabuchi Y, Takasaki I, Doi T, Ishii Y, Sakai H, Kondo T: Genetic networks responsive to sodium butyrate in colonic epithelial cells. FEBS letters. 2006, 580 (13): 30353041. 10.1016/j.febslet.2006.04.048.
 41.
Floyd RV, Wray S, MartínVasallo P, Mobasheri A: Differential cellular expression of fxyd1 (phospholemman) and fxyd2 (gamma subunit of na, katpase) in normal human tissues: a study using high density human tissue microarrays. Annals of AnatomyAnatomischer Anzeiger. 2010, 192 (1): 716. 10.1016/j.aanat.2009.09.003.
 42.
Samet I, Han J, Jlaiel L, Sayadi S, Isoda H: Olive (olea europaea) leaf extract induces apoptosis and monocyte/macrophage differentiation in human chronic myelogenous leukemia k562 cells: Insight into the underlying mechanism. Oxidative medicine and cellular longevity. 2014, 2014:
 43.
Cierniewski CS, PapiewskaPajak I, Malinowski M, SacewiczHofman I, Wiktorska M, Kryczka J, Wysocki T, Niewiarowska J, Bednarek R: Thymosin β4 regulates migration of colon cancer cells by a pathway involving interaction with ku80. Annals of the New York Academy of Sciences. 2010, 1194 (1): 6071. 10.1111/j.17496632.2010.05480.x.
 44.
Damm F, Thol F, Hollink I, Zimmermann M, Reinhardt K, van den HeuvelEibrink M, Zwaan CM, de Haas V, Creutzig U, Klusmann J: Prevalence and prognostic value of idh1 and idh2 mutations in childhood aml: a study of the amlbfm and dcog study groups. Leukemia. 2011, 25 (11): 17041710. 10.1038/leu.2011.142.
 45.
Zgheib C, Zouein FA, Kurdi M, Booz GW: Chronic treatment of mice with leukemia inhibitory factor does not cause adverse cardiac remodeling but improves heart function. European cytokine network. 2012, 23 (4): 191197.
 46.
Perry C, Pick M, Podoly E, GilboaGeffen A, Zimmerman G, Sklan E, BenShaul Y, Diamant S, Soreq H: Acetylcholinesterase/c terminal binding protein interactions modify ikaros functions, causing t lymphopenia. Leukemia. 2007, 21 (7): 14721480. 10.1038/sj.leu.2404722.
 47.
Sasaki H, Nishikata I, Shiraga T, Akamatsu E, Fukami T, Hidaka T, Kubuki Y, Okayama A, Hamada K, Okabe H: Overexpression of a cell adhesion molecule, tslc1, as a possible molecular marker for acutetype adult tcell leukemia. Blood. 2005, 105 (3): 12041213.
 48.
Toh Y, Nicolson GL: The role of the mta family and their encoded proteins in human cancers: molecular functions and clinical implications. Clinical & experimental metastasis. 2009, 26 (3): 215227. 10.1007/s1058500892338.
 49.
Guan X, Yang J, Zhu N, Wang Y, Li R, Zheng Z: [gene expression differences between high and low metastatic cells of adenoid cystic carcinoma]. Zhonghua kou qiang yi xue za zhi= Zhonghua kouqiang yixue zazhi= Chinese journal of stomatology. 2004, 39 (2): 118121.
 50.
Carlet M, Janjetovic K, Rainer J, Schmidt S, PanzerGrümayer R, Mann G, Prelog M, Meister B, Ploner C, Kofler R: Expression, regulation and function of phosphofructokinase/fructosebiphosphatases (pfkfbs) in glucocorticoidinduced apoptosis of acute lymphoblastic leukemia cells. BMC cancer. 2010, 10 (1): 63810.1186/1471240710638.
 51.
Meyer C, Kowarz E, Yip SF, Wan TSK, Chan TK, Dingermann T, Chan LC, Marschalek R: A complex¡ i¿ mll¡/i¿ rearrangement identified five years after initial mds diagnosis results in outofframe fusions without progression to acute leukemia. Cancer genetics. 2011, 204 (10): 557562. 10.1016/j.cancergen.2011.10.001.
 52.
