Here, we predict gene expression from epigenetic features based on public data available through the Epigenome Roadmap Project. This rich new dataset includes samples from primary tissues, which to our knowledge have not previously been studied in this context. Specifically, we used computational machine learning algorithms on five histone modifications to predict gene expression in a variety of samples. Our models reveal a high predictive accuracy, especially in cell cultures, with predictive ability dependent on sample type and anatomy. The relative importance of each histone mark feature varied across samples. We localized each histone mark signal to its relevant region, revealing that chromatin state enrichment varies greatly between histone marks. Our results provide several novel insights into epigenetic regulation of transcription in new contexts.