You are currently browsing the category archive for the ‘RNA-Seq’ category.
[Update July 15, 2016: A preprint describing sleuth is available on BioRxiv]
Today my student Harold Pimentel released the beta version of his new RNA-Seq analysis method and software program called sleuth. A sleuth for RNA-Seq begins with the quantification of samples with kallisto, and together a sleuth of kallistos can be used to analyze RNA-Seq data rigorously and rapidly.
Why do we need another RNA-Seq program?
A major challenge in transcriptome analysis is to determine the transcripts that have changed in their abundance across conditions. This challenge is not entirely trivial because the stochasticity in transcription both within and between cells (biological variation), and the randomness in the experiment (RNA-Seq) that is used to determine transcript abundances (technical variation), make it difficult to determine what constitutes “significant” change.
Technical variation can be assessed by performing technical replicates of experiments. In the case of RNA-Seq, this can be done by repeatedly sequencing from one cDNA library. Such replicates are fundamentally distinct from biological replicates designed to assess biological variation. Biological replicates are performed by sequencing different cDNA libraries that have been constructed from repeated biological experiments performed under the same (or in practice near-same) conditions. Because biological replicates require sequencing different cDNA libraries, a key point is that biological replicates include technical variation.
In the early days of RNA-Seq a few papers (e.g. Marioni et al. 2008, Bullard et al. 2010) described analyses of technical replicates and concluded that they were not really needed in practice, because technical variation could be predicted statistically from the properties of the Poisson distribution. The point is that in an idealized RNA-Seq experiment counts of reads are multinomial (according to abundances of the transcripts they originate from), and therefore approximately Poisson distributed. Their variance is therefore approximately equal to the mean, so that it is possible to predict the variance in counts across technical replicates based on the abundance of the transcripts they originate from. There is, however, one important subtlety here: “counts of reads” as used above refers to the number of reads originating from a transcript, but in many cases, especially in higher eukaryotes, reads are frequently ambiguous as to their transcript of origin because of the presence of multi-isoform genes and genes families. In other words, transcript counts cannot be precisely measured. However, the statement about the Poisson distribution of counts in technical replicates remain true when considering counts of reads by genomic features because then reads are no longer ambiguous.
This is why, in so-called “count-based methods” for RNA-Seq analysis, there is an analysis only at the gene level. Programs such as DESeq/DESeq2, edgeR and a literally dozens of other count-based methods first require counting reads across genome features using tools such as HTSeq or featureCounts. By utilizing read counts to genomic features, technical replicates are unnecessary in lieu of the statistical assumption that they would reveal Poisson distributed data, and instead the methods focus on modeling biological variation. The issue of how to model biological variation is non-trivial because typically very few biological replicates are performed in experiments. Thus, there is a need for pooling information across genes to obtain reliable variance estimates via a statistical process called shrinkage. How and what to shrink is a matter of extensive debate among statisticians engaged in the development of count-based RNA-Seq methods, but one theme that has emerged is that shrinkage approaches can be compatible with general and generalized linear models, thus allowing for the analysis of complex experimental designs.
Despite these accomplishments, count-based methods for RNA-Seq have two major (related) drawbacks: first, the use of counts to gene features prevents inference about the transcription of isoforms, and therefore with most count-based methods it is impossible to identify splicing switches and other isoform changes between conditions. Some methods have tried to address this issue by restricting genomic features to specific exons or splice junctions (e.g. DEXSeq) but this requires throwing out a lot of data, thereby reducing power for identifying statistically significant differences between conditions. Second, because of the fact that in general it is mathematically incorrect to estimate gene abundances by adding up counts to their genomic region. One consequence of this, is that it is not possible to accurately measure fold change between conditions by using counts to gene features. In other words, count-based methods are problematic even at the gene-level and it is necessary to estimate transcript-level counts.
While reads might be ambiguous as to exactly which transcripts they originated from, it is possible to statistically infer an estimate of the number of reads from each transcript in an experiment. This kind of quantification has its origin in papers of Jiang and Wong, 2009 and Trapnell et al. 2010. However the process of estimating transcript-level counts introduces technical variation. That is to say, if multiple technical replicates were performed on a cDNA library and then transcript-level counts were to be inferred, those inferred counts would no longer be Poisson distributed. Thus, there appears to be a need for performing technical replicates after all. Furthermore, it becomes unclear how to work within the shrinkage frameworks of count-based methods.
There have been a handful of attempts to develop methods that combine the uncertainty of count estimates at the transcript level with biological variation in the assessment of statistically significant changes in transcript abundances between conditions. For example, the Cuffdiff2 method generalizes DESeq while the bitSeq method relies on a Bayesian framework to simultaneously quantify abundances at the transcript level while modeling biological variability. Despite showing improved performance over count-based methods, they also have significant shortcomings. For example the methods are not as flexible as those of general(ized) linear models, and bitSeq is slow partly because it requires MCMC sampling.
Thus, despite intensive research on both statistical and computational methods for RNA-Seq over the past years, there has been no solution for analysis of experiments that allows biologists to take full advantage of the power and resolution of RNA-Seq.
The sleuth model
The main contribution of sleuth is an intuitive yet powerful model for RNA-Seq that bridges the gap between count-based methods and quantification algorithms in a way that fully exploits the advantages of both.
To understand sleuth, it is helpful to start with the general linear model:
Here the subscript t refers to a specific transcript, is a vector describing transcript abundances (of length equal to the number of samples), is a design matrix (of size number of samples x number of confounders), is a parameter vector (of size the number of confounders) and is a noise vector (of size the number of samples). In this model the abundances are normally distributed. For the purposes of RNA-Seq data, the may be assumed to be the logarithm of the counts (or normalized counts per million) from a transcript, and indeed this is the approach taken in a number of approaches to RNA-Seq modeling, e.g. in limma-voom. A common alternative to the general linear model is the generalized linear model, which postulates that some function of has a distribution with mean equal to where g is a link function, such as log, thereby allowing for distributions other than the normal to be used for the observed data. In the RNA-Seq context, where the negative binomial distribution may make sense because it is frequently a good distribution for modeling count data, the generalized model is sometimes preferred to the standard general model (e.g. by DESeq2). There is much debate about which approach is “better”.
In the sleuth model the in the general linear model are modeled as unobserved. They can be thought of us the unobserved logarithms of true counts for each transcript across samples and are assumed to be normally distributed. The observed data is the logarithm of estimated counts for each transcript across samples, and is modeled as
where the vector parameterizes a perturbation to the unobserved . This can be understood as the technical noise due to the random sequencing of fragments from a cDNA library and the uncertainty introduced in estimating transcript counts.
The sleuth model incorporates the assumptions that the response error is additive, i.e. if the variance of transcript t in sample i is then where the variance and the variance . Intuitively, sleuth teases apart the two sources of variance by examining both technical and biological replicates, and in doing so directly estimates “true” biological variance, i.e. the variance in biological replicates that is not technical. In lieu of actual technical replicates, sleuth takes advantage of the bootstraps of kallisto which serve as accurate proxies.
In a test of sleuth on data simulated according to the DESeq2 model we found that sleuth significantly outperforms other methods:
In this simulation transcript counts were simulated according to a negative binomial distribution, following closely the protocol of the DESeq2 paper simulations. Reference parameters for the simulation were first estimated by running DESeq2 on a the female Finnish population from the GEUVADIS dataset (59 individuals). In the simulation above size factors were set to be equal in accordance with typical experiments being performed, but we also tested sleuth with size factors drawn at random with geometric mean of 1 in accordance with the DESeq2 protocol (yielding factors of 1, 0.33, 3, 3, 0.33 and 1) and sleuth still outperformed other methods.
There are many details in the implementation of sleuth that are crucial to its performance, e.g. the approach to shrinkage to estimate the biological variance . A forthcoming preprint, together with Nicolas Bray and Páll Melsted that also contributed to the project along with myself, will provide the details.
Exploratory data analysis with sleuth
One of the design goals of sleuth was to create a simple and efficient workflow in line with the principles of kallisto. Working with the Shiny web application framework we have designed an html interface that allows users to interact with sleuth plots allowing for real time exploratory data analysis.
The sleuth Shiny interface is much more than just a GUI for making plots of kallisto processed data. First, it allows for the exploration of the sleuth fitted models; users can explore the technical variation of each transcript, see where statistically significant differential transcripts appear in relationship to others in terms of abundance and variance and much more. Particularly useful are interactive features in the plots. For example, when examining an MA plot, users can highlight a region of points (dynamically created box in upper panel) and see their variance breakdown of the transcripts the points correspond to, and the list of the transcripts in a table below:
The web interface contains diagnostics, summaries of the data, “maps” showing low-dimensional representations of the data and tools for analysis of differential transcripts. The interactivity via Shiny can be especially useful for diagnostics; for example, in the diagnostics users can examine scatterplots of any two samples, and then select outliers to examine their variance, including the breakdown of technical variance. This allows for a determination of whether outliers represent high variance transcripts, or specific samples gone awry. Users can of course make figures showing transcript abundances in all samples, including boxplots displaying the extent of technical variation. Interested in the differential transcribed isoform ENST00000349155 of the TBX3 gene shown in Figure 5d of the Cuffdiff2 paper? It’s trivial to examine using the transcript viewer:
One can immediately see visually that differences between conditions completely dominate both the technical and biological variation within conditions. The sleuth q-value for this transcript is 3*10^(-68).
