Dimension reduction with gene expression data using targeted variable importance measurement
© Wang and van der Laan; licensee BioMed Central Ltd. 2011
Received: 2 June 2011
Accepted: 29 July 2011
Published: 29 July 2011
When a large number of candidate variables are present, a dimension reduction procedure is usually conducted to reduce the variable space before the subsequent analysis is carried out. The goal of dimension reduction is to find a list of candidate genes with a more operable length ideally including all the relevant genes. Leaving many uninformative genes in the analysis can lead to biased estimates and reduced power. Therefore, dimension reduction is often considered a necessary predecessor of the analysis because it can not only reduce the cost of handling numerous variables, but also has the potential to improve the performance of the downstream analysis algorithms.
We propose a TMLE-VIM dimension reduction procedure based on the variable importance measurement (VIM) in the frame work of targeted maximum likelihood estimation (TMLE). TMLE is an extension of maximum likelihood estimation targeting the parameter of interest. TMLE-VIM is a two-stage procedure. The first stage resorts to a machine learning algorithm, and the second step improves the first stage estimation with respect to the parameter of interest.
We demonstrate with simulations and data analyses that our approach not only enjoys the prediction power of machine learning algorithms, but also accounts for the correlation structures among variables and therefore produces better variable rankings. When utilized in dimension reduction, TMLE-VIM can help to obtain the shortest possible list with the most truly associated variables.
Gene expression microarray data are typically characterized by large quantities of variables with unknown correlation structures [1, 2]. This high dimensionality has presented us challenges in analyzing the data, especially when correlations among variables are complex. Including many variables in standard statistical analyses can easily cause problems such as singularity and overfitting, and sometimes is not even doable. To manage this problem, the dimensionality of the data will often be reduced in the first step. There are multiple ways to achieve this goal. One is to select a subset of genes based on certain criteria such that this subset of genes is believed to best predict the outcome. This gene selection strategy is typically based on some univariate measurement related to the outcome, such as t-test and rank test [3, 4]. Another strategy is to use a weighted combination of genes of lower dimension to represent the total variation of the data. Representative approaches are principle component analysis (PCA)  and partial least squares (SLR) [6–9]. Machine learning algorithms such as LASSO [10, 11] and Random Forest  have embedded capacity to select variables while simultaneously making predictions, and can be used to accommodate high dimensional microarray data.
As always, there is no one-size-fits-all solution to this problem, and one often needs to resort to a mix-and-match strategy. The univariate-measurement based gene selection is a very popular approach in the field. It is fast and scales easily to the dimension of the data. The output is usually stable and easy to understand, and fulfills the objectives of the biologists to directly pursue interesting findings. However, it often relies on over-simplified models. For instance, the univariate analysis evaluates every gene in isolation of others, with the unrealistic assumption of independence among genes. As a result, it carries a lot of noise and the selected genes are often highly correlated, which themselves create problems in subsequent analysis. Also, due to the practical limit of the size of the gene subset, real informative genes with weaker signals will be left out. In contrast, PCA/PLS constructs a few gene components as linear combinations of all genes in a dataset. This "Super Gene" approach assumes that the majority of the variation in the dataset can be explained by a small number of underlying variables. One then uses these gene components to predict the outcome. These approaches can better handle the dependent structure of genes and their performances are quite acceptable . But it is harder to interpret gene components biologically, and to assess the effect of individual genes one needs to look at the weight coefficients of the linear combination. Machine learning algorithms are very attractive variable selection tools to deal with large quantities of genes. They are prediction algorithms with embedded abilities to select gene subsets. However, whether or not a gene is chosen by a learning algorithm may not be the best measurement of its importance. Machine learning algorithms are constructed to achieve an optimal prediction accuracy, which often overlooks the importance of each variable. Consequently, small changes in data or tuning parameters may result in big changes in variable rankings and the the selected gene subsets are instable. For example, Random Forest, a tree-based non-parametric method, has a variable importance measurement that greatly contributes to its popularity. This measurement is sensitive to the parameter choices of trees in the presence of high correlations among variables, because different sets of variables can produce nearly unchanged prediction accuracy [14, 15]. Another example is LASSO -- one of the most popular regularization algorithms. Assuming a sparse signal, LASSO handles the high dimensionality problem by shrinking the coefficients of most variables towards zero . A recent implementation of LASSO is in the GLMNET R package . The package uses a coordinate descent algorithm and can finish an analysis of 20,000 variables within a few seconds. To us, its result is somewhat sensitive to the choice of the penalizing parameter λ. Different λ s may result in gene subsets with little overlapping. In the mean time, variable importance measurements are not readily available in LASSO. One can simply rank genes by their coefficients, but this can be quite subtle. Although permutation tests may be used to derive p-values, how