BioBits

The Road to Genomic Based Cancer Care

2011-12-17T04:54:00.000-08:00

A blog post on clinical-based sequencing at www.massgenomics.org (The Great Divide: Cancer Genomics and Clinical Care) recently caught my eye. The author (Dan Koboldt at Washington University in St. Louis) described the efforts of academic cancer centers, such as Johns Hopkins, to screen all incoming patients for “clinically actionionable” sequence mutations. Dan applauded the efforts, but made a compelling case that they did not go far enough. Why not go for whole genome or at least whole-exome?

Dan’s post was November 23, 2011. Much has happened since then.

First, Arul Chinnaiyan and colleagues at the University of Michigan published the initial results of their Michigan Oncology Sequencing Project (MI-ONCOSEQ) protocol. The Michigan group tested their protocol on two tumor-derived xenografts, and two patients and performed whole genome sequencing of the tumor, whole exome sequencing of tumor and normal, and RNA-Seq. Turn-around time for all genomic profiling was approximately 3-4 weeks. Genomic results were then interpreted and debated by a multidisciplinary Sequencing Tumor Board (STB), consisting of experts in clinical oncology, pathology, cancer biology, bioethics, bioinformatics and clinical genetics. The STB was specifically charged with interpreting the genomic findings, and determining if the patient could benefit from any current or future clinical trials. If you want to understand how genomic-based medicine could actually be done in the very near future, this is a must read.

Second, on December 6, the NHGRI announced $40 million in funding over the next four years to support “Clinical Sequencing Exploratory Research Projects”. Five research projects were funded, each focusing on the clinical implications of next-generation sequencing. For example, how can hospitals integrate genomic data into medical records, how does genomic data affect the course of treatment, and how do patients process and understand genomic data? Funding will go to Baylor College of Medicine, Brigham and Woman's Hospital, Children's Hospital of Philadelphia, University of North Carolina, and University of Washington, Seattle. In one particularly compelling example, UNC will be performing whole exome sequencing on ~750 patients, primarily focusing on younger cancer patients and patients with a strong family history of cancer.

Both of these developments are real, not science fiction, and very encouraging. MI-ONCOSEQ is one of the most concrete examples of how next-generation sequencing could be integrated into clinical care and benefit real patients. The plans for the NHGRI-funded projects could also provide a real model for future genomic-based cancer care. When you add these developments to the growing list of new industry initiatives that aim to perform or analyze next-generation cancer genomics data (including Knome, Personal Genome Diagnostics, and the KEW Group), the future of clinical-based cancer genomics appears very promising. This is tempered, however by the reality that much of cancer genomics has yet to be translated into new drugs, new drug combinations and new clinical trials. The great divide in cancer care may therefore not necessarily be sequencing. This will probably happen sooner than we all think. Rather, the great divide is likely to be translating new genomic data into new treatments and new clinical trials, and extracting the tiny bits of a person’s genome that are actually useful and relevant to making them better. That will be much harder, and there is much work to be done. So, get back to work.

Relevant Links:

Sameek Roychowdhury et al. Personalized Oncology Through Integrative High-Throughput Sequencing: A Pilot Study. Science Translational Medicine, Nov. 30, 2011. First results from the Michigan MI-ONCOSEQ protocol for genome-based cancer care.

UNC scientists funded to study genome sequencing in clinical settings. Describes UNC’s plans for clinical sequencing. One of five recipients of the NHGRI Clinical Sequencing Exploratory Research Project.

NHGRI broadens sequencing program focus on inherited diseases, medical applications. Press release from NHGRI, describing the Clinical Sequencing Exploratory Research Project

The Great Divide: Cancer Genomics and Clinical Care. Blog post, making the case for routine whole genome sequencing in cancer care.

Making Genomics Routine in Cancer Care. Profiles the KEW Group (co-founded by Raju Kucherlapati at Harvard Medical School) and their plans to integrate genomics into cancer care across a network of cancer centers.

A new piece of the PARP Puzzle

2011-08-31T12:50:00.000-07:00

A recent paper by Gelman et. al. offers compelling new evidence for the effective use of PARP inhibitors in treating ovarian cancer. PARP inhibitors have been previously identified as effective in ovarian cancer patients with germline BRCA1 and BRCA2 mutations. However, only ~16% of ovarian patients harbor a BRCA germline defect. The study by Gelman et. al. therefore set out to identify if patients without germline defects --- possibly with other DNA repair defects --- would respond to the PARP inhibitor Olaparib.

