We introduce deep convolutional neural networks (CNNs) as powerful tools for regulatory genomics. Chromputer is a method for predicting chromatin state, chromatin marks and transcription factor (TF) binding from multiple genomic signals (including DNA sequence, chromatin accessibility and nucleosome positioning) to achieve high predictive accuracy, within and across different cell-types. DeepLIFT is the first comprehensive interpretation engine for deep neural networks in genomics and epigenomics that overcomes the limitations of commonly used approaches for interpreting neural network models in genomics, such as in-silico mutagenesis or filter nullification, that are computationally expensive and potentially provide misleading results. DeepLIFT analysis of Chromputer models reveals predictive features such as sequence motifs and TF footprints that are automatically learned from raw data and produces much cleaner and broader sequence motifs than traditional approaches, and enables deconvolving heterogeneity of regulatory sequence grammars.