to perform the permutation is a tricky matter due to selection of tuning parameters. For small p-values, it is still computationally infeasible. In this paper, inspired by concepts of counterfactual effects from the causal inference literature, we propose a targeted variable importance measurement [18, 19] to rank genes and reduce the dimensionality of the dataset. Counterfactuals are usually defined in the context of treatment to disease. It is the outcome a patient would have had a treatment been assigned differently, with everything else held the same. Hence counterfactuals are "counter"-fact and apparently impossible to be observed. But it can be estimated statistically. Suppose that we have an outcome Y, a binary treatment A, and the confounding variables W of A, and we have worked out correctly an estimate of the conditional expectation of Y given A and W. A common way to estimate the counterfactual effect of A is to compute the difference between the and the for every observation and then average over all observations, referred to as the G-computation method . Although counterfactuals may not be completely relevant to gene microarray data, thinking about the data in this way is very helpful for us to assess the importance of a gene. Our VIM definition uses the concepts of counterfactuals and the estimation framework is built on the methodology of targeted maximum likelihood estimation (TMLE) . By tailoring this recently developed technique specifically to gene expression data, we hope to introduce to the community an alternative strategy to carry out gene selection in addition to current methods. Our approach takes the advantage of prediction power of learning algorithms while targeting at the individual importance of each variable. Its mathematical property has been studied in , and we will focus on its application. In brief, our approach consists of two-stages. In the first stage, we predict the outcome given all genes. In the second stage, we improve the first stage by modeling the mechanism between an individual gene and its confounding variables. Both stages can be very flexible ranging from using univariate analysis to refined learning algorithms. When machine learning algorithms are used, we have the flexibility to determine how to make predictions without restricting ourselves to explicit models and distributions. In the meanwhile, as in the case of the univariate analysis, we return to a simple and well interpretable measure of the importance of each gene. This importance measurement is derived in the presence of the confounding variables of a gene, and hence can help to exploit the redundant information among correlated genes. It is generally also more stable than the variable importance produced by machine learning algorithms. In addition, our approach provides a simple way for statistical inference based on asymptotic theories, and is well suited for the exploratory analysis of microarray data.
where f(W) is a function of W. With this parameterization, we have Ψ(a) = βa. We can then view β as an index of the VIM of A. In the above model, the only assumption we make is the linearity of A. The definition of the VIM of A is closely related to the definition of the counterfactual effect in causal inference . Although β can not be directly interpreted as an causal effect without proper assumptions , it serves well as a surrogate of the magnitude of the causal relationship between the outcome and a gene. The motivation of this parameterization is that by selecting more causally related genes, the resulting prediction function will be better generalized to new experiments with the same causal relation between the outcome Y and A, but a different joint distribution of W. If in a next experiment, the technology or the sampling population is somewhat different, but the causal mechanism is still the same, then a prediction function that uses the correlates of the true causal variables will perform poorly while a prediction function using the true causal variables will still perform nicely. This idea will be illustrated in our simulations.
Our goal is to estimate β. In , this estimation problem was addressed in the framework of targeted maximum likelihood estimation (TMLE). TMLE is an estimating equation and efficient estimation theory based methodology , and is particularly useful when it comes to semiparametric models. Estimators from the traditional method such as MLE perform well for parametric models, however, they are generally biased relative to their variances especially when the model space is large. This is because the MLE focuses on doing a good job on the estimation of the whole density rather than on the parameter itself. TMLE is designed to achieve an optimal trade-off between the bias and the variance of the estimator. It uses an MLE framework, but instead of estimating the overall density, TMLE targets on the parameter of interest and produces estimators minimally affected by changes of the nuisance parameters in a model. In Additional File 1 we provide a brief overview of this methodology with a demo simulation example. The formal mathematical formulation of TMLE can be found in the original paper by van der Laan and Rubin . The implementation of TMLE to estimate β is fairly simple and consists of two stages. First, we estimate E(Y |A, W) without any parametric restriction. We then regress the residual of Y and E(Y |A = 0, W) onto βA to conform with our semiparametric regression model. This will yield an initial estimator of β and fitted values of E(Y |A, W), denoted by and the . In the second stage, we update these initial estimates in a direction targeted at β. This involves regressing the residuals of Y and the fitted on the clever covariate A - E(A|W). The E(A|W) evaluates the confounding of A with W, and we name it the "gene confounding mechanism". It needs to be estimated if unknown. Let us denote the coefficient before the clever covariate as ε. The updated TMLE estimate of β is , where ε n is the estimated value of ε. The variance estimate of β can be computed from its efficient influence curve. Below is a step-by-step implementation of our algorithm, and we refer to it as the TMLE-VIM procedure.
