Sun is on my face ...a beautiful day without you.be apartmore quotes

# differences: enlightening

In Silico Flurries: Computing a world of snow. Scientific American. 23 December 2017

# data visualization + art

To view the art you'll need a pair of red-blue 3D glasses.
The data will stand out—and you will too.

# BD Genomics stereoscopic art exhibit — AGBT 2017

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

Our art exhibit at AGBT 2017 asked new school questions in old school ways.

The final art pieces are the outcome of a long process of exploration, experimentation and more than a few dead-ends. Here I'll take you through the process of parsing, exploring and drawing the data and finding a story.

Early sketches, looking for a data encoding.

The first step was to identify how to present the theme of "differences" in the exhibit. We wanted to draw attention to the fact that the extent to which we can answer questions about biological states depends on how accurate and precise the measurements are. Initially, I thought that this might be a useful theme—highlighting both biological and technical variation.

Early sketches, shaping the data.

In parallel, I thought about different functional sources of variation and the kinds of questions that the data might be used to answer. I narrowed them down to differences that cause disease, differences that cause disease progression and differences that allow for a variety of normal function.

Early sketch, attempting an organic, cellular look.

We also considered the idea of showing differences due to purely experimental error, which are fluctuations due to the technology not necessarily the biology.

## single-cell gene expression

For each of the difference scenarios, the input data set were single-cell gene expression counts.

$cell1 sample cell_type gene1 count cell1 sample cell_type gene2 count ... cell2 sample cell_type gene1 count cell2 sample cell_type gene2 count ...$

Each cell was identified by its sample (e.g. blood normal vs tumor) and type (e.g. B cell, T cell, etc). Typically, we had counts for about 500 genes of 100's or 1000's of cells of a given type and sample.

## dimensional reduction

The transcriptome of each cell is a point in high-dimensional space—one dimension for each gene for which we have a count. To find a projection of the data onto the page, we dimensionally reduced the data using tSNE (t-distributed stochastic neighbour embedding) into either 2 or 4 dimensions.

$# 2 dimensions (x,y) cell1 sample cell_type tsne_x tsne_y cell2 sample cell_type tsne_x tsne_y ... # 4 dimensions (x,y,u,v) cell1 sample cell_type tsne_x tsne_y tsne_u tsne_v cell2 sample cell_type tsne_x tsne_y tsne_u tsne_v ...$

When the cells are drawn as points based on their tSNE coordinates, typically they will cluster both by type and disease status. They cluster by type because the gene expression profile for each cell type is different—this is what causes cells to have different function.

Transcripts of the cells in the healthy vs disease data set shown by their 2-dimensional tSNE coordinates. Each point represesents a cell. On the left cells are colored by type and on the right by normal (black) or disease (orange) state. The black dashed lines highlightes the classical monocyte cluster. (zoom)

What we're looking for here is cells that cluster by disease state within a given cell type cluster, like the classical monocytes highlighted in the figure above. What this happens, we can say that there's something fundamentally different for this cell type between the normal and diseased states. For cell types that don't have a differential expression in disease, we see more of a random mixing of the normal and disease cell populations within their clusters.

## exploring encodings

I like to start by exploring ways to map the data onto the page. Typically, at this stage I try a lot of different approaches and many take the shortest route to the trash.

The focus of our story is "differences". So, I'll be looking for visuals that capture the Gestalt of a difference. Because we're aiming for an equal mix of art and data visualization, it's not critical that we can judge the differences quantitatively—ability to make qualitative assessments will suffice—but it is important that something obviously appears to be different, both within one panel and across panels.

In the search for an encoding, two basic questions have to be addressed: how cells are to be (a) placed and (b) represented on the page. For example, placement could be systematic, such as on a grid, with order based on some property such as total gene count. With this approach, we can achieve a tiling that covers the full canavs.

Or, placement could be based on properties such as tSNE coordinates. In this case, we settle in not populating the canvas evenly and hope that the cells fall into groups in a meaningful way.

Below is one attempt at a tiling encoding. The cells are represented by a grid of squares and each square shows a comparison of gene counts in disease vs normal for four genes. This layout can be adapted into triangles (3 genes) or hexagons (6 genes).

Encoding the counts of 4 genes.
Comparison of 400 normal vs diseased CD4 cells. (zoom)

The genes can be chosen based on ones that are known to be of clinical significance or, as below, identified from the data set as ones that have the largest change in expression between normal and disease. Cells can be ordered based on similarity in counts between normal and disease.

Comparison of 400 normal vs diseased CD4 cells for (A) genes that had the largest increase in count in disease and (b) genes that had the largest decrease count in disease. The position of the triangle vertex along each square encodes log2 of count ratio. Cells are in decreasing ordered of similarity in counts between normal and disease. (zoom)

I then tried using the 4-dimensional $(x,y,u,v)$ tSNE coordinates to represent each cell, still sticking to a tiling of cells. Cells are drawn in a row-dominant order and those with more similar tSNE coordinates are drawn closer together. For example, for the circular encoding, each concentric circle radius is $(x,x+y,x+y+u,x+y+u+v)$.

