Abstract Next generation high-throughput sequencing (NGS) is poised to replace array-based technologies as the experiment of choice for measuring RNA expression levels. Several groups have demonstrated the power of this new approach (RNA-seq), making significant and novel contributions and simultaneously proposing methodologies for the analysis of RNA-seq data. In a typical experiment, millions of short sequences (reads) are sampled from RNA extracts and mapped back to a reference genome. The number of reads mapping to each gene is used as proxy for its corresponding RNA concentration. A significant challenge in analyzing RNA expression of homologous genes is the large fraction of the reads that map to multiple locations in the reference genome. Currently, these reads are either dropped from the analysis, or a naive algorithm is used to estimate their underlying distribution. In this work, we present a rigorous alternative for handling the reads generated in an RNA-seq experiment within a probabilistic model for RNA-seq data; we develop maximum likelihood-based methods for estimating the model parameters. In contrast to previous methods, our model takes into account the fact that the DNA of the sequenced individual is not a perfect copy of the reference sequence. We show with both simulated and real RNA-seq data that our new method improves the accuracy and power of RNA-seq experiments.