Among the maps, users can examine PCA projections onto any pair of components, allowing for rapid exploration of the structure of the data. Thus, with kallisto and sleuth raw RNA-Seq reads can be converted into a complete analysis in a matter of minutes. Experts will be able to generate plots and analyses in R using the sleuth library as they would with any R package. We plan numerous improvements and developments to the sleuth interface in the near future that will further facilitate data exploration; in the meantime we welcome feedback from users.
How to try out sleuth
Since sleuth requires the bootstraps and quantifications output by kallisto we recommend starting by running kallisto on your samples. The kallisto program is very fast, processing 30 million reads on a laptop in a matter of minutes. You will have to run kallisto with bootstraps- we have been using 100 bootstraps per sample but it should be possible to work with many fewer. We have yet to fully investigate the minimum number of bootstraps required for sleuth to be accurate.
To learn how to use kallisto start here. If you have already run kallisto you can proceed to the tutorial for sleuth. If you’re really eager to see sleuth without first learning kallisto, you can skip ahead and try it out using pre-computed kallisto runs of the Cuffdiff2 data- the tutorial explains where to obtain the data for trying out sleuth.
For questions, suggestions or help see the program websites and also the kallisto-sleuth user group. We hope you enjoy the tools!
About one and a half years ago I wrote a blog post titled “GTEx is throwing away 90% of their data“. The post was, shall we say, “direct”. For example, in reference to the RNA-Seq quantification program Flux Capacitor I wrote that
Using Flux Capacitor is equivalent to throwing out 90% of the data!
I added that “the methods description in the Online Methods of Montgomery et al. can only be (politely) described as word salad” (after explaining that the methods underlying the program were never published, except for a brief mention in that paper). I referred to the sole figure in Montgomery et al. as a “completely useless” description of the method (and showed that it contained errors). I highlighted the fact that many aspects of Flux Capacitor, its description and documentation provided on its website were “incoherent”. Can we agree that this description is not flattering?
The claim about “throwing out 90% of the data” was based on benchmarking I reported on in the blog post. If I were to summarize the results (politely), I would say that the take home message was that Flux Capacitor is junk. Perhaps nobody had really noticed because nobody cared about the program. Flux Capacitor was literally being used only by the authors of the program (and their affiliates, which turned out to include the ENCODE, GENCODE, GEUVADIS and GTEx consortiums). In fact, when I wrote the blog post, I don’t think the program had ever been benchmarked or compared to other tools. It was, after all, unpublished and besides, nobody reads consortium papers. However after I blogged a few others decided to include Flux Capacitor in their benchmarks and the conclusions reached were the same as mine: Flux Capacitor is junk and Flux Capacitor is junk. Of course some people objected to my blog post when it came out, so it’s fun to be right and have others say so in print. But true vindication has come in the form of a citation to the blog post in a published paper in a journal! Specifically, in
C. Iannone, A. Pohl, P. Papasaikas, D. Soronellas, G.P. Vincent, M. Beato and J. Valcárcel, Relationship between nucleosome positioning and progesterone-induced alternative splicing in breast cancer cells, RNA 21 (2015) 360–374
the authors cite my blog post. They write:
Ummm…. wait… WHAT THE FLUX? The authors actually used Flux Capacitor for their analysis, and are citing my blog at https://liorpachter.wordpress.com/tag/flux-capacitor/ as the definitive reference for the program. Wait, what again?? They used my blog post as a reference for the method??? This is like [[ readers are invited to leave a comment offering a suitable analogy ]].
I’m not really sure what the authors can do at this point. They could publish an erratum and replace the citation. But with what? Flux Capacitor still hasn’t been published (!) Then there is the journal. Does any journal really think it is acceptable to list my blog as the citation for an RNA-Seq quantification tool that is fundamental for the results in a paper? (I’m flattered, but still…) Speaking of the journal, where were the reviewers? How could they not catch this? And the readers? The paper has been out since January… I have to ask: has anybody read it? Of course the biggest embarrassment here is the fact that there is a citation for Flux Capacitor at all. Why on earth are the authors using this discredited program??? Well maybe one answer is to be found in the acknowledgments section, where the group of a PI from the GTEx project is thanked. Actually, this PI was the last author on one of the recently published GTEx companion papers, which, I am sad to say… used Flux Capacitor (albeit with some quantifications performed with Cufflinks as well to demonstrate “robustness”). Why would GTEx be pushing for Flux Capacitor and insist on its use? We’ve come full circle to my GTEx blog post. By now I don’t even know what I think is the most embarrassing part of this whole story. So I thought I’d host a poll:
Today I posted the preprint N. Bray, H. Pimentel, P. Melsted and L. Pachter, Near-optimal RNA-Seq quantification with kallisto to the arXiv. It describes the RNA-Seq quantification program kallisto. [Update April 5, 2016: a revised version of the preprint has been published: Nicolas L. Bray, Harold Pimentel, Páll Melsted and Lior Pachter, Near-optimal probabilistic RNA-Seq quantification, Nature Biotechnology (2016), doi:10.1038/nbt.3519 published online April 4, 2016.]
The project began in August 2013 when I wrote my second blog post, about another arXiv preprint describing a program for RNA-Seq quantification called Sailfish (now a published paper). At the time, a few students and postdocs in my group read the paper and then discussed it in our weekly journal club. It advocated a philosophy of “lightweight algorithms, which make frugal use of data, respect constant factors and effectively use concurrent hardware by working with small units of data where possible”. Indeed, two themes emerged in the journal club discussion:
1. Sailfish was much faster than other methods by virtue of being simpler.
2. The simplicity was to replace approximate alignment of reads with exact alignment of k-mers. When reads are shredded into their constituent k-mer “mini-reads”, the difficult read -> reference alignment problem in the presence of errors becomes an exact matching problem efficiently solvable with a hash table.
We felt that the shredding of reads must lead to reduced accuracy, and we quickly checked and found that to be the case. In fact, in our simulations, we saw that Sailfish significantly underperformed methods such as RSEM. However the fact that simpler was so much faster led us to wonder whether the prevailing wisdom of seeking to improve RNA-Seq analysis by looking at increasingly complex models was ill-founded. Perhaps simpler could be not only fast, but also accurate, or at least close enough to best-in-class for practical purposes.
After thinking about the problem carefully, my (now former) student Nicolas Bray realized that the key is to abandon the idea that alignments are necessary for RNA-Seq quantification. Even Sailfish makes use of alignments (of k-mers rather than reads, but alignments nonetheless). In fact, thinking about all the tools available, Nick realized that every RNA-Seq analysis program was being developed in the context of a “pipeline” of first aligning reads or parts of them to a reference genome or transcriptome. Nick had the insight to ask: what can be gained if we let go of that paradigm?
By April 2014 we had formalized the notion of “pseudoalignment” and Nick had written, in Python, a prototype of a pseudoaligner. He called the program kallisto. The basic idea was to determine, for each read, not where in each transcript it aligns, but rather which transcripts it is compatible with. That is asking for a lot less, and as it turns out, pseudoalignment can be much faster than alignment. At the same time, the information in pseudoalignments is enough to quantify abundances using a simple model for RNA-Seq, a point made in the isoEM paper, and an idea that Sailfish made use of as well.
Just how fast is pseudoalignment? In January of this year Páll Melsted from the University of Iceland came to visit my group for a semester sabbatical. Páll had experience in exactly the kinds of computer science we needed to optimize kallisto; he has written about efficient k-mer counting using the bloom filter and de Bruijn graph construction. He translated the Python kallisto to C++, incorporating numerous clever optimizations and a few new ideas along the way. His work was done in collaboration with my student Harold Pimentel, Nick (now a postdoc with Jacob Corn and Jennifer Doudna at the Innovative Genomics Initiative) and myself.
The screenshot below shows kallisto being used on my 2012 iMac desktop to build an index of the human transcriptome (5 min 8 sec), and then quantify 78.6 million GEUVADIS human RNA-Seq reads (14 min). When we first saw these results we thought they were simply too good to be true. Let me repeat: The quantification of 78.6 million reads takes 14 minutes on a standard desktop using a single CPU core. In some tests we’ve gotten even faster running times, up to 15 million reads quantified per minute.
The results in our paper indicate that kallisto is not just fast, but also very accurate. This is not surprising: underlying RNA-Seq analysis are the alignments, and although kallisto is pseudoaligning instead, it is almost always only the compatibility information that is used in actual applications. As we show in our paper, from the point of view of compatibility, the pseudoalignments and alignments are almost the same.
Although accuracy is a primary concern with analysis, we realized in the course of working on kallisto that speed is also paramount, and not just as a matter of convenience. The speed of kallisto has three major implications:
1. It allows for efficient bootstrapping. All that is required for the bootstrap are reruns of the EM algorithm, and those are particularly fast within kallisto. The result is that we can accurately estimate the uncertainty in abundance estimates. One of my favorite figures from our paper, made by Harold, is this one:
It is based on an analysis of 40 samples of 30 million reads subsampled from 275 million rat RNA-Seq reads. Each dot corresponds to a transcript and is colored by its abundance. The x-axis shows the variance estimated from kallisto bootstraps on a single subsample while the y-axis shows the variance computed from the different subsamples of the data. We see that the bootstrap recapitulates the empirical variance. This result is non-trivial: the standard dogma, that the technical variance in RNA-Seq is “Poisson” (i.e. proportional to the mean) is false, as shown in Supplementary Figure 3 of our paper (the correlation becomes 0.64). Thus, the bootstrap will be invaluable when incorporated in downstream application and we are already working on some ideas.