Cells are represented by a set of concentric circles sized by the cell's 4-dimensional tSNE coordinates. For each coordinate, I tried (A) randomized color palette and (B) Fixed color palette. Optionally, the cell can be sized based on its total gene count. For a given cell, the next cell in the grid is one that has the most similar tSNE coordinates. (zoom)
Cells are represented by a set of nested triangles sized by the cell's 4-dimensional tSNE coordinates. For each coordinate, I tried (A) randomized color palette and (B) Fixed color palette. Optionally, (C) the cell can be sized based on its total gene count. For a given cell, the next cell in the grid is one that has the most similar tSNE coordinates. (zoom)

Playing with colors and shapes makes for interesting tilings.

Cells are represented by a set of nested triangles sized by the cell's 4-dimensional tSNE coordinates. For each coordinate, I tried (A) randomized color palette and (B) Fixed color palette. Optionally, (C) the cell can be sized based on its total gene count. For a given cell, the next cell in the grid is one that has the most similar tSNE coordinates. (zoom)

As much as all these attempts have pretty shapes and colors, it's hard to point your finger at any part of these encodings and say, "Here's a difference worth investigating."

As well, by placing cells on a grid makes for a rigid represntation. We agreed that we needed something that looked a little more organic. Could the way the cells were drawn look like actual cells?

## clusters, networks and tesselation

We started looking at drawing the cells based on their 2-dimensional tSNE coordinates. In this representation, cells that are closer together on the plane have more similar transcription profiles.

Transcripts of the cells in the healthy vs tumor and in tumor vs metastasis data set shown by their 2-dimensional tSNE coordinates. Each point represesents a cell, color either by cell type (top) or normal, tumor or metastasis state (bottom). (zoom)

With this approach, it was really easy to draw attention to differences—either by cell type (which we wanted to do in the normal function case) or disease status. We also felt that this would be a familiar approach and one that showed populations of cells at the resolution of single cells.

One of the data sets that we identified early was the internal validation set, which captured the variability between operators and technical replication. Below are the results of six experiments, each done by a different operator on two different days. Here, we don't expect (and we don't see) any clustering based on days or operators. In the end, we chose not make use of this data set in the final exhibit and focus instead on clinically relevant differences.

Transcripts of normal cells in a series of internal validation trials performed by different operators on different days. There's no difference in the distribution of cells relative to either operator or day. The clusters that you see are the different cell types. (zoom)

We then explored ways in which the tSNE coordinates could be used to derive different representations, such as a network or tesselation.

Transcripts of the cells in the healthy vs tumor shown by their 2-dimensional tSNE coordinates (left). Nearest normal and tumor cells of a given cell type are connected into a minimum spanning tree graph (right). (zoom)
Transcripts of the cells in the healthy vs tumor. Cells are originally placed by tehir 2-dimensional tSNE coordinates and then a Voronoi tesselation is applied to generate polygons that bound the cells (one polygon per cell). The cell positions are then shifted to be closer to the center of each polygon and cell size adjusted to fit within the polygon. (zoom)

By hiding the cells and filling the polygons based on cell type and mixing the neighbouring polygon colors, we get a

The Voronoi tesselation 'gems' shown in the figure above with polygon colors based on cell types. Adjacent polygon colors are blended together to provide smoother transition across clusters. The two figures represent two different experiments from the normal vs disease data set. (zoom)

A more organic feel can be achieved by perturbing the edges of both the cells and polygons. We were happy enough with this representation to use it on the introductory panel in the exhibit.

An organic perturbation to the Voronoi tesselation of the normal vs disease data set. Diseased cells have their polygons and cell shapes filled. (zoom)
An organic perturbation to the Voronoi tesselation of the normal vs disease data set. Diseased cells have their polygons and cell shapes filled. (zoom)

The distance between a given tumor cell and its nearest normal cell of the same type can be emphasized by encoding the distance with color, such as below.

Cells placed based on Voronoi tesselation with polygons hidden. Cell color is either grey (normal) or yellow-red (tumor), where it maps the distance to the nearest normal cell of the same type. (zoom)

At this point we decided to explore mapping this difference by the 3rd dimension, literally.

Cells placed based on Voronoi tesselation with polygons hidden. Image is stereoscopic (requires red-blue 3d glasses) and depth towards the viewer encodes the distance between a tumor cell and its nearest normal neighbour of the same type. (zoom)

The stereoscopic image shown above is a quick prototype generated in Illustrator. The final images were rendered in a full 3D environment using Cinema4D.

VIEW ALL

# Curse(s) of dimensionality

Tue 05-06-2018
There is such a thing as too much of a good thing.

We discuss the many ways in which analysis can be confounded when data has a large number of dimensions (variables). Collectively, these are called the "curses of dimensionality".

Nature Methods Points of Significance column: Curse(s) of dimensionality. (read)

Some of these are unintuitive, such as the fact that the volume of the hypersphere increases and then shrinks beyond about 7 dimensions, while the volume of the hypercube always increases. This means that high-dimensional space is "mostly corners" and the distance between points increases greatly with dimension. This has consequences on correlation and classification.