Obtain the initial estimator and . Use your favorite algorithm here, for example, linear regression, LASSO, Random Forest, etc.
Obtain the g n (W) estimate for the gene confounding mechanism E(A|W ). As in the case of Q n (0), a broad spectrum of algorithms can be used. In this paper, we use LASSO (in the GLMNET R package) for its optimal speed.
- 3.Compute the "clever covariate":
Fit regression .
- 5.Update the initial estimate with
- 6.Compute the variance estimate for according to its efficient influence curve:
- 7.Construct the test statistic:
T(A) follows the standard Gaussian distribution under the null hypothesis β = 0 when the sample size n goes to infinity.
The TMLE estimator is a consistent estimator of β when either the or the g n (W ) is consistent. When the is consistent, it is also asymptotically efficient. The derivations of the clever covariate, the efficient influence curve, the TMLE estimate and its mathematical properties can be found in  and . Upon the construction of the test statistic, a p-value can be calculated for the adjusted marginal effect of A and used as an index of the variable importance.
In the application to dimension reduction, for each variable in the dataset, we compute a TMLE-VIM p-value. We then reduce our variable space based on these p-values. There are two notions. First, in principle, a separate initial estimator should be fitted for every gene A by forcing A as a term in the algorithm used. This can become quite time consuming. To solve the problem, instead of estimating E(Y |A, W) for each A, we obtain a grand estimate G n (V) for E(Y |V ). Here V represents all variables in the dataset. Then for every A in V, we carry out the regression Y ~ βA with the offset G n (A = 0) to get β n (0) and . Second, when obtaining the g n (W), we want to be attentive to how closely W is correlated with A. The independence between W and A results in zero adjustment to the initial estimator, while a complete association causes β to be unidentifiable. A simple option is to use all variables less than a pre-defined correlation with A as W. In , they authors suggest 0.7 as a conservative threshold. Instead of applying a universal cutoff, we can also set individualized correlation threshold for each A. Below we provide a data adaptive procedure to do it. One first defines a sequence of correlation cutoffs δ. For each choice of δ, one computes the corresponding TMLE p-value for A. One then sets a p-value threshold λ, and chooses the maximum δ that has produced a p-value less than λ. The degree of the protection is determined with the value of λ. In general, the smaller the λ is, the more the protection. The value of λ can be either fixed a priori or chosen by cross validation. We refer to it as the TMLE-VIM(λ) procedure. It allows us to adjust for the confounding in the dataset adaptively and flexibly, and protect the algorithm against the harm from high correlations among variables. It works best when many variables are closely correlated in a complex way. However, it does require more time to run, especially when λ needs to be chosen by cross validation. In many cases, a universal cutoff of 0.7 will work fine. In Additional File 1 we provide the mathematical formulation of the TMLE-VIM(λ) procedure. Once we have all the variables ranked by their p-values. the candidate list can be truncated by either applying a p-value threshold or taking the top k ranked variables. Both of them are sound practices. In our simulations and data analysis, we usually truncate the list at a p-value threshold 0.05.
Results and Discussion
Two prediction algorithms, LASSO and D/S/A (Deletion/Substitution/Addition) , were used. D/S/A searches through the variable space and selects the best subset of covariates by minimizing the cross validated residual sum of squares. In our simulations, LASSO and D/S/A predictions are often similar. We used D/S/A in simulation I as it provides convenience to count what variables are included in the prediction model. LASSO was used in simulation II for its faster speed. We also used multivariate linear regression (MVR) as a comparison to machine learning algorithms when applicable.
where j indexes the j-th A, and e i is a normal error with mean 0 and variance . Each A j was correlated with m W s, and hence the total number of W s is m w = 25 m. A j and its associated W s were jointly sampled from a multivariate Gaussian distribution with mean 0 and variance-covariance matrix S, where S is a correlation matrix with an exchangeable correlation coefficient ρ. This simulation scheme resulted in 25 independent clusters of covariates. Within each cluster, the covariates are correlated at level ρ.
