The individual sample heterogeneity is one of the biggest obstacles in biomarker identification for complex diseases such as cancers. Current statistical models to identify differentially expressed genes between disease and control groups often overlook the substantial human sample heterogeneity. Meanwhile, traditional nonparametric tests lose detailed data information and sacrifice the analysis power, although they are distribution free and robust to heterogeneity. Here, we propose an empirical likelihood ratio test with a mean-variance relationship constraint (ELTSeq) for the differential expression analysis of RNA sequencing (RNA-seq). As a distribution-free nonparametric model, ELTSeq handles individual heterogeneity by estimating an empirical probability for each observation without making any assumption about read-count distribution. It also incorporates a constraint for the read-count overdispersion, which is widely observed in RNA-seq data. ELTSeq demonstrates a significant improvement over existing methods such as edgeR, DESeq, t-tests, Wilcoxon tests and the classic empirical likelihood-ratio test when handling heterogeneous groups. It will significantly advance the transcriptomics studies of cancers and other complex disease.