Martin Krzywinski / Canada's Michael Smith Genome Sciences Centre / mkweb.bcgsc.ca Martin Krzywinski / Canada's Michael Smith Genome Sciences Centre / mkweb.bcgsc.ca - contact me Martin Krzywinski / Canada's Michael Smith Genome Sciences Centre / mkweb.bcgsc.ca on Twitter Martin Krzywinski / Canada's Michael Smith Genome Sciences Centre / mkweb.bcgsc.ca - Lumondo Photography Martin Krzywinski / Canada's Michael Smith Genome Sciences Centre / mkweb.bcgsc.ca - Pi Art Martin Krzywinski / Canada's Michael Smith Genome Sciences Centre / mkweb.bcgsc.ca - Hilbertonians - Creatures on the Hilbert CurveMartin Krzywinski / Canada's Michael Smith Genome Sciences Centre / mkweb.bcgsc.ca - Pi Day 2020 - Piku
Love itself became the object of her love.Jonathan Safran Foercount sadnessesmore quotes

design: intriguing


The Outbreak Poems — artistic emissions in a pandemic


data visualization + art
The Personal OncoGenomics Program (POG) is a research initiative to study the impact of embedding genomic sequencing into real-time treatment planning for BC patients with metastatic cancers. Based out of the BC Cancer Research Centre and the GSC, POG is a large world-class clinical research collaboration of BC Cancer oncologists, pathologists and other clinical staff, researchers and technical personnel throughout BC healthcare facilities.
Interested in more art based on the POG570 cohort from the Personal OncoGenomics Program? Check out our 5-year POG anniversary posters and desktops.

Nature Cancer Cover — April 2020 issue

Pan-cancer genomic landscapes of advanced tumors after therapy

Pleasance, E., Titmuss, E., Williamson, L. et al. (2020) Pan-cancer analysis of advanced patient tumors reveals interactions between therapy and genomic landscapes. Nat Cancer 1:452–468.

Art is science in love.
— E.F. Weisslitz

Martin Krzywinski / Martin Krzywinski @MKrzywinski mkweb.bcgsc.ca
Desktop image based on our design of the Nature Cancer April 2020 cover accompanying Pan-cancer genomic landscapes of advanced tumors after therapy. (download desktops)
Martin Krzywinski / Martin Krzywinski @MKrzywinski mkweb.bcgsc.ca
Desktop image based on our design of the Nature Cancer April 2020 cover accompanying Pan-cancer genomic landscapes of advanced tumors after therapy. (download desktops)

The design of the cover was inspired by Christian Stolte's DNA portraits from personal genomic data and PrintMyDNA. I've always loved Christian's style–respect the data but add playful flair. I am grateful for his allowing me to apply his approach to this design.

behind the design

respect the data, keep the eye interested

Every sample from the POG570 cohort corresponds to an individual patient and this is reflected in the design. The design is more compelling when samples are spatially distinct.

A core principle behind the design is intentionally managing and encouraging variation. If we have too much visual variation (in the extreme case, the data generation mechanism is a uniform random distribution), the eye sees nothing interesting. Although things are changing, there's nothing to lock onto. On the other hand, if we don't have enough variation, then the eye reacts with the same indifference.

To keep the eye happy (at least, our eyes), one approach is to have the shapes in a design split two or three categories. For example, the design has small ellipse systems and large ellipse systems. These are easy to spot and the eye can immediately begin to make some sense of what it sees, even though it may not yet know the reasons behind the patterns. Ideally, there should be a few cases that border on two categories, just to keep the category boundaries slightly ambiguous.

Within each category, there should be enough visual surprises that the eye wants to categorize further. Here is your opportunity to challenge it and vary things just enough so that this task isn't easy. The eye wants to group by shape and color similarity (Gestalt principles) and, to keep it challenged (but not frustrated or overwhelmed), make the first 40% of this grouping easy, the next 40% challenging and the last 20% impossible.

Below are three scenarios in increasing order of subjective interestingness. Too little or too much isn't as effective—find the Goldilocks zone. The "too interesting" case has ellipse properties sampled from a uniform distribution—it's always useful to see what your design looks like with random data so that you can figure out whether your data set has any kind of personality.

less interesting
Martin Krzywinski / Martin Krzywinski @MKrzywinski mkweb.bcgsc.ca
Design based on ellipse angle that is proportional to relative count. (zoom)
more interesting
Martin Krzywinski / Martin Krzywinski @MKrzywinski mkweb.bcgsc.ca
The final design, based on ellipse angle that is inversely proportional to relative count. (zoom)
too interesting
Martin Krzywinski / Martin Krzywinski @MKrzywinski mkweb.bcgsc.ca
Ellipse angles and sizes sampled from a uniform randomly distribution. (zoom)

Practically, these considerations are retrofitted into a design once you narrow down the approach. They may not help you decide what to do but they're excellent at helping you evaluate what you've done.

