Bioconductor Newsletter

posted by Valerie Obenchain, October 2015

The Bioconductor newsletter is a quarterly review of core infrastructure developments, community projects and future directions. We cover topics of general interest as well as those with the greatest impact on the software. This quarter the Bioconductor core team relocated from Seattle, Washington to Buffalo, New York and welcomed two new team members. Two new forums for workflows and analysis pipelines were introduced: the recently launched Bioconductor F1000Research Channel and the new section for R / Bioconductor Workflows in the BMC Source Code for Biology and Medicine journal. New features and functions are summarized for the infrastructure packages as well as a handful of contributed packages that have been especially active over the past devel cycle.


Bioconductor relocates to Roswell Park Cancer Institute

In September the Bioconductor core team relocated from Seattle, Washington to Buffalo, New York. The new home institution is Roswell Park Cancer Institute (RPCI) which is part of the Buffalo Niagara Medical Complex. The Complex covers about 120 acres in downtown Buffalo and is a consortium of hospitals, health care facilities and educational institutions. The campus is growing rapidly with current employee numbers at 12000 estimated to reach 17000 by 2017. RPCI has close ties with SUNY University at Buffalo.

The change in physical location also brought some staffing changes. Marc Carlson, Sonail Arora and Paul Shannon left the project at the end of August. All were instrumental in the design and implementation of AnnotationHub (and annotations in general), tools for biological network analysis, educational material and many other areas. A big thanks to them for their many contributions.

In Buffalo we are pleased to welcome new team members Jim Java and Brian Long. Jim formerly worked as a biostatistician for the Gynecologic Oncology Group at RPCI where he analyzed clinical trial data. He also worked as a software engineer for several companies on projects ranging from point-of-sale software to embedded systems. He has a Ph.D. in computer science (focused on natural language processing), and MA degrees in biostatistics and English literature. His research interests include scientific and statistical computing, natural language processing (especially authorship identification), linguistics, climate science, gun violence, ethics, and literature.

Brian’s work in software development has focused on relational database & NoSQL storage, UI/UX design, security and scale. Most recently, he’s been working with a team to develop a Platform-as-a-Service (similar to Google App Engine). The platform offering allows application developers to deploy components to on-demand virtual infrastructure and supports a large number of programming languages via Apache Thrift. Most of the work was not done as open source but the code has been released on GitHub. Feel free to check it out!

Reproducible Research

F1000Research channel launched

The Bioconductor home page has a new link for the F1000 Research Channel.

F1000Research is an open science publishing platform with the following goals:

The motivation for a Bioconductor F1000Research channel was a forum for task-oriented workflows that cover a current, genome-scale analysis problem from start to finish. In contrast to package vignettes, these workflows combine resources from several different packages, demonstrating integrative analysis and modeling techniques. The end goal is to make it easier for Bioconductor users to navigate the software offerings by adapting these workflows to quickly arrive at a solution for their specific problem.

This effort has been led by Wolfgang Huber, Vince Carey, Sean Davis, Kasper Daniel Hansen and Martin Morgan.

Source Code for Biology and Medicine hosts new R / Bioconductor Workflows section

Levi Waldron is an Assistant Professor of Biostatistics at the CUNY School of Public Health in New York City. He serves on the Bioconductor technical advisory board and is currently heads up the group developing new infrastructure for the analysis of multi-assay genomic experiments. Levi recently became a section editor for the BMC Source Code for Biology and Medicine Journal and earlier this month posted an announcement calling for R / Bioconducor workflows submissions to the new section. The journal is open access and aims to publish source code for distribution and use to advance biological and medical research. Manuscripts are considered on all aspects of workflow for information systems, decision support systems, client user networks, database management, and data mining.

The workflows should address a problem of general interest and be easily adaptable by other researchers. Implementation can be as a workflow on the Bioconductor web site, an Amazon or docker image or other cross-platform supported approach. More details are available in the author instructions.

