Theoretical characterisation of strand cross-correlation in ChIP-seq

BACKGROUND

Strand cross-correlation profiles are used for both peak calling pre-analysis and quality control (QC) in chromatin immunoprecipitation followed by sequencing (ChIP-seq) analysis. Despite its potential for robust and accurate assessments of signal-to-noise ratio (S/N) because of its peak calling independence, it remains unclear what aspects of quality such strand cross-correlation profiles actually measure.

RESULTS

We introduced a simple model to simulate the mapped read-density of ChIP-seq and then derived the theoretical maximum and minimum of cross-correlation coefficients between strands. The results suggest that the maximum coefficient of typical ChIP-seq samples is directly proportional to the number of total mapped reads and the square of the ratio of signal reads, and inversely proportional to the number of peaks and the length of read-enriched regions. Simulation analysis supported our results and evaluation using 790 ChIP-seq data obtained from the public database demonstrated high consistency between calculated cross-correlation coefficients and estimated coefficients based on the theoretical relations and peak calling results. In addition, we found that the mappability-bias-correction improved sensitivity, enabling differentiation of maximum coefficients from the noise level. Based on these insights, we proposed virtual S/N (VSN), a novel peak call-free metric for S/N assessment. We also developed PyMaSC, a tool to calculate strand cross-correlation and VSN efficiently. VSN achieved most consistent S/N estimation for various ChIP targets and sequencing read depths. Furthermore, we demonstrated that a combination of VSN and pre-existing peak calling results enable the estimation of the numbers of detectable peaks for posterior experiments and assess peak calling results.

CONCLUSIONS

We present the first theoretical insights into the strand cross-correlation, and the results reveal the potential and the limitations of strand cross-correlation analysis. Our quality assessment framework using VSN provides peak call-independent QC and will help in the evaluation of peak call analysis in ChIP-seq experiments.

Paired-end mappability of transposable elements in the human genome

Though transposable elements make up around half of the human genome, the repetitive nature of their sequences makes it difficult to accurately align conventional sequencing reads. However, in light of new advances in sequencing technology, such as increased read length and paired-end libraries, these repetitive regions are now becoming easier to align to. This study investigates the mappability of transposable elements with 50 bp, 76 bp and 100 bp paired-end read libraries. With respect to those read lengths and allowing for 3 mismatches during alignment, over 68, 85, and 88% of all transposable elements in the RepeatMasker database are uniquely mappable, suggesting that accurate locus-specific mapping of older transposable elements is well within reach.

False positives in trans-eQTL and co-expression analyses arising from RNA-sequencing alignment errors

Sequence similarity among distinct genomic regions can lead to errors in alignment of short reads from next-generation sequencing. While this is well known, the downstream consequences of misalignment have not been fully characterized.  We assessed the potential for incorrect alignment of RNA-sequencing reads to cause false positives in both gene expression quantitative trait locus (eQTL) and co-expression analyses. Trans-eQTLs identified from human RNA-sequencing studies appeared to be particularly affected by this phenomenon, even when only uniquely aligned reads are considered. Over 75% of trans-eQTLs using a standard pipeline occurred between regions of sequence similarity and therefore could be due to alignment errors. Further, associations due to mapping errors are likely to misleadingly replicate between studies. To help address this problem, we quantified the potential for "cross-mapping'' to occur between every pair of annotated genes in the human genome. Such cross-mapping data can be used to filter or flag potential false positives in both trans-eQTL and co-expression analyses. Such filtering substantially alters the detection of significant associations and can have an impact on the assessment of false discovery rate, functional enrichment, and replication for RNA-sequencing association studies.

False positives in trans-eQTL and co-expression analyses arising from RNA-sequencing alignment errors

Sequence similarity among distinct genomic regions can lead to errors in alignment of short reads from next-generation sequencing. While this is well known, the downstream consequences of misalignment have not been fully characterized.  We assessed the potential for incorrect alignment of RNA-sequencing reads to cause false positives in both gene expression quantitative trait locus (eQTL) and co-expression analyses. Trans-eQTLs identified from human RNA-sequencing studies appeared to be particularly affected by this phenomenon, even when only uniquely aligned reads are considered. Over 75% of trans-eQTLs using a standard pipeline occurred between regions of sequence similarity and therefore could be due to alignment errors. Further, associations due to mapping errors are likely to misleadingly replicate between studies. To help address this problem, we quantified the potential for "cross-mapping'' to occur between every pair of annotated genes in the human genome. Such cross-mapping data can be used to filter or flag potential false positives in both trans-eQTL and co-expression analyses. Such filtering substantially alters the detection of significant associations and can have an impact on the assessment of false discovery rate, functional enrichment, and replication for RNA-sequencing association studies.

A Statistical Framework for the Analysis of ChIP-Seq Data

Chromatin immunoprecipitation followed by sequencing (ChIP-Seq) has revolutionalized experiments for genome-wide profiling of DNA-binding proteins, histone modifications, and nucleosome occupancy. As the cost of sequencing is decreasing, many researchers are switching from microarray-based technologies (ChIP-chip) to ChIP-Seq for genome-wide study of transcriptional regulation. Despite its increasing and well-deserved popularity, there is little work that investigates and accounts for sources of biases in the ChIP-Seq technology. These biases typically arise from both the standard pre-processing protocol and the underlying DNA sequence of the generated data. We study data from a naked DNA sequencing experiment, which sequences non-cross-linked DNA after deproteinizing and shearing, to understand factors affecting background distribution of data generated in a ChIP-Seq experiment. We introduce a background model that accounts for apparent sources of biases such as mappability and GC content and develop a flexible mixture model named MOSAiCS for detecting peaks in both one- and two-sample analyses of ChIP-Seq data. We illustrate that our model fits observed ChIP-Seq data well and further demonstrate advantages of MOSAiCS over commonly used tools for ChIP-Seq data analysis with several case studies.