Chen C, Zhou Z, Ross JS, Zhou W, Dong JT: The amplified wwp1 gene is a potential molecular target in breast cancer. International journal of cancer. 2007, 121 (1): 8087. 10.1002/ijc.22653.
 53.
Zangrando A, Dell'Orto MC, te Kronnie G, Basso G: Mll rearrangements in pediatric acute lymphoblastic and myeloblastic leukemias: Mll specific and lineage specific signatures. BMC medical genomics. 2009, 2 (1): 3610.1186/17558794236.
 54.
Sung CO, Kim SC, Karnan S, Karube K, Shin HJ, Nam DH, Suh YL, Kim SH, Kim JY, Kim SJ, et al: Genomic profiling combined with gene expression profiling in primary central nervous system lymphoma. Blood. 2011, 117 (4): 12911300. 10.1182/blood201007297861.
 55.
Delmolino LM, Saha P, Dutta A: Multiple mechanisms regulate subcellular localization of human cdc6. Journal of Biological Chemistry. 2001, 276 (29): 2694726954. 10.1074/jbc.M101870200.
 56.
Glud SZ, Sørensen AB, Andrulis M, Wang B, Kondo E, Jessen R, Krenacs L, Stelkovics E, Wabl M, Sering E, et al: A tumorsuppressor function for nfatc3 in tcell lymphomagenesis by murine leukemia virus. Blood. 2005, 106 (10): 35463552. 10.1182/blood2005020493.
 57.
Seimiya M, Bahar R, Wang Y, Kawamura K, Tada Y, Okada S, Hatano M, Tokuhisa T, Saisho H, Watanabe T, et al: Clast5/stra13 is a negative regulator of b lymphocyte activation. Biochemical and biophysical research communications. 2002, 292 (1): 121127. 10.1006/bbrc.2002.6605.
 58.
de Leval L, Rickman DS, Thielen C, de Reynies A, Huang YL, Delsol G, Lamant L, Leroy K, Brièere J, Molina T, et al: The gene expression profile of nodal peripheral tcell lymphoma demonstrates a molecular link between angioimmunoblastic tcell lymphoma (aitl) and follicular helper t (tfh) cells. Blood. 2007, 109 (11): 49524963. 10.1182/blood200610055145.
 59.
Lin YW, Aplan PD: Gene expression profiling of precursor tcell lymphoblastic leukemia/lymphoma identifies oncogenic pathways that are potential therapeutic targets. Leukemia. 2007, 21 (6): 12761284. 10.1038/sj.leu.2404685.
Acknowledgements
This work was supported by the National Research Foundation of Korea(NRF) grants funded by the Korea government(MSIP) (No. 2012R1A3A2026438 and 20080062618).
Declarations
Publication charges for this work was funded by the National Research Foundation of Korea(NRF) grants funded by the Korea government(MSIP) (No. 2012R1A3A2026438 and 20080062618).
This article has been published as part of BMC Genomics Volume 15 Supplement 10, 2014: Proceedings of the 25th International Conference on Genome Informatics (GIW/ISCBAsia): Genomics. The full contents of the supplement are available online at http://0www.biomedcentral.com.brum.beds.ac.uk/bmcgenomics/supplements/15/S10.
Author information
Affiliations
Corresponding author
Additional information
Competing interests
The authors declare that they have no competing interests.
Authors' contributions
WY and TP designed the method. WY performed the analyses. WY and TP interpreted the results and wrote the manuscript.
Rights and permissions
This article is published under license to BioMed Central Ltd. This is an Open Access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/4.0), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited. The Creative Commons Public Domain Dedication waiver (http://creativecommons.org/publicdomain/zero/1.0/) applies to the data made available in this article, unless otherwise stated.
About this article
Cite this article
Yu, W., Park, T. AucPR: An AUCbased approach using penalized regression for disease prediction with highdimensional omics data. BMC Genomics 15, S1 (2014). https://0doiorg.brum.beds.ac.uk/10.1186/1471216415S10S1
Published:
DOI: https://0doiorg.brum.beds.ac.uk/10.1186/1471216415S10S1
Keywords
 AUC
 highdimensional data
 penalized regression
 ROC curve