Altman, N. & Krzywinski, M. (2018) Points of significance: Curse(s) of dimensionality Nature Methods 15:399–400.

# Statistics vs Machine Learning

Tue 03-04-2018
We conclude our series on Machine Learning with a comparison of two approaches: classical statistical inference and machine learning. The boundary between them is subject to debate, but important generalizations can be made.

Inference creates a mathematical model of the datageneration process to formalize understanding or test a hypothesis about how the system behaves. Prediction aims at forecasting unobserved outcomes or future behavior. Typically we want to do both and know how biological processes work and what will happen next. Inference and ML are complementary in pointing us to biologically meaningful conclusions.

Nature Methods Points of Significance column: Statistics vs machine learning. (read)

Statistics asks us to choose a model that incorporates our knowledge of the system, and ML requires us to choose a predictive algorithm by relying on its empirical capabilities. Justification for an inference model typically rests on whether we feel it adequately captures the essence of the system. The choice of pattern-learning algorithms often depends on measures of past performance in similar scenarios.

Bzdok, D., Krzywinski, M. & Altman, N. (2018) Points of Significance: Statistics vs machine learning. Nature Methods 15:233–234.

Bzdok, D., Krzywinski, M. & Altman, N. (2017) Points of Significance: Machine learning: a primer. Nature Methods 14:1119–1120.

Bzdok, D., Krzywinski, M. & Altman, N. (2017) Points of Significance: Machine learning: supervised methods. Nature Methods 15:5–6.

# Happy 2018 $\pi$ Day—Boonies, burbs and boutiques of $\pi$

Wed 14-03-2018

Celebrate $\pi$ Day (March 14th) and go to brand new places. Together with Jake Lever, this year we shrink the world and play with road maps.

Streets are seamlessly streets from across the world. Finally, a halva shop on the same block!

A great 10 km run loop between Istanbul, Copenhagen, San Francisco and Dublin. Stop off for halva, smørrebrød, espresso and a Guinness on the way. (details)

Intriguing and personal patterns of urban development for each city appear in the Boonies, Burbs and Boutiques series.

In the Boonies, Burbs and Boutiques of $\pi$ we draw progressively denser patches using the digit sequence 159 to inform density. (details)

No color—just lines. Lines from Marrakesh, Prague, Istanbul, Nice and other destinations for the mind and the heart.

Roads from cities rearranged according to the digits of $\pi$. (details)

The art is featured in the Pi City on the Scientific American SA Visual blog.

Check out art from previous years: 2013 $\pi$ Day and 2014 $\pi$ Day, 2015 $\pi$ Day, 2016 $\pi$ Day and 2017 $\pi$ Day.

# Machine learning: supervised methods (SVM & kNN)

Thu 18-01-2018
Supervised learning algorithms extract general principles from observed examples guided by a specific prediction objective.

We examine two very common supervised machine learning methods: linear support vector machines (SVM) and k-nearest neighbors (kNN).

SVM is often less computationally demanding than kNN and is easier to interpret, but it can identify only a limited set of patterns. On the other hand, kNN can find very complex patterns, but its output is more challenging to interpret.

Nature Methods Points of Significance column: Machine learning: supervised methods (SVM & kNN). (read)

We illustrate SVM using a data set in which points fall into two categories, which are separated in SVM by a straight line "margin". SVM can be tuned using a parameter that influences the width and location of the margin, permitting points to fall within the margin or on the wrong side of the margin. We then show how kNN relaxes explicit boundary definitions, such as the straight line in SVM, and how kNN too can be tuned to create more robust classification.

Bzdok, D., Krzywinski, M. & Altman, N. (2018) Points of Significance: Machine learning: a primer. Nature Methods 15:5–6.

Bzdok, D., Krzywinski, M. & Altman, N. (2017) Points of Significance: Machine learning: a primer. Nature Methods 14:1119–1120.

# Human Versus Machine

Tue 16-01-2018
Balancing subjective design with objective optimization.

In a Nature graphics blog article, I present my process behind designing the stark black-and-white Nature 10 cover.

Nature 10, 18 December 2017

# Machine learning: a primer

Thu 18-01-2018
Machine learning extracts patterns from data without explicit instructions.

In this primer, we focus on essential ML principles— a modeling strategy to let the data speak for themselves, to the extent possible.

The benefits of ML arise from its use of a large number of tuning parameters or weights, which control the algorithm’s complexity and are estimated from the data using numerical optimization. Often ML algorithms are motivated by heuristics such as models of interacting neurons or natural evolution—even if the underlying mechanism of the biological system being studied is substantially different. The utility of ML algorithms is typically assessed empirically by how well extracted patterns generalize to new observations.

Nature Methods Points of Significance column: Machine learning: a primer. (read)

We present a data scenario in which we fit to a model with 5 predictors using polynomials and show what to expect from ML when noise and sample size vary. We also demonstrate the consequences of excluding an important predictor or including a spurious one.

Bzdok, D., Krzywinski, M. & Altman, N. (2017) Points of Significance: Machine learning: a primer. Nature Methods 14:1119–1120.