Evolutionary Computation and Data Mining in Biology

For over 15 years, members of the computer science, machine learning, and data mining communities have gathered in a beautiful European location each spring to share ideas about biologically-inspired computation.  Stemming from the work of John Holland who pioneered the field of genetic algorithms, multiple approaches have been developed that exploit the dynamics of natural systems to solve computational problems.  These algorithms have been applied in a wide variety of fields, and to celebrate and cross-pollinate ideas from these various disciplines the EvoStar event co-locates five conferences at the same venue, covering genetic programming (EuroGP), combinatorial optimization (EvoCOP), music, art, and design (EvoMUSART), multidisciplinary applications (EvoApplications), and computational biology (EvoBIO).  EvoStar 2013 will be held in Vienna, Austria on April 3-5, and is always expertly coordinated by the wonderful Jennifer Willies from Napier University, UK. Multiple research groups from the US and Europe will attend to present their exciting work in these areas.

Many problems in bioinformatics and statistical analysis use what are considered “greedy” algorithms to fit parameters to data – that is, they settle on a nearby collection of parameters as the solution and potentially miss a global best solution.  This problem is well-known in the computer science community for toy problems like bin packing or the knapsack problem.  In human genetics, related problems are partitioning complex pedigrees or selecting maximally unrelated individuals from a dataset, and can also appear when maximizing likelihood equations.

EvoBIO focuses on using biologically-inspired algorithms (like genetic algorithms) to improve performance for many bioinformatics tasks.  For example, Stephen and I have both applied these methods for analysis of genetic data using neural networks, and for forward-time genetic data simulation (additional details here).

EvoBIO is very pleased to be sponsored by BMC Biodata Mining, a natural partner for this conference.  I recently wrote a blog post for BioMed Central about EvoBIO as well.  Thanks to their sponsorship, the winner of the EvoBIO best paper award will receive free publication in Biodata Mining, and runners-up will receive 25% discount off the article processing charge.

So, if you are in the mood for a new conference and would like to see and influence some of these creative approaches to data analysis, consider attending EvoSTAR -- We'd love to see you there!

Software Carpentry Bootcamp at University of Virginia

A couple of weeks ago I, with the help of others here at UVA, organized a Software Carpentry bootcamp, instructed by Steve Crouch, Carlos Anderson, and Ben Morris. The day before the course started, Charlottesville was racked by nearly a foot of snow, widespread power outages, and many cancelled incoming flights. Luckily our instructors arrived just in time, and power was (mostly) restored shortly before the boot camp started. Despite the conditions, the course was very well-attended.

Software Carpentry's aim is to teach researchers (usually graduate students) basic computing concepts and skills so that they can get more done in less time, and with less pain. They're a volunteer organization funded by Mozilla and the Sloan foundation, and led this two-day bootcamp completely free of charge to us.

The course started out with a head-first dive into Unix and Bash scripting, followed by a tutorial on automation with Make, concluding the first day with an introduction to Python. The second day covered version control with git, Python code testing, and wrapped up with an introduction to databases and SQL. At the conclusion of the course, participants offered near-universal positive feedback, with the git and Make tutorials being exceptionally popular.

Software Carpentry's approach to teaching these topics is unlike many others that I've seen. Rather than lecturing on for hours, the instructors inject very short (~5 minute) partnered exercises between every ~15 minutes of instruction in 1.5 hour sessions. With two full days of intensive instruction and your computer in front of you, it's all too easy to get distracted by an email, get lost in your everyday responsibilities, and zone out for the rest of the session.  The exercises keep participants paying attention and accountable to their partner.

All of the bootcamp's materials are freely available:

Unix and Bash: https://github.com/redcurry/bash_tutorial
Python Introduction: https://github.com/redcurry/python_tutorial
Git tutorial: https://github.com/redcurry/git_tutorial
Databases & SQL: https://github.com/bendmorris/swc_databases
Everything else: http://users.ecs.soton.ac.uk/stc/SWC/tutorial-materials-virginia.zip

Perhaps more relevant to a broader audience are the online lectures and materials available on the Software Carpentry Website, which include all the above topics, as well as many others.

