A comparative study on feature selection methods for drug discovery

Feature selection is frequently used as a preprocessing step to machine learning. The removal of irrelevant and redundant information often improves the performance of learning algorithms. This paper is a comparative study of feature selection in drug discovery. The focus is on aggressive dimensionality reduction. Five methods were evaluated, including information gain, mutual information, a chi2-test, odds ratio, and GSS coefficient. Two well-known classification algorithms, Na?ve Bayesian and Support Vector Machine (SVM), were used to classify the chemical compounds. The results showed that Na?ve Bayesian benefited significantly from the feature selection, while SVM performed better when all features were used. In this experiment, information gain and chi2-test were most effective feature selection methods. Using information gain with a Na?ve Bayesian classifier, removal of up to 96% of the features yielded an improved classification accuracy measured by sensitivity. When information gain was used to select the features, SVM was much less sensitive to the reduction of feature space. The feature set size was reduced by 99%, while losing only a few percent in terms of sensitivity (from 58.7% to 52.5%) and specificity (from 98.4% to 97.2%). In contrast to information gain and chi2-test, mutual information had relatively poor performance due to its bias toward favoring rare features and its sensitivity to probability estimation errors.     

变量选择中的正交变换法及苯酚和苯胺类衍生物色谱比移值的预测

正交变换法是变量选择的一种可行方法,但该种方法非常依赖于正交变换过程中变量的排序,侧重比较了不同排序方法,其中,后退法可以得到较好的结果。文中采用此种方法对由苯酚及苯胺类化合物所衍生的变量进行了正交变换,并对上述化合物的色谱比移值进行了预测。同时,与前进选择法、后退剔除法和逐步回归法几种传统方法进行了比较,得到了有启示性的结果。     

A comparison of different chemometrics approaches for the robust classification of electronic nose data

Accurate detection of certain chemical vapours is important, as these may be diagnostic for the presence of weapons, drugs of misuse or disease. In order to achieve this, chemical sensors could be deployed remotely. However, the readout from such sensors is a multivariate pattern, and this needs to be interpreted robustly using powerful supervised learning methods. Therefore, in this study, we compared the classification accuracy of four pattern recognition algorithms which include linear discriminant analysis (LDA), partial least squares-discriminant analysis (PLS-DA), random forests (RF) and support vector machines (SVM) which employed four different kernels. For this purpose, we have used electronic nose (e-nose) sensor data (Wedge et al., Sensors Actuators B Chem 143:365–372, 2009). In order to allow direct comparison between our four different algorithms, we employed two model validation procedures based on either 10-fold cross-validation or bootstrapping. The results show that LDA (91.56 % accuracy) and SVM with a polynomial kernel (91.66 % accuracy) were very effective at analysing these e-nose data. These two models gave superior prediction accuracy, sensitivity and specificity in comparison to the other techniques employed. With respect to the e-nose sensor data studied here, our findings recommend that SVM with a polynomial kernel should be favoured as a classification method over the other statistical models that we assessed. SVM with non-linear kernels have the advantage that they can be used for classifying non-linear as well as linear mapping from analytical data space to multi-group classifications and would thus be a suitable algorithm for the analysis of most e-nose sensor data.     

Operon prediction based on SVM

The operon is a specific functional organization of genes found in bacterial genomes. Most genes within operons share common features. The support vector machine (SVM) approach is here used to predict operons at the genomic level. Four features were chosen as SVM input vectors: the intergenic distances, the number of common pathways, the number of conserved gene pairs and the mutual information of phylogenetic profiles. The analysis reveals that these common properties are indeed characteristic of the genes within operons and are different from that of non-operonic genes. Jackknife testing indicates that these input feature vectors, employed with RBF kernel SVM, achieve high accuracy. To validate the method, Escherichia coli K12 and Bacillus subtilis were taken as benchmark genomes of known operon structure, and the prediction results in both show that the SVM can detect operon genes in target genomes efficiently and offers a satisfactory balance between sensitivity and specificity.