9.3 Factors that affect ChIP-seq experiment and analysis quality
9.3.1 Antibody specificity
Antibody specificity is a term which refers to how strongly an antibody binds to its preferred target, with respect to everything else present in the cell. It is the paramount measure influencing the successful execution of a ChIP experiment. Antibodies can bind multiple proteins with the same affinity. This is called antibody cross-reactivity. If an antibody cross-reacts with multiple proteins, the results of a ChIP experiment will be ambiguous. Instead of finding where our protein binds to the DNA, we will get a superposition of binding of multiple proteins. Such data are impossible to analyze correctly, and will produce false conclusions. There are many experimental procedures for validating antibody specificities, and an antibody should pass multiple tests in order to be considered valid. The exact recommendations are listed by the ENCODE consortium (Landt, Marinov, Kundaje, et al. 2012).
Every time we are analyzing a new ChIP-seq experiment, we have to take our time to convince ourselves that all of the appropriate experimental controls were performed to validate the antibody specificity (Wardle and Tan 2015).
9.3.2 Sequencing depth
Variation in sequencing depth is the first systematic technical bias we encounter in ChIP-seq experiments. Namely, different samples will contain different number of sequenced reads. Different sequencing depth influences our ability to detect enriched regions, and complicates comparisons between samples (Jung, Luquette, Ho, et al. 2014). The statistical procedure of removing the influence of sequencing depth on the quantification is called depth scaling; we calculate a scaling factor which is used to multiply the signal strength before the comparison. There are multiple methods for normalization, and each method comes with its assumptions. Scale normalization is done by dividing the read counts (in certain genomic locations) by the total amount of sequenced reads. This method presumes that the ChIP efficiency worked equally well in all studied conditions. Because the ChIP efficiency differs in different antibodies, it is often unsuitable for comparisons of ChIP-seq experiments done on different proteins. Robust normalization tries to locate genomic regions which do not change between different biological conditions (regions where the protein is constantly bound), and then uses the sum of the reads in those regions as the scaling factor. This method presumes that we can reliably identify regions which do not change (Shao, Zhang, Yuan, et al. 2012). Background normalization presumes that the genome can be split into two categories: background regions and true signal regions. It then uses the number of reads in the background regions to define the scaling factor (Liang and Keleş 2012). External normalization uses external reference for normalization; we add known amounts of chromatin from a distant species, or artificial spike-ins which are then used as a scaling reference. This is used when we think there are global changes in the biding profiles between two biological conditions – very large changes in the signal profile (Bonhoure, Bounova, Bernasconi, et al. 2014).
The choice of normalization method depends on the type of analysis (Angelini, Heller, Volkinshtein, et al. 2015); if we want to quantitatively compare the abundance of different histone marks in different cell types, we will need the different normalization procedure than if we want to compare TF binding in the same setting.
9.3.3 PCR duplication
The amounts of the DNA obtained after the ChIP experiment are quite often lower than the minimal amount which can be sequenced. Polymerase chain reaction (PCR) is a procedure used for amplification of DNA fragments. It is used to increase the amount of DNA in our sample prior to sequencing. PCR is a stochastic procedure, meaning that the results of each PCR reaction cannot be predicted. Due to its stochastic nature, PCR can be a significant source of variability in the ChIP-seq experiments (Aird, Ross, Chen, et al. 2011; Benjamini and Speed 2012; Teng and Irizarry 2016). A quality control is necessary to check whether all of our samples have the same sequence properties, i.e. the same enrichment of dinucleotides, such as CpG. If the samples differ in their sequence properties, that means we have to account for them during the analysis (Teng and Irizarry 2017).
9.3.4 Biological replicates
Biological replicates are independently executed ChIP-seq experiments from different samples, corresponding to the same biological conditions. They are indispensable for estimating ChIP quality, and give us an estimate of the variability in the experiment which we can expect due to unknown biological variables. Without biological replicates, it is statistically impossible to compare ChIP-seq samples from different biological conditions, because we do not know whether the observed changes are a result of the inherent biological variability (the source of which we do not understand), or they result from the change in the biological condition (different tissue or transcription factor used in the experiment). If we encounter an experimental setup which does not include biological replicates, we should be extremely skeptical about all conclusions derived from such analysis.
9.3.5 Control experiments
There are three types of control experiments which can be performed to control for known and unknown experimental biases:
Input control: Sequencing of genomic DNA without the immunoprecipitation step.
IgG control: Using a polyclonal mixture of non-specific IgG antibodies instead of a specific antibody.
Knockout control: Performing the ChIP experiment in a biological system which does not contains our protein of interest (i.e. in a cell line where the transcription factor was knocked out) (Krebs, Schmidt, Goren, et al. 2014).
Each type of control experiment controls for a certain set of experimental biases.