We capped the course at 50, and had 95 register within a day of opening registration, so we'll likely do this again in the future. I sit in countless meetings where faculty lament how nearly all basic science researchers enter grad school or their postdoc woefully unprepared for this brave new world of data-rich high-throughput science. Self-paced online learning works well for some, but if you're in a department or other organization that could benefit from a free, on-site, intensive introduction to the topics listed above, I highly recommend contacting Software Carpentry and organizing your own bootcamp.

Finally, when organizing an optional section of the course, we let participants vote whether they preferred learning number crunching with NumPy, or SQL/databases; SQL won by a small margin. However, Katherine Holcomb in UVACSE has graciously volunteered to teach a two-hour introduction to NumPy this week, regardless of whether you participated in the boot camp (although some basic Python knowledge is recommended). This (free) short course is this Thursday, March 21, 2-4pm, in the same place as the bootcamp (Brown Library Classroom in Clark Hall). Sign up here.

Comparing Sequence Classification Algorithms for Metagenomics

Metagenomics is the study of DNA collected from environmental samples (e.g., seawater, soil, acid mine drainage, the human gut, sputum, pus, etc.). While traditional microbial genomics typically means sequencing a pure cultured isolate, metagenomics involves taking a culture-free environmental sample and sequencing a single gene (e.g. the 16S rRNA gene), multiple marker genes, or shotgun sequencing everything in the sample in order to determine what's there.

A challenge in shotgun metagenomics analysis is the sequence classification problem: i.e., given a sequence, what's it's origin? I.e., did this sequence read come from E. coli or some other enteric bacteria? Note that sequence classification does not involve genome assembly - sequence classification is done on unassembled reads. If you could perfectly classify the origin of every sequence read in your sample, you would know exactly what organisms are in your environmental sample and how abundant each one is.

The solution to this problem isn't simply BLAST'ing every sequence read that comes off your HiSeq 2500 against NCBI nt/nr. The computational cost of this BLAST search would be many times more expensive than the sequencing itself. There are many algorithms for sequence classification. This paper examines a wide range of the available algorithms and software implementations for sequence classification as applied to metagenomic data:

Bazinet, Adam L., and Michael P. Cummings. "A comparative evaluation of sequence classification programs." BMC Bioinformatics 13.1 (2012): 92.

In this paper, the authors comprehensively evaluated the performance of over 25 programs that fall into three categories: alignment-based, composition-based, and phylogeny-based. For illustrative purposes, the authors constructed a "phylogenetic tree" that shows how each of the 25 methods they evaluated are related to each other:

Figure 1: Program clustering. A neighbor-joining tree that clusters the classification programs based on their similar attributes.

The performance evaluation was done on several different datasets where the composition was known, using a similar set of evaluation criteria (sensitivity = number of correct assignments / number of sequences in the data; precision = number of correct assignments/number of assignments made). They concluded that the performance of particular methods varied widely between datasets due to reasons like highly variable taxonomic composition and diversity, level of sequence representation in underlying databases, read lengths, and read quality. The authors specifically point out that just because some methods lack sensitivity (as they've defined it), they are still useful because they have high precision. For example, marker-based approaches (like Metaphyler) might only classify a small number of reads, but they're highly precise, and may still be enough to accurately recapitulate organismal distribution and abundance.

Importantly, the authors note that you can't ignore computational requirements, which varied by orders of magnitude between methods. Selection of the right method depends on the goals (is sensitivity or precision more important?) and the available resources (time and compute power are never infinite - these are tangible limitations that are imposed in the real world).

