I collaborated with Scientific American to create a data graphic for the September 2014 issue. The graphic compared the genomes of the Denisovan, bonobo, chimp and gorilla, showing how our own genomes are almost identical to the Denisovan and closer to that of the bonobo and chimp than the gorilla.
Here you'll find Hilbert curve art, a introduction to Hilbertonians, the creatures that live on the curve, an explanation of the Scientific American graphic and downloadable SVG/EPS Hilbert curve files.
In the blog post I argue that the genome is not a blueprint—a common metaphor that doesn't leave room for appreciating the complexity of the genome—and talk about the process of creating the figure.
The graphic shows the differences between the genome sequence of human and each of Denisovan, chimp, bonobo and gorilla. Differences are measured by the fraction of bases in the gene regions of human sequence that do not align to the other genome.
The approximately 1 Gb of sequence of gene regions (most introns are included) is divided into 2,047 bins which are mapped onto the Hilbert curve as circles.
The color of the circle, which represents about 500 kb of sequence, encodes the fraction of unaligned bases.
The original color scheme submitted for production was derived from the yellow-orange-red Brewer palette.
There's more than one way to do it.
The approach taken by the graphic is one of the simplest—this is why it was chosen. It's easy to understand and easy to explain. On the other hand, the answer depends on the state of the sequence resources for each species (especially bonobo, whose sequence assembly is in version 1) and completely overlooks the functional implications of these differences.
The real goal of identifying differences, a relatively superficial problem, is to find the subset of differences that make a difference, which is a deep problem.
For example, if someone told you that Vancouver, Canada and Sydney, Australia were 85% similar, you would likely assume that (a) this metric isn't that useful to you unless it aligns to your priorities in how city similarities should be judged, (b) other metrics would give different answers, and (c) some parts of Sydney are nothing like Vancouver while others might be identical. This goes the same for genomes, except that cities are easier to figure out since we built them ourselves.
The differences will be scattered throughout the genome and will take many forms: single base changes, small insertions or deletions, inversions, copy number changes, and so on. In parts critical to basic cell function we expect no differences (e.g. insulin gene exons) while in genes that are rapidly evolving we expect to see some differences.
A comparison of protein coding genes reveals approximately 500 genes showing accelerated evolution on each of the gorilla, human and chimpanzee lineages, and evidence for parallel acceleration, particularly of genes involved in hearing.
— Insights into hominid evolution from the gorilla genome sequence by Scally et al.
Parts of the genome that don't impact function are going to accumulate differences at a background rate of mutation.
Any single-number statistic that compares two genomes is necessarily going to be a gross approximation. Such numerical measures should be taken as a starting point and at best as some kind of average that hides all of the texture in the data.
Statements like "the 1% difference" are incomplete because they do not incorporate an uncertainty. If you see four separate reports claiming a 1%, 2%, 5% and 7% difference, this does not necessarily mean that we cannot agree. It means that the error in our measurement is large. You might venture a guess that the answer is somewhere in the range 1–7% (at the very least).
While confidence intervals and error bars are a sine qua non in physical sciences, assessing uncertainty in life sciences is a lot more difficult. To assess the extent of biological variation, which will add to the uncertainty in our result, we need to collecting data from independent biological samples. Often this is too expensive or not practical.
To provide a sober and practical guide to statistics for the busy biologist, Naomi Altman and myself write the Points of Significance column in Nature Methods. These kinds of resources are needed as long as errors persist in the translation between statistical analysis and conclusions (e.g. `5 sigma` and P values).
We don't yet have a full handle on individual levels of genomic variation, especially for non-human primates for which we have a single and incomplete genome. Even for humans, although we have resources like dbSNP, which catalogue individual variation, it is common to use the canonical human reference sequence for analysis. This reference sequence is only a single instance of a human genome (in fact, parts of it are derived from different individuals).
As a result, many of the reported values (and certainly almost all that make it to popular media) are without any confidence limits and thus are likely to be interpreted as fact rather than as an estimate. This causes all sorts of problems—two compatible estimates can easily (but wrongly) be interpreted as incompatible facts.