Infrastructure Updates

Build machines to cloud

Over the past weeks Dan has been busy transferring package builds, the single package builder and the web site to the cloud. The move provides more flexible login access as well as machine configuration. Everything should look the same from the outside so user interaction with these resources will not change.

This has been a huge effort and is almost complete. Thanks Dan!

HTS core package stack

Bioconductor encourages software reuse and aims for a flexible, integrated set of infrastructure packages. One consequence of interrelationship is that a change in a low-level package can affect packages downstream.

Throughout a devel cycle new classes are added and code is reorganized which may cause the dependency hierarchy to change. We do our best to identify and fix packages affected by these changes. Modified packages and their dependencies are committed to svn/git and should propagate together through the nightly builds and become available via biocLite() the following day.

Hervé recently added a graphic to S4Vectors/inst/doc/ that depicts the High Throughput Sequencing (HTS) core package stack. Knowledge of this hierarchy is useful when developing new S4 classes and methods. It also benefits leading-edge developers working directly from svn (vs biocLite()). A package installed from svn may have updated dependencies that have not yet propagated through the build system and must be installed (in order) by hand.

The package stack file will be updated as the core packages change. This is the current snapshot.

as of August 2015

                    |     |
                    v     v
       GenomicFeatures   BSgenome
                    |     |
                    v     v
                  |         |
                  v         v
 SummarizedExperiment   Rsamtools
                |       |      |
                v       v      v
            GenomicRanges   Biostrings
                    |          |
                    v          v
           GenomeInfoDb   XVector
                    |     |
                    v     v

New functions

All are available in the devel branch, Bioconductor 3.2:

Activity in contributed packages

As complement to the section on new features and functions in the infrastructure packages we want to highlight significant changes made in contributed packages. Not all packages keep current NEWS files so it can be tricky to determine which packages have added new features. One way of gauging active development is the number of svn/git commits over a period of time.

As of September 22, these packages all had 50+ commits since the April 2015 release: CopywriteR, systemPipeR, ComplexHeatmap, derfinderHelper, ggtree, RnBeads and cogena.

A few authors said the commits were due to maintenance and the package had not changed much since the last release. Other packages did change significantly and the authors have summarized the changes below. Comments have been lightly edited for length.

ComplexHeatmap Author: Zuguang Gu

This package provides a framework to combine and visualize multiple heat maps. Combined maps can be flexibly annotated and decorated (add graphics post-generation). The oncoPrint() function offers a compact means of visualizing genomic alteration events.

ggtree Authors: Guangchuang Yu and Tommy Tsan-Yuk Lam

A phylogenetic tree viewer with extensive capabilities for adding node-level (taxa) annotation layers. Extends the graph grammar and infrastructure in ggplot and ggplot2.

Both gheatmap() and msaplot() add a new layer to tree view and can be transformed to circular form by adding +coord_polar(theta="y") to the grammar.

cogena Authors: Zhilong Jia and Michael Barnes

cogena is a workflow for co-expressed gene-set enrichment analysis.

systemPipeR Author: Thomas Girke

systemPipeR provides infrastructure for building and running automated analysis workflows for a wide range of next generation sequence (NGS) applications.




Why vectorize?

When looking for ways to optimize R code one of the most common suggestions is to ‘vectorize’. A ‘vectorized’ function in R is one that has a loop-like construct written in a compiled language with a light R wrapper. There are a couple of qualities that give these functions their performance advantage.

Data are passed to a ‘vectorized’ function as a vector instead of individual elements. Vectors in R are ‘typed’ meaning all elements must be of the same data type. Passing a whole vector to the complied code reduces the amount of work R has to do to interpret data type. Second, the loop computation is done in a complied language such as C, C++ or FORTRAN.