The study included 91 patients (65 with ovarian cancer and 26 with breast cancer), and found that a portion of ovarian cancer patients without germline BRCA defects did indeed respond. Specifcally:

"confirmed objective responses were seen in seven (41%; 95% CI 22-64) of 17 patients with BRCA1 or BRCA2 mutations and 11 (24%; 14-38) of 46 without mutations."

This is quite an exciting finding -- it means that PARP inhibitors could be used to treat a much larger subset of ovarian cancer patients -- and not just those with BRCA germline defects. Of course, these are still early clinical findings, and further clinical trials will be needed, and are no doubt already underway.

This study also dovetails well with the recent TCGA ovarian findings, indicating that BRCA or Homologous Recombination (HR) defects may be present in up to 50% of high-grade serous ovarian patients. What exactly do the non-BRCA responders have in common? Do they share a common genomic event, such as BRCA1 hypermethylation, a somatic BRCA mutation, or some other as yet unknown genomic event? And, can the HR genomic events identified by TCGA be used to stratify responders and non-responders?

As stated to by Gelman, et. al, repeat biopsies of many patients in this study were taken, and whole transciptome sequencing is already underway. We are therefore one step closer to answering many of these questions and discovering genomic biomarkers predictive of PARP response.

References:

Gelmon KA, Tischkowitz M, Mackay H, et. al. Olaparib in patients with recurrent high-grade serous or poorly differentiated ovarian carcinoma or triple-negative breast cancer: a phase 2, multicentre, open-label, non-randomised study. Lancet Oncol. 2011 Sep;12(9):852-61. Epub 2011 Aug 19. [Abstract].

Cancer Genome Atlas Research Network. Integrated genomic analyses of ovarian carcinoma. Nature. 2011 Jun 29;474(7353):609-15. [Abstract].

Branching with Mercurial: A (Sandbox) Tutorial

2011-08-26T10:22:00.000-07:00

Mercurial is a great distributed source control management tool. According to many, one of its great strengths lies in its ability to do branching.

There are a number of good posts on the subject:

Managing releases and branchy development, from Mercurial: The Definitive Guide, by Bryan O'Sullivan;
Mercurial Workflows: Stable & Default, by Steve Losh;
Branching, from the Mercurial web site.

If you are like me, however, you probably don't want your first experimental steps in branching done on a real mercurial repository. Much better to play around in a safe sandbox, where you can try things out, and freely experiment.

To that end, I have put together a short sandbox tutorial that shows how to do branching with a local mercurial repository that you can create from scratch.

In the tutorial, I follow Steve Losh's recommended set-up of having just two branches: default, which is the branch for new development; and stable, which is the branch for stable code, which requires occasional, but high-priority bug fixes.

Getting Started:

Create a new repository from scratch:

>mkdir sandbox

>cd sandbox

>hg init

Create a new file: hello_world.txt. Its contents:

hello1
hello2
hello3

Add and commit the file:

>hg add hello_world.txt
>hg commit -m "Initial commit" hello_world.txt 

Creating a stable branch:

A few days go by, and our code base is ready for a major release. We want to keep this code base "stable", so that we can add quick bug fixes later. So, we branch and commit:

>hg branch stable

marked working directory as branch stable 

>hg commit -m "Created stable branch"

To see all the available branches, and verify our change went into effect:

>hg branches

stable    1:1c25f1208528 
default   0:3afd60d8ec4a 

Applying a bug fix to the stable branch, and merging with default:

A few more days go by. Development continues on the default branch. We discover a major bug, and have to push it out quickly. We first get the stable branch code:

> hg update -C stable

1 files updated, 0 files merged, 0 files removed, 0 files unresolved

and, again verify what branch we are working on:

> hg branch

stable

We now make our major bug fix. We modify hello_world.txt as follows:

hello1: major bug fix
hello2
hello3

We commit the change:

>hg commit -m "Fixed bug" hello_world.txt

This is an important bug fix, so we need to merge the change in stable back into default.

First, we switch to the default branch and do a complete update:

>hg update -C default

 1 files updated, 0 files merged, 0 files removed, 0 files unresolved

Now merge with the stable branch:

>hg merge stable
1 files updated, 0 files merged, 0 files removed, 0 files unresolved
(branch merge, don't forget to commit)

And, commit the changes:

hg commit -m "Merging with stable branch"

The bug fix has now been applied to both stable and default, and we are done.

Major Releases:

There is only one wrinkle left to work out. A few more weeks go by... The default branch is now ready for release. We now want the default branch to be the new stable branch. What to do? We check out the stable branch and merge in all the changes from the default branch.

