Semi-supervised Classification of Breast Cancer Expression Profiles Using Neural Networks

In classification tasks of biological data, there are usually fewer labeled than unlabeled samples because labeling samples is costly or time-consuming. In addition, labeled data sets can be re-used in different contexts as additional unlabeled data sets. For example, when searching the Gene Expression Omnibus (GEO) repository for microarray data sets of drug sensitivity and resistance experiments, the largest one has 2,522 samples, but the median has only 12 samples. In machine learning in general, utilizing unlabeled data in classification tasks is called semi-supervised learning. Artificial neural networks can be used to pre-train on unlabeled data before fine-tuning via back-propagation with labeled data. Such artificial neural networks enabling deep learning have gained attention since around 2010, since when they have been among the best-performing algorithms in visual object recognition. We measured accuracies in the task of classifying tissue taken from breast cancer patients at reductive surgery as chemotherapy-resistant or -sensitive. Different data sets were constructed by subsampling from GEO data set GSE25055 and GSE25065. Using these data sets, we compared classification accuracy of the neural networks autoencoder, Restricted Boltzmann Machine, Deep Belief Network (DBN) and support vector machine (SVM), and Transductive SVM (TSVM). Training was done both in supervised and semi-supervised mode. For the neural networks, we tried several different network architectures. Smoothing the validation set accuracies obtained during training iterations to alleviate low sample numbers helped in model selection of the best classifier. We also investigated the effect of different normalization procedures on the classification accuracy. The data were normalized with either RMA or MAS5, followed by either no batch-effect correction or Combat batch-effect correction. Only MAS5 profited from added Combat batch-effect correction, but normalization with RMA alone yielded the best classification accuracy. We were particularly interested whether classification accuracies improve when adding unlabeled samples in semi-supervised learning. Overall, neural networks and support vector machines performed similar. We found a slight improvement of classification accuracy when the number of unlabeled samples presented to DBN and TSVM was increased to the maximal number of samples in our data sets. However, this effect was only observed when the learning algorithms were presented the expression values of all 22,283 genes, not just the 500 most variable genes