Vectorized functions call .C, .Call, .Primitive or .Internal in the source code. One example is base::which()

> which
function (x, arr.ind = FALSE, useNames = TRUE)
    wh <- .Internal(which(x))
    if (arr.ind && !is.null(d <- dim(x)))
        arrayInd(wh, d, dimnames(x), useNames = useNames)
    else wh
<bytecode: 0x3699c90>
<environment: namespace:base>

An concrete example of how these functions can improve performance is this post on the support site where Hervé helped the author of the ChIPseeker package identify a bottleneck in ChIPseeker:::getFirstHitIndex(). The solution was to replace a call to sapply() (i.e., R-level iteration) with two vectorized functions, duplicated() and which(). This elegant one-liner reduced the algorithm from quadratic in time to linear in time. Nice!

Hervé took a classic approach to improving the code: simplifying the original support site example (thanks!) to one that illustrated the problem in a timely fashion; stepping through the code a line at a time and noticing that one function call took inordinately long; understanding the small section of code that caused problems; and identifying a vectorized alternative.

We may not all see such drastic improvements but take home is that these functions are worth knowing about and implementing when possible.

The ellipsis

In R, the ellipsis (…) is used to pass a variable number of arguments to a function. A common question from developers writing their own S4 generics is when (and why) to include the ellipsis in the function signature. Hervé answered a recent post on Bioc-devel about this topic. Main ideas are summarized here but for the full details you’ll want to read the post.

Defining generics and methods instead of ordinary functions for the getters and setters of an S4 object is generally considered good practice. When introducing a new generic, the recommendation is to keep the signature as “generic” as possible so methods can add arguments that are specific to the objects they deal with.

     function(object, ...) standardGeneric("parameters")

Another case is that of having extra arguments such as a modifier or toggle precede the ellipsis in the generic.

     function(object, method="auto", ...)

The purpose of signature="object" is to limit dispatch to object only. Defining sensible default values (e.g., method=”auto” or verbose=FALSE) is also important to encourage consistent, user-friendly behavior across methods.

Going the direction of more restrictive, an example where the ellipsis was intentionally omitted is the organism() generic recently added to BiocGenerics. This is a straightforward getter and we want to enforce this and encourage methods to stick to that very simple contract. If someone comes up with a use-case where extra arguments are needed for their method then the ellipsis can be added to the generic.

While ... is often appropriate in the signature of generics, methods are often best written to include only the arguments actually used. In this way, invalid arguments are not silently ignored:

.A = setClass("A", representation(x="integer"))
setMethod("enrichment", "A", function(object, method="auto", ...) {
    cat("enrichment:", class(object), "; method:", method, "\n")
enrichment(.A(), mehods="special")    # typo silently ignored
## enrichment: A ; method: auto

setMethod("enrichment", "A", function(object, method="auto") {
    cat("enrichment:", class(object), "\nmethod:", method, "\n")
enrichment(.A(), mehods="special")
## Error in .local(object, method, ...) :
## unused argument (mehods = "special")

The key when designing generics is to identify or anticipate the greatest common factor across methods and formalize that at the level of the generic. When in doubt, better to underestimate than overestimate. For methods, use the most restrictive signature possible.

Project Statistics

Website traffic

The following compares the number of sessions and new users from the third quarter of 2015 (July 1 - September 25) with the third quarter of 2015. Sessions are broken down by new and returning visitors. New visitors correspond to the total new users.

Third Quarter Website Traffic 2015 vs 2014
Sessions: Total 28.05% increase (339,991 vs 265,507)
Sessions: Returning Visitor 34.03% increase (227,171 vs 169,498)
Sessions: New Visitor 17.50% increase (112,820 vs 96,009)

Statistics generated with Google Analytics.

Package downloads and new submissions

The number of unique IP downloads of software packages for July, August and September of 2015 were 35349, 33600 and 29953 respectively. For the same time period in 2014, numbers were 36362, 36118 and 47918. Numbers must be compared by month (vs sum) because some IPs are the same between months. See the web site for a full summary of download stats.

As of September 25, a total of 54 software packages have been added in the third quarter of 2015 bringing counts to 1078 in devel (Bioconductor 3.2) and 1024 in release (Bioconductor 3.1).

News, Events and Courses

See the events page for a listing of all courses and conferences.

Please send comments or questions to Valerie.