Every lab wants to produce high-quality, reproducible data. But when that data is destined for use by the whole community as part of an international consortium, there is an even greater incentive to ensure the highest standards. A new paper in G3: Genes|Genomes|Genetics defines a critical step for the success of widely-used gene regulation experiments.
Cheryl Keller is an Associate Research Professor in Ross Hardison’s lab at Penn State University, which studies blood cell development and how cells make the transition from pluripotent stem cells to fully differentiated blood cells. They also provided data for ENCODE, a project to catalogue the functions of the non-coding parts of the genome. To identify protein binding sequences in the genome, the researchers ran assays called chromatin immunoprecipitation and sequencing (ChIP-seq). Keller noticed that the lab was getting inconsistent results from their ChIP-seq assays, despite using the same protocols and reagents. They needed to figure out why.
“This whole project really started out as a troubleshooting experiment for our own purposes as a production lab for ENCODE 3,” Keller says. “We wanted to be able to generate the high-quality data that we said we could generate, and to try to be more consistent.”
Researchers use ChIP-seq to identify DNA sequences where a particular protein binds. First, the protein is crosslinked to the DNA, and then the chromatin is broken into small segments for sequencing. Next, an antibody against the protein of interest is used to pull out the protein and whatever DNA fragments it’s bound to. The crosslinking is reversed, the protein is removed from the sample, and the DNA fragments are sequenced. Bound sequences show up as “peaks” in the genome browser, with the height of the peak depending on the number of copies sequenced.
Keller had noticed a lot of unexplained variation in the results of the ChIP-seq experiments even when using the same antibody, which is often the major determinant of ChIP-seq success. “If I did several ChIPs with the same antibody, why didn’t they all work?” she says. “I know the antibody works, I know the protein is bound in those particular cells. Why are some working and some not?”
The researchers suspected the problem might lie with their sonicator, the machine that breaks the chromatin into fragments using sound waves. The device was old, its probe studded with pockmarks, so they invested in a brand-new machine. Yet they still had some problems with inconsistent results. They also heard from other labs that others were having similar issues with poor reproducibility in their ChIP-seq data.
“We decided to try to figure out whether if we optimized sonication, it would lead to more reproducible results,” Keller explains.
To separate antibody problems from issues in other parts of the protocol, Keller’s lab tested the ChIP-seq assay using two well-studied antibodies to the transcriptional regulators CTCF and TAL1. “One of the reasons we chose CTCF and TAL1 for our troubleshooting experiments is because there have been a number of publications on these factors and we knew, generally speaking, where these proteins were going to be binding,” Keller explains. “We had some basis for assessing our experiments.”
By systematically sonicating cells for different amounts of time, the researchers demonstrated that longer sonication generated smaller average fragment sizes, as expected. Then they looked at how the average fragment size affected the sequencing results.
Because they knew which bound sequences should turn up, the researchers were able to categorize how well the assay performed in each case. They classified each data set as “pass,” “low pass,” or “fail,” depending on how well the set of peaks matched what was expected. If most expected peaks were present, that dataset would “pass,” if some peaks were missing or not as strong as they should be, that might get classified as “low pass,” and a dataset with few to no peaks would be a “fail.”
The researchers found the level of sonication and the average chromatin length had pronounced impacts on the quality of ChIP-seq signals. Too much sonication consistently reduced data quality, while the impact of undersonication differed between transcription factors. “While most of the CTCF experiments look pretty good, we had a bit of a different story with the TAL1,” Keller says. When average fragment sizes were in the mid-range, the datasets passed, but the quality declined with the largest or smallest fragment sizes.
Excess sonication may lead to the disruption of the very DNA-protein interactions the experiment is designed to identify, while fragments that are too long may allow for 3D configurations that hide the binding site. “It’s possible that when the chromatin fragments are larger, the epitope is sequestered and the antibody doesn’t have good access to it,” Keller suggests.
While in these experiments fragments in the 200-250 bp range work the best, Keller emphasizes that it is unlikely that there is a universal best fragment size. “The take home is that it should be determined empirically for each factor, for each cell line,” she says. “Ideally, you’d do a titration and determine what is the optimal shearing size for your factor of interest.”
This optimization step has benefits beyond overall confidence in the data. Keller’s team took advantage of their improved procedures to produce ChIP-seq datasets for rare cell types, which are challenging experiments where every sample is precious and there is no room for error. Their work suggests that careful control of chromatin shearing will improve the success of not only individual labs but the entire field.