Presentation on theme: "Algorithms for Multisample Read Binning"— Presentation transcript:
1 Algorithms for Multisample Read Binning Student: Gabriel IlieMajor advisor: Ion MăndoiuAssociate Advisor: Sanguthevar RajasekaranAssociate Advisor: Yufeng WuUniversity of ConnecticutNovember 2013
2 OutlineMotivationPrevious approachesAlgorithmsResultsOngoing work
3 Human microbiomeThe microorganisms inhabiting our bodies comprise the microbiome.Microbes outnumber our own cells by 10 to 1 [Wooley et al, 2010].Diabetes, obesity, cancer and even attractiveness to mosquitoes seem to correlate with changes in our microbiome [Turnbaugh et al, 2010].To better understand our own condition we also need to understand the composition of the microbial communities inhabiting our bodies and how they interact not just between themselves but also with their habitat (e.g. the host organism).
4 Single organism studies they rely on clonal culturesmost microorganism cannot be cultivatedsuffer from amplification biasmicrobes do not live in single species communitiesmembers of these communities interact not just with each other but also with their habitats, which includes the host organismData obtained from clonal cultures is highly biased and does not capture a true picture of microbial life.
5 TranscriptomicsGenomes only provide information about the potential function of organisms.Having a gene does not mean that this gene is also expressed inside the host or during a particular condition.The transcriptome is the set of all RNA molecules produced in one or a population of cells.In order to understand the physiology of the microorganisms we need to know their transcriptome.
6 MetatranscriptomicsThe transcriptome of a community is the union over the transcriptomes of each of its members.In metatranscriptomic studies:bulk RNA is extracted directly from environmental samplesthe RNA is reverse-transcribedthe resulting RNA-Seq libraries are sequencedMetatranscriptomic studies are essential in order to understand the physiology of the microbiome.
7 Challenges of working with metatranscriptomic data The volume of sequencing data is several orders of magnitude larger than single organisms.Reads can come from hundreds of different species, each with a different abundance level.In addition to having a range in the abundance levels of the microorganisms, genes are expressed at drastically different levels.Genes usually have multiple isoforms.Metatranscriptomics has to deal with all of the challenges of metagenomics (1 and 2) plus some extra challenges (3 and 4), therefore algorithms devised for metagenomic data can also be applied to metatranscriptomic data.
10 OutlineMotivationPrevious approachesAlgorithmsResultsOngoing work
11 Clustering reads into bins Analysis of environmental samples is difficult.To simplify the assembly process, many metagenomic tools have been developed to cluster the reads into bins (i.e. species).Algorithms developed for binning metagenomic reads can also be applied to (meta)transcriptomic reads (bins represent transcripts instead of species).
12 Types of reads binning algorithms Genome dependentCompostBin(2008)Metacluster(2012)DNA composition patterns.G+C content, dinucleotide frequencies vary amongst species.Drawbacks:achieve reasonable performance only for long reads (800~1000 bp, [Wu et al, 2011])NGS technologies produce short readsGenome independentAbundanceBin(2011)MultiBin(2011)K-mer frequencies are usually linearly proportional to a genome’s abundance.Sufficiently long k-mers are usually unique.Works with short sequencing reads.Drawbacksgroup together reads from different species if they have close abundance levelsdo not perform well on species with low abundances
13 Metacluster (2012) Two round unsupervised binning algorithm: the first round clusters high-abundance readsfilter reads with 16-mer that appear less than T timesthe reads are grouped based on shared 36-mersthe groups are clustered using 5-mer distributionsthe second round clusters the remaining reads (low- abundance)those with unique 16-mers are discardedthe reads are grouped based on shared 22-mersthe groups are clustered using 4-mer distributionsAdvantages:better binning of reads from low abundance speciesuses both techniques: k-mer abundances and DNA composition patterns
14 MultiBin (2011)Processes multiple samples (N > 1) of the same microbial community.Clusters the reads into b bins (b is the number of species).Binning algorithm:all reads are pooled togethera graph G=(V, E) is generatedV is the set of readsedges connect reads with substantial overlap (50 bp)greedily partition the vertices into a set, the tags, s.t. each read is either a tag or affiliated with one which substantially overlaps itcluster the set of tags using Vt=(ct1,ct2, …, ctN), cti=number of reads from sample i which substantially overlap tag teach non-tag read is assigned to the same bin as its affiliated tagb needs to be known in advance or estimated somehow.Uses abundance differences from any of the samples to tell low abundance species apart.The algorithm is quadratic in the total number of readsBinning algorithm:all reads are pooled togethera graph G=(V, E) is generatedV is the set of readsedges connect reads with substantial overlap (50 bp)greedily find a maximal independent set in G, the set of tagseach read is now either a tag or affiliated with a tag which substantially overlaps itVt=(ct1,ct2, …, ctN), cti=number of reads from sample i which substantially overlap tag tperform k-medoids clustering on the set of tagsassign each non-tag read to thebin which holds its affiliated tag
