BackgroundComputational modeling transcription factor (TF) sequence specificity is an important research topic in regulatory genomics. A systematic comparison of 26 algorithms to learn TF-DNA binding specificity in in vitro protein-binding microarray (PBM) data was published recently, but the quality of those examined PBMs was not evaluated completely.ResultsHere, new quality-control parameters such as principal component analysis (PCA) ellipse is proposed to assess the data quality for either single or paired PBMs. Additionally, a biophysical model of TF-DNA interactions including adjacent dinucleotide interdependence was implemented in a new program - BayesPI2, where sparse Bayesian learning and relevance vector machine are used to predict unknown model parameters. Then, 66 mouse TFs from the DREAM5 challenge were classified into two groups (i.e. good vs. bad) based on the paired PBM quality-control parameters. Subsequently, computational methods to model TF sequence specificity were evaluated between the two groups.ConclusionResults indicate that both the algorithm performance and the predicted TF-binding energy-level of a motif are significantly influenced by PBM data quality, where poor PBM data quality is linked to specific protein domains (e.g. C2H2 DNA-binding domain). Especially, the new dinucleotide energy-dependent model (BayesPI2) offers great improvement in testing prediction accuracy over the simple energy-independent model, for at least 21% of analyzed the TFs.Electronic supplementary materialThe online version of this article (doi:10.1186/1471-2105-15-289) contains supplementary material, which is available to authorized users.