Correlating gene expression levels with transcription factor binding sites facilitates identification of key transcription factors from transcriptome data.

Identification of key transcription factors from transcriptome data by correlating gene expression levels with transcription factor binding sites is important for transcriptome data analysis. In a typical scenario, we always set a threshold to filter the top ranked differentially expressed genes and top ranked transcription factor binding sites. However, correlation analysis of filtered data can often result in spurious correlations. In this study, we tested four methods for creating the gene expression inputs (ranked gene list) in the correlation analysis: star coordinate map transformation (START), expression differential score (ED), preferential expression measure (PEM), and the specificity measure (SPM). Then, Kendall's tau correlation statistical algorithms implementing the standard (STD), LINEAR, MIX-LINEAR, DENSITY-CURVE, and MIXED-DENSITY-CURVE weighting methods were used to identify key transcription factors. ED was identified as the optimal method for creating a ranked gene list from filtered expression data, which can address the "unable to detect negative correlation" fallacy presented by other methods. The MIXED-DENSITY-CURVE was the most sensitive for identifying transcription factors from the gene set and list in which only the top proportion was correlated. Ultimately, 644 transcription factor candidates were identified from the transcriptome data of 1,206 cell lines, six of which were validated by wet lab experiments. The Jinzer and Flaver software implementing these methods can be obtained from http://www.thua45/cn/flaver under a free academic license.