15 OutlineMotivationPrevious approachesAlgorithmsResultsOngoing work
16 Our approachWe propose a novel method for unsupervised abundance-based multiple samples reads binning algorithmin brief, for N>1 samples:we split the reads into k-mers and count how many times they appear in each samplewe run a sample-by-sample error removal algorithmwe pool the k-mer counts together to get N-dimensional count vectorswe run the error removal algorithm once more on all of the k-merswe use the structure of the de Bruijn graph defined by the k-mers and the counts to partition the graph into paths or chordless cycles (putative pseudo- exons)
17 Pseudo-exonsPseudo-exons are defined as substrings whose k-mers and (k+1)- mers appear in the same transcripts with the same multiplicity.T1T2T3T4ABCDABCAABCEFFEDExon signatures:A: T1x1 T2x2 T3x1B: T1x1 T2x1 T3x1C: T1x1 T2x1 T3x1D: T1x1 T3x1E: T3x1 T4x1F: T3x1 T4x1Pseudo-exons:ABCDEFNotice that exons E and F form different pseudo-exons because in T4 they are in reverse order compared to T3.
18 K-mer counting we count k-mers and (k+1)-mers using Jellyfish (2011) Jellyfish was designed for shared memory parallel computers with more than one core, it uses several lock-free data structures and multi-threading to count k-mers much faster than other toolsformally, for a given value of k, we count the number of occurrences of all k-mers in each of the N samplesour algorithms assume we have strand nonspecific data, therefore the counts of complementary k-mers are summed together as they are indistinguishablewe store k-mers in canonical form (the smaller value lexicographically between a k-mer and its reverse complement)we combine the counts over all the samples into a list of N- dimensional vectorsthe maximum value for k supported by Jellyfish is 31
19 De Bruijn graphWe construct the de Bruijn graph; vertices are k-mers and edges are (k+1)-mers.Because we store vertices in canonical form, each vertex represents two k-mers: itself and its reverse-complement.We add an edge between any two k-mers if and only if there is a (k+1)-mer in the reads such that its prefix of length k in canonical form matches one of the vertices, while its suffix of length k in canonical form matches the other one.We will sometimes use vertices to refer to k-mers and edges to refer to (k+1)-mers.
20 De Bruijn graphACGCGTCGCGCGVerticesACGCGCCGAATCEdgesACGCCGCGACGAATCGCGATCGATCGATThe lists of k-mers and (k+1)-mers define an implicit representation of the de Bruijn graph, therefore we don’t need to construct it explicitly.Relative to the canonical form of a vertex, for each edge we can say whether it is incoming or outgoing:if the the k-mer matches the 5’ end of either forms of an edge then that is an outgoing edgeelse it is an incoming edge
21 Error removalA common approach is to remove k-mers which have counts lower than t >= 1.We found that even for t = 1, we lose too much information, because removing unique k-mers compromises the results for ultra- low abundance transcripts.We found that “tip removal” and “bubble removal” give much better results [Zerbino et al, Velvet, 2008].These methods use the structure of the de Bruijn graph instead of coverage information to remove k-mers affected by sequencing errors.
22 Tip and bubble error removal When a read contains a sequencing errorthe first few k-mers may be correct, until they start to overlap the position where the error occurredthis creates a branch going out of the “correct” paththis new branch will either end in a leaf (creating a “tip”), or if the read is long enough, the k-mers will stop overlapping the error and the branch will merge back into the path (creating a “bubble”)A “tip” is a chain of nodes that is disconnected on one endwe expect the majority of tips to have a maximum length of 2kremoving tips is straightforward; we remove all tips which have a length up to some thresholdremoving a tip does not disrupt the connectivity of the graphImplementing “bubble” removal is still an ongoing work.
23 Partitioning the de Bruijn graph From the de Bruijn graph we want to extract putative pseudo-exons.These putative pseudo-exons, if we ignore self-edges, correspond to paths or chordless cycles in the de Bruijn graph.We use the structure of the graph and the vectors with the counts to do the partitioning.