2. It is not just the kallisto quantification that is fast; the index building, and even compilation of the program are also easy and quick. The implication for biologists is that RNA-Seq analysis now becomes interactive. Instead of “freezing” an analysis that might take weeks or even months, data can be explored dynamically, e.g. easily quantified against different transcriptomes, or re-quantified as transcriptomes are updated. The ability to analyze data locally instead of requiring cloud computation means that analysis is portable, and also easily secure.
3. We have found the fast turnaround of analysis helpful in improving the program itself. With kallisto we can quickly check the effect of changes in the algorithms. This allows for much faster debugging of problems, and also better optimization. It also allows us to release improvements knowing that users will be able to test them without resorting to a major computation that might take months. For this reason we’re not afraid to say that some improvements to kallisto will be coming soon.
As someone who has worked on RNA-Seq since the time of 32bp reads, I have to say that kallisto has personally been extremely liberating. It offers freedom from the bioinformatics core facility, freedom from the cloud, freedom from the multi-core server, and in my case freedom from my graduate students– for the first time in years I’m analyzing tons of data on my own; because of the simplicity and speed I find I have the time for it. Enjoy!
When I was an undergraduate at Caltech I took a combinatorics course from Rick Wilson who taught from his then just published textbook A Course in Combinatorics (co-authored with J.H. van Lint). The course and the book emphasized design theory, a subject that is beautiful and fundamental to combinatorics, coding theory, and statistics, but that has sadly been in decline for some time. It was a fantastic course taught by a brilliant professor- an experience that had a profound impact on me. Though to be honest, I haven’t thought much about designs in recent years. Having kids changed that.
A few weeks ago I was playing the card game Colori with my three year old daughter. It’s one of her favorites.
The game consists of 15 cards, each displaying drawings of the same 15 items (beach ball, boat, butterfly, cap, car, drum, duck, fish, flower, kite, pencil, jersey, plane, teapot, teddy bear), with each item colored using two of the colors red, green, yellow and blue. Every pair of cards contains exactly one item that is colored exactly the same. For example, the two cards the teddy bear is holding in the picture above are shown below:
The only pair of items colored exactly the same are the two beach balls. The gameplay consists of shuffling the deck and then placing a pair of cards face-up. Players must find the matching pair, and the first player to do so keeps the cards. This is repeated seven times until there is only one card left in the deck, at which point the player with the most cards wins. When I play with my daughter “winning” consists of enjoying her laughter as she figures out the matching pair, and then proceeds to try to eat one of the cards.
An inspection of all 15 cards provided with the game reveals some interesting structure:
Every card contains exactly one of each type of item. Each item therefore occurs 15 times among the cards, with fourteen of the occurrences consisting of seven matched pairs, plus one extra. This is a type of partially balanced incomplete block design. Ignoring for a moment the extra item placed on each card, what we have is 15 items, each colored one of seven ways (v=15*7=105). The 105 items have been divided into 15 blocks (the cards), so that b=15. Each block contains 14 elements (the items) so that k=14, and each element appears in two blocks (r=2). If every pair of different (colored) items occurred in the same number of cards, we would have a balanced incomplete block design, but this is not the case in Colori. Each item occurs in the same block as 26 (=2*13) other items (we are ignoring the extra item that makes for 15 on each card), and therefore it is not the case that every pair of items occurs in the same number of blocks as would be the case in a balanced incomplete block design. Instead, there is an association scheme that provides extra structure among the 105 items, and in turn describes the way in which items do or do not appear together on cards. The association scheme can be understood as a graph whose nodes consist of the 105 items, with edges between items labeled either 0,1 or 2. An edge between two items of the same type is labeled 0, edges between different items that appear on the same card are labeled 1, and edges between different items that do not appear on the same card are labeled 2. This edge labeling is called an “association scheme” because it has a special property, namely the number of triangles with a base edge labeled k, and other two edges labeled i and j respectively is dependent only on i,j and k and not on the specific base edge selected. In other words, there is a special symmetry to the graph. Returning to the deck of cards, we see that every pair of items appears in the same card exactly 0 or 1 times, and the number depends only on the association class of the pair of objects. This is called a partially balanced incomplete block design.
The author of the game, Reinhard Staupe, made it a bit more difficult by adding an extra item to each card making the identification of the matching pair harder. The addition also ensures that each of the 15 items appears on each card. Moreover, the items are permuted in location on the cards, in an arrangement similar to a latin square, making it hard to pair up the items. And instead of using 8 different colors, he used only four, producing the eight different “colors” of each item on the cards by using pairwise combinations of the four. The yellow-green two-colored beach balls are particularly difficult to tell apart from the green-yellow ones. Of course, much of this is exactly the kind of thing you would want to do if you were designing an RNA-Seq experiment!
Instead of 15 types of items, think of 15 different strains of mice. Instead of colors for the items, think of different cellular conditions to be assayed. Instead of one pair for each of seven color combinations, think of one pair of replicates for each of seven cellular conditions. Instead of cards, think of different sequencing centers that will prepare the libraries and sequence the reads. An ideal experimental setup would involve distributing the replicates and different cellular conditions across the different sequencing centers so as to reduce batch effect. This is the essence of part of the paper Statistical Design and Analysis of RNA Sequencing Data by Paul Auer and Rebecca Doerge. For example, in their Figure 4 (shown below) they illustrate the advantage of balanced block designs to ameliorate lane effects:
Figure 4 from P. Auer and R.W. Doerge’s paper Statistical Design and Analysis of RNA Sequencing Data.
Of course the use of experimental designs for constructing controlled gene expression experiments is not new. Kerr and Churchill wrote about the use of combinatorial designs in Experimental Design for gene expression microarrays, and one can trace back a long chain of ideas originating with R.A. Fisher. But design theory seems to me to be a waning art insofar as molecular biology experiments are concerned, and it is frequently being replaced with biological intuition of what makes for a good control. The design of good controls is sometimes obvious, but not always. So next time you design an experiment, if you have young kids, first play a round of Colori. If the kids are older, play Set instead. And if you don’t have any kids, plan for an extra research project, because what else would you do with your time?
Nature Publishing Group claims on its website that it is committed to publishing “original research” that is “of the highest quality and impact”. But when exactly is research “original”? This is a question with a complicated answer. A recent blog post by senior editor Dorothy Clyde at Nature Protocols provides insight into the difficulties Nature faces in detecting plagiarism, and identifies the issue of self plagiarism as particularly problematic. The journal tries to avoid publishing the work of authors who have previously published the same work or a minor variant thereof. I imagine this is partly in the interests of fairness, a service to the scientific community to ensure that researchers don’t have to sift through numerous variants of a single research project in the literature, and a personal interest of the journal in its aim to publish only the highest level of scholarship.
On the other hand, there is also a rationale for individual researchers to revisit their own previously published work. Sometimes results can be recast in a way that makes them accessible to different communities, and rethinking of ideas frequently leads to a better understanding, and therefore a better exposition. The mathematician Gian-Carlo Rota made the case for enlightened self-plagiarism in one of his ten lessons he wished he had been taught when he was younger:
After getting my degree, I worked for a few years in functional analysis. I bought a copy of Frederick Riesz’ Collected Papers as soon as the big thick heavy oversize volume was published. However, as I began to leaf through, I could not help but notice that the pages were extra thick, almost like cardboard. Strangely, each of Riesz’ publications had been reset in exceptionally large type. I was fond of Riesz’ papers, which were invariably beautifully written and gave the reader a feeling of definitiveness.
As I looked through his Collected Papers however, another picture emerged. The editors had gone out of their way to publish every little scrap Riesz had ever published. It was clear that Riesz’ publications were few. What is more surprising is that the papers had been published several times. Riesz would publish the first rough version of an idea in some obscure Hungarian journal. A few years later, he would send a series of notes to the French Academy’s Comptes Rendus in which the same material was further elaborated. A few more years would pass, and he would publish the definitive paper, either in French or in English. Adam Koranyi, who took courses with Frederick Riesz, told me that Riesz would lecture on the same subject year after year, while meditating on the definitive version to be written. No wonder the final version was perfect.
Riesz’ example is worth following. The mathematical community is split into small groups, each one with its own customs, notation and terminology. It may soon be indispensable to present the same result in several versions, each one accessible to a specific group; the price one might have to pay otherwise is to have our work rediscovered by someone who uses a different language and notation, and who will rightly claim it as his own.
The question is: where does one draw the line?
I was recently forced to confront this question when reading an interesting paper about a statistical approach to utilizing controls in large-scale genomics experiments:
J.A. Gagnon-Bartsch and T.P. Speed, Using control genes to corrected for unwanted variation in microarray data, Biostatistics, 2012.
