Separating DIP-Seq Signals through Independent Component Analysis - Arabidopsis thaliana

DNA immunoprecipitation (DIP-seq) is the primary method to study protein-DNA binding interactions in vitro by sequencing the fragments of DNA that bind to the target protein of choice. Because the DIP-seq protocol is prone to sequencing artifacts, the detection of “true” enrichment peaks is difficult. Current methods either employ coverage comparison to a control sample, which requires arbitrary thresholds to allow for error, or directly modelling the control sample using probabilistic models, which often fails since the artifacts are non-random.

To improve the detection of true enrichment in DIP-seq, we propose the usage of independent component analysis (ICA), an optimization algorithm for blind-source separation. The ICA algorithm detects enrichment by minimizing the mutual information between independent samples while also maximizing non-gaussianity. This primal-dual optimization doesn’t rely on arbitrary thresholds or inappropriate background models, leading to more robust results. We applied this ICA method to Arabidopsis thaliana chromosome 4 DNA during endoreplication and sequenced at 4 sequential time points. Results from ICA were compared to a log-fold enrichment analysis and validated by microarray. ICA returned fewer false positives than the log-fold while also having higher specificity.