ChIPseq QC

I am speaking at the CHI Next-Gen Sequencing conf in Providence 9.26.2010 (Sunday short course). My topic is going to be about the role of bioinformatics in QC for NG sequencing, with examnples from ChIPseq, where I have the most experience.

My main point is that the informatics team works hardest on experiments that produce poor data - or where the data contradict the investigator's expectations. When the experiment is beautiful, then you can use your automated (or semi-automated) pipeline, and hand over the analyzed data with a standard report. For a transcription factor type ChIPseq, the standard result is a set of peaks with p-value and fold change vs. an input DNA, annotated by distance to the nearest gene's Transcription Start Site. If pressed, we can deliver this about 2 days after the sequencing run is complete. For an epigenomics type ChIPseq (histone methylation, acetylation, etc) we deliver both peaks vs input DNA and some type of fold-change for each peak comparing one biological condition vs. another.

However, we spend a lot more time squabbling about runs with high PCR duplication, weird artifacts, low yield, peaks in the input DNA lane, etc. To deal with this, we have been developing a variety of tools to quantify overall data quality in a ChIPseq run. We are looking as the overall clustering of mapped reads on the Reference genome (average spacing of adjacent/overlapping reads), as well as coverage at various depths. Some of these metrics make intersting graphs, but we have not completely pinned down their predictive power for understanding the data.

We have recently been playing with selecting sets of genes based on external data such as gene exprssion values from microarray or RNAseq experiments, and looking at the aggregate profile of reads mapped near the TSS of groups of genes that are upregulated, downregulated, unchanged, etc. By combining reads for a bunch of genes, we get smooter curves and you can actually say fairly clearly that upregualted genes have (or do NOT have) a change in histone methylation near the TSS as compared with downreg or unchanged genes.

Correlation does not equal causation

I am not a biostatistican, but I can play on on Blogger. Attention science news writers: "Correlation does not equal causation." Can we institute a simple electroshock penalty for those who cite such fraudulent statistics?

Check this out:

"new study finds that older women who use multivitamins may be more likely than non-users to develop breast cancer"

http://www.reuters.com/article/idUSTRE62S4F520100329

Is there any chance of a sampling bias in this study? Any chance at all that the population of elderly women who take multivitamins may be different in any health parameters from the population that finds no need for such supplements? Furthermore, is there any chance that the study subjects taking the vitamins are more likely to be examined more frequently and therefore more likely to have cancer detected?

This reminds me of the brilliant study (cited from the grocery store tabloids by my sister-in-law) that diet soda makes you fat, because a study found that people who drink diet soda were more likely to be overweight than those who do not drink it.

Esther Dyson disrespects geneticists

I was reading the NY Times on Sunday (3/21/2010). I do that. In the Business section there was an article about the relatively poor sales of the direct-to-consumer genetic tests offered by the startup company 23andMe. Other competitors in this market have done even worse. Several scientists interviewed in the article said basically that there is very little predictive medical value to these SNP profile tests. Then they interviewed Esther Dyson, who is apparently on the board of directors of 23andMe. She provided this clunker of a quote:

Ms. Dyson called it “appallingly paternalistic,” to think consumers could not interpret genetic information without help of a doctor. “People can understand statistics about baseball,” she said, “and I think they ought to understand statistics about genetics.”

Does this not trivialize all of the work of medical geneticists, biostatisticians, and bioinformaticians. Is our work really no more challenging than interpreting baseball statistics? Gee thanks Ms. Dyson.

Targeted Resequencing

Targeted Resequencing is one area of the DNA sequencing landscape that has not yet been revolutionized by Next-Gen technologies.

Targeted resequencing typically investigates a few genes (or a few dozen) across large populations. The largest portion of the effort involves lots of PCR to collect all the exons — or in some projects entire gene regions, then sequencing each amplification product, while keeping track of which PCR product comes from which individual. Even a small project - 10 genes, with 10 exons each, on 100 individuals means 10,000 PCR reactions, and 10,000 sequencing reactions (while keeping accurate track of 10,000 different DNA fragments and avoiding cross-contamination).