A cornerstone in the logic and methodology of biology is the notion of a “control”. For example, when testing the result of a drug on patients, a subset of individuals will be given a placebo. This is done to literally control for effects that might be measured in patients taking the drug, but that are not inherent to the drug itself. By examining patients on the placebo, it is possible to essentially cancel out uninteresting effects that are not specific to the drug. In modern genomics experiments that involve thousands, or even hundreds of thousands of measurements, there is a biological question of how to design suitable controls, and a statistical question of how to exploit large numbers of controls to “normalize” (i.e. remove unwanted variation) from the high-dimensional measurements.
Formally, one framework for thinking about this is a linear model for gene expression. Using the notation of Gagnon-Bartsch & Speed, we have an expression matrix of size (m samples and n genes) modeled as
Here X is a matrix describing various conditions (also called factors) and associated to it is the parameter matrix that records the contribution, or influence, of each factor on each gene. is the primary parameter of interest to be estimated from the data Y. The are random noise, and finally Z and W are observed and unobserved covariates respectively. For example Z might encode factors for covariates such as gender, whereas W would encode factors that are hidden, or unobserved. A crucial point is that the number of hidden factors in W, namely k, is not known. The matrices and record the contributions of the Z and W factors on gene expression, and must also be estimated. It should be noted that X may be the logarithm of expression levels from a microarray experiment, or the analogous quantity from an RNA-Seq experiment (e.g. log of abundance in FPKM units).
Linear models have been applied to gene expression analysis for a very long time; I can think of papers going back 15 years. But They became central to all analysis about a decade ago, specifically popularized with the Limma package for microarray data analysis. In an important paper in 2007, Leek and Storey focused explicitly on the identification of hidden factors and estimation of their influence, using a method called SVA (Surrogate Variable Analysis). Mathematically, they described a procedure for estimating k and W and the parameters . I will not delve into the details of SVA in this post, except to say that the overall idea is to first perform linear regression (assuming no hidden factors) to identify the parameters and to then perform singular value decomposition (SVD) on the residuals to identify hidden factors (details omitted here). The resulting identified hidden factors (and associated influence parameters) are then used in a more general model for gene expression in subsequent analysis.
Gagnon-Bartsch and Speed refine this idea by suggesting that it is better to infer W from controls. For example, house-keeping genes that are unlikely to correlate with the conditions being tested, can be used to first estimate W, and then subsequently all the parameters of the model can be estimated by linear regression. They term this two-step process RUV-2 (acronym for Remote Unwanted Variation) where the “2” designates that the procedure is a two-step procedure. As with SVA, the key to inferring W from the controls is to perform singular value decomposition (or more generally factor analysis). This is actually clear from the probabilistic interpretation of PCA and the observation that what it means to be a in the set of “control genes” C in a setting where there are no observed factors Z, is that
That is, for such control genes the corresponding parameters are zero. This is a simple but powerful observation, because the explicit designation of control genes in the procedure makes it clear how to estimate W, and therefore the procedure becomes conceptually compelling and practically simple to implement. Thus, even though the model being used is the same as that of Leek & Storey, there is a novel idea in the paper that makes the procedure “cleaner”. Indeed, Gagnon-Bartsch & Speed provide experimental results in their paper showing that RUV-2 outperforms SVA. Even more convincing, is the use of RUV-2 by others. For example, in a paper on “The functional consequences of variation in transcription factor binding” by Cusanovitch et al., PLoS Genetics 2014, RUV-2 is shown to work well, and the authors explain how it helps them to take advantage of the controls in experimental design they created.
There is a tech report and also a preprint that follow up on the Gagnon-Bartsch & Speed paper; the tech report extends RUV-2 to a four step method RUV-4 (it also provides a very clear exposition of the statistics), and separately the preprint describes an extension to RUV-2 for the case where the factor of interest is also unknown. Both of these papers build on the original paper in significant ways and are important work, that to return to the original question in the post, certainly are on the right side of “the line”
The wrong side of the line?
The development of RUV-2 and SVA occurred in the context of microarrays, and it is natural to ask whether the details are really different for RNA-Seq (spoiler: they aren’t). In a book chapter published earlier this year:
D. Risso, J. Ngai, T.P. Speed, S. Dudoit, The role of spike-in standards in the normalization of RNA-Seq, in Statistical Analysis of Next Generation Sequencing Data (2014), 169-190.
the authors replace “log expression levels” from microarrays with “log counts” from RNA-Seq and the linear regression performed with Limma for RUV-2 with a Poisson regression (this involves one different R command). They call the new method RUV, which is the same as the previously published RUV, a naming convention that makes sense since the paper has no new method. In fact, the mathematical formulas describing the method are identical (and even in almost identical notation!) with the exception that the book chapter ignores Z altogether, and replaces with O.
To be fair, there is one added highlight in the book chapter, namely the observation that spike-ins can be used in lieu of housekeeping (or other control) genes. The method is unchanged, of course. It is just that the spike-ins are used to estimate W. Although spike-ins were not mentioned in the original Gagnon-Bartsch paper, there is no reason not to use them with arrays as well; they are standard with Affymetrix arrays.
My one critique of the chapter is that it doesn’t make sense to me that counts are used in the procedure. I think it would be better to use abundance estimates, and in fact I believe that Jeff Leek has already investigated the possibility in a preprint that appears to be an update to his original SVA work. That issue aside, the book chapter does provide concrete evidence using a Zebrafish experiment that RUV-2 is relevant and works for RNA-Seq data.
The story should end here (and this blog post would not have been written if it had) but two weeks ago, among five RNA-Seq papers published in Nature Biotechnology (I have yet to read the others), I found the following publication:
D. Risso, J. Ngai, T.P. Speed, S. Dudoit, Normalization of RNA-Seq data using factor analysis of control genes or samples, Nature Biotechnology 32 (2014), 896-902.
This paper has the same authors as the book chapter (with the exception that Sandrine Dudoit is now a co-corresponding author with Davide Risso, who was the sole corresponding author on the first publication), and, it turns out, it is basically the same paper… in fact in many parts it is the identical paper. It looks like the Nature Biotechnology paper is an edited and polished version of the book chapter, with a handful of additional figures (based on the same data) and better graphics. I thought that Nature journals publish original and reproducible research papers. I guess I didn’t realize that for some people “reproducible” means “reproduce your own previous research and republish it”.
At this point, before drawing attention to some comparisons between the papers, I’d like to point out that the book chapter was refereed. This is clear from the fact that it is described as such in both corresponding authors’ CVs.
How similar are the two papers?
Final paragraph of paper in the book:
Internal and external controls are essential for the analysis of high-throughput data and spike-in sequences have the potential to help researchers better adjust for unwanted technical effects. With the advent of single-cell sequencing , the role of spike-in standards should become even more important, both to account for technical variability  and to allow the move from relative to absolute RNA expression quantification. It is therefore essential to ensure that spike-in standards behave as expected and to develop a set of controls that are stable enough across replicate libraries and robust to both differences in library composition and library preparation protocols.
Final paragraph of paper in Nature Biotechnology:
Internal and external controls are essential for the analysis of high-throughput data and spike-in sequences have the potential to help researchers better adjust for unwanted technical factors. With the advent of single-cell sequencing27, the role of spike-in standards should become even more important, both to account for technical variability28 and to allow the move from relative to absolute RNA expression quantification. It is therefore essential to ensure that spike- in standards behave as expected and to develop a set of controls that are stable enough across replicate libraries and robust to both differences in library composition and library preparation protocols.
Abstract of paper in the book:
Normalization of RNA-seq data is essential to ensure accurate inference of expression levels, by adjusting for sequencing depth and other more complex nuisance effects, both within and between samples. Recently, the External RNA Control Consortium (ERCC) developed a set of 92 synthetic spike-in standards that are commercially available and relatively easy to add to a typical library preparation. In this chapter, we compare the performance of several state-of-the-art normalization methods, including adaptations that directly use spike-in sequences as controls. We show that although the ERCC spike-ins could in principle be valuable for assessing accuracy in RNA-seq experiments, their read counts are not stable enough to be used for normalization purposes. We propose a novel approach to normalization that can successfully make use of control sequences to remove unwanted effects and lead to accurate estimation of expression fold-changes and tests of differential expression.
Abstract of paper in Nature Biotechnology:
Normalization of RNA-sequencing (RNA-seq) data has proven essential to ensure accurate inference of expression levels. Here, we show that usual normalization approaches mostly account for sequencing depth and fail to correct for library preparation and other more complex unwanted technical effects. We evaluate the performance of the External RNA Control Consortium (ERCC) spike-in controls and investigate the possibility of using them directly for normalization. We show that the spike-ins are not reliable enough to be used in standard global-scaling or regression-based normalization procedures. We propose a normalization strategy, called remove unwanted variation (RUV), that adjusts for nuisance technical effects by performing factor analysis on suitable sets of control genes (e.g., ERCC spike-ins) or samples (e.g., replicate libraries). Our approach leads to more accurate estimates of expression fold-changes and tests of differential expression compared to state-of-the-art normalization methods. In particular, RUV promises to be valuable for large collaborative projects involving multiple laboratories, technicians, and/or sequencing platforms.