And always: experiment and try not to go with your first idea.

spectrum of mutations

To explain the design, I'll use one of the 570 samples—a B-cell lymphoma—as an example. The method is the same for the other 569.

Using the genomic sequence of each sample, we first tabulate the number of mutations. These are classified into 7 classes: 6 kinds of single nucleotide variants (SNV: T>C, T>G, T>A, C>G, C>A, C>T) and indels (insertions or deletions). The use of the term SNV should be distinguished from SNP—typically SNP (single nucleotide polymorphism) is used to describe changes due to natural variation in the population (blue eyes, can roll your tongue, etc) but SNVs are somatic variants found in tumours.

Each sample has many other properties and a metric ton of data that describe it. We wanted to pick something that was easy to explain. For example, more categories of mutations are possible but our returns would diminish with each category.

The counts of these mutations are used to create a mutation spectrum composed of 7 ellipses. This is shown on the left of the legend.

Martin Krzywinski / Martin Krzywinski @MKrzywinski mkweb.bcgsc.ca
The legend on the poster explains the encoding. (zoom)

ellipse: a mutation class

Individual ellipses represent a class of mutations. The color of the ellipse is based on its mutation class: SNVs are colored and indels are grey. This color scheme is used for the outline of the ellipse.

The median and maximum counts across the samples for each mutation class are shown below.

data %>% group_by(mutation_class)
     %>% summarise(median=median(count),max=max(count))
  class  median     max
  T.G       361   15912
  T.A       586   25997
  C.G       578   31900
  T.C       850   34438
  C.A       974   65867
  indel     422  146329
  C.T      1995  397807

The ellipse fill also uses this color but at an opacity that is a function of the number of days between the time of advanced disease diagnosis and biopsy (`\sqrt{t}` mapped onto [0,1]). Sequencing of the sample was performed shortly after the biopsy.

mutation class context

Each of the 6 SNV classes (excluding indels) and are divided into 96 contexts based on what is either side of the mutated base. For example, TTC > TAC is a T>A change with a T on one side and a C on the other.

Below is the full profile of the SNV mutations (excluding indels) for the case used in the legend above. This signature is typical of lymphoid cell hypermutation—a phenomena by which B cells produce many distinct antibodies—and of alteration in polymerase activity.

Martin Krzywinski / Martin Krzywinski @MKrzywinski mkweb.bcgsc.ca
The legend on the poster explains the encoding. (zoom)

In contrast, the profile below is of a "standard" rectal carcinoma defined as "broadly colorectal" in terms of genes that are driving it. The signature itself, however, is interesting because it was induced by the treatment the patient received before being sequenced.

Martin Krzywinski / Martin Krzywinski @MKrzywinski mkweb.bcgsc.ca
The legend on the poster explains the encoding. (zoom)

ellipse layers

The ellipses for a sample are stacked based on the counts of mutations in their class.

The class with the most counts goes on the bottom—in our sample this is the C>A SNVs of which we have 10,422—and the class with the fewest counts goes on top—the indels of which the sample has 17.

Below are each of the 7 layers that make up the final design.

Martin Krzywinski / Martin Krzywinski @MKrzywinski mkweb.bcgsc.ca
Layer #1 of ellipses. (zoom)
Martin Krzywinski / Martin Krzywinski @MKrzywinski mkweb.bcgsc.ca
Layer #2 of ellipses. (zoom)
Martin Krzywinski / Martin Krzywinski @MKrzywinski mkweb.bcgsc.ca
Layer #3 of ellipses. (zoom)
Martin Krzywinski / Martin Krzywinski @MKrzywinski mkweb.bcgsc.ca
Layer #4 of ellipses. (zoom)
Martin Krzywinski / Martin Krzywinski @MKrzywinski mkweb.bcgsc.ca
Layer #5 of ellipses. (zoom)
Martin Krzywinski / Martin Krzywinski @MKrzywinski mkweb.bcgsc.ca
Layer #6 of ellipses. (zoom)
Martin Krzywinski / Martin Krzywinski @MKrzywinski mkweb.bcgsc.ca
Layer #7 of ellipses. (zoom)

The first layer, made up of mutations with the fewest counts, is mostly T>G and C>G SNVs with a few indels. A few blue ellipses are from samples for which the class with fewest counts were the T>A SNVs.