Simulations were run for combinations of:
m = (10, 20) corresponding to m w = (250, 500);
σ e = (1, 5, 10);
and ρ = (0.1, 0.3, 0.5, 0.7, 0.9).
For each combination, we simulated a training set of 500 data points and a testing set of 5000 data points. The training set was used to obtain the prediction model while the testing set was used to calculate the L2 risk. We also calculated a cross-examined L2 risk using a testing set with a ρ other than that of the training set. This is to demonstrate that by identifying more causally related variables, TMLE-VIM is robust to the change of the joint distribution among the covariates A s and W s. In specific, for each prediction model obtained from a training set, we calculated the L 2 risk on the testing set generated with ρ = 0.1 regardless of what ρ was used to generate the training set. As a benchmark, we also used univariate regression in parallel with TMLE-VIM to reduce the dimensionality of the dataset, denoted with UR-VIM. Once the variable importance was calculated, we cut short the variable list using a p-value threshold 0.05. Each combination was replicated 10 times and results took the average.
The simulation I results
m w = 250
m w = 500
0.2341 ; 0.2251
0.2436 ; 0.2351
0.4035 ; 0.3784
0.4230 ; 0.3943
0.2202 ; 0.2297
0.2231 ; 0.2247
0.2341 ; 0.2307
0.2027 ; 0.1975
0.2425 ; 0.3115
0.1169 ; 0.1285
0.4883 ; 0.5959
0.1268 ; 0.1217
0.3599 ; 0.5872
0.1307 ; 0.2545
0.8001 ; 0.9093
0.1740 ; 0.2976
0.2262 ; 0.7248
-0.1364 ; 0.2805
0.9390 ; 0.9885
-0.5498 ; 0.2657
R r = (UR-VIM risk - TMLE-VIM risk)/ UR risk: the proportion of the risk reduction of TMLE-VIM relative to the UR-VIM risk. It measures by how much TMLE-VIM outperforms UR-VIM. The bigger the number, the more the advantage.
R A = TMLE-VIM N A / UR-VIM N A : the ratio of the number of A s (N A ) in the TMLE-VIM list to the number of A s in the UR-VIM list.
R W = TMLE-VIM N W / UR-VIM N W : the ratio of the number of Ws (N W ) in the TMLE-VIM list to the number of W s in the UR-VIM list.
RRDSA = TMLE-VIM P A / UR-VIM P A : the ratio of the proportion of A s (P A ) in the final D/S/A prediction model resulted from the TMLE-VIM procedure to that from the UR-VIM. It measures the relative chance of arriving at a truly associated variable in the final model through the path of TMLE-VIM, referenced to the UR-VIM.
The R r was calculated on two different testing sets. One is the testing set generated with the same ρ as the corresponding training set, and we refer it to "testing set (a)"; the other is the testing set generated with ρ = 0.1, and we refer it to "testing set (b)". Testing set (a) shares the same correlation structure as the training set, while in testing set (b) all the variables are essentially independent of each other. Testing set (b) is a simple representation of the scenario that when a new experiment is conducted the overall joint distribution of the covariates changes while the causal mechanism remains the same. In Table 1, the bold R r was calculated on testing set (a), and the Italic R r was on testing set (b). We make a few points here about Table 1:
The proportion of the risk reduction (R r ) of the TMLE-VIM relative to the UR-VIM is typically more than 20% for the MVR prediction and 10% for the D/S/A prediction. In some cases, the risk reduction of the MVR can be very significant. For example, when m w = 500 and ρ = 0.7, the TMLE-VIM risk is close to only half of the UR-VIM risk. TMLE-VIM tends to deliver more advantages when m w = 500 than when m w = 250. When the correlation coefficient ρ increases, the TMLE-VIM performs increasingly better than the UR-VIM for the MVR prediction. For the D/S/A prediction, small or large ρ s seem to benefit most from the TMLE-VIM. For intermediate ρ, the benefit is still there but reduced. We believe that how much the risk can be reduced by the TMLE-VIM is a combination of factors such as the number of A s and W s in the reduced candidate list, the correlation structures among covariates and the internal optimization procedures of D/S/A. The advantage of the TMLE-VIM over the UR-VIM does seem to be more significant on the testing set (b) than the testing set (a), in support of our hypothesis that by identifying more causally related variables the TMLE-VIM results generalize better to new experiments.
