Accurately modeling the DNA sequence preferences of transcription factors and predicting their genomic binding sites are key problems in regulatory genomics. These efforts have long been frustrated by the limited availability and accuracy of TF binding site motifs. Today, protein binding microarray (PBM) experiments and chromatin immunoprecipitation followed by sequencing (ChlP-seq) experiments are generating unprecedented high-resolution data on in vitro and in vivo TF binding. This paper will present a flexible new discriminative framework for representing and learning TF binding preferences using these massive data sets. Support vector regressions models were trained with a novel string kernel on PBM data to learn the mapping from probe sequences to binding intensities. Results confirm that discriminative sequence models presented here significantly outperform existing motif discovery algorithms, and it is found that ChlP-trained models greatly improved TF occupancy prediction over PBM-trained models, suggesting distinct in vivo sequence information. Finally, discriminative chromatin models using histone modification ChlP-seq data were trained and results show that models combining sequence and chromatin signatures strongly outperformed using either one alone. This work establishes effective new techniques for analyzing next generation sequencing data sets to study the interplay of chromatin and sequence in TF binding in the human genome.