As we go down the layers, we encounter classes with progressively more counts. In layer 2, blue ellipses(T>G SNVs) begin to appear. In layer 3, we start to see green (T>C SNVs) and orange (C>A SNVs) ellipses.

By the time we reach layer 7, where the most common mutations are, these are almost all C>T SNVs, with a few samples having indels (grey) as the most common class.

The outline on the ellipse gets thicker towards the bottom of the stack—this is a fixed progression that does not depend on the data.

The order of the layers could have been the other way: the most frequent mutations on top. And I'm sure that we could have made a go of it. As is, the least frequent mutations get an ellipse with a thin outline and these sit on ellipses with thicker outlines. This way, we're limit the ammount of a stroke that is occluded by strokes drawn above it.

ellipse size and angle

Whereas the layering tells us how the mutation classes are ranked within a sample, the size of the ellipse is related to the rank within a layer.

For a given ellipse we first find its layer. For our example sample, let's say we're looking at the layer 2 ellipse (486 C>G SNVs). We take all the mutation counts in that layer (remember, these are going to be of whatever class is the 2nd most rare in a sample) and map the 486 count onto a sigmoid curve made.

Martin Krzywinski / Martin Krzywinski @MKrzywinski mkweb.bcgsc.ca
Mapping a mutation count onto ellipse size with a sigmoid curve. For a given layer, a sigmoid mapping is constructed between all the counts in the layer and the interval [0,1]. Our example sample's count of 486 is mapped onto 0.34. (zoom)
data %>% group_by(layer) %>% mutate(a = sigmoid(count,SoftMax=TRUE))

In this layer, the sample with the smallest ellipse is the one with the fewest counts in that layer and the sample with the largest ellipse is the one with the most counts (7,585). The median and maximum count values for each layer are shown below.

data %>% group_by(layer) %>% summarise(median=median(count),max=max(count))

  layer median    max
      1   284    4201
      2   398    7585
      3   562   10714
      4   704   18077
      5   870   34438
      6  1196   57499
      7  2016  397807

The sigmoid mapping defines the major axis of the ellipse (`a`) with `b=a/2` thus a fixed eccentricity of `e = \sqrt{1-b^2/a^2} = \sqrt{3}/2`. The eccentricities of all the ellipses is the same. For a given ellipse size, the count may be different. This depends on its layer's sigmoid mapping as shown below.

Martin Krzywinski / Martin Krzywinski @MKrzywinski mkweb.bcgsc.ca
The sigmoid mapping of counts to major ellipse axis for all layers. The mutation classes in the legend are ordered by their maximum counts across all samples. (zoom)

The angle of the ellipse is `\theta = kN/n` where `n` is the ellipse's mutation class count, `N` is the total number of mutations in a sample and `k` is a magic sauce factor.

Note that the angle is inversely proportional to the count. This was done to avoid having the ellipses in the first layer (fewest counts) all at similar angles. By making the angle proportional to `1/n` the angle variation is increased and the design looks substantially better. We played around with how things looked for various values of `k` and picked one that looked best to our eyes.

If we made the angle proportional to the relative count, `\theta = \pi n /N`, the design would look very ridig and unintersting. The images below show this—notice that all ellipses in the first layer are essentially horizontal (small angle) because their relative counts are small (median 0.05). Similarly, all the ellipses in the last layer (most counts) are closer to vertical because the median of this layer's proportional count is about 0.33.

data %>% group_by(layer) %>% summarise(median=median(count/total))

  layer  median
      1  0.0502
      2  0.0676
      3  0.0919
      4  0.115
      5  0.142
      6  0.188
      7  0.331
Martin Krzywinski / Martin Krzywinski @MKrzywinski mkweb.bcgsc.ca
Angles of ellipses in the first layer if the angle were proportional to the relative count. Boring! (zoom)
Martin Krzywinski / Martin Krzywinski @MKrzywinski mkweb.bcgsc.ca
Angles of ellipses the the 7th layer if the angle were proportional to the relative count. Boring! (zoom)

Notice that the ellipses have no absolute scale: every variable is scaled, either linearly, inversely or sigmoidally. I like relative scalings—once you split the data into sensible groups, relative scalings allow you to ask questions within a group.

When creating artistic data designs, explore different encodings, even if they break some rules. Always know what rules you're bending, breaking or ignoring.