First, update to the stable branch:

>hg update -C stable

Then, merge everything from default into stable:

>hg merge default 

and, commit the changes:

hg commit -m "Merging in default"

Repeat as necessary. Hopefully, the sandbox set-up allows you enough freedom to play around before working on a real repository.

The Unsung Heroes of GO

2011-08-26T06:35:00.000-07:00

The Gene Ontology (GO) is so pervasive in bioinformatics, that I imagine most of us simply take it for granted. However, it's instructive to realize that GO has only been around since 1999, and far from being a static ontology reflecting unchanging biology, GO has continually evolved since its inception to keep pace with new biological research.

A recent paper by Leonelli et. al. drives this point home, describing how the GO ontology has evolved over time. The paper describes multiple "ontology shifts", including dealing with biological anomalies, expanding biological scope, and mirroring scientific advance --- all of which need to be integrated and managed in order to keep GO up to date. This is serious, hard work, requiring outreach to many in the biological community. Imagine getting a roomful of biologists to agree on the meaning of "gametogenesis", and you can start to see where the real work of ontology development takes place.

So, the next time you run GO enrichment analysis, or any of the other dozens of GO tools, think of all the hard work involved and give a little thanks to the unsung heroes of GO.

References:

Leonelli S, Diehl AD, Christie KR, Harris MA, Lomax J., How the Gene Ontology Evolves. BMC Bioinformatics. 2011 Aug 5;12(1):325. [Abstract] [Full Text].

ExpressionPlot: New web-based tool for analyzing RNA-Seq Data

2011-08-23T04:08:00.000-07:00

Genome Biology has new paper on ExpressionPlot: a web-based tool for analyzing RNA-Seq data.

ExpressionPlot includes a back-end processing pipeline for importing and processing raw RNA-Seq and Affymetrix microarray based data, plus a front-end web user interface for analyzing and visualizing the results.

You can check out the web interface via the prototype server, which is pre-loaded with a number of RNA-Seq and Affymetrix data sets. More details and installation instructions are available at: http://www.expressionplot.com/.

Notably, the authors submitted ExpressionPlot to iDEA (Illumina’s Data Excellence Award) Challenge for "Most Creative Visualization", but the award was eventually awarded to GenomeView.

References:

Friedman BA, Maniatis T. ExpressionPlot: A web-based framework for analysis of RNA-Seq and microarray gene expression data. Genome Biol. 2011 Jul 28;12(7):R69. [Abstract] [Full Text].

Project Achilles and the Search for Essential Genes in Cancer

2011-08-17T11:50:00.000-07:00

A recent PNAS paper by Cheung et. al. reports the latest findings from the Broad Institute's Project Achilles. The goal of the project is to use high-throughput RNAi screening to systematically identify genes essential for proliferation and survival in multiple cancer cell lines.

There are (obviously) many details involved in the RNAi screen developed for Project Achilles, and you find a complete description in the first PNAS paper. However, the simplest explanation is that the screen takes one cell line at a time, systematically knocks down one gene at a time (via RNAi), and then measures what effect (if any) this specific knock-down has on cell-proliferation. If a knock-down results in a dramatic decrease in cell-proliferation, one can infer that the target gene is functionally important or essential to the cancer phenotype. Significantly, one can then also combine the complete set of functional data with primary tumor data, e.g. mutation or copy number data, to hone in on those primary genomic alterations most likely to drive tumorigenesis.

The PNAS paper reports findings for 102 human cancer cell lines, each of which was screened with 54,020 shRNAs, targeting a total of 11,195 genes. Complete data from the project is available via the Broad's Integrative Genomics Portal.

Other than the data sets themselves, which will be hugely important for helping to interpret future cancer genomic studies, the most significant findings of the paper relate to ovarian cancer. Specifically, by combining RNAi data with primary genomic data from TCGA, the group made the novel observation that PAX8 is both frequently amplified in serous ovarian cancer, and essential to proliferation in multiple ovarian cancer cell cells.

References:

Cheung, et. al, Systematic investigation of genetic vulnerabilities across cancer cell lines reveals lineage-specific dependencies in ovarian cancer. (PNAS, July 11, 2011) [Abstract].

Background on RNAi: Five questions for David Root: RNA Interference explained.

