IPLS-LDA: An Improved Partial Least Square Discriminant Analysis for Heterogeneous Transcriptomics and Metabolomics Data Analysis

Supervised machine learning (SML) is an approach that learns from training data with known category membership to predict the unlabeled test data. There are many SML approaches in the literature and most of them use a linear score to learn its classifier. However, these approaches fail to elucidate biodiversity from heterogeneous biomedical data. Therefore, their prediction accuracies become low. Partial Least Square Linear Discriminant Analysis (PLS-LDA) is widely used in gene expression (GE) and metabolomics datasets for predicting unlabelled test data. Nevertheless, it also does not consider the non-linearity and heterogeneity pattern of the datasets. Hence, in this study, an improved PLS-LDA (IPLS-LDA) was developed by capturing the heterogeneity of datasets through an unsupervised hierarchical clustering approach. In our approach a non-linear score was calculated by combining all the linear scores obtained from the clustering method. The performance of IPLS-LDA was investigated in a comparison with six frequently used SML methods (SVM, LDA, KNN, Naive Bayes, RF, PLS-LDA) using one simulation data, one colon cancer gene expression data (GED) and one lung cancer metabolomics datasets. The resultant IPLS-LDA predictor achieved accuracy 0.841 using 10-fold cross validation in colon cancer data and accuracy 0.727 from two independent metabolomics data analysis. In both the cases IPLS-LDA outperformed other SML predictors. The proposed algorithm has been implemented in an R package, Uplsda was given in the https://github.com/snotjanu/UplsLda.