Abstract of Gagnon-Bartsch & Speed paper that already took credit for a “new” method called RUV:
Microarray expression studies suffer from the problem of batch effects and other unwanted variation. Many methods have been proposed to adjust microarray data to mitigate the problems of unwanted variation. Several of these methods rely on factor analysis to infer the unwanted variation from the data. A central problem with this approach is the difficulty in discerning the unwanted variation from the biological variation that is of interest to the researcher. We present a new method, intended for use in differential expression studies, that attempts to overcome this problem by restricting the factor analysis to negative control genes. Negative control genes are genes known a priori not to be differentially expressed with respect to the biological factor of interest. Variation in the expression levels of these genes can therefore be assumed to be unwanted variation. We name this method “Remove Unwanted Variation, 2-step” (RUV-2). We discuss various techniques for assessing the performance of an adjustment method and compare the performance of RUV-2 with that of other commonly used adjustment methods such as Combat and Surrogate Variable Analysis (SVA). We present several example studies, each concerning genes differentially expressed with respect to gender in the brain and find that RUV-2 performs as well or better than other methods. Finally, we discuss the possibility of adapting RUV-2 for use in studies not concerned with differential expression and conclude that there may be promise but substantial challenges remain.
Many figures are also the same (except one that appears to have been fixed in the Nature Biotechnology paper– I leave the discovery of the figure as an exercise to the reader). Here is Figure 9.2 in the book:
The two panels appears as (b) and (c) in Figure 4 in the Nature Biotechnology paper (albeit transformed via a 90 degree rotation and reflection from the dihedral group):
Basically the whole of the book chapter and the Nature Biotechnology paper are essentially the same, down to the math notation, which even two papers removed is just a rehashing of the RUV method of Gagnon-Bartsch & Speed. A complete diff of the papers is beyond the scope of this blog post and technically not trivial to perform, but examination by eye reveals one to be a draft of the other.
Although it is acceptable in the academic community to draw on material from published research articles for expository book chapters (with permission), and conversely to publish preprints, including conference proceedings, in journals, this case is different. (a) the book chapter was refereed, exactly like a journal publication (b) the material in the chapter is not expository; it is research, (c) it was published before the Nature Biotechnology article, and presumably prepared long before, (d) the book chapter cites the Nature Biotechnology article but not vice versa and (e) the book chapter is not a particularly innovative piece of work to begin with. The method it describes and claims to be “novel”, namely RUV, was already published by Gagnon-Bartsch & Speed.
Below is a musical rendition of what has happened here:
“An entertaining freshness… Tic Tac!” This is Ferrero‘s tag line for its most successful product, the ubiquitous Tic Tac. And the line has stuck. As WikiHow points out in how to make your breath fresh: first buy some mints, then brush your teeth.
One of the amazing things about Tic Tacs is that they are sugar free. Well… almost not. As the label explains, a single serving (one single Tic Tac) contains 0g of sugar (to be precise, less than 0.5g, as explained in a footnote). In what could initially be assumed to be a mere coincidence, the weight of a single serving is 0.49g. It did not escape my attention that 0.50-0.49=0.01. Why?
To understand it helps to look at the labeling rules of the FDA. I’ve reproduced the relevant section (Title 21) below, and boldfaced the relevant parts:
TITLE 21–FOOD AND DRUGS
CHAPTER I–FOOD AND DRUG ADMINISTRATION
DEPARTMENT OF HEALTH AND HUMAN SERVICES
SUBCHAPTER B–FOOD FOR HUMAN CONSUMPTION
(c) Sugar content claims –(1) Use of terms such as “sugar free,” “free of sugar,” “no sugar,” “zero sugar,” “without sugar,” “sugarless,” “trivial source of sugar,” “negligible source of sugar,” or “dietarily insignificant source of sugar.” Consumers may reasonably be expected to regard terms that represent that the food contains no sugars or sweeteners e.g., “sugar free,” or “no sugar,” as indicating a product which is low in calories or significantly reduced in calories. Consequently, except as provided in paragraph (c)(2) of this section, a food may not be labeled with such terms unless:
(i) The food contains less than 0.5 g of sugars, as defined in 101.9(c)(6)(ii), per reference amount customarily consumed and per labeled serving or, in the case of a meal product or main dish product, less than 0.5 g of sugars per labeled serving; and
(ii) The food contains no ingredient that is a sugar or that is generally understood by consumers to contain sugars unless the listing of the ingredient in the ingredient statement is followed by an asterisk that refers to the statement below the list of ingredients, which states “adds a trivial amount of sugar,” “adds a negligible amount of sugar,” or “adds a dietarily insignificant amount of sugar;” and
(iii)(A) It is labeled “low calorie” or “reduced calorie” or bears a relative claim of special dietary usefulness labeled in compliance with paragraphs (b)(2), (b)(3), (b)(4), or (b)(5) of this section, or, if a dietary supplement, it meets the definition in paragraph (b)(2) of this section for “low calorie” but is prohibited by 101.13(b)(5) and 101.60(a)(4) from bearing the claim; or
(B) Such term is immediately accompanied, each time it is used, by either the statement “not a reduced calorie food,” “not a low calorie food,” or “not for weight control.”
It turns out that Tic Tacs are in fact almost pure sugar. Its easy to figure this out by looking at the number of calories per serving (1.9) and multiplying the number of calories per gram of sugar (3.8) by 0.49 => 1.862 calories. 98% sugar! Ferrero basically admits this in their FAQ. Acting completely within the bounds of the law, they have simply exploited an arbitrary threshold of the FDA. Arbitrary thresholds are always problematic; not only can they have unintended consequences, but they can be manipulated to engineer desired outcomes. In computational biology they have become ubiquitous, frequently being described as “filters” or “pre-processing steps”. Regardless of how they are justified, thresholds are thresholds are thresholds. They can sometimes be beneficial, but they are dangerous when wielded indiscriminately.
There is one type of thresholding/filtering in used in RNA-Seq that my postdoc Bo Li and I have been thinking about a bit this year. It consists of removing duplicate reads, i.e. reads that map to the same position in a transcriptome. The motivation behind such filtering is to reduce or eliminate amplification bias, and it is based on the intuition that it is unlikely that lightning strikes the same spot multiple times. That is, it is improbable that many reads would map to the exact same location assuming a model for sequencing that posits selecting fragments from transcripts uniformly. The idea is also called de-duplication or digital normalization.
Digital normalization is obviously problematic for high abundance transcripts. Consider, for example, a transcripts that is so abundant that it is extremely likely that at least one read will start at every site (ignoring the ends, which for the purposes of the thought experiment are not relevant). This would also be the case if the transcript was twice as abundant, and so digital normalization would prevent the possibility for estimating the difference. This issue was noted in a paper published earlier this year by Zhou et al. The authors investigate in some detail the implications of this problem, and quantify the bias it introduces in a number of data sets. But a key question not answered in the paper is what does digital normalization actually do?
To answer the question, it is helpful to consider how one might estimate the abundance of a transcript after digital normalization. One naive approach is to just count the number of reads after de-duplication, followed by normalization for the length of the transcript and the number of reads sequenced. Specifically if there are n sites where a read might start, and k of the sites had at least one read, then the naive approach would be to use the estimate suitably normalized for the total number of reads in the experiment. This is exactly what is done in standard de-duplication pipelines, or in digital normalization as described in the preprint by Brown et al. However assuming a simple model for sequencing, namely that every read is selected by first choosing a transcript according to a multinomial distribution and then choosing a location on it uniformly at random from among the sites, a different formula emerges.
Let X be a random variable that denotes the number of sites on a transcript of length n that are covered in a random sequencing experiment, where the number of reads starting at each site of the transcript is Poisson distributed with parameter c (i.e., the average coverage of the transcript is c). Note that
The maximum likelihood estimate for c can also be obtained by the method of moments, which is to set
from which it is easy to see that
This is the same as the (derivation of the) Jukes-Cantor correction in phylogenetics where the method of moments equation is replaced by yielding , but I’ll leave an extended discussion of the Jukes-Cantor model and correction for a future post.
The point here, as noticed by Bo Li, is that since by Taylor approximation, it follows that the average coverage can be estimated by . This is exactly the naive estimate of de-duplication or digital normalization, and the fact that as means that blows up, at high coverage hence the results of Zhou et al.
Digital normalization as proposed by Brown et al. involves possibly thresholding at more than one read per site (for example choosing a threshold C and removing all but at most C reads at every site). But even this modified heuristic fails to adequately relate to a probabilistic model of sequencing. One interesting and easy exercise is to consider the second or higher order Taylor approximations. But a more interesting approach to dealing with amplification bias is to avoid thresholding per se, and to instead identify outliers among duplicate reads and to adjust them according to an estimated distribution of coverage. This is the approach of Hashimoto et al. in a the paper “Universal count correction for high-throughput sequencing” published in March in PLoS One. There are undoubtedly other approaches as well, and in my opinion the issue will received renewed attention in the coming year as the removal of amplification biases in single-cell transcriptome experiments becomes a priority.