Bioinformatics Challenges for Pesonalized Medicine

2011-07-14T17:38:00.000-07:00

A recent review by Fernald et. al. describes key bioinformatics challenges for personalized medicine. The key challenges break down into four categories, and in the discussion below, I have recast these categories as questions:

How can we better process large-scale genomic data? Next generation sequencing technology may be one of the most profound enablers of personalized medicine, but there remain significant challenges in processing read data. Challenges include accurate mapping and alignment of short-read sequences; identification of indels, copy number variants and structural variants; and improved quality control metrics to adjust with the inherent errors rates associated with next-gen sequencing technology.

How can we better predict the functional effect of specific genomic variations? Key challenges include developing better computational methods to predict the functional consequences of genomic variants, as well as better methods for prioritizing gene candidates for experimental validation.

How can we better translate genomic insights into disease-specific insights? Key challenges include use of genome wide association study (GWAS) approaches to identify patient response to drugs, integrating multiple genomic data types, and integrating biological pathway data.

How do we get all this to real patients? Here, the authors note that many physicians are currently unprepared to incorporate personal genomic information into regular clinical practice. However, they also not some significant advances, in particular the use of SNP testing to determine dosage of the anti-coagulant drug warfarin.

The road to personalized medicine is certainly fraught with challenges. We may soon be able to economically sequence the full genome of every patient, but bioinformatics has a long way to go to make full sense of this data.

One minor quibble with this paper if that it pays scant attention to cancer genomics, one of the few areas in which we are already seeing and are very likely to see additional examples in personalized medicine.

Chromothripsis

2011-07-05T07:47:00.000-07:00

This is a fascinating Cell paper from the Sanger Institute, published in January 2011. As depicted in the graphical abstract (shown at right), the authors find evidence for a new type of genomic rearrangement, referred to as chromothripsis: (Greek, chromos for chromosome; thripsis, shattering into pieces).

In chromothripsis, a single catastrophic event causes the shattering of one of more chromosomes. Cells then attempt to repair the shattered chromosome, resulting in 10s to 100s of stitched DNA fragments (many of which are detected up by paired-end RNA sequencing). That cells are able to survive such a catostrophic event would appear very unlikely, but a few cells clearly do, and the authors provide evidence that these cells not only survive, but have some type of selective advantage that promotes cancer development.

The current model of cancer development assumes a model of "gradualism" whereby cells acquire mutations through a series of progressive steps. In the case of chromothripsis, however, we have a new model, whereby cells acquire multiple genomic alterations in a single catastrophic step. This is quite new. However, according to the authors, chromothripsis may not be all that prevalent. For example, they find evidence of chromothripsis in 2%-3% of cells lines and primary tumor samples, but a much larger 25% of bone cancers.

The paper describes the major steps in first identifying chromothripsis. This includes an initial paired-end RNA-seq analysis of a patient with Chronic Lympocytic Leukemia (CLL), and the identification of a strange pattern of rearrangements. The observations included multiple genomic rearrangements, all confined to a single chromosomal region, chromosome 4q, and alteration of copy number states between 1 and 2 only.

The authors next searched for a similar pattern of genomic rearrangement in a large panel of cell lines and primary tumors, and selected a few cell lines for further RNA-seq analysis. The paper also includes a Monte Carlo simulation providing simulated evidence that the large-scale rearrangements observed would be very unlikely to occur under the current model of "gradual" rearrangements.

Using RNA-Seq to Identify Fusion Genes in Breast Cancer

2011-07-01T06:24:00.000-07:00

The Institute for Molecular Medicine Finland has recently published a paper in Genome Biology on using RNA-Seq to identify fusion genes in breast cancer.

As noted by the authors, fusion genes have been identified as important oncogenic alterations in leukemias, lyphompas, sarcomas, and prostate cancer. The role of fusion genes in breast cancer has not, however, been fully explored.

Paired-end RNA sequencing is a relatively new and useful tool in systematically identifying fusion genes in cell lines and primary cancer samples. In this technique, one only sequences expressed genes, and one sequences 36 to 100 basepairs of both ends of 200 to 500 base pair long molecules. One then maps each end to the genome. If one end maps to one gene, and the other end maps to another gene, and these genes are a sufficient distance apart (e.g. they are not directly adjacent), and you have multiple reads to support this, you have evidence of a gene fusion (see figure above).

The Finland study focused on well-known breast cancer cell lines, including MCF-7, BT-474, and KPL-4. Each cell line was analyzed via RNA-Seq, reads were mapped to the genome, putative novel gene fusions were identified and many gene fusions were then validated via RT-PCR. One novel gene fusion, involving VAPB-IKZF3 (highlighted at left) was also knocked down via RNA interference, resulting in a direct impact on cell growth.

