Wednesday, July 16, 2014

Trinity RNA-Seq

In January a site went live called readingroom.info. I suspect due to the timing and the subject matter it was a student project. The idea was to write summaries - no more than 500 words - of scientific papers and allow people to comment on and discuss them. I thought it was a neat idea. They had some ideas for incentivizing writers and so forth, but I didn't have time to contribute anything until this summer, by which time the authors had apparently lost interest. I sent in a summary of a paper, but after several weeks it had not been approved by the moderators. Maybe it just wasn't that good! 

The paper is Trinity: reconstructing a full-length transcriptome without a genome from RNA-Seq data

This is what I sent them:

The reconstruction of a transcriptome from the short reads generated by RNA-Seq techniques presents many challenges, particularly in the absence of an existing reference genome with which to compare the reads. Challenges include: uneven coverage of the various transcripts; (ii) uneven coverage inside each transcript; (iii) sequencing errors in highly expressed transcripts; (iv) transcripts encoded by adjacent loci can overlap and thus can be erroneously fused to form a chimeric transcript; (v) data structures need to accommodate multiple transcripts per locus, owing to alternative splicing; and (vi) sequences that are repeated in different genes introduce ambiguity. The Trinity pipeline leverages several properties of transcriptomes in its assembly procedure: it uses transcript expression to guide the initial transcript assembly procedure in a strand-specific manner, it partitions RNA-Seq reads into sets of disjoint transcriptional loci, and it traverses each of the transcript graphs systematically to explore the sets of transcript sequences that best represent variants resulting from alternative splicing or gene duplication by exploiting pairs of RNA-Seq reads. The series of steps performed by the pipeline correctly reconstructs a significant percentage of the transcripts without relying on the existence of a reference genome.

A major data structure used the pipeline is the de Bruijn graph. A de Bruijn graph places each k-mer in a node, and has connected nodes if the k-mers are identical in all but the first or last position. While an efficient structure for representing heavily overlapping sequences, there are challenges in the usage of these graphs: (i) efficiently constructing this graph from large amounts (billions of base pairs) of raw data; (ii) defining a suitable scoring and enumeration algorithm to recover all plausible splice forms and paralogous transcripts; and (iii) providing robustness to the noise stemming from sequencing errors and other artifacts in the data. Sequencing errors would introduce false nodes to the graph, potentially resulting in a great deal of wasted memory.

The Trinity pipeline consists of the following steps: it first analyzes the short reads to create a dictionary of all sequences of length 25 in the reads, indexing the locations where each sequence may be found. After removing likely errors, the unique k-mers are recombined, starting with the most frequently occurring sequences and extending the combination until no more k-mers can be matched. Each contig is then added to a cluster based on potential alternative spliced transcripts or otherwise unique portions of paralogous genes. Then, a de Bruijn graph is generated from each cluster with the weight of each edge assigned from the number of k-mers in the original read set that support the connection. In the final phase, a merge-and-prune operation on each graph, for error correction, is performed, followed by an enumeration of potential paths through the graph with a greater likelihood placed on paths with greater read support.

The authors built transcriptomes from both original data and reference sets, having a great deal of success in either case.

Tuesday, July 08, 2014

The Slim protocol

I wrote earlier about Fitnesse and Slim style integration testing.

There are only a few commands associated with Slim tables, but the precise meaning of the command is left to the Slim server. For example, according to the Slim documentation, the Import instruction:

causes the <path> to be added to the search path for fixtures. In java <path> gets added to the CLASSPATH. In .NET, the <path> would name a dll. 

This explains why import tests are always green. You can pass them any random string and it's just added to the path. In WaferSlim, the concept is extended slightly: if the string has a slash or backslash in it, it's considered a path. Otherwise, it's considered a file and is added to the list of files searched for classes. Let's take an example. Create a file called fixtures.py with the following code:

class SomeDecisionTable:
def setInput(self, x):
self.x = str(x)

def getOutput(self):
return int(self.x) + 1

There's an odd Unicode issue that I think requires context variables to be saved as strings. I don't know if Fitnesse, Slim or Waferslim is responsible for that. At any rate, it's simple enough: set the input to an integer and the output will be the integer plus one.

Save that file. Now the Import table will take two lines, one with the path, one with the file:

!|import |
|/path/to/fixtures|
|fixtures|

Now WaferSlim knows where to find our code, so let's take a look at another table.

