Ritornello: High fidelity control-free chip-seq peak calling

Kelly Stanton, Jiaqi Jin, Sherman M. Weissman, Yuval KlugerbioRxiv 2015 · DOI

TL;DR

This paper introduces Ritornello, a novel ChIP-seq peak calling algorithm that significantly improves accuracy and reduces costs by eliminating the need for a matched DNA input control. By leveraging digital signal processing techniques, Ritornello effectively identifies true transcription factor binding sites while robustly filtering out artifacts, even in low-quality experimental data.

What Problem Does This Paper Address?

Chromatin immunoprecipitation followed by sequencing (ChIP-seq) is a crucial technique for identifying where proteins bind to DNA across the entire genome, which is essential for understanding gene regulation. However, existing ChIP-seq analysis methods face several challenges:

  • False Positives: Many algorithms generate "artifactual peaks" that don't represent actual binding events, leading to misinterpretations.
  • Poor Localization: Precisely pinpointing the exact binding sites can be difficult, especially when multiple binding events occur close together.
  • High Cost: A significant hurdle is the requirement for a matched total DNA input control sample for each ChIP-seq experiment. This control is used to distinguish true binding signals from background noise and artifacts, but it effectively doubles the sequencing cost and limits the number of experiments that can be performed.

These limitations hinder the widespread adoption and comprehensive application of ChIP-seq for biological discovery.

Methods

Ritornello tackles these challenges with a multi-pronged approach rooted in digital signal processing (DSP) and statistical modeling:

  • Fragment Length Distribution (FLD) Deconvolution: Instead of relying on a simple average fragment length, Ritornello derives a sample-specific empirical FLD from single-end reads. This is achieved through a deconvolution process that leverages cross-correlation analysis of reads on opposing DNA strands, particularly in regions with uniform read coverage. This detailed FLD is a key component for accurate peak prediction.
  • Matched Filtering: Ritornello employs a matched filter, a DSP technique, to efficiently identify candidate peak regions. This filter is designed to closely match the expected shape of true binding signals, as characterized by the derived FLD and a probabilistic model of read distribution around binding sites.
  • Artifact Removal via Cross-Correlation: A novel classifier is trained to distinguish true binding events from "read length artifacts." These artifacts, often arising from mapping issues or sequencing biases, exhibit characteristic patterns in cross-correlation analysis. Ritornello identifies these artifacts by looking for a "phantom peak" at the read length in the cross-correlation, which is absent in true binding events.
  • Probabilistic Modeling and Deconvolution of Binding Intensities: Ritornello models read coverage around candidate peaks using a generalized linear model (Poisson GLM). This allows for the deconvolution of binding intensities from overlapping peaks and background noise, enabling accurate quantification of binding at each site.
  • Likelihood Ratio Test for Significance: Candidate peaks are then tested for statistical significance using a likelihood ratio test, comparing a null model (no binding) to an alternative model (binding present).

Crucially, Ritornello's approach is designed to function without a matched DNA input control, addressing the cost and complexity barrier.

Datasets

The authors evaluated Ritornello using 14 transcription factor ChIP-seq experiments from the ENCODE project, each with two biological replicates. Matched DNA input or IgG controls were available for these experiments, allowing for direct comparison with other peak callers. No new datasets were introduced for training or benchmarking beyond these publicly available ENCODE samples.

Key Insights and Ideas

  • Control-Free Peak Calling is Feasible and Superior: The most significant insight is that accurate ChIP-seq peak calling can be achieved without a matched DNA input control. Ritornello demonstrates that by employing sophisticated signal processing and modeling, it can outperform methods that rely on such controls, effectively reducing experimental costs.
  • Fragment Length Distribution is Key: Accurately characterizing the entire fragment length distribution, not just its mean, is critical for precise peak calling. Ritornello's deconvolution method for deriving this distribution from single-end reads is a major innovation.
  • Read Length Artifacts are a Major Source of False Positives: The paper clearly identifies and characterizes "read length artifacts" as a primary cause of false positives in ChIP-seq data. These artifacts, often related to mapping errors, can mimic true binding signals and are a key target for Ritornello's filtering mechanisms.
  • DSP Techniques Offer Powerful Solutions: The successful application of DSP concepts like matched filtering and deconvolution highlights their utility in biological data analysis, particularly for signal extraction and noise reduction in complex genomic datasets.
  • Low-Quality Data Can Be Rescued: Ritornello's robustness allows it to identify true binding events even in ChIP-seq samples that would typically be discarded due to low quality (as assessed by standard metrics like NSC and RSC).

Key Results

  • Improved Reproducibility: Ritornello achieved the largest number of reproducible peaks in 7 out of 14 ENCODE experiments compared to MACS2 and GEM.
  • Comparable or Better Performance: Ritornello generally performed equally or better than ENCODE-recommended peak callers (MACS2 and GEM) in terms of peak calling accuracy and signal quality.
  • Reduced False Positives: Ritornello demonstrated a significantly lower false positive rate, particularly when compared to MACS2 and GEM run without their matched controls.
  • Higher Motif Enrichment: Ritornello showed a higher motif occurrence rate in its top reproducible peaks compared to MACS2 and GEM in 16 out of 28 samples, suggesting it identifies more biologically relevant binding events.
  • Effective Rescue of Low-Quality Samples: Ritornello successfully identified true binding events in ChIP-seq samples that were deemed low quality by ENCODE standards, which would typically lead to discarding the data.
  • Cost Reduction: By eliminating the need for a matched DNA input control, Ritornello effectively halves the sequencing cost for ChIP-seq experiments.

Limitations and Future Work

  • Point-Source Peaks: The authors acknowledge that Ritornello is primarily designed for detecting "point-source peaks," such as those generated by transcription factor binding events. For "broad-source peaks," like those associated with epigenetic modifications, they recommend other peak callers.
  • Future Work: The authors suggest that Ritornello's ability to rescue low-quality samples could be valuable for strengthening findings from higher-quality samples. Further exploration of its application to other types of ChIP-seq experiments (e.g., histone modifications) could be a future direction, although the current design is optimized for point-source binding.

Summary updated

Abstract

Abstract With the advent of next generation high-throughput DNA sequencing technologies, omics experiments have become the mainstay for studying diverse biological effects on a genome wide scale. ChIP-seq is the omics technique that enables genome wide localization of transcription factor binding or epigenetic modification events. Since the inception of ChIP-seq in 2007, many methods have been developed to infer ChIP target binding loci from the resultant reads after mapping them to a reference genome. However, interpreting these data has proven challenging, and as such these algorithms have several shortcomings, including susceptibility to false positives due to artifactual peaks, poor localization of binding sites, and the requirement for a total DNA input control which increases the cost of performing these experiments. We present Ritornello, a new approach with roots in digital signal processing (DSP) that addresses all of these problems. We show that Ritornello generally performs equally or better than the peak callers tested and recommended by the ENCODE consortium, but in contrast, Ritornello does not require a matched total DNA input control to avoid false positives, effectively decreasing the sequencing cost to perform ChIP-seq.
BibTeX
@misc{stanton2015ritornello,
  author = {Kelly Stanton and Jiaqi Jin and Sherman M. Weissman and Yuval Kluger},
  title = {{Ritornello: High fidelity control-free chip-seq peak calling}},
  year = {2015},
  doi = {10.1101/034090},
  howpublished = {bioRxiv},
}

Loading related content...

Figure