Effect of using varying negative examples in transcription factor binding site predictions

Identifying transcription factor binding sites computationally is a hard problem as it produces many false predictions. Combining the predictions from existing predictors can improve the overall predictions by using classification methods like Support Vector Machines (SVM). But conventional negative examples (that is, example of non-binding sites) in this type of problem are highly unreliable. In this study, we have used different types of negative examples. One class of the negative examples has been taken from far away from the promoter regions, where the occurrence of binding sites is very low, and another one has been produced by randomization. Thus we observed the effect of using different negative examples in predicting transcription factor binding sites in mouse. We have also devised a novel cross-validation technique for this type of biological problem.