Identifying transcription factor target genes (TFTGs) is a vital step towards understanding regulatory mechanisms of gene expression. Methods for the de novo identification of TFTGs are generally based on screening for novel DNA binding sites. However, experimental screening of new binding sites is a technically challenging, laborious and time-consuming task, while computational methods still lack accuracy. We propose a novel systematic computational approach for predicting TFTGs directly on a genome scale. Utilizing gene co-expression data, we modeled the prediction problem as a 'yes' or 'no' classification task by converting biological sequences into novel reverse-complementary position-sensitive n-gram profiles and implemented the classifiers with support vector machines. Our approach does not necessarily predict new DNA binding sites, which other studies have shown to be difficult and inaccurate. We applied the proposed approach to predict auxin-response factor target genes from published Arabidopsis thaliana co-expression data and obtained satisfactory results. Using ten-fold cross validations, the area under curve value of the receiver operating characteristic reaches around 0.73.