Most R A values are slightly higher than 1 while the R W values are much smaller. This indicates that on average, in the TMLE-VIM list, the number of correctly identified A s is slightly higher than that in the UR-VIM list, while the number of falsely associated Ws is much less. It is especially the case when the correlations are high among variables. The low counts of false positives is a major contributing factor that the prediction made on the TMLE-VIM candidate list is better than that on the UR-VIM.
As to the number of A s that are finally made into the D/S/A prediction model, the TMLE-VIM in most cases displays a slight advantage over the UR-VIM. A closer look reveals that the variables included in the D/S/A model only differs by one or two between the TMLE-VIM and the UR-VIM. But the prediction risk has a measurable difference. This probably implies that every single variable counts in making good predictions in these simulations.
When ρ = 0.9, the situation seems to be losing its track. The TMLE-VIM did worse than the UR-VIM in terms of correctly identified variables as well as the prediction risk of the testing set (a). Considering the high correlations among variables, this could possibly be attributed to the overfitting in the g n (W ). Indeed, in , the authors showed that TMLE deteriorates when adjusting for variables with correlation coefficients beyond 0.7. However, the RR DSA indicates that the chance of including a correct variable in the final D/S/A model based on the TMLE-VIM list is higher than that on the UR-VIM. Further looking into the data, we found out that the number of A s that made into the D/S/A model from the TMLE-VIM list is actually greater than that from the UR-VIM, while the number of W s is much less. Henceforth, the D/S/A model built on the TMLE-VIM list is closer to truth, but somehow its prediction is worse than the model built on the UR-VIM list. This seems to suggest that when provided with the UR-VIM list, the D/S/A has offset its model for the missed A s from highly correlated W s, while for the TMLE-VIM, this can not be done since there are not many W s in the list. It is the same reason that the UR-VIM underperforms the TMLE-VIM on the testing set (b) when those surrogates of A s were lost. For the MVR, although the TMLE-VIM shows a dominant advantage over the UR-VIM with respect to the prediction accuracy, the TMLE-VIM only identified 77% (when m w = 250) and 57% (when m w = 500) of the A s identified by the UR-VIM. The better prediction is merely due to the fact that the MVR breaks down when too many variables entered the model. This is particularly the case when m w = 500.
Simulation II examines the TMLE-VIM on larger-scale datasets with much more complex correlation structures. The simulation consists of 500 samples and 1000 variables. We used a correlation matrix derived from the top 800 genes in a real dataset published in . For these genes, the median absolute correlation coefficient was centered at 0.26, the 1st/3rd quartile being 0.16/0.37, and the maximum as high as 0.9977. Hence, simulation II tried to mimic the correlation structure in this real data set. The outcome Y was generated from two different models using 20 A s. One is a linear model, and the other is polynomial.
The simulation II results (p-value)
The simulation II results (top 100)
We also carried out the TMLE-VIM(λ) procedure with LASSO as the initial estimator, allowing the data select the correlation cutoff for variables to be adjusted in the g n (W ). Results are also reported in Table 2 and Table 3. TMLE-VIM(λ) identified more A s but also more W s, and the prediction accuracy is only slightly improved. On the other hand, the correlations among the selected top 100 variables are quite small. It seems by data adaptively adjusting for the correlation levels in the g n (W), TMLE-VIM(λ) returns a more independent set of genes. The actual risks and standard errors are contained in Additional File 2.
Breast cancer patients are often put on chemotherapy after the surgical removal of the tumor. However not all patients will respond to chemotherapy, and proper guidance for selecting the optimal regimen is needed. Gene expression data have the potential for such predictions, as studied in . The dataset from  contains the gene expression profiling on 22283 genes for 133 breast cancer patients. The outcome is the pathological complete response (pCR). This is a binary response associated with long-term cancer free survival. There are also 13 clinical variables collected in the dataset including the ER (estrogen receptor) status, which is a very significant clinical indicator for chemotherapy response.