As mentioned above, digital normalization/de-duplication is just one of many thresholds applied in a typical RNA-Seq “pipeline”. To get a sense of the extent of thresholding, one need only scan the (supplementary?) methods section of any genomics paper. For example, the GEUVADIS RNA-Seq consortium describe their analysis pipeline as follows:
“We employed the JIP pipeline (T.G. & M.S., data not shown) to map mRNA-seq reads and to quantify mRNA transcripts. For alignment to the human reference genome sequence (GRCh37, autosomes + X + Y + M), we used the GEM mapping suite24 (v1.349 which corresponds to publicly available pre-release 2) to first map (max. mismatches = 4%, max. edit distance = 20%, min. decoded strata = 2 and strata after best = 1) and subsequently to split-map (max.mismatches = 4%, Gencode v12 and de novo junctions) all reads that did not map entirely. Both mapping steps are repeated for reads trimmed 20 nucleotides from their 3′-end, and then for reads trimmed 5 nucleotides from their 5′-end in addition to earlier 3′-trimming—each time considering exclusively reads that have not been mapped in earlier iterations. Finally, all read mappings were assessed with respect to the mate pair information: valid mapping pairs are formed up to a maximum insert size of 100,000 bp, extension trigger = 0.999 and minimum decoded strata = 1. The mapping pipeline and settings are described below and can also be found in https://github.com/gemtools, where the code as well as an example pipeline are hosted.”
This is not a bad pipeline- the paper shows it was carefully evaluated– and it may have been a practical approach to dealing with the large amount of RNA-Seq data in the project. But even the first and seemingly innocuous thresholding to trim low quality bases from the ends of reads is controversial and potentially problematic. In a careful analysis published earlier this year, Matthew MacManes looked carefully at the effect of trimming in RNA-Seq, and concluded that aggressive trimming of bases below Q20, a standard that is frequently employed in pipelines, is problematic. I think his Figure 3, which I’ve reproduced below, is very convincing:
It certainly appears that some mild trimming can be beneficial, but a threshold that is optimal (and more importantly not detrimental) depends on the specifics of the dataset and is difficult or impossible to determine a priori. MacManes’ view (for more see his blog post on the topic) is consistent with another paper by Del Fabbro et al. that while seemingly positive about trimming in the abstract, actually concludes that “In the specific case of RNA-Seq, the tradeoff between sensitivity (number of aligned reads) and specificity (number of correctly aligned reads) seems to be always detrimental when trimming the datasets (Figure S2); in such a case, the modern aligners, like Tophat, seem to be able to overcome low quality issues, therefore making trimming unnecessary.”
Alas, Tic Tac thresholds are everywhere. My advice is: brush your teeth first.
I was recently reading the latest ENCODE paper published in PNAS when a sentence in the caption of Figure 2 caught my attention:
“Depending on the total amount of RNA in a cell, one transcript copy per cell corresponds to between 0.5 and 5 FPKM in PolyA+ whole-cell samples according to current estimates (with the upper end of that range corresponding to small cells with little RNA and vice versa).”
Although very few people actually care about ENCODE, many people do care about the interpretation of RNA-Seq FPKM measurements and to them this is likely to be a sentence of interest. In fact, there have been a number of attempts to provide intuitive meaning for RPKM (and FPKM) in terms of copy numbers of transcripts per cell. Even though the ENCODE PNAS paper provides no citation for the statement (or methods section explaining the derivation), I believe its source is the RNA-Seq paper by Mortazavi et al. In that paper, the authors write that
“…absolute transcript levels per cell can also be calculated. For example, on the basis of literature values for the mRNA content of a liver cell [Galau et al. 1977] and the RNA standards, we estimated that 3 RPKM corresponds to about one transcript per liver cell. For C2C12 tissue culture cells, for which we know the starting cell number and RNA preparation yields needed to make the calculation, a transcript of 1 RPKM corresponds to approximately one transcript per cell. “
This statement has been picked up on in a number of publications (e.g., Hebenstreit et al., 2011, van Bakel et al., 2011). However the inference of transcript copies per cell directly from RPKM or FPKM estimates is not possible and conversion factors such as 1 RPKM = 1 transcript per cell are incoherent. At the same time, the estimates of Mortazavi et al. and the range provided in the ENCODE PNAS paper are informative. The “incoherence” stems from a subtle issue in the normalization of RPKM/FPKM that I have discussed in a talk I gave at CSHL, and is the reason why TPM is a better unit for RNA abundance. Still, the estimates turn out to be “informative”, in the sense that the effect of (lack of) normalization appears to be smaller than variability in the amount of RNA per cell. I explain these issues below:
Why is the sentence incoherent?
RNA-Seq can be used to estimate transcript abundances in an RNA sample. Formally, a sample consists of n distinct types of transcripts, and each occurs with different multiplicity (copy number), so that transcript i appears times in the sample. By “abundance” we mean the relative amounts where . Note that and . Suppose that for some j. The corresponding is therefore where . The question is what does this value correspond to in RPKM (or FPKM).
RPKM stands for “reads per kilobase of transcript per million reads mapped” and FPKM is the same except with “fragment” replacing read (initially reads were not paired-end, but with the advent of paired-end sequencing it makes more sense to speak of fragments, and hence FPKM). As a unit of measurement for an estimate, what FPKM really refers to is the expected number of fragments per kilboase of transcript per million reads. Formally, if we let be the length of transcript i and define then abundance in FPKM for transcript i is abundance measured as . In terms of , we obtain that
The term in the denominator can be considered a kind of normalization factor, that while identical for each transcript, depends on the abundances of each transcript (unless all lengths are equal). It is in essence an average of lengths of transcripts weighted by abundance. Moreover, the length of each transcript should be taken to be taken to be its “effective” length, i.e. the length with respect to fragment lengths, or equivalently, the number of positions where fragments can start.
The implication for finding a relationship between FPKM and relative abundance constituting one transcript copy per cell is that one cannot. Mathematically, the latter is equivalent to setting in the formula above and then trying to determine . Unfortunately, all the remaining are still in the formula, and must be known in order to calculate the corresponding FPKM value.
The argument above makes clear that it does not make sense to estimate transcript copy counts per cell in terms of RPKM or FPKM. Measurements in RPKM or FPKM units depend on the abundances of transcripts in the specific sample being considered, and therefore the connection to copy counts is incoherent. The obvious and correct solution is to work directly with the . This is the rationale of TPM (transcripts per million) used by Bo Li and Colin Dewey in the RSEM paper (the argument for TPM is also made in Wagner et al. 2012).
Why is the sentence informative?
Even though incoherent, it turns out there is some truth to the ranges and estimates of copy count per cell in terms of RPKM and FPKM that have been circulated. To understand why requires noting that there are in fact two factors that come into play in estimating the FPKM corresponding to abundance of one transcript copy per cell. First, is M as defined above to be the total number of transcripts in a cell. This depends on the amount of RNA in a cell. Second are the relative abundances of all transcripts and their contribution to the denominator in the formula.
The best paper to date on the connection between transcript copy numbers and RNA-Seq measurements is the careful work of Marinov et al. in “From single-cell to cell-pool transcriptomes: stochasticity in gene expression and RNA splicing” published in Genome Research earlier this year. First of all, the paper describes careful estimates of RNA quantities in different cells, and concludes that (at least for the cells studied in the paper) amounts vary by approximately one order of magnitude. Incidentally, the estimates in Marinov et al. confirm and are consistent with rough estimates of Galau et al. from 1977, of 300,000 transcripts per cell. Marinov et al. also use spike-in measurements are used to conclude that in “GM12878 single cells, one transcript copy corresponds to ∼10 FPKM on average.”. The main value of the paper lies in its confirmation that RNA quantities can vary by an order of magnitude, and I am guessing this factor of 10 is the basis for the range provided in the ENCODE PNAS paper (0.5 to 5 FPKM).
In order to determine the relative importance of the denominator in I looked at a few RNA-Seq datasets we are currently examining. In the GEUVADIS data, the weighted average can vary by as much as 20% between samples. In a rat RNA-Seq dataset we are analyzing, the difference is a factor of two (and interestingly very dependent on the exact annotation used for quantification). The point here is that even the denominator in does vary, but less, it seems, than the variability in RNA quantity. In other words, the estimate of 0.5 to 5 FPKM corresponding to one transcript per cell is incoherent albeit probably not too far off.
One consequence of all of the above discussion is that while differential analysis of experiments can be performed based on FPKM units (as done for example in Cufflinks, where the normalization factors are appropriately accounted for), it does not make sense to compare raw FPKM values across experiments. This is precisely what is done in Figure 2 of the ENCODE PNAS paper. What the analysis above shows, is that actual abundances may be off by amounts much larger than the differences shown in the figure. In other words, while the caption turns out to contain an interesting comment the overall figure doesn’t really make sense. Specifically, I’m not sure the relative RPKM values shown in the figure deliver the correct relative amounts, an issue that ENCODE can and should check. Which brings me to the last part of this post…
What is ENCODE doing?
