Mix²: An Efficient Software Tool for Accurate Gene Isoform Quantification from RNA-Seq Data

Quantification of RNA transcripts with RNA-Seq is inaccurate due to positional fragment bias, which is not represented appropriately by current statistical models of RNA-Seq data. We developed the Mix² (rd. ”mixquare”) model, which uses a mixture of probability distributions to model the transcript specific positional fragment bias. The parameters of the Mix² model are efficiently trained with the Expectation Maximization (EM) algorithm resulting in simultaneous estimates of the transcript abundances and transcript specific positional biases. We compare the Mix² model to state-of-the-art RNA-Seq quantification methods eXpress, RSEM, PennSeq and Cufflinks on synthetic data and real RNA-Seq data from the Microarray Quality Control (MAQC) consortium. On the MAQC data the Mix² model achieves improved correlation to qPCR concentration measurements with an increase in R² value between 4% and 50%. The Mix² model yields furthermore repeatable concentration estimates across technical replicates with an increase in R² between 8% and 47% and leads to more accurate detection of differential expression with an increase in true positive rate between 74% and 378% for a false positive rate of 5%. We also use the Mix² model to investigate biases in the MAQC data. This reveals 5 dominant biases which deviate from the common assumption of a uniform fragment distribution. We developed a highly efficient implementation of the Mix² model in C++, which is slightly faster than RSEM and considerably faster than eXpress and Cufflinks by a factor of 13 and 70, respectively. The Mix² software is available for free for academic researchers at https://secure.lexogen.com/solo/products/ProductOption.aspx?ProdOptionID=1030.