Original Data Set

The data used in this work are the time-series microarray data that are extracted from a large liver xenobiotic and pharmacological response database of Iconix Biosciences. The data are publicly available at GEO http://www.ncbi.nlm.nih.gov/geo under accession number GSE8858. The initial data set consists of 1695 individual animal studies and 5288 microarrays. GE Healthcare/Amersham Biosciences CodeLink UniSet Rat I Bioarray, layout EXP5280X2-584, layout EXP5280X2-613 and layout EXP5280X2-648 containing about 10000 probes was used to analyze the global gene expression in the livers of compound-treated rats. Only treatments with gene expression data of day 1, 3 and 5 were involved in our analysis, including 402 treatments with 306 compounds.

Data Construction

First, we get a list of 10399 common probe sets between GE Healthcare/Amersham Biosciences CodeLink UniSet Rat I Bioarray, layout EXP5280X2-584, layout EXP5280X2-613 and layout EXP5280X2-648. Secondly, the gene expression profiles of 402 treatments on day 1, 3 and 5 were obtained from corresponding 3563 microarrays by averaging the duplicated experiments. Then, the control probe sets and probe sets without GenBank Accession number were excluded. The probe sets with more than 30% missing value were also excluded. This yields a subset of 9852 probes. After probe filtering, the missing expression data were imputed using nearest neighbor averaging. Finally, we normalized the expression data of 402 treatments on day 1, 3 and 5 using quantile method.

Thus expression data of 9852 genes of each day (day 1, 3 or 5) were involved in our study, producing 9852*3 = 29556 features for each of the 402 samples. Each sample is to be allocated into the 34 categories listed in Table 1, with the allowance of multiple entries into the categories, using the 29556 features.

Minimum Redundancy Maximum Relevance Feature Selection

Minimum-Redundancy-Maximum-Relevance (mRMR) [28] is a widely used method for feature selection. The goal of mRMR is to select a feature subset that can best characterize the statistical property of a target classification variable, subject to the constraint that these features are mutually as dissimilar to each other as possible, but marginally as similar to the classification variable as possible.

The feature which has maximum relevance with the target variable and minimum redundancy within the features is defined as a “good” feature. Mutual information (MI) is defined to describe both relevance and redundancy:(1)Where and are two vectors; is the joint probabilistic density; and are the marginal probabilistic densities.

The whole vector set is defined as , The selected vector set with vectors is defined as , and the to-be-selected vector set with vectors is defined as . Relevance of a feature in can be calculated by Eq (2):(2)Here is a classification variable.

Redundancy of a feature in with all the features in can be calculated by Eq (3):(3)mRMR function maximize relevance and minimize redundancy by integrating Eq (2) and Eq (3):(4)After the pre-evaluation procedure, a feature set S is provided:(5)the feature index reflects the evaluations for feature. The feature which fits the Eq(4) better will be added to the set S earlier. For example, If a<b, f_a is considered to be better than f_b.

Prediction Model

With the mRMR selected features, Nearest Neighbor Algorithm (NNA) [29] is used to classify the data into the above mentioned categories. NNA allocates a new data into categories by comparing the features of the data with the features of those that have known categories. The similarity between two vectors p_x, p_y is defined as [25]:(6)where is the inner product of p_x and p_y, and ||p|| is the module of vector p. p_x and p_y are considered to be more similar if D(p_x,p_y) is smaller.

Traditionally, NNA chooses to classify the new pattern p_t into the class of its nearest neighbor which has the smallest D(p_n,p_t). That is:(7)where N represents the number of training samples.

Because this research is about multi-target classification i.e. a data can belong to more than one category, the prediction model needs to be adjusted to cope with the multi-target problem. In the prediction of multi-targets, if D(p_m,p_t)<D(p_n,p_t), it means that p_t is closer to p_m than to p_n. Thus we rank the predicted classes of each drug data as:(8)From Eq. (8), we can get a list with the most likely class (defined as order-1 response) to be in the first position, and the second likely class (defined as order-2 response) to be in the second position, and so on.