The goal of the study is to select a set of genes that best predict the clinical response pCR. The first step is to reduce the number of genes worth of consideration, and we applied both UR-VIM and TMLE-VIM (with Q(0) = UR and Q(0) = LASSO) for this purpose. For the TMLE-VIM(Q(0) = LASSO), the was estimated by LASSO using the top 5000 ranked genes. We then took all the genes with the FDR-adjusted p-values less than 0.005 , as suggested in the original paper, and upon them we built a predictor using the Random Forest (tuning parameters mtry = number of variables/3, ntree = 3000 and nodesize = 1). The clinical covariates were treated in the same way as genes. To prevent the algorithm from breaking down, we only adjusted for the confounder with correlation coefficients less than 0.7 with A in the g n (W ). We carried out a 10-fold honest cross validation. We divided the dataset into 10 subsets. Each subset was regarded as a validation set and the rest as the training set. We reperformed the entire analysis, i.e. VIM calculation → dimension reduction → Random Forest classifier, on all 10 training sets and predicted the outcome of the validation set using the classifier built on the training samples. We can then use these cross validated predictions to assess the true classification accuracy of our algorithm.
The analysis result of the breast cancer dataset
Num. of genes in the candidate list
C.V. classification accuracy
Corr. level among the top 100 genes
The TMLE-VIM(Q(0) = UR) and the TMLE-VIM(Q(0) = LASSO) results are quite similar to each other regardless of the adequate difference between the initial estimators. It seems the modeling of the g n (W ) had played a significant role and steered away the initial univariate estimates. Further investigation found out that genes in the UR-VIM list are highly correlated with the clinical indicator ER status, while the TMLE-VIM genes are not. Consequently, the TMLE-VIM genes are less correlated to each other than the UR-VIM genes. Looking at the first 100 ranked genes, the absolute median of the correlation coefficients for the UR-VIM is 0.43, while for the TMLE-VIM, it is about half of that number. Although the input variables to the Random Forest are different, The cross validated (CV) classification accuracy are quite similar among these three methods. We also passed all 22,000 genes to Random Forest and looked at its variable importance measurement. The Random Forest VIM (RF-VIM) is more similar to the UR-VIM: about 50% of them overlap but only a few overlap with the TMLE-VIM. The RF-VIM genes are also highly correlated with the ER status, albeit the less severity than the UR-VIM. Its OOB classification accuracy (0.7669) is comparable with all three other methods.
In summary, the UR-VIM and RF-VIM seemed to have identified genes that are strong predictors of the clinical variable ER status. The ER status is a strong indicator of the outcome pCR. Hence, the final prediction accuracy still seems quite good. The TMLE-VIM has identified a list of genes of which a small proportion is strong predictors of ER status and others are not associated with the ER status. Its prediction accuracy is slightly better than that of the UR-VIM and RF-VIM.
We have shown in this paper with extensive simulations that the TMLE based variable importance measurement can be incorporated into a dimension reduction procedure to improve the quality of the list of the candidate variables. It requires an initial estimator and a gene confounding mechanism estimate g n (W). A consistent ensures the consistency and the efficiency of the TMLE estimate. When is not consistent, a correct specification of g n (W) can still produce consistent estimates while that estimate will not be efficient any more. We generally recommend to do as good a job as we can on obtaining the , as a better means both a smaller bias and a smaller variance. Nevertheless, algorithms as simple as univariate regression are also valid choices, and in this case, we will rely solely on the goodness of g n (W ). The computation of g n (W ) directly affects the speed of the TMLE-VIM, as it has to be redone for every variable. Hence, one may want to choose an approach that is reasonably fast. In our study, we chose the GLMNET R Package as our primary tool to get g n (W ), and it worked very well. In practice, one needs to balance the resources used for the initial estimator and the gene confounding mechanism. With a proper design of the two estimating stages, TMLE-VIM is a fairly fast procedure. It is also worth mentioning that the TMLE-VIM can sometimes be sensitive to the overfitting in the , and hence, caution needs to be exercised when choosing an aggressive algorithm.
A popular dimension reduction approach is the principle component analysis (PCA). The PCA computation does not involve the outcome, and so it could be less powerful when prediction is the primary goal. Its output is a linear combination of all the genes. Though not a gene selection approach, we still carried it out on our simulation I data as an interesting comparison to our approach. PCA demonstrates an intermediate performance with respect to the UR-VIM and the TMLE-VIM on small p-value cutoffs. This means a few top components carry all the prediction power. When the p-value cutoff is increased, and more components enter the candidate list, its results became quite unsatisfying. When the correlation structure changes among the genes, PCA has done a poor predicting job. The PCA results are contained in Additional File 3.