The next step, of course will be to determine if these novel fusion genes exist in primary breast tumors, and with what frequency. But, this paper provides an interesting insight into many things to come, and we are very likely to soon see comprehensive search for gene fusions via RNA-seq added to many large-scale projects, including the TCGA.

g:Profiler: 2011 Update

2011-06-30T10:35:00.000-07:00

g:Profiler is a web-based tool for analyzing lists of genes. A recent paper in Nucleic Acids Research describes its 2011 update. It's always great to see software updates in bioinformatics, as many projects lose funding or lose key people, or both, and never really progress further. Not so for g:Profiler, which has added a number of new features and data sets.

The core feature of g:Profiler remains its g:GOSt Gene Group Functional Profiling feature. This feature enables you to upload a set of genes, e.g. differentially expressed genes from a microarray study, and identify enriched gene sets (determined via a hypergeometic test) across multiple resources, including Gene Ontology, KEGG or Reactome pathways, etc.

The 2011 update also includes an interesting network enrichment feature, which determines whether your input gene list is enriched for interacting genes, providing evidence that your input gene list contains specific network modules. The network enrichment test is novel in that it identifies "core" genes, defined as those genes in the input list which interact with each other v. "neighborhood" genes, defined as all partners of core genes, which are not in the original input list. A tally of all interacting genes in the input list is then compared to a tally of all interacting genes in the neighborhood, and a hypergeometric test is then used to determine statistical enrichment.

g:Profiler also includes a very useful ID conversion service, a programmatic web service, and an R package.

PathScan: New Gene Set Analysis Method from Wash U

2011-06-29T06:18:00.000-07:00

Wash U. has recently published a new method, PathScan for determining the mutational significance of gene sets in cancer.

The method belongs to the broad category of gene set analysis methods in cancer genomics, whereby one takes observed mutations in cancer and compares these mutations against predefined gene sets, as defined by, e.g. KEGG or MSigDB. One then determines whether each predefined gene set is enriched for mutations.

In contrast to other gene set analysis methods, PathScan has two distinct features:

First, it takes into account the variability in gene length. This is biologically grounded in the fact that larger genes contain more mutations under the null hypothesis of random mutations.

Second, PathScan does not use a "tally" based or "pool-based" technique. In these techniques, one tallies all the mutations within a gene set across all samples, and performs one statistical test. Rather PathScan performs a statistical test on each sample individually, and then integrates the results of each test into an overall p-value.

The authors make a statistical case for the second feature, but not necessarily a biological one. And, here I wish the authors went into more detail about the biological basis of their approach. For example, if one performs a statistical test on a single sample, and a pathway is identified as significant, this implies that multiple genes within the pathway are mutated in this one sample.

By this logic, PathScan focuses of identifying a specific class of pathway alterations --- for example, a single gene mutation is not enough to provide tumorigenic advantage, but 2 or 3 mutations are. By focusing on this class of alterations, PathScan can easily miss other types of alterations, e.g. where one dominant mutation is enough to cause tumorigenic advantage, or where 2 or 3 low-frequency, but mutually exclusive mutations predominate, any one of which can result it tumorigenic advantage.

A more thorough description of the classes of pathway alterations that would be identified and potentially missed by PathScan would definitely clarify the paper. The paper would also benefit from clarifying how the Background Mutation Rate (BMR) is estimated. Is it estimated from all samples? Or, does the method use a sample-specific BMR? If it is estimated from all samples, the sample-specific statistical test would appear flawed. However, if it is a sample-specific BMR, this is another novel feature that should more clear be highlighted.

PathScan software is available as Perl code, and available from Wash U.

Three Takes on Cloud Computing and Genomics

2011-06-28T05:56:00.000-07:00

If you want to get a good take on cloud computing, the best option is to simply head over to Amazon Web Services, create an account, and poke around within the free usage tier.

For greater context on cloud computing and genomics, check out these three recent papers:

The case for cloud computing in genomic informatics (May 2010): A frequently cited paper by Lincoln Stein (Ontario Institute for Cancer Research), describing how advances in next-gen sequencing threaten the conventional genomics informatics ecosystem, and why cloud computing may provide a feasible and cost-effective option. Paper also includes a frequently referenced figure illustrating historical trends in storage prices v. DNA sequencing costs.

Cloud computing and DNA data race (July 2010): A concise overview of cloud computing, from the makers of Crossbow, a cloud-compute pipeline for genome analysis. Includes a concise overview of the MapReduce algorithm, and the best line opener of the bunch: "In the race between DNA sequencing and computer speed, sequencing is winning by a mile."

