Transcription Factor-Target Gene Mapping Enhanced By Integrating Motif Search, Function Annotation And Expression Data

Transcriptional regulation is essential for all eukaryotes. Defining regulatory networks, linking transcription factors (TF) to target genes, is of fundamental importance to biology. Developing predictive models for TF-target mapping is critical to test current understandings and to propose new hypotheses. Nevertheless, predicting target genes remains challenging because TF binding motifs are often short, degenerated and widely-spread. Moreover, the binding mechanisms are generally more complex than recognizing a specific sequence. Thus, apart from leveraging the motif identification accuracy, it may be helpful to integrate alternative knowledge depicting the relationship between the TF and the targets. Herein we developed an integrated TF-target mapping strategy and examined its performance. The candidate genes were predicted by a sum of three factors: the enrichment and quality of the putative motifs, identified via an optimized position weighted matrix (PWM) score over phylogenetically conserved promoter regions; a Gene Ontology-based semantic functional relevancy measure; and a causal relationship measure between the gene and the TF derived from expression profiles. The evaluation was conducted using 52 transcription factors covering most of the known, PWM-available TFs in higher eukaryotes, and their total of 1315 curated target genes. The integrated strategy achieved a considerably higher accuracy with an area under receiver operating characteristic curve (AUC) of 0.67, compared to the commonly-adopted method relying on solely a motif enrichment score (AUC of 0.56). In particular, optimizing the PWM score in phylogenetically conserved promoters increased both sensitivity and specificity; functional relevancy and causal correlation further lowered the false positives. The results illustrate the feasibility of integrating multiple knowledge sources to improve TF-target mapping. The presented strategy is readily scalable to genome-wide, and can be applied along with other inference tools to assist the regulatory network reengineering.