The genetic analysis of high-throughput phenotypes

The genetic analysis of
high-throughput phenotypes

Karl Broman

Biostatistics & Medical Informatics, University of Wisconsin – Madison

kbroman.org github.com/kbroman @kwbroman

slides: bit.ly/Texas2015

Intercross

Genome scan for QTL

Genome-scale phenotypes

Why?

People embark on high-throughput phenotyping projects for a variety of different reasons.

They may just want to learn more things, or work together with others, sharing the costs of gathering and genotyping individuals and developing a coordinated phenotyping protocol. And one may identify important relationships among traits.

In mouse genetics, we’re particularly interested in speeding the effort to identify genes underlying complex clinical traits, and to improve our understanding of the underlying disease etiology.

In many cases, the high-throughput phenotyping is a potentially less-expensive proxy for the real trait of interest. This is particularly the case for electronic medical records (EMRs), which often seem more related to medical billing than to medical research.

Or we may just have no idea what to measure (or in what tissue to measure it), and so we try everything, just as we measure genotype genome-wide rather than at specific candidate genes.

In agriculture (and to some extent with humans), the primary interest may be in prediction the future course of disease, or to get an early understanding of a trait that can’t be measured until much later. (For example, in plants, early growth as a predictor of final yield.)

But more is not necessarily better

Challenges: diagnostics

kbroman.wordpress.com/2012/04/25/microarrays-suck

Challenges: metadata

What the heck is “`FAD_NAD SI 8.3_3.3G`”?

Challenges: Organizing & automating

The phenotypes (and samples) typically come in batches. Each batch will need a bit of curation (e.g., diagnostics).

Our ideas of what analyses to perform generally evolve over the course of a project. With a single phenotype, this is pretty easy: you just revise your scripts and re-run.

With many batches of many phenotypes, you may need to go back and re-do things. How to keep track of what was done for which phenotypes?

I suggest having a table that describes, for each phenotype or batch of phenotype, what needs to be done. And the results should include some record of how they were obtained. And then you run some uber-script that checks to see what needs to be done or re-done and ships off the computations and collects the results.

Animations: click on the “genotypes” box to have more batches of individuals displayed. Then click on the “phenotypes” box to have more batches of phenotypes displayed.

Challenges: Scale of the results

The ridiculome

Kim et al (2007) PNAS 104:20274–20279

Causal?

Multivariate analysis

Composite phenotypes

Shaffer et al. (2013) J Dent Res 92:32-37

Share more data, sooner.

Training

Data analysis and visualization

Project and data organization

Data Carpentry

Software engineering

Software Carpentry

To be able to manage and study high-throughput phenotyping, biologists and statisticians and others are going to need to improve their practices, from visualization to programming.

Statistics courses often focus on small-scale, nicely-behaved data. The curriculum needs to be modernized for our modern needs. Jenny Bryan’s Stat 545 course at the University of British Columbia is a really good example. Also look at the data science Coursera courses from Brian Caffo, Jeff Leek, and Roger Peng, at Johns Hopkins.

Even more, we all need to learn to better manage our large data projects. Data Carpentry is an important effort along these lines.

And we need to improve not just our programming skills but really our software engineering skills: program organization, documentation, version control, and testing. Software Carpentry provides provides important courses on these topics, for scientists.

Interactive graphics

A major part of my effort to conquer the high-throughput phenotyping problems is to develop interactive data visualization that enable a deeper and more rapid exploration of the data and analysis results.

Heatmaps of things like correlation matrices can be quite useful, but they are often hard to read. It’s nice to be able to read off the values by hovering.

Even better: to be able to click on a pixel and view the corresponding scatterplot.

This is a set of genes whose expression values are associated with genotype at a common locus. In the scatterplot, the points are colored by genotype at that locus.

With high-dimensional data, data visualizations will necessarily involve compressed summaries. But with interactive data visualizations, one can have immediate access to the underlying details.

code | data

All graphs can be improved with interactivity

While my efforts towards interactive graphs were motivated by the needs for high-dimensional data, I’ve come to the conclusion that basically all graphs could be improved with at least a bit of interactivity. You at least want tooltips.

We know we should not rely solely on summary statistics (like the LOD score), but if it takes an extra step to plot the underlying details, we tend not to do it, or at least not so much as we should.

The curves at the top are the same LOD curves we saw before. The LOD score is a measure of association between genotype and phenotype.

Click on a chromosome in the top panel to see a detailed view in the lower-left panel. Click on a marker in the lower-left panel to see effect estimates and raw data on the right. Hover over things on the right to get some further information.

code | data

Genome scan for a longitudinal phenotype

Here we consider a phenotype measured over time, and perform QTL analysis with each time point individually, to seek genetic loci that affect the response.

It’s a lot of work to make sense of these results without an interactive graph.

The top-left plot is a “heat map” of the LOD scores for each time point at each genomic position. LOD scores are also colored to indicate the sign of the QTL effect, with red indicating that BB lines have larger phenotype values and blue indicating that AA lines have larger phenotype values. We consider only those (position, time) pairs with LOD > 1.

When you hover over a point in the top-left plot, the LOD curves for the corresponding time are shown below, and the phenotype averages and estimated QTL effect (across time) for the corresponding genomic position are shown to the right.

Click on a point to show pointwise confidence bands on the QTL effect (± 2 SE). (We require a mouse click, as otherwise the graph was painfully slow to refresh.)

code | data

Interactivity is great for teaching

code | data

Interactive eQTL plot

Here’s my most fancy plot. But when I showed this to my collaborator, his first question was, “Where’s my favorite gene?” I needed to add a search box. Moral: These tools are not just for playing around; they need to address the scientific questions.

The top-left image shows inferred eQTL (sites in genome where genotype is associated with the expression of some gene). The y-axis corresponds to the genomic location of probes on a gene expression array. The x-axis corresponds to marker positions. Dots indicate that the genotypes at a particular marker are associated with the expression of a particular probe.

Hover over a point in the top-right to see all eQTL for a given probe. Click on a point to see the LOD curves across the genome for that probe: the association between that probe’s expression and genotype at each position in the genome.

Click on a marker in the lower LOD curve plot to see the corresponding phenotype-vs-genotype plot in the top-right.

code

D3 is awesome!

You just need to learn html, css, svg, and javascript.

And don’t forget .enter()

http://mbostock.github.io/d3/talk/20111018/collision.html

I’m using D3, a javascript library for data visualization. (D3 stands for “data-driven documents.”)

This is one of the images that convinced me that I wanted to learn D3. There are lots of additional examples at the D3 gallery.

There’s a ton to learn, and it’s rather forbidding (and much of it is dull).

My most common mistake is to forget to include .enter(). No error is given, but nothing shows up in the plot.

R/qtlcharts

Summary

Challenges

Organizing analyses
Making sense of the enormous pile of results
Inferring networks; breaking down to small networks
Joint consideration of multiple traits; form composite traits

Share more data, sooner
Needed upgrades in training

Data analysis and visualization
Data management
Software engineering

Interactive graphics require effort, but they

Facilitate exploration
Are great collaborative tools
Enable summaries with access to the details
Are great for teaching
Need to be tailored to the data and questions

R/qtlcharts, Broman (2015) Genetics 199:359–361

Acknowledgments

Alan Attie Mark Keller		Biochemistry, UW–Madison
Brian Yandell		Statistics and Horticulture, UW–Madison
Christina Kendziorski Aimee Teo Broman		Biostatistics & Medical Informatics, UW–Madison
Eric Schadt		Mount Sinai
Danielle Greenawalt Amit Kulkarni		Merck & Co., Inc.

Śaunak Sen		Epidemiology & Biostatistics, UC-San Francisco

Edgar Spalding Candace Moore Logan Johnson		Botany, UW-Madison

slides: bit.ly/Texas2015

kbroman.org

github.com/kbroman

@kwbroman

kbroman.org/qtlcharts