Having realized the possible issue with RPKM comparisons in Figure 2, I took a look at Figure 3 to try to understand whether there were potential implications for it as well. That exercise took me to a whole other level of ENCODEness. To begin with, I was trying to make sense of the x-axis, which is labeled “biochemical signal strength (log10)” when I realized that the different curves on the plot all come from different, completely unrelated x-axes. If this sounds confusing, it is. The green curves are showing graphs of functions whose domain is in log 10 RPKM units. However the histone modification curves are in log (-10 log p), where p is a p-value that has been computed. I’ve never seen anyone plot log(log(p-values)); what does it mean?! Nor do I understand how such graphs can be placed on a common x-axis (?!). What is “biochemical signal strength” (?) Why in the bottom panel is the grey H3K9me3 showing %nucleotides conserved decreasing as “biochemical strength” is increasing (?!) Why is the green RNA curves showing conservation below genome average for low expressed transcripts (?!) and why in the top panel is the red H3K4me3 an “M” shape (?!) What does any of this mean (?!) What I’m supposed to understand from it, or frankly, what is going on at all ??? I know many of the authors of this ENCODE PNAS paper and I simply cannot believe they saw and approved this figure. It is truly beyond belief… see below:
All of these figures are of course to support the main point of the paper. Which is that even though 80% of the genome is functional it is also true that this is not what was meant to be said , and that what is true is that “survey of biochemical activity led to a significant increase in genome coverage and thus accentuated the discrepancy between biochemical and evolutionary estimates… where function is ascertained independently of cellular state but is dependent on environment and evolutionary niche therefore resulting in estimates that differ widely in their false-positive and false-negative rates and the resolution with which elements can be defined… [unlike] genetic approaches that rely on sequence alterations to establish the biological relevance of a DNA segment and are often considered a gold standard for defining function.”
The ENCODE PNAS paper was first published behind a paywall. However after some public criticism, the authors relented and paid for it to be open access. This was a mistake. Had it remained behind a paywall not only would the consortium have saved money, I and others might have been spared the experience of reading the paper. I hope the consortium will afford me the courtesy of paywall next time.
Last Monday some biostatisticians/epidemiologists from Australia published a paper about a “visualization tool which may allow greater understanding of medical and epidemiological data”:
H. Wand et al., “Quilt Plots: A Simple Tool for the Visualisation of Large Epidemiological Data“, PLoS ONE 9(1): e85047.
A brief look at the “paper” reveals that the quilt plot they propose is a special case of what is commonly known as a heat map, a point the authors acknowledge, with the caveat that they claim that
” ‘heat maps’ require the specification of 21 arguments including hierarchical clustering, weights for reordering the row and columns dendrogram, which are not always easily understood unless one has an extensive programming knowledge and skills. One of the aims of our paper is to present ‘‘quilt plots’’ as a useful tool with simply formulated R-functions that can be easily understood by researchers from different scientific backgrounds without high-level programming skills.”
In other words, the quilt plot is a simplified heat map and the authors think it should be used because specifying parameters for a heat map (in R) would require a terrifying skill known as programming. This is of course all completely ridiculous. Not only does usage of R not require programming skill, there are also simplified heat map functions in many programming languages/computation environments that are as simple as the quilt plot function.
The fact that a paper like this was published in a journal is preposterous, and indeed the authors and editor of the paper have been ridiculed on social media, blogs and in comments to their paper on the PLoS One website.
Wand et al. do have one point… those 21 parameters are not an entirely trivial matter. In fact, the majority of computational biologists (including many who have been ridiculing Wand) appear not to understand heat maps themselves, despite repeatedly (ab)using them in their own work.
What are heat maps?
In the simplest case, heat maps are just the conversion of a table of numbers into a grid with colored squares, where the colors represent the magnitude of the numbers. In the quilt plot paper that is the type of heat map considered. However in applications such as gene expression analysis, heat maps are used to visualize distances between experiments.
Heat maps have been popular for visualizing multiple gene expression datasets since the publication of the “Eisengram” (or the guilt plot?). So when my student Lorian Schaeffer and I recently needed to create a heat map from RNA-Seq abundance estimates in multiple samples we are analyzing with Ryan Forster and Dirk Hockemeyer, we assumed there would be a standard method (and software) we could use. However when starting to look at the literature we quickly found 3 papers with 4 different opinions about which similarity measure to use:
- In “The evolution of gene expression levels in mammalian organs“, Brawand et al., Nature 478 (2011), 343-348, the authors use two different distance measures for evaluating the similarity between samples. They use the measure where is Spearman’s rank correlation (Supplementary figure S2) and also the Euclidean distance (Supplementary figure S3).
- In “Differential expression analysis for sequence count data“, Anders and Huber, Genome Biology 11 (2010), R106, a heat map tool is provided that is based on Euclidean distance but differs from the method in Brawand et al. by first performing variance stabilization on the abundance estimates.
- In “Evolutionary Dynamics of Gene and Isoform Regulation in Mammalian Tissues“, Merkin et al., Science 338 (2012), a heat map is displayed based on the Jensen-Shannon metric.
There are also the folks who don’t worry too much and just try anything and everything (for example using the heatmap.2 function in R) hoping that some distance measure produces the figure they need for their paper. There are certainly a plethora of distance measures for them to try out. And even if none of the distance measures provide the needed figure, there is always the opportunity to play with the colors and shading to “highlight” the desired result. In other words, heat maps are great for cheating with what appears to be statistics.
We wondered… what is the “right” way to make a heat map?
Consider first the obvious choice for measuring similarity: Euclidean distance. Suppose that we are estimating the distance between abundance estimates from two RNA-Seq experiments, where for simplicity we assume that there are only three transcripts (A,B,C). The two abundance estimates can be represented by 3-tuples and such that both and . If and , then the Euclidean distance is given by . This obviously depends on and , a dependence that is problematic. What has changed between the two RNA-Seq experiments is that transcript has gone from being the only one transcribed, to not being transcribed at all. It is difficult to justify a distance metric that depends on the relative changes in and . Why, for example, should be closer to than to ?
The Jensen-Shannon divergence, defined for two distributions and by
where and is the Kullback-Leibler divergence, is an example of a distance measure that does not have this problem. For the example above the JSD is always (regardless of and ). However the JSD is not a metric (hence the term divergence in its name). In particular, it does not satisfy the triangle inequality (which the Euclidean distance does). Interestingly, this defect can be rectified by replacing JSD with the square root of JSD (the JSD metric). Formal proofs that the square root of JSD is a metric were provided in “A new Metric for Probability Distributions” by Dominik Endres and Johannes Schindelin (2003), and separately (and independently) in “A new class of metric divergences on probability spaces and its applicability in statistics” by Ferdinand Österreicher and Igor Vajda (2003). The paper “Jensen-Shannon Divergence and Hilbert space embedding” by Bent Fuglede and Flemming Topsøe (2004) makes clear the mathematical origins for this result by showing that the square root of JSD can be isometrically embedded into Hilbert space (as a logarithmic spiral)
The 2-simplex with contour lines showing points equidistant
from the probability distribution (1/3, 1/3, 1/3) for the JSD metric.
The meaning of the JSD metric is not immediately apparent based on its definition, but a number of results provide some insight. First, the JSD metric can be approximated by Pearson’s distance (Equation (7) in Endres and Schindelin). This relationship is confirmed in the numerical experiments of Sung-Hyuk Cha (see Figure 3 in “Comprehensive survey on distance/similarity measures between probability distance functions“, in particular the close relationship between JSD and the probabilistic symmetric ). There are also information theoretic and physical interpretations for the JSD metric stemming from the definition of JSD in terms of Kullback-Leibler divergence.
In “Transcript assembly and quantification by RNA-Seq reveals unannotated transcripts and isoform switching during cell differentiation“, Trapnell et al., Nature Biotechnology 28 (2010), we used the JSD metric to examine changes to relative isoform abundances in genes (see, for example, the Minard plot in Figure 2c). This application of the JSD metric makes sense, however the JSD metric is not a panacea. Consider Figure 1 in the Merkin et al. paper mentioned above. It displays a heat map generated from 7713 genes (genes with singleton orthologs in the five species studied). Some of these genes will have higher expression, and therefore higher variance, than others. The nature of the JSD metric is such that those genes will dominate the distance function, so that the heat map is effectively generated only from the highly abundant genes. Since there is typically an (approximately) exponential distribution of transcript abundance this means that, in effect, very few genes dominate the analysis.
I started thinking about this issue with my student Nicolas Bray and we began by looking at the first obvious candidate for addressing the issue of domination by high variance genes: the Mahalanobis distance. Mahalanobis distance is an option in many heat map packages (e.g. in R), but has been used only rarely in publications (although there is some history of its use in the analyses of microarray data). Intuitively, Mahalanobis distance seeks to remedy the problem of genes with high variance among the samples dominating the distance calculation by appropriate normalization. This appears to have been the aim of the method in the Anders and Huber paper cited above, where the expression values are first normalized to obtain equal variance for each gene (the variance stabilization procedure). Mahalanobis distance goes a step further and better, by normalizing using the entire covariance matrix (as opposed to just its diagonal).
Intuitively, given a set of points in some dimension, the Mahalanobis distance is the Euclidean distance between the points after they have been transformed via a linear transformation that maps an ellipsoid fitted to the points to a sphere. Formally, I think it is best understood in the following slightly more general terms:
Given an expression matrix (rows=transcripts, columns=experiments), let be the matrix consisting of projections of onto its principal components, and denote by the distance between projection of points i and j onto the kth component, i.e. . Let be the singular values. For some , define the distance
When and then the distance D defined above is the Mahalanobis distance.
The Mahalanobis ellipses. In this figure the distance shown is from every point to the center (mean of the point) rather than between pairs of points. Mahalanobis distance is sometimes defined in this way. The figure is reproduced from this website. Note that the Anders-Huber heat map produces distances looking only at the variance in each direction (in this case horizontal and vertical) which assumes that the gene expression values are independent, or equivalently that the ellipse is not rotated.