Usually, the reduced set of variables will serve as the input of a prediction algorithm to build a model. Such algorithms used in this article include MVR, LASSO, and D/S/A. We have noticed that in most of our simulations, the MVR prediction often achieves a similar risk as LASSO and D/S/A on the TMLE-VIM reduced set of variables. It suggests that further variable selection may not be necessary for the TMLE-VIM candidate list, and we can use simpler algorithms to get a good prediction. In fact, the TMLE-VIM can go beyond the scope of dimension reduction. It can be iteratively applied to the data until it converges to a list of several variables that are most likely to be causal to the outcome. In this case, one may want to use the Super Learner  as the prediction algorithm, which works more effectively with the TMLE-VIM. The Super Learner is an ensemble learner that combines predictions from multiple candidate learners with optimal weights. It has been shown in  that the Super Learner performs asymptotically equal to or better than any of its candidate learners. The Super Learner allows the data to objectively blend results from different algorithms rather than relying on a single algorithm chosen subjectively by an analyst. Hence it enjoys a greater flexibility to explore the model space and usually produces reasonable predictions consistently across a wide variety of datasets, and serves as a very good prediction algorithm for the TMLE-VIM. On the other hand, it is also more computationally demanding.
TMLE-VIM is a quite general approach. Besides gene expression data, TMLE-VIM can also be applied to genetic mapping problems. The genome-wide association studies (GWAS) can involve more than a million of genetic markers. In this case, only the univariate analysis seems to be feasible of ranking every marker. With the TMLE-VIM procedure, we can run more complex algorithms on a subset of top ranked markers, taking it as the initial estimator, and then evaluate every single marker. The variable importance of each marker is thus obtained through a multi-marker approach and being adjusted for its confounder. However, the GWAS in human beings is usually case-control data, and the current TMLE-VIM needs to be extended to accommodate such outcomes.
The authors want to thank Cathy Tugulus for sharing her codes and her helpful comments on this work. The authors also thank the reviewers for their precious appraisal of the earlier version of this manuscript. This work was by NIH R01 AI074345. The authors declare no conflicts of interest.
- West M, Blanchette C, Dressman H, Huang E, Ishida S, Spang R, Zuzan H, Olson JA, Marks J, J N: Predicting the Clinical Status of Human Breast Cancer using Gene Expression Profiles. Proc Natl Acad Sci USA 2001, 98: 11462–11467. 10.1073/pnas.201162998PubMed CentralView ArticlePubMedGoogle Scholar
- Dudoit S, Fridlyand J, Speed T: Comparison of discrimination methods for the classification of tumors using gene expression data. JASA 2002, 97: 77–87.View ArticleGoogle Scholar
- Hedenfalk I, Duggan D, Chen Y, Radmacher M, Bittner M, Simon R, Meltzer P, Gusterson B, Esteller M, Rafield M, Yakhini Z, A BD, Dougherty E, Kononen J, Bubendorf L, Fehrle W, Pittaluga S, Gruvberger D, Loman N, Johannsson O, Olsson H, Wilfond B, Sauter G, Kallioniemi O, Borg A, Trent J: Gene expression profiles in hereditary breast cancer. New England Journal of Medicine 2001, 244: 539–548.View ArticleGoogle Scholar
- Dettling M, Buhlmann P: Boosting for tumor classification with gene expression data. Bioinformatics 2003, 19: 1061–1069. 10.1093/bioinformatics/btf867View ArticlePubMedGoogle Scholar
- Ghosh D: Singular value decomposition regression modeling for classification of tumors from microarray experiments. Proceedings of the Pacific Symposium on Biocomputing 2002, 18–29.Google Scholar
- Nguyen DV, Rocke DM: Tumor classification by partial least squares using microarray gene expression data. Bioinformatics 2002, 18: 39–50. 10.1093/bioinformatics/18.1.39View ArticlePubMedGoogle Scholar
- Nguyen DV, Rocke DM: Multi-class cancer classification via partial least squares with gene expression profiles. Bioinformatics 2002, 18: 1216–1226. 10.1093/bioinformatics/18.9.1216View ArticlePubMedGoogle Scholar