Translational bioinformatics in the cloud: an affordable alternative (August 2010). An apples-to-apples comparison of doing an expression quantitative trait loci (eQTL) analysis on a local compute cluster v. Amazon EC2. Illustrates just how hard it is to do a cost-comparison. It's relatively easy to determine the cost of Amazon EC2, as you get an itemized bill. But, estimating the cost of a local computer cluster is actually quite involved, and requires that you take into account many difficult to measure factors, including staff time, electricity, back-up power systems, etc.

The genesis and evolution of high-grade serous ovarian cancer

2011-06-27T05:48:00.000-07:00

In anticipation of the TCGA ovarian paper, scheduled for release in Nature on June 30, 2011, check out David Bowtell's recent piece in Nature Reviews Center on The genesis and evolution of high-grade serous ovarian cancer. It provides an excellent background and context for understanding the TCGA study.

A summary of the main points discussed in the paper:

worldwide, more than 100,000 women die each year of ovarian cancer (the 5th leading cause of cancer death).
two-thirds of ovarian cancer deaths are associated with high-grade serous subtype (the focus of TCGA profiling).
TP53 mutation is a primary hallmark of high-grade serous.
homologous recombination repair defects (including, but not limited to BRCA mutations) are a secondary hallmark of high-grade serous.
high-grade serous tumors are marked by pronounced genomic instability.
recent studies have identified four robust mRNA expression sub-types of high-grade serous.
understanding extent of BRCA pathway defects has important clinical implications for use of PARP inhibitors.
"a comprehensive catalog of all possible mechanisms of inactivation of the HR DNA repair in a large series of HG-SOC is needed." (something that TCGA focused on, and will be included in June 30 paper).
many genomic parallels between high-grade serous and basal-like breast cancer, including BRCA dsyfunction, TP53 loss, and chromosomal instability.

Creating Integrated PDF Reports in R

2011-05-27T07:41:00.000-07:00

This R-Hack shows how to create integrated reports of both plots and results of statistical tests in one integrated PDF.

It works by using the gplots library. The textplot() command displays text output in a graphics window or a previously opened PDF device. The sinkplot() command diverts standard text output to a graphics device, such as a previously opened PDF device.

To use sinkplot(), you call sinkplot("start"). After this, any commands which would have been displayed to the screen will be diverted to the graphical device. You then call sinkplot("plot") to signal the end of diversion.

When you prepend all this with a call to pdf([file name]), you can divert everything into one integrated, multi-page PDF. This enables you to create integrated reports of your statistical work, and I find it is considerably easier than any of the other more sophisticated options, e.g. Sweave.

A complete, bare bones example is shown below. It generates this PDF:


#!/usr/bin/Rscript --no-save
# Illustrates How to Create an Integrated PDF Report in R
# Ethan Cerami

library ("gplots")

# Save all results to External PDF
pdf("report.pdf")

# Create Title Page for Report via textplot() command
title = "Sleep Report\n\n"
line1 = "This report includes:  \n\n"
line2 = "1.  Box Plot\n"
line3 = "2.  T-Test"
cover = paste (title, line1, line2, line3)

# Output via textplot
textplot (cover, col="darkblue", valign="top", cex=1.5)

# Create the Box Plot
plot(extra ~ group, data = sleep, main="Sleep Data")

# Perform a t-test
# Start capture
sinkplot("start")
t.test(extra ~ group, data = sleep)
noquote ("If p<0.05, we can conclude that the two groups are different.")
sinkplot("plot",col="darkblue", valign="top", cex=0.8)
title("t-test")

# End PDF
dev.off()

Plotting with Python (move over R)

2011-05-19T13:28:00.000-07:00

R is a great platform for statistics. But, I still find the syntax non-intuitive, and can never seem to "just do" anything without lots of referencing to old code or googling. Contrast this to Python, which is so simple and spare. Even if you can't remember how to do something, you think of the simplest way to do it, and that's probably the correct syntax in Python.

With that in mind, I have been trying out matplotlib as an alternative for all the statistics plots that I would normally do in R. It is beautiful. Simple. And, much easier than R.

Tip: if you are on Mac OS X, and want an easy way to install matplotlib, check out Scipy Superpack for Mac OSX.

Here is example code from the getting started guide:

# Do a simple plot
import matplotlib.pyplot as plt
plt.plot([1,2,3,4], [1,4,9,16], 'ro')
plt.axis([0, 6, 0, 20])
plt.show()

This plots x v. y, specifies the points as red circles, and sets the x/y axis:

That was easy enough to convince me.