Here's what the Slim protocol tells us: it can send Make instructions and Call instructions. A Make instruction tells the server to create an object; a Call instruction tells it to call a method on that object. I'd hoped to go through the Fitnesse code to determine exactly how that works, but I didn't. We'll take it on faith that the first line of a table sends a Make instruction and further lines send Call instructions. So, in order to make calls into our object, we write:

!|SomeDecisionTable|
|input|get output?|
|1    |2          |
|10   |11         |

From the first line of the table, Fitnesse sends a Make: SomeDecisionTable command to the Slim server. Because of our Import statements, the server searches the /path/to/fixtures directory for the fixtures.py file, finds the SomeDecisionTable class, and instantiates it. 

The second line of the table tells Fitnesse what Call statements to make. It will call setInput and then getOutput. Each further line of the table gives arguments for the input and output. So the sequence of events from the Slim server's perspective is as follows:

Make: SomeDecisionTable
Call: setInput(1)
Call: getOutput()
Call: setInput(10)
Call: getOutput()

Click the test button now and everything should run green.

Sunday, July 06, 2014

Fitnesse and Slim

Fitnesse is a rather nice testing tool. It's been around since what seems like the beginning of the integration testing movement. It's based around HTML tables that are translated into application code. The wiring that translates tables into code are called fixtures. Since Fitnesse was written in Java, it was mostly useful for testing Java code, although many translations and tricks were devised to allow tests of other languages. Fitnesse also includes a wiki to make the process of creating the test tables easier.

A few years after Fitnesse was introduced, an alternate translation tool, SLIM, was created. The idea behind Slim was to allow testers to implement tests in their language of choice, with the communication between Fitnesse and the tests taking place over a socket. This allows any application that implements the correct protocol to run tests written in Fitnesse tables.

Several applications were created to support the protocol in various languages, including RubySlim for Ruby and WaferSlim for Python. I had occasion to write a set of integration tests recently, so I thought it would be simple to set up a Slim server to get the tests running. Turned out, I was wrong.

For my tests, I wanted to set up the simplest thing that could possibly work. So what is the minimum requirement for a Slim test page? If you look at the RubySlim documentation you end up with a page that looks something like this:
    !define TEST_SYSTEM {slim}
    !define TEST_RUNNER {rubyslim}
    !define COMMAND_PATTERN {rubyslim}
    !path your/ruby/fixtures
 
    !|import|
    |<ruby module of fixtures>|
    |SomeDecisionTable|
    |input|get output?|
    |1    |2          |

What all these things do is not clearly explained. Click the test button; you get a bunch of cryptic error messages. So what do they mean?

Well, the first line is pretty clear: if you want to use Slim, you need to set the test system. For the rest of it, we pretty much need to go directly to the source code. It turns out that the heart of Slim is in the Java class ProcessBuilder. In the CommandRunner class of Fitnesse we find:

ProcessBuilder processBuilder = new ProcessBuilder(command);

The command argument is no more than the COMMAND_PATTERN variable defined on the page, with a single argument appended, a port number. So if you wanted to use WaferSlim (a Python Slim server), you might say:

!define COMMAND PATTERN {python /home/benfulton/slim-init.py}

(Since we need to understand how Slim works, let's get the WaferSlim source from Github rather than installing it via Pip. The slim-init source is from this gist.)

So, after downloading WaferSlim and the code from the gist, we should be able to get the Simplest Possible Page to work. Here it is:

!define TEST_SYSTEM {slim}
!define COMMAND_PATTERN {python /home/benfulton/slim-init.py }

!|import |
|RubySlim|
|app|

Click the test button, and you should get a green test.

If you don't, you may have a Slim versioning error. Go to protocol.py and change the version number from 0.1 to 0.3. Or Python may not be in your path, or the init-script may not be found. Or - here's a tricky one - you put a space before the python command. If you did that, the first command that gets passed to the ProcessBuilder class will be an empty string, and it won't run. Still, this will run green without too much effort.

But wait! What's the use of this import statement that went green before we wrote anything?


Sunday, June 02, 2013

Suffix Arrays

In our quest to win a million dollars, we're trying to Index the Human Genome. A typical index, like in the back of a book, is composed of an entry, say "Sheep", and a list of page numbers, say "88,121,265". But when we went to try this idea on the 3,000,000,000 characters in the genome, we found that it took up significantly more space than the genome itself, which was unworkable. Surely we can find a more efficient solution to the problem. (Or, we could throw up our hands, call it "Big Data", and put a supercomputer to work on it. Let's not do that.)

Actually, there's a fair amount of redundancy in an index. If there are three instances of the word "Sheep" in our book, and we add a fourth instance of it to the index, we're increased the number of sheep, so to speak, by 33%. If we do that for every word in the book we'll add a significant chunk to the overall size.