Input control is the most frequent type of control performed. It shows the differential susceptibility of genomic regions to the ChIP-procedure. Due to the hierarchical structure of chromatin, different genomic regions have different propensities for cross-linking, sonication, and immunoprecipitation. This causes an uneven probability of observing DNA fragments originating from different genomic regions. Because different cell types (cell lines, and cancer cell lines), have different chromatin structure, ChIP samples will show a cell-type-specific bias in observed enrichment profiles. An important note to consider is that the input control is basically a reduced whole genome sequencing experiment, while the ChIP enriches for only a subset of genomic regions. If both ChIP and Input samples are sequenced to the same depth (same number of reads), the background distribution in the input sample will be under sampled. It is recommended to sequence the input sample deeper than the ChIP sample (Chen, Negre, Li, Mieczkowska, Slattery, Liu, Zhang, Kim, He, Zieba, and others 2012).
IgG control uses a soup of nonspecific antibodies to control for background binding. In principle, the antibodies should be isolated from the same batch of serum which was used to create the specific antibody (used for ChIP). It should, in theory, give a background profile of non-specific binding. The proper control, is however, seldom available. Additionally, because the antibodies are unspecific, the amount of precipitated DNA will be low, and the samples will require additional rounds of PCR amplification.
KO control is a ChIP experiment performed in the biological system where the native protein is not present. Such an experiment profiles the non-specific binding of the antibody to other proteins, and directly to the DNA. The primary, and only, concern is that the perturbation caused by the knock-out (or knock-down), changes the cell so much, that the ChIP profile is not comparable to the original cell. This is the most accurate type of control experiment, however, it is frequently technically challenging to perform if the cells are not viable after the knock-out, or if the knock-out is impossible to perform.
9.3.6 Using tagged proteins
If an antibody of sufficient quality is not available, it is possible to resort to constructs where the protein of interest gets engineered with a ChIP-able tag. The proper control for such experiments is to perform the ChIP in the cell line containing the engineered protein, and without the protein. It must be noted that the tagging procedure can change the binding preferences of the protein, and therefore the experimental conclusions.
Aird, Ross, Chen, Danielsson, Fennell, Russ, Jaffe, Nusbaum, and Gnirke. 2011. “Analyzing and Minimizing PCR Amplification Bias in Illumina Sequencing Libraries.” Genome Biol 12 (2): R18. https://doi.org/10.1186/gb-2011-12-2-r18.
Angelini, Heller, Volkinshtein, and Yekutieli. 2015. “Is This the Right Normalization? A Diagnostic Tool for ChIP-Seq Normalization.” BMC Bioinformatics 16 (May): 150. https://doi.org/10.1186/s12859-015-0579-z.
Benjamini, and Speed. 2012. “Summarizing and Correcting the GC Content Bias in High-Throughput Sequencing.” Nucleic Acids Res 40 (10): e72. https://doi.org/10.1093/nar/gks001.
Bonhoure, Bounova, Bernasconi, et al. 2014. “Quantifying ChIP-Seq Data: A Spiking Method Providing an Internal Reference for Sample-to-Sample Normalization.” Genome Res 24 (7): 1157–68. https://doi.org/10.1101/gr.168260.113.
Chen, Negre, Li, et al. 2012. “Systematic Evaluation of Factors Influencing Chip-Seq Fidelity.” Nature Methods 9 (6): 609–14.
Jung, Luquette, Ho, et al. 2014. “Impact of Sequencing Depth in ChIP-Seq Experiments.” Nucleic Acids Res 42 (9): e74. https://doi.org/10.1093/nar/gku178.
Krebs, Schmidt, Goren, et al. 2014. “Optimization of Transcription Factor Binding Map Accuracy Utilizing Knockout-Mouse Models.” Nucleic Acids Res 42 (21): 13051–60. https://doi.org/10.1093/nar/gku1078.
Landt, Marinov, Kundaje, et al. 2012. “ChIP-Seq Guidelines and Practices of the ENCODE and modENCODE Consortia.” Genome Res 22 (9): 1813–31. https://doi.org/10.1101/gr.136184.111.
Liang, and Keleş. 2012. “Normalization of ChIP-Seq Data with Control.” BMC Bioinformatics 13 (August): 199. https://doi.org/10.1186/1471-2105-13-199.
Shao, Zhang, Yuan, Orkin, and Waxman. 2012. “MAnorm: A Robust Model for Quantitative Comparison of ChIP-Seq Data Sets.” Genome Biol 13 (3): R16. https://doi.org/10.1186/gb-2012-13-3-r16.
Teng, and Irizarry. 2017. “Accounting for Gc-Content Bias Reduces Systematic Errors and Batch Effects in Chip-Seq Data.” Genome Research. https://doi.org/10.1101/gr.220673.117.
Teng, and Irizarry. 2016. “Accounting for GC-Content Bias Reduces Systematic Errors and Batch Effects in ChIP-Seq Peak Callers.” bioRxiv, January. http://biorxiv.org/content/early/2016/11/30/090704.
Wardle, and Tan. 2015. “A Chip on the Shoulder? Chromatin Immunoprecipitation and Validation Strategies for Chip Antibodies [Version 1; Referees: 2 Approved].” F1000Research 4 (235). https://doi.org/10.12688/f1000research.6719.1.