Non-canonical protein-DNA interactions identified by ChIP are not artifacts

Background: ChIP-chip and ChIP-seq are widely used methods to map protein-DNA interactions on a genomic scale in vivo. Waldminghaus and Skarstad recently reported, in this journal, a modified method for ChIP-chip. Based on a comparison of our previously-published ChIP-chip data for Escherichia coli sigma(32) with their own data, Waldminghaus and Skarstad concluded that many of the sigma(32) targets identified in our earlier work are false positives. In particular, we identified many non-canonical sigma(32) targets that are located inside genes or are associated with genes that show no detectable regulation by sigma(32). Waldminghaus and Skarstad propose that such non-canonical sites are artifacts, identified due to flaws in the standard ChIP methodology. Waldminghaus and Skarstad suggest specific changes to the standard ChIP procedure that reportedly eliminate the claimed artifacts. Results: We reanalyzed our published ChIP-chip datasets for sigma(32) and the datasets generated by Waldminghaus and Skarstad to assess data quality and reproducibility. We also performed targeted ChIP/qPCR for sigma(32) and an unrelated transcription factor, AraC, using the standard ChIP method and the modified ChIP method proposed by Waldminghaus and Skarstad. Furthermore, we determined the association of core RNA polymerase with disputed sigma(32) promoters, with and without overexpression of sigma(32). We show that (i) our published sigma(32) ChIP-chip datasets have a consistently higher dynamic range than those of Waldminghaus and Skarstad, (ii) our published sigma(32) ChIP-chip datasets are highly reproducible, whereas those of Waldminghaus and Skarstad are not, (iii) non-canonical sigma(32) target regions are enriched in a sigma(32) ChIP in a heat shock-dependent manner, regardless of the ChIP method used, (iv) association of core RNA polymerase with some disputed sigma(32) target genes is induced by overexpression of sigma(32), (v) sigma(32) targets disputed by Waldminghaus and Skarstad are predominantly those that are most weakly bound, and (vi) the modifications to the ChIP method proposed by Waldminghaus and Skarstad reduce enrichment of all protein-bound genomic regions. Conclusions: The modifications to the ChIP-chip method suggested by Waldminghaus and Skarstad reduce rather than increase the quality of ChIP data. Hence, the non-canonical sigma(32) targets identified in our previous study are likely to be genuine. We propose that the failure of Waldminghaus and Skarstad to identify many of these sigma(32) targets is due predominantly to the lower data quality in their study. We conclude that surprising ChIP-chip results are not artifacts to be ignored, but rather indications that our understanding of DNA-binding proteins is incomplete.