The Next-Gen approach would amplify the genomic regions in larger chunks, combine all of the chunks from one individual together, then run the library prep protocol (fragment, attach linkers, etc). So how does this play out in reality?

I read a paper in Genome Biology yesterday (Harismendy et al http://genomebiology.com/2009/10/3/R32) about targeted sequencing. They looked at six genes, which were covered by 28 large PCR amplicons (all exons plus some introns) which ranged in size from 3 to 14 kb, for a total of 266 kb of genomic DNA. These PCR products were then combined, and used in the sample prep protocols for 454, ABI SOLID, and Illumina GA sequencing. The same genes also were sequenced by standard Sanger methods using 273 short PCR reactions (88 kb).

Overall, the NG seq methods showed distinct bias favoring the ends of PCR products, and required very high coverage (34-fold, 110-fold and 101-fold for Roche 454, Illumina GA, and ABI SOLiD, respectively) to achieve a 10% false positive rate - false negative rates were much lower.

Lets talk about costs. Sanger sequencing costs from $3-10 per sample. I've got an Internet offer here for $4 per reaction, so lets use that for this study:

Sanger: $4 x 273 = $1092 per individual
Illumina is about $1000 per sample plus about $300 per sample for the library prep kit.

So I think they are about the same.

However, the Next Gen methods come out far ahead if you multiplex a group of individuals together in the same sequencing reaction. This is not possible with Sanger methods since the sequence is read from the average of a large number of molecules. Then the question becomes how deep can you multiplex while still producing enough reads from each individual research subject to achieve the depth of coverage needed? Our Illumina GAII currently produces about 2 million (usable) 35 bp reads per lane, but we are ramping up toward 5 million 50 bp reads with the latest upgrades.

2 M X 35 bp = 70 M bases
5 M X 50 bp = 250 M bases

So for 250 kb X 100x coverage = 25 M bp

So it looks like the current generation of NG machines do have a cost advantage over Sanger methods if you include 8, 10, or 12 X multiplexing. Improved accuracy and reduced sampling bias (sample prep methods) could bring down the coverage requirements and increase the advantage of NG methods.

I'd really like to hear some other opinions about this issue. We are writing several grant proposals for projects like these and I need some convincing arguments.

—Stuart

Metagenomics of the effects of antibiotics on the human gut

The Pervasive Effects of an Antibiotic on the Human Gut Microbiota, as Revealed by Deep 16S rRNA Sequencing

Dethlefsen L, Huse S, Sogin ML, Relman DA
PLoS Biology Vol. 6, No. 11, e280 doi:10.1371/journal.pbio.0060280

A paper in PLOS Biology from the Relman lab investigates the effect of a treatment with the antibiotic ciprofloxacin on the bacteria in the intestine. They collected over 7,000 full-length 16S rDNA sequences (1100-1400 bp) by Sanger sequencing and over 900,000 reads (~250 bp) from 454 sequencing of the V3 and the V6 regions. 

There are many important results in this paper, but it is particularly relevant that 454 sequencing reveals more taxonomic variation with greater stability than traditional sequencing. In my own work, I have found that sequence variants that occur only once in the experiment cannot be used to differentiate samples. Deep sequencing reveals more taxa, and also reduces the frequency of singletons. A rare sequence variant (OTU) that occurs only once in the ~7000 full-length sequences occurs about 65 times in the 454 data set, providing more than enough "probability of detection" to be used for comparisons between samples. 

"This set of 7,208 sequences is among the largest datasets of full-length 16S rRNA sequences from the human microbiota (or any environment), the rarefaction curves for V6 and V3 tag pyrosequencing eventually rise higher and display more curvature toward the horizontal than the OTU0.01 curve. These features show that a single run of the [454] FLX sequencer targeting V6 or V3 tags from the human gut microbiota can reveal more taxa, and capture a larger proportion of the detectable taxa, than a more extensive effort directed toward full-length 16S rRNA clone sequencing."