Background: Physical interactions between transcription factors (TFs) are necessary for forming regulatory protein complexes and thus play a crucial role in gene regulation. Currently, knowledge about the mechanisms of these TF interactions is incomplete and the number of known TF interactions is limited. Computational prediction of such interactions can help identify potential new TF interactions as well as contribute to better understanding the complex machinery involved in gene regulation. Methodology: We propose here such a method for the prediction of TF interactions. The method uses only the primary sequence information of the interacting TFs, resulting in a much greater simplicity of the prediction algorithm. Through an advanced feature selection process, we determined a subset of 97 model features that constitute the optimized model in the subset we considered. The model, based on quadratic discriminant analysis, achieves a prediction accuracy of 85.39% on a blind set of interactions. This result is achieved despite the selection for the negative data set of only those TF from the same type of proteins, i.e. TFs that function in the same cellular compartment (nucleus) and in the same type of molecular process (transcription initiation). Such selection poses significant challenges for developing models with high specificity, but at the same time better reflects real-world problems. Conclusions: The performance of our predictor compares well to those of much more complex approaches for predicting TF and general protein-protein interactions, particularly when taking the reduced complexity of model utilisation into account. © 2011 Schmeier et al.