Wednesday, January 21, 2009

The Experts Agree

It depends what you are trying to do. That is the take home message in Genome Technology’s (GT) trouble-shooting guide on picking assembly and alignment algorithms for Next-Gen sequence data.

In the guide, the GT team asked nine Next-Gen sequencing and bioinformatics experts to answer six questions:
  1. How do you choose which alignment algorithm to use?
  2. How do you optimize your alignment algorithm for both high speed and low error rate?
  3. What approach do you use to handle mismatches or alignment gaps?
  4. How do you choose which assembly algorithm to use?
  5. Do you use mate-paired reads for de novo assembly? how?
  6. What impact does the quality of raw read data have on alignment or assembly? how do your algorithms enhance this?
Even a quick look at the questions shows us that many factors need to be considered in setting up a Next-Gen sequencing lab. Questions 1 and 4 point out that aligning sequences is different from assembling them. Other questions address issues related to the size of the data sets being compared, the quality of the data being analyzed, the kinds of information that can be obtained, and the computational approaches being used for different problems.

What the experts said

First, they all agree that different problems require different approaches and have different requirements. In the first question about which aligner to use, the most common response was “for what application and which instrument?” Fundamentally, SOLiD data are different from Illumina GA data which are different from 454 data. While the end results may all be sequences of A's, G's, C's, and T's; the data are derived in different ways because of the platform-specific twists in collecting the data (recall “Color Space, Flow Space, Sequence Space, or Outer Space). Not only are there platform-specific methods for interpreting raw data, multiple programs have been developed for each instrument with their own strengths and weaknesses in terms of speed, sensitivity, the kinds of data they use (color, base, or flow spaces, quality values, and paired end data), and the information that is finally produced. Hence, in addition to choosing a sequencing platform you also have to think about the sequencing application, or the kind of experiment, that will be performed. In gene expression studies, for example, an RNA-Seq experiment has different requirements in terms of aligning the data and interpreting the output than an experiment with Tag Profiling.

Overall the trouble-shooting guide discussed 17 total algorithms, eight for alignment, and nine for assembly (two of which were for Sanger methods). Even this selection wasn't a comprehensive list. When other sites [1, 2] and articles [3] are included and proprietary methods are factored in, over 20 algorithms are available. So what to do? Which is best?

That depends

Yes, the choice of algorithm ultimately depends on what you are trying to do. While we can agree that there is no best solution, we also know that is not a helpful response. What is needed is a way to test the suitability of different algorithms for different kinds of experiments and to represent data in standard ways so that the features of specific algorithms can be evaluated. Also, as this is a new field, standard requirements for how data should be aligned, defining a correct alignment, and what kinds of information are the most informative in describing alignments are still emerging. Some of the early programs are helping to define these requirements.

One program we've used, at Geopsiza, for identifying requirements is MAQ, a program for sequence alignment. As noted in previous blogs [MAQ attack], MAQ is a great general purpose tool. It provides comprehensive information about the data being aligned and details about alignments. MAQ works well for many applications including RNA-Seq, Tag Profiling, ChIP-Seq, and resequencing assays focused on SNP discovery. In performance tests, MAQ is slower than some of the newer programs, one of which is being developed by MAQ’s author, but MAQ is a good model for getting the right kinds of information, formatted in a decent way. Indeed MAQ was the most cited program in the GT guide.

Let’s return to the bigger issue. That is, how can we easily compare between algorithms? For that we need a system where one can easily define a standardized dataset and reference sequence, and have a platform where a new algorithm can be added and run from a common interface. Standard reports that present features of the alignments could then be used to compare programs and parameters.

The laboratory edition of GeneSifter supports these kinds of comparisons. The distributed system architecture allows one to quickly develop control scripts to run programs and format their output in figures and tables that make comparisons possible. With this kind of system in place, the challenges move from which program to run and how to run it, to how to get the right kinds of information and best display the data. To address these issues, Geospiza’s research and development team is working on projects focused on using technologies like HDF5 to create scalable standardized data models for storing information from alignment and assembly programs. Ultimately this work will make it easy to optimize Next-Gen sequencing applications and assays and compare between assorted programs.

3. Shendure J., Ji H., 2008. Next-generation DNA sequencing. Nat Biotechnol 26, 1135-1145.

No comments: