Martin Krzywinski / Genome Sciences Center / mkweb.bcgsc.ca Martin Krzywinski / Genome Sciences Center / mkweb.bcgsc.ca - contact me Martin Krzywinski / Genome Sciences Center / mkweb.bcgsc.ca on Twitter Martin Krzywinski / Genome Sciences Center / mkweb.bcgsc.ca - Lumondo Photography Martin Krzywinski / Genome Sciences Center / mkweb.bcgsc.ca - Pi Art Martin Krzywinski / Genome Sciences Center / mkweb.bcgsc.ca - Hilbertonians - Creatures on the Hilbert Curve
Tango is a sad thought that is danced.Enrique Santos Discépolothink & dancemore quotes

space: beautiful



UCD Computational and Molecular Biology Symposium, Dublin, Ireland. 1-2 Dec 2016.


visualization + design

Martin Krzywinski @MKrzywinski mkweb.bcgsc.ca
Cover image accompanying our article on mouse vasculature development. Biology turns astrophysical. PNAS 1 May 2012; 109 (18) (zoom, PNAS)

Creating the PNAS Cover

One of my goals in life, which I can now say has been accomplished, is to make biology look like astrophysics. Call it my love for the Torino Impact Hazard Scale.

Recently, I was given an opportunity to attend to this (admittedly vague) goal when Linda Chang from Aly Karsan's group approached me with some microscopy photos of mouse veins. I was asked to do "something" with these images for a cover submission to accompany the manuscript.

When people see my covers, sometimes they ask "How did you do that?" Ok, actually they never ask this. But being a scientist, I'm trained me to produce answers in anticipation of such questions. So, below, I show you how the image was constructed.

The image was published on the cover of PNAS (PNAS 1 May 2012; 109 (18))

Tools

Photoshop CS5, Nik Color Efex Pro 4, Alien Skin Bokeh 2 and a cup of coffee from a Rancilio Silvia.

source images

Below are a few of the images I had the option to work with. These are mouse embryonic blood vessels, with a carotid artery shown in the foreground with endothelial cells in green, vascular smooth muscle cells in red and the nuclei in blue.

Of course, as soon as I saw the images, I realized that there was very little that I needed to do to trigger the viewer's imagination. These photos were great!

Martin Krzywinski @MKrzywinski mkweb.bcgsc.ca
Mouse carotid arteries. (zoom)
Martin Krzywinski @MKrzywinski mkweb.bcgsc.ca
Mouse carotid arteries. (zoom)
Martin Krzywinski @MKrzywinski mkweb.bcgsc.ca
Mouse carotid arteries. (zoom)
Martin Krzywinski @MKrzywinski mkweb.bcgsc.ca
Mouse carotid arteries. (zoom)
Martin Krzywinski @MKrzywinski mkweb.bcgsc.ca
Mouse carotid arteries. (zoom)
Martin Krzywinski @MKrzywinski mkweb.bcgsc.ca
Mouse carotid arteries. (zoom)
Martin Krzywinski @MKrzywinski mkweb.bcgsc.ca
Mouse carotid arteries. (zoom)
Martin Krzywinski @MKrzywinski mkweb.bcgsc.ca
Mouse carotid arteries. (zoom)
Martin Krzywinski @MKrzywinski mkweb.bcgsc.ca
Mouse carotid arteries. (zoom)

memories of star trek

Immediately I thought of two episodes of Star Trek (original series): Doomsday Machine and the Immunity Syndrome, as well as of images from the Hubble Telescope.

Martin Krzywinski @MKrzywinski mkweb.bcgsc.ca
Enterprise is about to be consumed by a horror tube: a planet killer! (The Doomsday Machine)
Martin Krzywinski @MKrzywinski mkweb.bcgsc.ca
Enterprise heads into a giant amoeba. Who eats whom? I'll let you guess. (The Immunity Syndrome)
Martin Krzywinski @MKrzywinski mkweb.bcgsc.ca
Orion nebula (M42) as seen by the Hubble telescope. (zoom)

I though it would be pretty easy to make the artery images look all-outer-spacey. They already looked it.

centerpiece image

And then I saw the image below.

Martin Krzywinski @MKrzywinski mkweb.bcgsc.ca
A particularly spectacular image of a mouse carotid artery. I'm thinking 10 on the Torino scale. (zoom)

constructing the cover

background

The background was created from the two images shown here. The second image was sampled three times, at different rotations.

Martin Krzywinski @MKrzywinski mkweb.bcgsc.ca
Images used for background. (zoom)
Martin Krzywinski @MKrzywinski mkweb.bcgsc.ca
Images used for background. (zoom)
Martin Krzywinski @MKrzywinski mkweb.bcgsc.ca
Layer composition for background elements. (zoom)

The channel mixer was used to remove the green channel and leave only red and blue.

Martin Krzywinski @MKrzywinski mkweb.bcgsc.ca
Background elements for PNAS cover image. (zoom)

middle ground

The next layer was composed of what looked like ribbons of blue gas. This was created by sampling the oval shapes from the source images. Here the red channel was a great source for cloud shapes, and this was the only channel that was kept. The hue was shifted to blue and a curve adjustment was applied to increase the contrast.

Martin Krzywinski @MKrzywinski mkweb.bcgsc.ca
First set of middle ground elements, before adjustments. (zoom)
Martin Krzywinski @MKrzywinski mkweb.bcgsc.ca
First set of middle ground elements, after channel adjustments. (zoom)
Martin Krzywinski @MKrzywinski mkweb.bcgsc.ca
Second set of middle ground elements. (zoom)
Martin Krzywinski @MKrzywinski mkweb.bcgsc.ca
Layer composition for middle ground elements. (zoom)

When the foreground and middle ground elements were combined, the result was already 40 parsecs away.

Martin Krzywinski @MKrzywinski mkweb.bcgsc.ca
Background and foreground elements for PNAS cover image. (zoom)

foreground

The foreground was created from the spectacular comet-like image of a mouse artery. Very little had to be done to make this element look good. It already looked good.

I applied a little blur using Alien Skin's Bokeh 2 to narrow the apparent depth of field, masked out elements at the bottom of the image and removed some of the green channel. The entire blue channel was removed altogether (this gave the tail of the comet a mottled, flame-like appearance).

Martin Krzywinski @MKrzywinski mkweb.bcgsc.ca
Foreground element, before adjustments. (zoom)
Martin Krzywinski @MKrzywinski mkweb.bcgsc.ca
Foreground element, after channel adjustments. (zoom)
Martin Krzywinski @MKrzywinski mkweb.bcgsc.ca
Layer composition for foreground element. (zoom)

post processing

Martin Krzywinski @MKrzywinski mkweb.bcgsc.ca
Initial composition of background, middle ground and foreground elements. (zoom)
Martin Krzywinski @MKrzywinski mkweb.bcgsc.ca
40% localized application of Nik's Tonal Contrast (Color Efex 4 plugin) to increase structure in red channel. (zoom)
Martin Krzywinski @MKrzywinski mkweb.bcgsc.ca
50% blend with Nik's Pro Contrast (Color Efex 4 plugin). (zoom)

And here we have the final image.

Martin Krzywinski @MKrzywinski mkweb.bcgsc.ca
Final PNAS cover. Spacey! (zoom)

upcoming

VIEW ALL

news + thoughts

Model Selection and Overfitting

Tue 13-09-2016

With four parameters I can fit an elephant and with five I can make him wiggle his trunk. —John von Neumann.

By increasing the complexity of a model, it is easy to make it fit to data perfectly. Does this mean that the model is perfectly suitable? No.

When a model has a relatively large number of parameters, it is likely to be influenced by the noise in the data, which varies across observations, as much as any underlying trend, which remains the same. Such a model is overfitted—it matches training data well but does not generalize to new observations.

Martin Krzywinski @MKrzywinski mkweb.bcgsc.ca
Nature Methods Points of Significance column: Model Selection and Overfitting (read)

We discuss the use of training, validation and testing data sets and how they can be used, with methods such as cross-validation, to avoid overfitting.

Altman, N. & Krzywinski, M. (2016) Points of Significance: Model Selection and Overfitting. Nature Methods 13:703-704.

Background reading

Lever, J., Krzywinski, M. & Altman, N. (2016) Points of Significance: Classifier evaluation. Nature Methods 13:603-604.

Lever, J., Krzywinski, M. & Altman, N. (2016) Points of Significance: Logistic regression. Nature Methods 13:541-542.

...more about the Points of Significance column

Classifier Evaluation

Tue 13-09-2016

It is important to understand both what a classification metric expresses and what it hides.

We examine various metrics use to assess the performance of a classifier. We show that a single metric is insufficient to capture performance—for any metric, a variety of scenarios yield the same value.

Martin Krzywinski @MKrzywinski mkweb.bcgsc.ca
Nature Methods Points of Significance column: Classifier Evaluation (read)

We also discuss ROC and AUC curves and how their interpretation changes based on class balance.

Lever, J., Krzywinski, M. & Altman, N. (2016) Points of Significance: Classifier evaluation. Nature Methods 13:603-604.

Background reading

Lever, J., Krzywinski, M. & Altman, N. (2016) Points of Significance: Logistic regression. Nature Methods 13:541-542.

...more about the Points of Significance column

Happy 2016 `\pi` Approximation, roughly speaking

Sun 24-07-2016

Today is the day and it's hardly an approximation. In fact, `22/7` is 20% more accurate of a representation of `\pi` than `3.14`!

Time to celebrate, graphically. This year I do so with perfect packing of circles that embody the approximation.

By warping the circle by 8% along one axis, we can create a shape whose ratio of circumference to diameter, taken as twice the average radius, is 22/7.

If you prefer something more accurate, check out art from previous `\pi` days: 2013 `\pi` Day and 2014 `\pi` Day, 2015 `\pi` Day, and 2016 `\pi` Day.

Logistic Regression

Tue 13-09-2016

Regression can be used on categorical responses to estimate probabilities and to classify.

The next column in our series on regression deals with how to classify categorical data.

We show how linear regression can be used for classification and demonstrate that it can be unreliable in the presence of outliers. Using a logistic regression, which fits a linear model to the log odds ratio, improves robustness.

Martin Krzywinski @MKrzywinski mkweb.bcgsc.ca
Nature Methods Points of Significance column: Logistic regression? (read)

Logistic regression is solved numerically and in most cases, the maximum-likelihood estimates are unique and optimal. However, when the classes are perfectly separable, the numerical approach fails because there is an infinite number of solutions.

Lever, J., Krzywinski, M. & Altman, N. (2016) Points of Significance: Logistic regression. Nature Methods 13:541-542.

Background reading

Altman, N. & Krzywinski, M. (2016) Points of Significance: Regression diagnostics? Nature Methods 13:385-386.

Altman, N. & Krzywinski, M. (2015) Points of Significance: Multiple Linear Regression Nature Methods 12:1103-1104.

Altman, N. & Krzywinski, M. (2015) Points of significance: Simple Linear Regression Nature Methods 12:999-1000.

...more about the Points of Significance column