As an example, look at the phylogenetic trees in the figure below. Without incorporating uncertainty, the top tree presents a fixed and deceptive state of what we know about the uncertainty in what we know.
Recent work has shed some light on the uncertainty in determining population split times. The two trees in the figure above are generated from the data in the table below, from 2012 Generation times in wild chimpanzees and gorillas suggest earlier divergence times in great ape and human evolution Proc Natl Acad Sci U S A 109 (39) 15716-15721.
Notice that the human/chimp/gorilla split time uncertainty overlaps the human/chimp split.
The addition of uncertainty is the inevitable consequence of making multiple measurements and upgraded analytical models. It is a blessing not a curse.
That our genome is "similar" to that of the chimp, bonobo and gorilla is not in dispute. How to classify and quantify the differences is an active field of research, a process that often looks like a dispute.
We have been sequencing quickly and cheaply for less than 10 years. It's amazing how much we've been able to understand in such a short period of time. Genome sequencing (or some kind of genotyping) is now routinely done in the treatment of cancer. It is not long before a medical diagnosis will include an assessment of the full genome sequence.
As we sequence more and reflect more, we expect to change our minds. In fact, this is why we do science: so that our minds are changed.
Scientists engage the public in the process of scientific inquiry, testing and observation by way of reports in popular science media and newspapers. Understanding these reports requires that one holds as a core value to process of science and its outcomes. Groups with different agendas and a fundamentally different epistemology hijaack observations such as "In 30% of the genome, gorilla is closer to human or chimpanzee than the latter are to each other." (from the gorilla sequence paper) in an attempt to argue that our evolutionary models are sinfully wrong. They don't understand the implications of the uncertainty in our measurements (e.g. phylogenetic tree figure above) and have world outlooks that are impervious to the impact of observation.
It is certain that these genomes hold more surprises for us, but not in the way these groups claim.
Is our science incomplete? Absolutely. How do we address this? We do more science.
The genome is not blueprint. It's also absolutely not a recipe, which is promulgated by people who agree that it is not a blueprint. I explain my view of this here and why I think these analogies have disasterous effects on the public understanding of how their genomes (i.e. their bodies) work.
Sometimes metaphors are wonderful and they help expand our mind.
I, a universe of atoms, an atom in the universe.
Other times they are like jailors, keeping us from having productive thoughts.
Genomics: the big blueprint
You might argue that "blueprint" is one of the closest words in meaning, so its use is justified. The trouble is that its actually very far in meaning.
Consider the following figure.
A blueprint shows you "what". A genome doesn’t encode "what". It doesn’t even encode “how”. Nor does it encode "from what". It encodes "with what", which is several degrees removed from "what". I promise that this will make sense shortly.
The reason why the blueprint analogy is pernicious is that it makes it sound like once the genome sequence is known, the rest easily follows. The reality is that these days the genome sequence is easily determined and the rest follows with great effort (or never) (see The $1,000 genome, the $100,000 analysis? by Elaine Mardis).
I'm going to try to motivate you that the analogy is false by an example. Suppose that you wanted to build a house but instead of getting blueprints from the architect, you received this strange drawing.
You’d be right to be confused—welcome to genome science. This house’s genome looks a lot like a set of tools and bears no resemblance to the house itself. The genome tells you nothing about (a) what the function of each tool is, (b) the effect of the tools form to its function (e.g. what are the many ways in which a hammer can diverge from its original shape before it ceases to be useful), (c) what the tools act on (this is why above I said "with what" rather than "from what"), (d) how the tools act together, and importantly (e) what the tools are used to build.
This is more of a way a genome works. It encodes the protein enzymes that make biochemical reactions possible at room temperature. In the house example, the tools encoded by the genome (e.g. saw, hammer) can be thought of automatically doing their job when they’re in the presence of the correct material (wood, nail). This is in analogy to enzymes, which mediate reactions when in physical proximity of chemical substrates.
