A structure-based approach to predicting in vitro transcription factor-DNA interaction

Summary form only given. Understanding the mechanism of transcriptional regulation remains to be an inspiring stage of molecular biology. Within the popular methods for modeling TFBS, position-specific weight matrix and k-mer based approaches have gained great success. However, both approaches fail to consider the structural properties of a binding site. Recently, a novel TFBS modeling and predicting approach is presented by Bauer et al. (2010), where the sequence-specific chemical and structural features of DNA are applied. However, the in vivo protein-DNA interactions observed in ChIP-chip assays, which were used in this study, are not necessarily direct, as some TFs tend to interact with DNAs extensively through other partners. Therefore, an evaluation on a proper in vitro dataset would be more appropriate to reveal the benefit of such physicochemical features in modeling TF-DNA interactions. Recently, in vitro protein-binding microarray experiment has greatly improved the understanding of transcription factor-DNA interaction. It is a high-throughput experiment used to measure the in vitro binding affinity of a given TF to the sequences on the probe array. Because typical confounding factors such as transcription co-factors present in ChIP-based experiments are eliminated, PBM data provide an excellent information source to develop structural models for TF-DNA interactions. On the other hand, directly mapping of the 3-mer or 4-mer based meta-features to the candidate DNA binding sequences as in their work may not reflect the TF-DNA binding nature, since a TFBS is usually an 8 to 12 base-pair. As a result, conventionally machine-learning algorithms, which rely on well-structured feature vector and label pairs, may not work well in modeling PBM data. In this paper we propose a novel approach to predicts in vitro transcription factor binding based on the structural properties of DNA using a so-called multiple-instance learning algorithm. Compared to conventional (single-instance based) learning algorithms, our multi-instance learning-based algorithm does not require the knowledge of the actual binding site within a candidate probe sequence, yet can still take full advantage of the physicochemical properties in modeling and predicting TF-DNA interactions. Evaluation on an in vitro protein binding microarray data of twenty mouse TFs shows that our new model performs significantly better than several k-mer or structure-based single-instance learning algorithms. It indicates that combining multi-instance learning and structural properties of DNA has promising potential for studying biological regulatory networks.