This paper was first received at BMC Bioinformatics a year ago, and since then many new methods for sequence classification have been published. Further, this paper only evaluates methods for classification of unassembled reads, and does not evaluate methods that rely on metagenome assembly (that's the subject of another much longer post, but check out Titus Brown's blog for lots more on this topic).

Overall, this paper was a great demonstration of how one might attempt to evaluate many different tools ostensibly aimed at solving the same problem but functioning in completely different ways.

Bazinet, Adam L., and Michael P. Cummings. "A comparative evaluation of sequence classification programs." BMC Bioinformatics 13.1 (2012): 92.

NetGestalt for Data Visualization in the Context of Pathways

Many of you may be familiar with WebGestalt, a wonderful web utility developed by Bing Zhang at Vanderbilt for doing basic gene-set enrichment analyses. Last year, we invited Bing to speak at our annual retreat for the Vanderbilt Graduate Program in Human Genetics, and he did not disappoint! Bing walked us through his new tool called NetGestalt.

NetGestalt provides users with the ability to overlay large-scale experimental data onto biological networks. Data are loaded using continuous and binary tracks that can contain either single or multiple lines of data (called composite tracks). Continuous tracks could be gene expression intensities from microarray data or any other quantitative measure that can be mapped to the genome.  Binary tracks are usually insertion/deletion regions, or called regions like ChIP peaks.  NetGestalt extends many of the features of WebGestalt, including enrichment analysis for modules within a biological network, and provides easy ways to visualize the overlay of multiple tracks with Venn diagrams.

Netgestalt provides a very nice interface for interacting with data. Extensive documentation on how to use it can be found here.  Bing and his colleagues also went the extra mile to create video tutorials on how to use their web tool, and walk you through an analysis of some tumor data.

http://www.netgestalt.org/

"Document Design and Purpose, Not Mechanics"

If you ever write code for scientific computing (chances are you do if you're here), stop what you're doing and spend 8 minutes reading this open-access paper:

Wilson et al. Best Practices for Scientific Computing. arXiv:1210.0530 (2012). (Direct link to PDF).

The paper makes a number of good points regarding software as a tool just like any other lab equipment: it should be built, validated, and used as carefully as any other physical instrumentation. Yet most scientists who write software are self-taught, and haven't been properly trained in fundamental software development skills. 

The paper outlines ten practices every computational biologist should adopt when writing code for research computing. Most of these are the usual suspects that you'd probably guess - using version control, workflow management, writing good documentation, modularizing code into functions, unit testing, agile development, etc. One that particularly jumped out at me was the recommendation to document design and purpose, not mechanics. 

We all know that good comments and documentation is critical for code reproducibility and maintenance, but inline documentation that recapitulates the code is hardly useful. Instead, we should aim to document the underlying ideas, interface, and reasons, not the implementation.

For example, the following commentary is hardly useful:

# Increment the variable "i" by one.

i = i+1

The real recommendation here is that if your code requires such substantial documentation of the actual implementation to be understandable, it's better to spend the time rewriting the code rather than writing a lengthy description of what it does. I'm very guilty of doing this with R code, nesting multiple levels of functions and vector operations:

# It would take a paragraph to explain what this is doing.

# Better to break up into multiple lines of code.

sapply(data.frame(n=sapply(x, function(d) sum(is.na(d)))), function(dd) mean(dd))

It would take much more time to properly document what this is doing than it would take to split the operation into manageable chunks over multiple lines such that the code no longer needs an explanation. We're not playing code golf here - using fewer lines doesn't make you a better programmer.

Best Practices for Scientific Computing

Scotty, We Need More Power! Power, Sample Size, and Coverage Estimation for RNA-Seq

Two of the most common questions at the beginning of an RNA-seq experiments are "how many reads do I need?" and "how many replicates do I need?". This paper describes a web application for designing RNA-seq applications that calculates an appropriate sample size and read depth to satisfy user-defined criteria such as cost, maximum number of reads or replicates attainable, etc. The power and sample size estimations are based on a t-test, which the authors claim, performs no worse than the negative binomial models implemented by popular RNA-seq methods such as DESeq, when there are three or more replicates present. Empirical distributions are taken from either (1) pilot data that the user can upload, or (2) built in publicly available data. The authors find that there is substantial heterogeneity between experiments (technical variation is larger than biological variation in many cases), and that power and sample size estimation will be more accurate when the user provides their own pilot data.