But suppose, rather than just give the user the page numbers where the word "Sheep" is, we provided them with with the line and word numbers as well. "Sheep" is on page 88, line 14, word six. Now, the index is in alphabetical order of course, so what we can do is simply eliminate all those redundant words from the index. So the index entry for "Sheep" would just say "88,14,6". Say readers want to find the word "Streams". They would go and find an entry "88,14,6" and look up the word in the book. Finding that it's "Sheep", they realize the word is later in the index, since "Streams" is alphabetically after "Sheep". They go to the next entry in the index, maybe "81,19,11", and look up that word in the book. It's "Streams", so they've found their word, and the index didn't require any of those annoying, redundant words in it!

Sheep by a Stream
Sheep. By a stream.
OK, not a very simple operation for a human reader. But easy enough for a computer. And since our genome doesn't have pages or lines, we could simply record the location of each individual 20-mer and put the locations in alphabetical order. We can even take it one step further: since 20 is just an arbitrary length that we chose, we'll remove it from the solution and just say that we'll take as many characters as we need to get a unique ordering of all of the strings. Notice that you might need to take all of the remaining characters of the genome to alphabetize it correctly, and if you do, you have a suffix of the genome. If you don't need all the characters, it doesn't matter if you add them or not, so you might as well, and therefore we have an array of the suffixes of the genome. A suffix array.

I won't go into the details of how to create such an index right now, but it can be done in relatively few lines of code. (One easy-to-use library is called SAIS.) Now, the simplest way to write out a suffix array is a text file of the alphabetized suffix indices, in ASCII format, one number per line. This, unfortunately, brings us straight back to the size problem - let's say eight bytes on average, per line, with one line per character in the genome, and we end up with 24 gigabytes worth of index. But at least it's a workable index. If we split the index into several files to keep it manageable, it even suggests a refinement of our attack on the overall problem. We'll see how another time.

(Update: I wrote an article on suffix array algorithms.)

Part I: A Million Dollars Up For Grabs
Part II: Analyzing DNA with BWA
Part III: Analyzing DNA Programmatically
Part IV: Indexing the Human Genome
Part V: Suffix Arrays

Sunday, May 19, 2013

Indexing the human genome

Last time I had decided that to efficiently analyze the reads, we had to make an index of the human genome. So how do we go about that?

Library of Congress Classification - Reading Room
Indexing at the Library of Congress
What we need is an efficient way to access any given substring in the genome. It's not quite the same as indexing a book; rather than determining the locations of "dynamo", "father", and "pseudopodia" in the book, we need to be able to find the location of EVERY substring. It's as if, in our book, we had to find instances of "dynamo", "ynamo", "namo" and so on. Not only that, but if the book had the sentence, "A dynamo has unlimited duration." we have to find instances of "namo h", "namo ha" "namo has" and so on.

So we can't just split the genome by word boundaries like we would for a book. Can we split it into even-sized chunks and index those? For example, could we choose a chunk size of five and split every ten characters into two index entries?

It won't work. For example, if the genome was "AGACTTGCTG", we might choose to index every five characters (called a 5-mer). This would give us two strings, "AGACT" and "TGCTG", which is fine, but if we come along later and try to search for "CTTGC", we're out of luck - that's not in our index. But it is in the string.

So we have to choose an index size and go through the genome character by character. In our ten-character genome, indexing by 5-mer, we get the strings:

AGACT
GACTT
ACTTG
CTTGC
TTGCT
TGCTG

(and a few shorter strings at the end, if we so desire.)

To be useful, we'll have to store our index as a dictionary of 5-mers to an array of integers, representing the locations in the genome where that string was found. For our sample, we have 30 characters worth of 5-mers, and just six integers to save, for a total of 54 bytes. What happens when we index the whole genome?

It's a fair assumption that every possible string will be in the genome, so we'll have 4or 1024 entries in our index. The number of values, though...we have to have an index value for every single character in the genome. If our genome was the book Moby Dick, we'd have around a million characters to index. Each index value would have, on average, about 1000 items in it. If we're really hoping to match all of our reads to each item in the index, we're going to go through 1000 entries for each read, which might be a bit slow. Unfortunately, our genome isn't Moby Dick. It's roughly the size of 6,000 Moby Dicks. Each index is going to be in just about six million locations. It can't possibly read all those at the speed we require.

Okay, so if we have too many items per index value, we just have to make our index larger. What if we do 10-mers, or even 20-mers? Well, 10-mers gives us 410 entries - call it a million. That means each index will be found in about 6,000 locations. Kind of a lot, but maybe doable. If we do 20-mers - well, 420 is 1,099,511,627,776. This is more on the lines of what we need in an index - it's well beyond the number of characters in the genome, so each index shouldn't show up in more than one or two locations. There's just one small problem: We now have six billion entries with twenty-character identifiers, and the space our index needs is now up to 120 gigabytes!

