MOTIVATION: RNA-Seq is a promising new technology for accurately measuring gene expression levels. Expression estimation with RNA-Seq requires the mapping of relatively short sequencing reads to a reference genome or transcript set. Because reads are generally shorter than transcripts from which they are derived, a single read may map to multiple genes and isoforms, complicating expression analyses. Previous computational methods either discard reads that map to multiple locations or allocate them to genes heuristically. RESULTS: We present a generative statistical model and associated inference methods that handle read mapping uncertainty in a principled manner. Through simulations parameterized by real RNA-Seq data, we show that our method is more accurate than previous methods. Our improved accuracy is the result of handling read mapping uncertainty with a statistical model and the estimation of gene expression levels as the sum of isoform expression levels. Unlike previous methods, our method is capable of modeling non-uniform read distributions. Simulations with our method indicate that a read length of 20-25 bases is optimal for gene-level expression estimation from mouse and maize RNA-Seq data when sequencing throughput is fixed.