- Huang X, W P: Linear regression and two-class classification with gene expression data. Bioinformatics 2003, 19: 2072–2978. 10.1093/bioinformatics/btg283View ArticlePubMedGoogle Scholar
- Boulesteix A: PLS Dimension reduction for classification with microarray data. Statistical applications in genetics and molecular biology 2004, 3: 1–33.View ArticleGoogle Scholar
- R T: Regression shrinkage and selection via the lasso. J Royal Statist Soc B 1996, 58: 267–288.Google Scholar
- Efron B, Hastie T, Johnstone I, R T: Least angle regression. Annals of Statistics 2004, 32: 407–499. 10.1214/009053604000000067View ArticleGoogle Scholar
- L B: Random forests. Machine Learning 2001, 45: 5–32. 10.1023/A:1010933404324View ArticleGoogle Scholar
- Dai JJ, Lieu L, Rocke D: Dimension Reduction for Classification with Gene Expression Microarray Data. Statistical Applications in Genetics and Molecular Biology 2006, 5: Article 6.View ArticleGoogle Scholar
- Strobl C, Boulesteix AL, Zeileis A, Hothorn : Bias in random forest variable importance measures: illustrations, sources and a solution. BMC Bioinformatics 2007, 8: 25. 10.1186/1471-2105-8-25PubMed CentralView ArticlePubMedGoogle Scholar
- Strobl C, Boulesteix AL, Kneib T, Augustin T, Zeileis A: Conditional variable importance for random forests. BMC Bioinformatics 2008, 9: 307. 10.1186/1471-2105-9-307PubMed CentralView ArticlePubMedGoogle Scholar
- Rosset S, Zhu J, Hastie T: Boosting as a regularized path to a maximum margin classifier. Journal of Machine Learning Research 2004, 5: 941–973.Google Scholar
- Friedman J, Hastie T, Tibshirani R: Regularization paths for generalized linear models via coordinate descent. Journal of Statistical Software 2010., 33:Google Scholar
- Yu Z, van der Laan MJ: Measuring treatment effects using semiparametric models. U.C. Berkeley Division of Biostatistics Working Paper Series 2003. [http://www.bepress.com/ucbbiostat/paper136]Google Scholar
- van der Laan MJ: Statistical inference for variable importance. Int J Biostat 2006, 2: Article 2.Google Scholar
- Robins JM: A new approach to causal inference in mortality studies with sustained exposure periods - application to control of the healthy worker survivor effect. Mathematical Modelling 1986, 7: 1393–1512. 10.1016/0270-0255(86)90088-6View ArticleGoogle Scholar
- van der Laan MJ, Rubin DB: Targeted maximum likelihood learning. Int J Biostat 2006, 2: Article 11.Google Scholar
- Tuglus C, van der Laan MJ: Targeted methods for biomarker discovery, the search for a standard. U.C. Berkeley Division of Biostatistics Working Paper Series 2008. [http://www.bepress.com/ucbbiostat/paper233]Google Scholar
- Rubin DB: Estimating Causal Effects of Treatments in Randomized and Nonrandomized Studies. Journal of Educational Psychology 1974, 66: 688–701.View ArticleGoogle Scholar
- Bickel PJ, Klaassen CAJ, Ritove Y, Wellner JA: Efficient and adaptive estimation for semiparametric models. Baltimore: The Johns Hopkins University Press; 1993.Google Scholar
- Sinisi SE, van der Laan MJ: Deletion/Substitution/Addition algorithm in learning with applications in genomics. Statistical Applications in Genetics and Molecular Biology 2004, 3: Article 18.View ArticleGoogle Scholar
- Hess KR, Anderson K, Symmans WF, Valero V, Ibrahim N, Mejia JA, Booser D, Theriault RL, Buzdar AU, Dempsey PJ, Rouzier R, Sneige N, Ross JS, Vidaurre T, Gomez HL, Hortobagyi GN, Pusztai L: predictor of sensitivity to preoperative chemotherapy with paclitaxel and fluorouracil, doxorubicin, and cyclophosphamide in breast cancer. J Clin Oncol 2006, 24: 4236–4244. 10.1200/JCO.2006.05.6861View ArticlePubMedGoogle Scholar
- Benjamini Y, Hochberg Y: Controlling the false discovery rate: a practical and powerful approach to multiple testing. J Royal Statist Soc B 1995, 57: 289–300.Google Scholar
- van der Laan MJ, Polley EC, Hubbard AE: Super Learner. Statistical Applications in Genetics and Molecular Biology 2007, 6: Article 25.View ArticleGoogle Scholar
- van der Laan MJ, Dudoit S, van der Vaart AW: The cross-validated adaptive epsilon-net estimator. Statistics and Decisions 2006, 24: 373–395. 10.1524/stnd.2006.24.3.373Google Scholar
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/2.0), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.