Maybe we could tweak and tune and find a sweet spot, but instead we'll try a different approach entirely to indexing. Next time.

Part I: A Million Dollars Up For Grabs
Part II: Analyzing DNA with BWA
Part III: Analyzing DNA Programmatically
Part IV: Indexing the Human Genome



Monday, May 13, 2013

Analyzing DNA programmatically

Last time I discussed using a program called BWA to try to determine if a given sequence was part of the human genome or not. It didn't seem to do a great job. How do we approach it programmatically?

Let's review the problem first. You have a machine that has analyzed some DNA, and it's given you some sequences it's found. Roughly 300,000 of them. Some with maybe only five characters, some with a few hundred. We call them "reads".

Genomes of Canis lupus major ups! familiaris : Assembled chromosomes
Recall that a DNA sequence, from a programming standpoint, consists of any number of any of four characters: A,C,G, and T. A human genome, in the end, is just a DNA sequence. So if the human sequence was "AAGGTTTCC" and your sequence is "AGGT", bing! You've got a match.

Now, here's the catch: a read, instead of being four characters, is reasonably a minimum of fifty characters. The human genome, instead of being nine characters, is roughly six billion characters. 6,000,000,000 characters. If your first thought is to fire up a text editor and do a search, you had better make sure that (a) the editor is capable of handling a six gigabyte file, and (b) that your computer has six gigs available to throw around (In 2013, a high-end laptop will probably have eight gigs, so just barely enough space to hold the genome and run a couple of other programs.)

Another, minor issue is that most genome files are formatted with sixty-character lines. If your sequence is split across two lines, a simple text search won't find it. You can preprocess the file to remove all the carriage returns, but if you do, your text editor had better be able to handle one line with 6 billion characters on it!

Your next idea: grep. Will grep find a 50-character match in a genome? I couldn't find a good way to do it reliably, since, again, genome files tend to be formatted into 60-character lines, and grep isn't really capable of finding a string that breaks across a line. Moreover, grep is a line-based tool, so odds are if you try preprocessing the line, grep will think (rightly) that you have a file with one six-billion character line in it, try to read in the line, and run out of memory.

Here's the other catch: Suppose you could get grep to work, and it managed to search the whole genome in three seconds. But we have 300,000 reads to get through! That's about 250 hours worth of search time and we're under a three-hour deadline. So we can't really afford to search the entire genome for each read.

So we'll write a script. A good processing plan might be this: index the genome, index the reads, then march down the two together to find our list of best prefixes. Sounds great!

But it turns out that even indexing a file the size of the human genome is a pain. I'll think about why next time.

Also see: A Million Dollars Up For Grabs

Saturday, March 23, 2013

Analyzing DNA with BWA

Last time I discussed a million-dollar challenge posted by Innocenture. The challenge is to analyze a series of DNA reads and determine the source. My professor suggested using a program named BWA, a Burrows-Wheeler Aligner. Burrows-Wheeler alignment is an algorithm for matching up two sequences which may not be the same length. Perhaps one sequence has two extra nucleotides in the middle but the entire rest of the sequences match. These kinds of sequences are very difficult to find naively, and which match is better can come down to a judgment call.

So the goal is to align each sequence in the example file to the human genome. BWA, for speed, asks that you index the target genome - the human genome, in this case - which takes a fair amount of time. Then, you pass it each sequence on the command line, with a command something like:

>bwa align AAGCTCTA human_genome

and it does its magic and returns the location in the genome of the best several matches.

So I worked up a Python script to analyze the input file and pass each sequence to BWA in that format, and let it run for a few hours. (I'm never sure how many because I never remember to change the machine's power settings to not switch itself off after a few minutes of no UI activity). But it eventually completed. I went to look at the results, and I found them very intriguiing. Although, according to the challenge, no less than 90% of the DNA was human, the BWA program only managed to match about 50%, about 150,000 reads.

I took a closer look at the BWA manual, and found this option:

-o INTMaximum number of gap opens [1]

If I understand this correctly, it means that if the alignment has more than one gap in it, BWA will discard it as not being a match. You can change the value of this parameter, but when I did, it seemed to slow down the analysis quite a bit. I didn't let it complete, but based on the portions I did run I suspect it would have gone over the three-hour limitation - at least on my workstation, which I'm sure is underpowered compared to the target hardware.

So I started to think about what sort of coding would need to be done to meet this challenge. Next time I'll think about Analyzing DNA Programmatically