Occupancy classification of position weight matrix-inferred Transcription Factor Binding Sites

Background: Computational prediction of Transcription Factor Binding Sites (TFBS) from sequence data alone is difficult and error-prone. Machine learning techniques utilizing additional environmental information about a predicted binding site (such as distances from the site to particular chromatin features) to determine its occupancy/functionality class show promise as methods to achieve more accurate prediction of true TFBS in silico. We evaluate the Bayesian Network (BN) and Support Vector Machine (SVM) machine learning techniques on four distinct TFBS data sets and analyze their performance. We describe the features that are most useful for classification and contrast and compare these feature sets between the factors. Results: Our results demonstrate good performance of classifiers both on TFBS for transcription factors used for initial training and for TFBS for other factors in cross-classification experiments. We find that distances to chromatin modifications (specifically, histone modification islands) as well as distances between such modifications to be effective predictors of TFBS occupancy, though the impact of individual predictors is largely TF specific. In our experiments, Bayesian network classifiers outperform SVM classifiers. Conclusions: Our results demonstrate good performance of machine learning techniques on the problem of occupancy classification, and demonstrate that effective classification can be achieved using distances to chromatin features. We additionally demonstrate that cross-classification of TFBS is possible, suggesting the possibility of constructing a generalizable occupancy classifier capable of handling TFBS for many different transcription factors.