Sequencing tech

The basic conclusions about sequencing tech were these:

1. Illumina is by far the "best" current technology in terms of breadth and depth of applications. 454 is waning in utility and it's simply not cost effective; PacBio is too error prone for widespread use; and Nanopore technologies are not yet available, much less proven. Illumina MiSeq looks particularly beguiling, with fast turnaround and long reads. One surprise to me was that the Ion Torrent was billed as being error prone and currently rather expensive.
2. Illumina is likely to remain "it" as far as RNAseq and shotgun metagenomics goes: depth of sampling is critical for these applications, and no other technology has anything like the deep sampling provided by Illumina.

Given the speakers and attendees we had this year, I think these are pretty robust conclusions for what's available today. 3 months from now? Who knows!? The sequencing field is moving awfully fast...

Bioinformatics needs

For me, of course, the bioinformatics needs ware just as interesting, if not more so. Bench biologists seem to finally be getting the idea that data generation isn't very interesting if you can't do a good analysis of the data (and those "good analyses" also need good downstream hypothesis generation & validation, but that's a separate point...)

My perspective (informed by both discussions with course faculty as well as results from my own research program -- yarr, these be forward looking statements, me matey!) is that several NGS problems have been essentially solved, modulo some important finer details. These include reference-free transcriptome assembly, isoform detection/extraction/analysis, and resequencing analysis of individuals. Good software exists for these tasks, and while I'm sure speedups and enhancements to software and algorithms will come, I doubt that these areas will be fruitful for serious, novel bioinformatics work. I also personally believe that efficient assembly of genomes, transcriptomes, and metagenomes (with one caveat -- see below) is solved by digital normalization, although quality of assembly will remain an issue for a while.

So... what's left? What's still really hard?

1. Differential gene expression. The biggest surprise from the course was that none of the faculty had good recommendations for a protocol or pipeline for determining differentially expressed genes. RSEM and other EM-based algorithms were mentioned but it turns out that head-to-head comparisons of these algorithms show poor agreement between them. All wrong? All right? Who knows?? We had a wide ranging discussion about this that led me to some thoughts about a possible solution, which is always nice...

   (In contrast, differential exon expression is much easier...)

2. Assembly of non-model genomes. There's a dark cloud hanging over genome assembly: polymorphism, thy name is mud. Repeats and polymorphism (aka heterozygosity and strain variation) are hard for current assemblers to resolve, and I have experienced the horrors of these problems myself in several projects.

3. Combining 454, Illumina, and other read technologies. It kind of baffles me that this is true, but as far as I can tell there is no good way to combine multiple read technologies for genome or transcriptome assembly. I can find lots of ad hoc protocols, of course, but everyone always seems to end with "...so this is what we use, and it kind of sucks." References to better solutions are welcome...

4. Recovery or inference of haplotypes. Short-insert data isn't good for the haplotype phasing problem, alas. Still very challenging.

5. Metagenome assembly of diverse environments. Digital normalization makes most assembly scaling challenges go away, by shifting the problem to one that scales with the diversity of your sample; I see no reason you can't do contig assembly of pretty much anything on a commodity machine with 512 GB-1 TB of RAM. The problem here is that soil and marine environments seem to have nigh-infinite diversity. This is where partitioning comes in handy, but we think we can only complete an assembly of about 1-2 Tbp of data with current holistic approaches.

6. Efficient error correction. Error correction of Illumina reads, PacBio, 454, etc. is a growth area -- especially efficient error correction.

7. Population sequencing. People insist on feeding mixed populations into sequencers, for both good and bad reasons. This is bioinformatically challenging to resolve in terms of recovering variants and distinguishing them from errors.

So, anyway, that's my 2 cents. I'm actually working on all seven of these, and I am pretty sure we can provide significant leverage on #2, #3, #5, and #6. Stay tuned!

I'd love to hear other thoughts on these weighty issues.

--titus