Neither wood nor nails—both essential materials for construction—appear anywhere in your home's genome. This directly translates into a biochemical example. We use sugar as a source of energy but the genome hints nothing at this—it only encodes the enzymes that act on sugar. Things are made more complex by the fact that the function of an enzyme is essentially impossible to predict without additional information, such as knowledge of functions of enzymes with similar characteristics.
You can probably imagine that the effect of changes in the home's genome is extraordinary difficult to predict. The figure below extends our example to that of your neighbour, which was recently observed to have collapsed. I’ll leave you to work out the mystery yourself.
So next time someone says that the genome is a blueprint, or that it is the "code of life", point out that it is merely the "code of tools" for life, which is the emerging property of a set of chemicals confined within a physical space.
The use of the Hilbert curve in genomics is not new. It appeared on the cover of Science in 2009 in connection to the 3-dimensional packing of the genome. It is an order 5 curve and just a flip of the curve I use in the Scientific American graphic. Here the corners of the curve have been smoothed out to give it a more organic and gooey feel.
At least one tool exists (HilbertVis) that allows you to wrap genomic data onto the curve.
2009 Visualization of genomic data with the Hilbert curve Bioinformatics 25 (10) 1231-1235.
I've used the Hilbert curve before to show the organization of genes in the genome. This figure shows the chromosome at a much higher resolution than would be possible if an ordinary line was used.
Because the Hilbert curve stretches the line into a square, it increases our ability to see details in data at higher resolution. In the figure below you can see distinct clumpiness in the organization of genes on the chromsome that is not representative of a purely random sampling.
Except for the Denisovan, the net alignments (e.g. human vs chimp net) from UCSC Genome browser were used for the analysis.
Gaps were intersected with human gene regions. For each gene, the region between the start of the first coding region and end of the last coding region was used.
The RefSeq gene annotation from the UCSC Genome table browser was used. The union of all 51,010 RefSeq gene records was used.
The gene region was taken as the extent of the gene's coding sequence (CDS), not just the exons within it.
For example, for the BRCA2 gene, the RefSeq entry is
tx cds BRCA2 NM_000059 chr13 + 32889616-32973809 32890597-32972907 exons exonstart exonend 27 32889616,32890558... 32889804,32890664...
This record's contribution was the region 32890597-32972907, shown in bold above.
The total size of the union of tx regions is 1.28 Gb (20,722 coverage elements), of cds regions as defined above is 0.99 Gb (24,931 coverage elements) and of exons is 74.5 Mb (225,404 coverage elements).
Assembly version: Feb 2009 (CRCh37/hg19)
30x sequence was aligned to the human genome at Max Planck (data portal).
Assembly version: Feb 2011 (panTro4)
Assembly version: May 2012 (panPan1).
At the moment this genome is available only on the test version of the browser.
Assembly version: Feb 2009 (CRCh37/hg19)
Assembly version: May 2011 (gorGor3.1/gorGor3)
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.
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.",
Just in time for the season, I've simulated a snow-pile of snowflakes based on the Gravner-Griffeath model.
Gravner, J. & Griffeath, D. (2007) Modeling Snow Crystal Growth II: A mesoscopic lattice map with plausible dynamics.
We introduce two common ensemble methods: bagging and random forests. Both of these methods repeat a statistical analysis on a bootstrap sample to improve the accuracy of the predictor. Our column shows these methods as applied to Classification and Regression Trees.
For example, we can sample the space of values more finely when using bagging with regression trees because each sample has potentially different boundaries at which the tree splits.
Random forests generate a large number of trees by not only generating bootstrap samples but also randomly choosing which predictor variables are considered at each split in the tree.
Krzywinski, M. & Altman, N. (2017) Points of Significance: Ensemble methods: bagging and random forests. Nature Methods 14:933–934.
Krzywinski, M. & Altman, N. (2017) Points of Significance: Classification and regression trees. Nature Methods 14:757–758.