It is interesting to note that D is defined even when , providing a generalization of Mahalanobis distance for high-dimensional data.
The cutoff p involves ignoring the last few principal components. The reason one might want to do this is that the last few principal components presumably correspond to noise in the data. Amplifying this noise and treating it the same as the signal is not desirable. This is because as p increases the denominators get smaller, and therefore have an increasing effect on the distance. So even though it makes sense to normalize by variance thereby allowing all genes to count the same, it is important to keep in mind that the last few principal components may be desirable to toss out. One way one could choose the appropriate threshold is by examination of a scree plot.
We’re still not completely happy with Mahalanobis distance. For example, unlike the Jensen-Shannon metric, it does not provide a metric over probability distributions. In functional genomics, almost all *Seq assays produce an output which is a (discrete) probability distribution (for example in RNA-Seq the output after quantification is a probability distribution on the set of transcripts). So making heat maps for such data seems to not be entirely trivial…
Does any of this matter?
The landmark Michael Eisen et al. paper “Cluster analysis and the display of genome-wide expression patterns“, PNAS 95 (1998), 14863–14868 describing the “Eisengram” was based on correlation as the distance measure between expression vectors. This has a similar problem to the issues we discussed above, namely that abundant genes are weighted more heavily in the distance measure, and therefore they define the characteristics of the heat map. Yet the Eisengram and its variants have proven to be extremely popular and useful. It is fair to ask whether any of the issues I’ve raised matter in practice.
Depends. In many papers the heat map is a visualization tool intended for a qualitative exploration of the data. The issues discussed here touch on quantitative aspects, and in some applications changing distance measures may not change the qualitative results. Its difficult to say without reanalyzing data sets and (re)creating the heat maps with different parameters. Regardless, as expression technology continues to transition from microarrays to RNA-Seq, the demand for quantitative results is increasing. So I think it does matters how heat maps are made. Of course its easy to ridicule Handan Wand for her quilt plots, but I think those guilty of pasting ad-hoc heat maps based on arbitrary distance measures in their papers are really the ones that deserve a public spanking.
P.S. If you’re going to make your own heat map, after adhering to sound statistics, please use a colorblind-friendly palette.
P.P.S. In this post I have ignored the issue of clustering, namely how to order the rows and columns of heat maps so that similar expression profiles cluster together. This goes along with the problem of constructing meaningful dendograms, a visualization that has been a major factor in the popularization of the Eisengram. The choice of clustering algorithm is just as important as the choice of similarity measure, but I leave this for a future post.
The paper “Genomic-scale capture and sequencing of endogenous DNA from feces” by George H. Perry, John C. Marioni, Páll Melsted and Yoav Gilad is literally full of feces. The word ‘fecal’ appears 100 times.
Poop jokes aside, the paper presents an interesting idea that has legs. Perry et al. show that clever use of Agilent’s SureSelect allows for capturing nuclear genomic regions from fecal DNA. Intellectually, it is the predecessor of T. Mercer et al.‘s “Targeted RNA sequencing reveals deep complexity of the human transcriptome“, Nature Biotechnology, 2011 (another paper I like and for which I wrote a research highlight). Even though the Perry et al. paper does not have many citations, the Mercer et al. paper does (although unfortunately the authors forgot to cite Perry et al., which I think they should have). In other words, the Perry et al. paper is not as well known as it ought to be, and this post is an attempt to rectify that.
The ‘*’ in sh*t in my title is for *Seq. At a high level, the Perry et al. paper shows how high-throughput sequencing technology can be leveraged to sequence deeply a single genome from among a community of metagenomes. For this reason, and for convenience, I henceforth will refer to the Perry et al. paper as the Sh*t-Seq paper. The “*” is inserted in lieu of the “i” not as censorship, but to highlight the point that the method is general and applies not only to sequencing nuclear genome from fecal DNA, but also as Mercer et al. shows, for targeted transcriptome sequencing (one can imagine also many other applications).
The Sh*t-Seq protocol is conceptually simple yet complicated in practice. DNA was captured using the Agilent SureSelect target enrichment system coupled to the Illumina single-end sequencing platform library prep protocol (of note is that the paper is from 2010 and the kits are from 2009). However because of the very small amount of DNA targeted (the authors claim ~1.8%), a number of adjustments to standard SureSelect capture / Illumina library prep had to be implemented. To the authors credit, the paragraphs in the section “DNA Capture” are exemplary in their level of detail and presumably greatly facilitate replicability. I won’t repeat all the detail here. However there are two steps that caught my attention as possibly problematic. First, the the authors performed substantial PCR amplification of the adapter-ligated fecal DNA. This affects the computational analysis they discuss later and leads to a computational step they implemented that I have some issues with (more on this later). Second, they performed two rounds of capture as one round was insufficient for capturing the needed material for sequencing. This necessitated additional PCR, which also is possibly problematic.
The samples collected were from six chimpanzees. This is a fairly small n but the paper is a proof of principle and I think this is sufficient. Both fecal and blood samples were collected allowing for comparison of the fecally derived nuclear DNA to the actual genome of the primates. In what is clearly an attempt to channel James Bond, they collected fecal samples (2 g of stool) within 1 hour of defecation in tubes containing RNALater and these were then “shaken vigorously” (not stirred).
The next part of the paper is devoted to computational analyses to confirm that the Sh*t-Seq protocol can in fact be used to target nuclear endogenous DNA. As a sanity check, mtDNA was considered first. They noted too much diversity to align reads to a reference genome with BWA, and opted instead for de novo assembly using ABySS. This is certainly overkill, possible only because of the high copy count of mitochondrion reads. But I suppose it worked (after filtering out all the low coverage ABySS sequence, which was presumably junk). One interesting idea given more modern RNA-Seq assembly tools would be to assemble the resulting reads with an RNA-Seq de novo assembler that allows for different abundances of sequences. Ideally, such an assembly should indicate naturally the sought after enrichment.
Next, nuclear DNA was investigated, specifically the X chromosome and chromosome 21. Here the analyses is very pre-2014. First, all multi-mapping reads were removed. This is not a good idea for many reasons, and I am quite certain that with the new GRCh38 (with alternate sequence representation for variant regions) it is a practice that will rapidly be phased out. I’d like to give Perry et al. the benefit of the doubt for making this mistake since they published in 2010, but their paper appeared 6 months after the Cufflinks paper so they could have, in principle, known better. Having said that, while I do think multi-mapping would have allowed them to obtain much stronger results as to the accuracy and extent of their enrichment by avoiding the tossing of a large number of reads, their paper does manage to prove their principle so its not a big deal.
The removal of multi-mapping reads was just the first step in a series of “filters” designed to narrow down the nuclear DNA reads to regions of the chimpanzee genome that could be argued to be unambiguously representative of the target. I won’t go into details, although they are all in the paper. As with the experimental methods, I applaud the authors on publishing reproducible methods, especially computational methods, with all details included. But there was a final red flag for me in the computational methods: namely the selection of a single unique fragment (at random) for each genomic (start) site for the purposes of calling SNPs. This was done to eliminate problems due to amplification biases, which is indeed a serious concern, but if heterozygous sites appear due to the PCR steps, then there ought to have been telltale signatures. For example, a PCR “SNP’ would, I think, appear only in reads specific to a single position, but not in other overlapping reads of the site. It would have been very helpful if they would have done a detailed analysis of this issue, rather than just pick a single read at random for each genomic (start) site. They kicked the can down the road.
Having removed multiple mapping reads, repetitive regions, low coverage regions, etc. etc. Dayenu, they ended up estimating a false positive rate for heterozygous sites (using the X chromosome in males) at 0.0007% for fecal DNA and 0.0010% for blood DNA. This led them to conclude that incorrectly-identified heterozygous sites in their study were 0.8%, 2.0%, 1.1%, and 2.7% for fecal DNA chromosome 21, fecal DNA chromosome X in females, blood DNA chromosome 21, and blood DNA chromosome X in females, respectively. Such good news is certainly the result of their extraordinarily stringent filtering, but I think it does prove that they were able to target effectively. They give further proof using PCR and Sanger sequencing of 20 regions.
I have a final nitpick and it relates to Figure 4. It is a a companion to Figure 3 which shows the Chimpanzee phylogeny for their samples based on the mtDNA. As expected in that figure, the fecal and blood samples cluster together. Figure 4 shows two phylogenies, one based on chr 21, the other on chr X. My issue here is with the way that distances were constructed. Its a technical point, but it looks like they used hamming distance, and I don’t think that makes a lot of sense, not to mention the fact that neighbor-joining does not seem like the appropriate algorithm for building a tree in this setting (I plan to blog about neighbor-joining shortly). But this is a methodological point not really relevant to the main result of the paper, namely proof of principle for targeted sequencing of endogenous DNA from fecal matter.
I think Sh*t-Seq has a future. The idea of targeted capture coupled to high-throughput sequencing has more than an economic rationale. It provides the possibility to probe the “deep field” as discussed in the previously mentioned review on targeted RNA-Seq. This is a general principle that should be more widely recognized.
And of course, dung is just cool. Happy new year!