24 Partitioning the de Bruijn graph assuming perfect data, finding the putative pseudo-exons would simply mean removing all incoming/outgoing edges out of vertices that have an in/out degree greater than 1because we have sequencing errors we need to distinguish between erroneous and real edgeswe have the following two casesif we have a correct edge and at least one wrong edge coming out of the same vertex, then we expect the abundance of the correct edge to dwarf the sums of the erroneous, therefore the technique described earlier would remove the wrong edges and keep the correct oneif we have two correct edges coming out of the same vertex we want to remove both of them, assuming none of the edges comes from ultra-low abundance transcripts, then we expect none of the edges to pass the ratio test and our algorithm should remove both of them
25 Partitioning the de Bruijn graph We have the following cases:If a vertex has indegree and outdegree equal to at most 1 (it is on a path), we do nothing.If a vertex has outdegree (indegree) greater than 1,then we remove all outgoing (incoming) edges from that vertexhowever, we keep the most abundant edge if the ratio between its abundance and the sum of the abundances of the other edges is higher than a threshold 0 < e < 1The value of e should be close to 1 (e.g. 0.97)
26 Our approachWe propose a novel method for unsupervised abundance-based multiple samples reads binning algorithmin brief, for N>1 samples:we split the reads into k-mers and count how many times they appear in each samplewe run a sample-by-sample error removal algorithmwe pool the k-mer counts together to get N-dimensional count vectorswe run the error removal algorithm once more on all of the k-merswe use the structure of the de Bruijn graph defined by the k-mers and the counts to partition the graph into paths or chordless cycles (putative pseudo- exons)
27 OutlineMotivationPrevious approachesAlgorithmsResultsOngoing work
28 Test data - error freeGNF Atlas [Su et al, 2004] is a dataset which contains information about the expression levels of a set of genes in several human tissuesFrom this dataset we extracted the expression levels of 19,371 genes in 10 human tissuesWe used only one isoform per geneWe simulated 30 million error free RNA-Seq paired- reads of length 50 from this dataset using a tool called Grinder (2012)
29 Test data - with sequencing errors Grinder was very useful for simulating the error free data, however when we wanted to introduce errors its long running time became an issue.Instead, we simulated sequencing errors by using the error free data.We simulated only one type of errors, substitutions, because these are the most common type found in Illumina datasets.We introduced substitutions into the error free reads with a probability of 0.1%, 0.5% and 1% per base.
30 K-mer counts in the simulated data transcriptsreads#30-mers#31-merserrork#correctk-mers#incorrect#missing k-merspercentage of incorect k-mers49,572,54349,611,6910%3049,512,83959,7043149,546,27965,4120.1%49,508,364109,380,69064,17968.84%49,540,750108,528,92570,94168,66%0.5%49,483,820404,415,62288,72389.1%49,510,299402,714,823101,39289.05%1%49,433,000708,460,569139,54393.48%49,446,451706,531,643165,24093.46%Because of ultra-low abundance transcripts, even in the error free data we have missing k-mersEven for an error rate of 0.1% we notice that the number of unique k- mer more than triples when compared to the error free data, 70% of which do not appear in the transcriptsThis shows the importance of error removal/correction algorithms
31 Efficiency of different error removal techniques Error removal methoderror#correct 31-mers#incorrect 31-mers#missing 31-mersnone0%49,546,27965,412non-unique in atleast 1 sample46,520,4863,091,2050.1%46,257,2106,906,0693,354,481remove tips <= 21 over the union of the samples49,502,47021,772,808109,221remove tips <= 60 over the union of the samples49,236,25512,120,069375,436remove tips <= 21 sample-by-sample and over the union of the samples49,127,45610,998,956484,235remove tips <= 60 sample-by-sample and over the union of the samples48,707,5675,304,983904,124
32 Results for the graph partitioning errorerror removaltechniqueedge removalthreshold#pseudo-exons>= 50bp#pseudo-exons >= 50bp anddo not have wrong 31-mers#30-mers%transcriptome covered bycol 50%none160,28549,299,66599.5%65,05149,239,33599.3%0.1%remove tips <= 60 over the union of the samples0.97416,85360,33537,815,25676.3%remove tips <= 60 sample-by-sample and over the union of the samples228,84177,01046,699,96394.2%The first row (green) uses all k-mers, the counts are computed from the transcripts.
33 OutlineMotivationPrevious approachesAlgorithmsResultsOngoing work
34 Ongoing workWe believe that bubble removal will help us get rid of most of the erroneous k-mers which are still present in the data after tip removal.We want to incorporate an error correction algorithm, SEECER (2013), to correct the erroneous reads before doing starting the k-mer counting.Currently our algorithms do not take into account strand strand-specific RNA-Seq data. Optimizing the algorithms to take advantage of this information, when available, represents another opportunity to improve the results of this approach.