This is true especially for cover designs, for which a more playful and interpretive approach is needed. There are more than enough accurate visualizations in the paper itself and a cover is usually no place for this.

let's see some exploration

Ultimately, the success of the design hinges on a combination of variation and uniformity and of symmetry and assymetry. Finding this balance is a kind of data exploration of its own.

Below are some of the experiments along the way. Notice that while each of these does show variation, there's a strong sense of uniformity across the panels. There are no surprises—after the first 10 ellipse sets (or so), each additional is more of the same.

Martin Krzywinski / Martin Krzywinski @MKrzywinski mkweb.bcgsc.ca
Twinkle fairies. (zoom)
Martin Krzywinski / Martin Krzywinski @MKrzywinski mkweb.bcgsc.ca
Exoplanetary neighbourhoods. (zoom)
Martin Krzywinski / Martin Krzywinski @MKrzywinski mkweb.bcgsc.ca
Back to the 80's. (zoom)
Martin Krzywinski / Martin Krzywinski @MKrzywinski mkweb.bcgsc.ca
Ghosts at night. (zoom)

VIEW ALL

news + thoughts

Cover of Nature Genetics August 2020

Mon 03-08-2020

Our design on the cover of Nature Genetics's August 2020 issue is “Dichotomy of Chromatin in Color” . Thanks to Dr. Andy Mungall for suggesting this terrific title.

Martin Krzywinski @MKrzywinski mkweb.bcgsc.ca
Dichotomy of Chromatin in Color. Nature Genetics, August 2020 issue. (read more)

The cover design accompanies our report in the issue Gagliardi, A., Porter, V.L., Zong, Z. et al. (2020) Analysis of Ugandan cervical carcinomas identifies human papillomavirus clade–specific epigenome and transcriptome landscapes. Nature Genetics 52:800–810.

Poster Design Guidelines

Wed 15-07-2020

Clear, concise, legible and compelling.

The PDF template is a poster about making posters. It provides design, typography and data visualiation tips with minimum fuss. Follow its advice until you have developed enough design sobriety and experience to know when to go your own way.

Martin Krzywinski @MKrzywinski mkweb.bcgsc.ca
Poster Design Guidelines — Clear, concise, legible and compelling..

The SEIRS model for infectious disease dynamics

Thu 18-06-2020

Realistic models of epidemics account for latency, loss of immunity, births and deaths.

We continue with our discussion about epidemic models and show how births, deaths and loss of immunity can create epidemic waves—a periodic fluctuation in the fraction of population that is infected.

Martin Krzywinski @MKrzywinski mkweb.bcgsc.ca
Nature Methods Points of Significance column: The SEIRS model for infectious disease dynamics. (read)

This column has an interactive supplemental component (download code) that allows you to explore epidemic waves and introduces the idea of the phase plane, a compact way to understand the evolution of an epidemic over its entire course.

Martin Krzywinski @MKrzywinski mkweb.bcgsc.ca
Nature Methods Points of Significance column: The SEIRS model for infectious disease dynamics. (Interactive supplemental materials)

Bjørnstad, O.N., Shea, K., Krzywinski, M. & Altman, N. (2020) Points of significance: The SEIRS model for infectious disease dynamics. Nature Methods 17:557–558.

Background reading

Bjørnstad, O.N., Shea, K., Krzywinski, M. & Altman, N. (2020) Points of significance: Modeling infectious epidemics. Nature Methods 17:455–456.

Gene Machines

Fri 05-06-2020

Shifting soundscapes, textures and rhythmic loops produced by laboratory machines.

In commemoration of the 20th anniversary of Canada's Michael Smith Genome Sciences Centre, Segue was commissioned to create an original composition based on audio recordings from the GSC's laboratory equipment, robots and computers—to make “music” from the noise they produce.

Martin Krzywinski @MKrzywinski mkweb.bcgsc.ca
Gene Machines by Segue. Now available on vinyl.

Virus Mutations Reveal How COVID-19 Really Spread

Mon 01-06-2020

Genetic sequences of the coronavirus tell story of when the virus arrived in each country and where it came from.

Our graphic in Scientific American's Graphic Science section in the June 2020 issue shows a phylogenetic tree based on a snapshot of the data model from Nextstrain as of 31 March 2020.

Martin Krzywinski @MKrzywinski mkweb.bcgsc.ca
Virus Mutations Reveal How COVID-19 Really Spread. Text by Mark Fischetti (Senior Editor), art direction by Jen Christiansen (Senior Graphics Editor), source: Nextstrain (enabled by data from GISAID).