Decision trees classify data by splitting it along the predictor axes into partitions with homogeneous values of the dependent variable. Unlike logistic or linear regression, CART does not develop a prediction equation. Instead, data are predicted by a series of binary decisions based on the boundaries of the splits. Decision trees are very effective and the resulting rules are readily interpreted.
Trees can be built using different metrics that measure how well the splits divide up the data classes: Gini index, entropy or misclassification error.
When the predictor variable is quantitative and not categorical, regression trees are used. Here, the data are still split but now the predictor variable is estimated by the average within the split boundaries. Tree growth can be controlled using the complexity parameter, a measure of the relative improvement of each new split.
Individual trees can be very sensitive to minor changes in the data and even better prediction can be achieved by exploiting this variability. Using ensemble methods, we can grow multiple trees from the same data.
Krzywinski, M. & Altman, N. (2017) Points of Significance: Classification and regression trees. Nature Methods 14:757–758.
Lever, J., Krzywinski, M. & Altman, N. (2016) Points of Significance: Logistic regression. Nature Methods 13:541-542.
Altman, N. & Krzywinski, M. (2015) Points of Significance: Multiple Linear Regression Nature Methods 12:1103-1104.
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: Model Selection and Overfitting. Nature Methods 13:703-704.
Lever, J., Krzywinski, M. & Altman, N. (2016) Points of Significance: Regularization. Nature Methods 13:803-804.
The artwork was created in collaboration with my colleagues at the Genome Sciences Center to celebrate the 5 year anniversary of the Personalized Oncogenomics Program (POG).
The Personal Oncogenomics Program (POG) is a collaborative research study including many BC Cancer Agency oncologists, pathologists and other clinicians along with Canada's Michael Smith Genome Sciences Centre with support from BC Cancer Foundation.
The aim of the program is to sequence, analyze and compare the genome of each patient's cancer—the entire DNA and RNA inside tumor cells— in order to understand what is enabling it to identify less toxic and more effective treatment options.
Principal component analysis (PCA) simplifies the complexity in high-dimensional data by reducing its number of dimensions.
To retain trend and patterns in the reduced representation, PCA finds linear combinations of canonical dimensions that maximize the variance of the projection of the data.
PCA is helpful in visualizing high-dimensional data and scatter plots based on 2-dimensional PCA can reveal clusters.
Altman, N. & Krzywinski, M. (2017) Points of Significance: Principal component analysis. Nature Methods 14:641–642.
Altman, N. & Krzywinski, M. (2017) Points of Significance: Clustering. Nature Methods 14:545–546.
To achieve a `k` index for a movement you must perform `k` unbroken reps at `k`% 1RM.
The expected value for the `k` index is probably somewhere in the range of `k = 26` to `k=35`, with higher values progressively more difficult to achieve.
In my `k` index introduction article I provide detailed explanation, rep scheme table and WOD example.
The effect is intriguing and facetious—yes, those are real words.
But these are not: necronology, abobionalism, gabdologist, and nonerify.
These places only exist in the mind: Conchar and Pobacia, Hzuuland, New Kain, Rabibus and Megee Islands, Sentip and Sitina, Sinistan and Urzenia.
And these are the imaginary afflictions of the imagination: ictophobia, myconomascophobia, and talmatomania.
And these, of the body: ophalosis, icabulosis, mediatopathy and bellotalgia.
Want to name your baby? Or someone else's baby? Try Ginavietta Xilly Anganelel or Ferandulde Hommanloco Kictortick.
When taking new therapeutics, never mix salivac and labromine. And don't forget that abadarone is best taken on an empty stomach.
And nothing increases the chance of getting that grant funded than proposing the study of a new –ome! We really need someone to looking into the femome and manome.
An exploration of things that are missing in the human genome. The nullomers.
Julia Herold, Stefan Kurtz and Robert Giegerich. Efficient computation of absent words in genomic sequences. BMC Bioinformatics (2008) 9:167