Here is a longer example that illustrates how to plot mRNA v. Methylation data from the TCGA ovarian project. Most of the code gets and parses the data from our cBio Cancer Genomics Portal, and the last few lines are focused on plotting:

Here is output for BRCA1:

Evaluating Amazon SimpleDB

2011-05-10T10:12:00.000-07:00

I am new to cloud computing in general, but have spent the past few weeks evaluating Amazon SimpleDB. Below are some of my initial impressions, and my current verdict.

The Context: Many Database Options to Choose From

Amazon Web Services provides a host of database options, including Amazon SimpleDB, Amazon Relational Database Service (RDS), and the ability to install and manage multiple databases, including MySQL, SQL Server, Oracle, and PostgreSQL via Amazon Machine Images (AMI). With so many options, it can be a bit overwhelming to figure out where exactly so start. Amazon does provide a short guide to help you better understand the options, and figure out which option might be best for your particular project. However, to really understand any of these options, there is no substitute for hands-on experience. Fortunately, you can get that for free via the Amazon Free Usage Tier.

Amazon SimpleDB: The Basics

First, imagine that you cannot create traditional relational database tables. Rather, you are restricted to putting and selecting items with key/value attribute pairs only, similar to a giant hashmap. Then, imagine that all your data is only accessible via web service calls --- you "put" data into some giant black box that is managed by Amazon, and that the black box has many complex details that are completely hidden from you. That is Amazon SimpleDB.

In a nutshell, Amazon SimpleDB is therefore a NoSQL database that runs within the Amazon cloud. You focus on putting and selecting data. Amazon takes care of managing the infrastructure, scaling to meet user demands, and managing all back-ups. For all that, Amazon charges you on a "pay-as-you-go" rate schedule.

Famously, NetFlix now uses Amazon SimpleDB, and has been quite open about challenges and lessons learned in transitioning to Amazon SimpleDB.

Amazon SimpleDB: Getting Started

I was able to get started with Amazon SimpleDB in just a few days. The documentation is excellent, and very easy to follow. Amazon even has a Javascript "Scratchpad", that enables you to interact with SimpleDB via a web user interface. You can therefore get started without having to write any code, and the scratchpad examples walk you through the steps of creating a new domain, putting data into a domain, and then getting it back.

When Simple turns Complicated

If you have millions of records you want to pre-load into Amazon SimpleDB, things take a decidedly complicated turn. In the world of relational databases, such as MySQL or Oracle, we have all come to expect simple, high performance command line tools, such as mysqlimport for importing large sets of data. As of now, there is no such equivalent for SimpleDB, and you have to roll your own. If you want high-performance, you also have to write some rather complex code, handling all sorts of issues, including data sharding (e.g. placing data into multiple domains), multi-threaded puts, and a measured way of backing off of Amazon SimpleDB if you throttle its service with too many requests at once.

This is a lot of work. And, even then, it's still not clear whether can get the write-performance you really need. For example, the high-performance example from Amazon claims a write-throughput of 300 requests per second. Pete Warden has also open sourced his Java code for multi-threaded parallel inserts, and has been able to achieve a write-throughput of 1,000 records / second. Depending on how much data you have, that may be fine. But, consider that if you have 100 million records, you are talking about 28 hours just to load the data (NetFlix claims that they have achieved a much higher write-throughput of 10K-11K / second, but they have not released all the details on how they did this).

The Verdict

Amazon SimpleDB has a lot of potential. I found the documentation straightforward, and was up and running in no time. However, if you have millions of records to import, be prepared to deal with lots of headaches. Until Amazon comes up with a simpler, faster way of doing bulk inserts, you are likely to be writing your own import code, and are very likely to be underwhelmed with the write-throughput provided by Amazon SimpleDB.

cBio Cancer Genomics Portal

2010-12-22T12:16:00.000-08:00

We are releasing an upgrade to the cBio Cancer Genomics Portal today. Many new features and new data sets: http://www.cbioportal.org/cgx/news.jsp.

Check it out and send us your feedback.

Advances in understanding cancer genomes through second-generation sequencing

2010-12-14T18:06:00.000-08:00

An excellent review paper covering the current state and future of next-generation sequencing in cancer genomics. Among the many useful topics covered:

cancer-specific considerations in sequencing, e.g. dealing with tumor heterogeneity, purity and ploidy.
pair-end sequencing, whole genome sequencing, RNA-seq.
computational issues related to alignment and mutation detection.
useful software for alignment, mutation calling, and visualization.