Jackknife Cross-Validation Method

Jackknife Cross-Validation Method [14], [18] is an effective and objective way to evaluate statistical predictions. Each sample in the data set is in turn knocked out and tested by the predictor trained by the other samples remaining in the data set. During the process, every sample is used not only for the training, but also for the testing. The prediction accuracy Q for overall samples was used to evaluate the performance of predictor:(9)where TP, TN, FP and FN stand for true positive, true negative, false positive and false negative, respectively.

Incremental Feature Selection (IFS)

mRMR only provides a list of features by sorting the features according to their importance to the prediction without telling how many fore features in the list should be selected. The fore features are selected by testing all possible feature sets, and choosing the feature set that achieves the best prediction rate. A possible feature subset S_i can be expressed by the following equation.(10)The initial feature subset is , and the last feature subset is which includes all the features. Jackknife test is then used to obtain the accurate prediction rates of all the feature subsets. The one that achieves the highest prediction accuracy is considered to be optimized feature set selected by IFS. We can plot a curve, called IFS curve, with index i as its x-axis and the overall accurate rate as its y-axis.

IFS Curves of the Drug Responses

Because a drug may have several pharmaceutical responses, Eq. 8 is used to rate all the available responses. We only take the first three responses for every drug. And more will be available if they are needed in a future research. The cumulated prediction accuracies of first one, two, and three responses using different number of features, are shown in Figure 1, evaluated by jackknife cross-validation test. The highest prediction accuracy of first order response was 63.9% with 141 features. The highest cumulated prediction accuracies of first two responses and first three responses were also achieved with these 141 features. The detailed information of the IFS procedure and these 141 features can be found in Table S1 and Table S2.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Figure 1. The IFS curve of first three responses prediction.

The order-1 response is the most possible response according to the prediction. The highest prediction accuracy of first order response was 63.9% with 141 features. The highest cumulated prediction accuracies of first two responses and first three responses were also achieved with these 141 features. The red color points represent the highest accuracy points of each kind of accuracy.

https://doi.org/10.1371/journal.pone.0008126.g001

IFS Feature Selection and the Prediction Accuracy

141 features are selected as the result according to the IFS curves. Using these 141 features, the highest prediction accuracy for the first order response is 63.9%, evaluated by jackknife cross-validation test. Unfortunately, the prediction accuracy is rather low, which might be due to the sparse data points in the high-dimensional feature space. More samples could be used in a future research to study how much the prediction accuracy is affected by the number of samples available for training and predicting the prediction model. And the biological relevance of these 141 features was explored by KEGG and GO category enrichment analysis.

The KEGG category enrichment analysis (see Table S3) shows that two of the 141 features, Cyp3a9 and Ephx1, involves in the pathway for the metabolism of xenobiotics by cytochrome P450. Cytochrome P450s (CYP), comprising a superfamily of heme-thiolate proteins, is the main metabolizing enzyme system for foreign compounds, including drugs, and has a primary role in organism protection against potential harmful assaults from the environment [30]. It is often used as biomarker to determine human exposure to environmental molecules or to predict the susceptibility to certain pathologies [31], [32].

The GO category enrichment analysis results (see Table S4, Table S5 and Table S6) show that many of these candidate biomarkers are involved in insulin signaling pathway. The insulin-mediated receptor tyrosine kinase (RTK) signaling pathways [33], [34] by downstream effectors such as phosphatidylinositol 3-kinase, mitogen activated protein kinase (MAPK), Akt/protein kinase B (PKB), mammalian target of rapamycin (mTOR), and the p70 ribosomal protein S6 kinase (p70S6 kinase) have been reviewed [35] in the regulation of drug metabolizing enzyme expression in response to insulin and growth factors. The term fatty acid metabolism, comprising genes such as fatty acid synthase, enoyl-CoA hydratase, acyl-CoA synthetase among others is also enriched. The liver is a major site for fatty acid and lipid metabolism, and several major classes of compounds appearing in the database (statins, fibrates, glitazones, estrogen receptor modulators and others) affect the lipid synthesis and degradation. Fatty acids are a major energy source and important constituents of membrane lipids, and they serve as cellular signaling molecules that play an important role in the etiology of the metabolic syndrome [36]. Some liver samples exhibited elevated triglyceride levels that were correlated with changes in the urinary associated with defective metabolism of fatty acids, confirmed by the in vitro experiments [37].