Spectacle: Faster and more accurate chromatin state annotation using spectral learning

Recently, a wealth of epigenomic data has been generated by biochemical assays and next-generation sequencing (NGS) technologies. In particular, histone modification data generated by the ENCODE project and other large-scale projects show specific patterns associated with regulatory elements in the human genome. It is important to build a unified statistical model to decipher the patterns of multiple histone modifications in a cell type to annotate chromatin states such as transcription start sites, enhancers and transcribed regions rather than to map histone modifications individually to regulatory elements. Several genome-wide statistical models have been developed based on hidden Markov models (HMMs). These methods typically use the Expectation-Maximization (EM) algorithm to estimate the parameters of the model. Here we used spectral learning, a stateof-the-art parameter estimation algorithm in machine learning. We found that spectral learning plus a few (up to five) iterations of local optimization of the likelihood outperforms the standard EM algorithm. We also evaluated our software implementation called Spectacle on independent biological datasets and found that Spectacle annotated experimentally defined functional elements such as enhancers significantly better than a previous state-of-the-art method. Spectacle can be downloaded from https://github.com/jiminsong/Spectacle .