Network fragility of cancer genes

2009-09-09T13:29:00.000-07:00

A recent paper by Davide Rambaldi et. al. on the network properties of cancer genes is deceptively concise, but quite interesting. Given that recent unbiased screenings of cancer genomes have revealed the heterogeneity of genes altered in individual tumors, the paper asks a relatively simple question: do these genes actually share more in common than meets the eye? More specifically, do these genes share common network properties or gene duplicability properties (e.g. do the genes tend to exist as singletons, or are they duplicated many times within the genome?)

The authors find that cancer genes tend to be both singletons and highly connected protein hubs. These properties suggest an "intrinsic fragility of cancer genes toward perturbations: gene dosage modifications of highly interconnected hubs (but arguably also sequence mutations) are likely to produce simultaneous effects on several processes."

Furthermore, since there are not many genes that share these two properties within the human genome, the authors go a step further to identify a set of 101 potential novel candidate cancer genes.

Minireview: Network medicine approach to human disease

2009-09-01T07:24:00.000-07:00

FEBS Letters has a review on the use of network analysis to understand and treat human disease. The first part of the review summarizes recent research on the role of network connectivity and its connection to human disease -- for example, somatic mutations in cancer are more likely to occur within protein hubs and to occur within the same sub-network. The second part summarizes recent research in using subnetworks as biomarkers -- for example, Trey Ideker's 2007 work in identifying discriminating subnetworks in breast cancer. The review concludes with more a speculative overview of network / pathway-centric therapeutics, an area which is currently receiving much attention, but lacking in concrete examples.

2nd Generation Yeast Two-Hybrid (Y2H) Networks

2009-08-31T06:11:00.000-07:00

The yeast two-hyrbid (Y2H) assay for detecting protein-protein interactions has been heavily criticized for its poor quality, and its potentially high rate of both false positives and false negatives. A 2008 Science article out of Marc Vidal's group at Harvard attempts to set the record straight. The paper is divided into three parts. In part I, the authors systematically compare the quality of interactions derived from Y2H, affinity purification followed by mass spectrometry (AP/MS), and literature curation, and conclude that "high-throughput Y2H data sets can be comparable in quality to literature curated information." In part II, the authors conduct a quality controlled, "second generation" Y2H screen in yeast that covers ~20% of all yeast binary interactions. In part III, the authors revisit the global network properties of the newly derived yeast network and discover that protein connectivity does not actually correctly with essentiality as previously described, but does correlate with genetic pleiotropy.

A curious side finding in the paper is that "literature-curated interactions seem prone to sociological and other inspection biases", a finding which is followed up in a subsequent paper.

High-Quality Binary Protein Interaction Map of the Yeast Interactome Network

For the full network, go to: Interactome Projects at Harvard CCSB

The Biology Of Ovarian Cancer

2009-08-28T09:16:00.000-07:00

Nature Reviews Cancer has an excellent review of ovarian cancer biology in its June 2009 issue. The review is quite timely, as the The Cancer Genome Atlas Project (TCGA) is now in full swing generating comprehensive genomic profiles for hundreds of ovarian tumor samples. A few highlights of the review:

Box 1 provides an overview of early-detection screening in ovarian cancer, and the importance such screening could have for improving the overall cure rate.

Box 2 describes the four histotypes of epithelial ovarian cancer (the TCGA project is focusing on only one of these: serous cystadenocarcinoma.)

Box 4 outlines the major signaling pathways altered in >50% of ovarian cancers. This includes PI3K, Src, and Jak-Stat. The main text also summarizes mutations and alterations in genes affecting cell proliferation, apoptosis, anoikis, autophagy, angiogenesis, cell motility, and invasion.

The article also summarizes recent clinical trials in ovarian cancer, including for example, new PI3K and AKT inhibitors now entering Phase I-II clinical trials.

The biology of ovarian cancer: new opportunities for translation

New NIH Director Collins Gets to Work

2009-08-26T05:55:00.000-07:00

Science Insider has a news article on Francis Collins, President Obama's newly appointed director of the NIH. The article briefly outlines Collins's five main themes for the NIH, including health care reform, translating research into medicine, global health, "empowering the biomedical research community", and "large biology" projects. The article also briefly touches on Collins's concern for NIH funding: "what happens when the $10.4 billion that NIH received as part of this year's stimulus package runs out in 2011."

http://blogs.sciencemag.org/scienceinsider/2009/08/collins-sets-fi.html