Inference from binary gene expression data

Microarrays provide a practical method for measuring the mRNA abundances of thousands of genes in a single experiment. Analysing such large dimensional data is a challenge which attracts researchers from many different fields and machine learning is one of them. However, the biological properties of mRNA such as its low stability, measurements being taken from a population of cells rather than from a single cell, etc. should make researchers sceptical about the high numerical precision reported and thus the reproducibility of these measurements. In this study we explore data representation at lower numerical precision, down to binary (retaining only the information whether a gene is expressed or not), thereby improving the quality of inferences drawn from microarray studies. With binary representation, we propose a solution to reduce the effect of algorithmic choice in the pre-processing stages.First we compare the information loss if researchers made the inferences from quantized transcriptome data rather than the continuous values. Classification, clustering, periodicity detection and analysis of developmental time series data are considered here. Our results showed that there is not much information loss with binary data. Then, by focusing on the two most widely used inference tools, classification and clustering, we show that inferences drawn from transcriptome data can actually be improved with a metric suitable for binary data. This is explained with the uncertainties of the probe level data. We also show that binary transcriptome data can be used in cross-platform studies and when used with Tanimoto kernel, this increase the performance of inferences when compared to individual datasets. In the last part of this work we show that binary transcriptome data reduces the effect of algorithm choice for pre-processing raw data. While there are many different algorithms for pre-processing stages there are few guidelines for the users as to which one to choose. In many studies it has been shown that the choice of algorithms has significant impact on the overall results of microarray studies. Here we show in classification, that if transcriptome data is binarized after pre-processed with any combination of algorithms it has the effect of reducing the variability of the results and increasing the performance of the classifier simultaneously