My only complaint, for all the reasons expressed in my previous blog post about why you shouldn't host things like this exclusively on your lab website, is that the code to run this analysis doesn't appear to be available to save, study, modify, maintain, or archive. When lead author Michele Busby leaves Gabor Marth's lab, hopefully the app doesn't fall into the graveyard of computational biology web apps.  Update 2/7/13: Michele Busby created a public Github repository for the Scotty code: https://github.com/mbusby/Scotty

tl;dr? There's a new web app that does power, sample size, and coverage calculations for RNA-seq, but it only works well if the pilot or public data you give it closely matches the actual data you'll collect. 

Paper: Busby, et al. "Scotty: A Web Tool For Designing RNA-Seq Experiments to Measure Differential Gene Expression." Bioinformatics (2013): 10.1093/bioinformatics/btt015.

Web app: http://euler.bc.edu/marthlab/scotty/scotty.php

Source code: https://github.com/mbusby/Scotty

The Pacific Symposium on Biocomputing 2013

For 18 years now, computational biologists have convened on the beautiful islands of Hawaii to present and discuss research emerging from new areas of biomedicine. PSB Conference Chairs Teri Klein (@teriklein), Keith Dunker, Russ Altman (@Rbaltman) and Larry Hunter (@ProfLHunter) organize innovative sessions and tutorials that are always interactive and thought-provoking. This year, sessions included Computational Drug Repositioning, Epigenomics, Aberrant Pathway and Network Activity, Personalized Medicine, Phylogenomics and Population Genomics, Post-Next Generation Sequencing, and Text and Data Mining. The Proceedings are available online here, and a few of the highlights are:

Cheng et al. examine various analytical methods for processing data from the Connectivity Map, a dataset of gene expression changes due to small molecule treatment. They compare methods for identifying drug-induced gene expression profiles to a benchmark based on the Anatomical Theraputic Chemical (ATC) system with the hope of discovering additional mechanisms of action.

Huang et al. developed a recursive K-means spectral clustering algorithm and applied this method to gene expression data from the Cancer Genome Atlas. It provides better cluster separation than traditional hierarchical clustering, and better execution time than similar K-means approaches.

Schrider et al. used pooled paired-end sequence data from multiple Drosophila melanogaster species along the eastern US coast to identify copy number variants under selective pressure. Many of the CNVs identified contain CYP enzymes likely influencing insecticide resistance. Schrider also pointed out in his talk that human salivary amylase (AMY1) has copy numbers that are differentiated across human populations due to differences in dietary starch content. Cool!

Verspoor et al. presented an awesome application of text mining to identify catalytic protein residues from the biomedical literature. Text mining tasks are always wrought with difficulties such as identifier ambiguity and resolution, or simply identifying the corpus of text needed for the task. Using Literature-Enhanced Automated Prediction of Functional Sites (LEAP-FS) and the Protein Data Bank (with Pubmed references), they compare their text mining approach to the Catalytic Site Atlas as a ‘silver standard’. Despite the difficulty, a simple classifier gives an accuracy around 70% (measured by F-statistic).

Also, my colleague Ting Hu presented her excellent work on statistical epistasis networks which use entropy-based measures to identify high-order interactions in genetic data. And in case you are interested, I’ll end by shamelessly listing our own publications in complex data analysis and rare-variant population structure (with Marylyn Ritchie), and performance of the Illumina Metabochip in Hispanic samples and high-throughput epidemiology (with Dana Crawford).

PSB is always a fantastic meeting – hope to see you in 2014!