Permalink
Find file
Fetching contributors…
Cannot retrieve contributors at this time
152 lines (115 sloc) 7.03 KB

Moore-supported projects that I'd like to highlight:

Re-analyzing the Marine Microbial Eukaryotic Transcriptome Sequencing Project

Ably driven forward by Lisa Cohen (graduate student) and Harriet Alexander (postdoc), we are re-analyzing the ~700 MMETSP transcriptome data sets generated about 5 years ago by the Moore Foundation.

Major highlights:

  • this effort should dramatically improve usability of these data sets by the field, both now (updated annotations, expression) and in the future (because we can update the annotations fairly easily).

  • we have a single script that will trim, assemble, annotate, and calculate expression levels for ~arbitrary numbers of RNAseq samples.

  • the reanalysis effort has led to a lot of insight into the problems of bio(medical) data reanalysis and sharing, which Harriet is writing up as a companion "op-ed" piece to Lisa's paper.

  • the community reaction has varied from "why bother?" to "oh please dear god yes", revealing a (slightly surprising) range of reactions to the basic premise that while the data may be the same, analysis tools improve over time. Because Harriet is a known member of the ocean biogeochem community (and hopefully because I am not viewed as evil by most people) I gather that people are cautiously optimistic about the effort.

Blog post: MMETSP re-assemblies

Metagenome assembly analysis

We are close to a final draft of a paper on a project led by Sherine Awad (postdoc, funded in part by Moore) to compare three metagenome assemblers on a mock metagenome community.

Major highlights:

  • all three assemblers more or less perform the same in terms of recovering content, but one assembler (MEGAHIT) is much, much faster and requires fewer resources.

  • contamination in the mock metagenome turns out to be the cause of a mismatch between assembled content and expected content; the assemblers appear to be working just fine.

Automated tools for transcriptomics

Camille Scott (graduate student) has developed two tools that are critical in our scaling-up efforts, shmlast and dammit:

dammit is a transcriptome annotator pipeline that is already achieving market dominance (because it actually works), while shmlast is a (maintainable, Python) reimplementation of crb-blast that is used in dammit for annotating transcripts.

Continued development of khmer and screed

Our "flagship" software package, khmer, and its companion project, screed, suffered from neglect for about a year after our 2.0 release and the publication of our F1000 Research paper (The khmer software package). This was due in some part to the departure of Michael Crusoe, who was our lead software engineer, and also in significant part to my inability to lead the project scientifically at the same time as I was trying to spin up the new lab at UC Davis.

In this past year, these issues were resolved by the arrival of Daniel Standage (postdoc, funded entirely on an NIH R01), and the hiring of Tim Head (software engineer/consultant). Daniel is building a new large-genome variant calling system on top of khmer called Kevlar, and as a result is pushing khmer library development in new directions. Tim is supporting the projects in the open source way: fixing bugs, helping with code review, and generally doing the things that grad students and postdocs find secondary to their research (but are nonetheless important to software integrity, adaptibility, and forward development).

khmer remains central to the lab and, in addition to being a suite of useful exploratory tools, is now maturing into a robust platform that several projects in the lab are relying on. Despite this I am still somewhat alarmed (at a strategic level) regarding the level of resources required to achieve even a minimally responsible level of scientific software development.

Major highlights:

  • We finally have a sustainable way to move forward software development in the lab: have a postdoc or grad student take point scientifically, and support them with a (part time) PhD-level software engineer. For our lab, a ratio of 1 half-time software engineer to 2-3 active grad/postdoc developers seems to work ok.

  • Despite an increasing number of "eyes" on the code, and a frankly rather exhaustive (and exhausting) testing regime, we continue to find minor and not-so-minor bugs in khmer. This suggests to me that plain ol' software bugs are likely to have a significant contribution to the repeatability/replication/reproducibility crisis. (See Dave Soergel's paper, "Rampant software errors may undermine scientific results")

  • Camille Scott (grad student) has made significant progress on streaming RNAseq assembly, which is laying the foundation for streaming analysis of very large continually arriving sequencing data sets.

  • Our authorship policy for F1000 Research broke some people: Pubwication of software papers, and authorship on them

Development of MinHash-based technologies

The most significant shift in the lab focus over the last year occurred with the publication of the mash paper from Adam Phillipy's lab (Ondov et al., 2016). The MinHash approach detailed in this paper turns out to directly solve or provide a basis for solutions to a wide range of critical problems, including metagenome comparison, indexing sequence databases by content, and streaming classification of sequencing data.

Major highlights:

  • we developed our own toolkit, sourmash, for exploring MinHash sketches, and published v1.0 in the Journal of Open Source Software: Sourmash in JOSS

  • Luiz Irber (grad) implemented fast searches of very large databases of MinHash sketches using Sequence Bloom Trees (a technique first published by Carl Kingsford's lab).

  • Luiz also implemented a distributed architecture for sketching very large sequence collections from Web-accessible archives (blog.luizirber.org/2016/12/28/soursigs-arch-1/). This is connected to Casey Greene and Carl Kingsford's sequence refinery project.

  • I used these tools to index all the microbial genomes and worked with Luiz to index and classify 400k public records from the NCBI Sequence Read Archive:

  • We initiated a collaboration with Blair Sullivan's group and Dominik Moritz from Jeff Heer's lab at the DDD Barnraising at MDIBL, to build tools combining contracted De Bruijn graphs, dominating sets, and MinHash-based search. We hope this will lead to some major advances in metagenome and RNAseq analysis in the coming year.