1 Velvet: Algorithms for De Novo Short Assembly Using De Bruijn Graphs March 12, 2008 Daniel R. Zerbino and Ewan Birney Presenter: Seunghak Lee

2 What is de Bruijn Graphs? “De Bruijn graph” is a directed graph An edge represents overlap between sequences of symbols V=(s 1, s 2, …, s m ) E={(v 1,v 2,…, v n ),(w 1,w 2,…,w n )):v 2 =w 1,v 3 =w 2, …, v n =w n-1 }

3 Introduction New sequencing techniques are commercially available (e.g. 454 Sequencing, Solexa) 454 Sequencing ~ 100 – 200bp Solexa ~ 30bp Algorithms whole genome shotgun (WGS) assembly are not suitable for short reads  Overlap graph with a node per read is extremely large  More ambiguous connections in assembly

4 Introduction (cont) Euler assembler (Pevzner 2001) used k-mer for a node of de Bruijn graphs Reads are mapped as a path through the de Brujin graph High redundancy does not affect the number of nodes “Velvet” effectively deals with experimental errors and repeats by using Brujin graphs with k-mers

5 De Bruijn Graphs - structure Structure

6 De Bruijn Graphs – structure (cont) Adjacent k-mers overlap by k-1 nucleotides Each node is attached to twin node  Reverse series of reverse complement k-mers  Overlap between reads from opposite strand Union of a node and its twin node is called a “block” Last k-mer overlaps with the first of its destination

7 De Bruijn Graphs – construction (cont) Construction Reads are hashed with predefined k-mer length Small k-mer → increase connectivity → more ambiguous repeats Large k-mer → increase specificity → decrease connectivity Determine k considering “sensitivity” and “specificity”

8 De Bruijn Graphs – construction (cont) For each k-mer, hash table records ID of the first read and its position Each k-mer is recorded with reverse complement Node is created if there is distinct interruption points Reads are traced through the graph Create a directed arc if necessary

9 De Bruijn Graphs – simplification Simplify the chains of blocks  No information loss If node A has only one outgoing arc to node B, and if node B has only one ingoing arc → merge AB

10 De Bruijn Graphs – error removal Velvet focuses on “topological features” of the graph First step: remove tips  Tip: chain of nodes disconnected on one end Use two criteria: (1) length and (2) minority count  Length: remove a tip if < 2k bp since two nearby errors can create a tip up to 2k bp error k k

11 De Bruijn Graphs – error removal (cont) Minority count: multiplicity m < n Starting from node B, going through the tip is an alternative to a more common path m n B tip A C

12 De Bruijn Graphs – error removal (cont) Second step: remove bubbles using Tour Bus Redundant paths start and end at the same nodes Bubbles are created by errors or biological variants such as SNP Bubble

13 De Bruijn Graphs – error removal (cont) 1.Detect redundant paths 2. Compare them using dynamic programming methods 3. If similar, merge them Tour Bus

14 De Bruijn Graphs – error removal (cont) Third step: remove erroneous connections Remove erroneous connections after Tour Bus algorithm Remove erroneous connections with basic coverage cutoff Genuine short nodes which cannot be simplified in the graph should have high coverage

15 Breadcrumb: resolution of repeats 1. Using read pairs, pair up the long nodes 2. Flag paired reads using unambiguous long nodes unambiguous long nodes

16 Breadcrumb: resolution of repeats 1. Using read pairs, pair up the long nodes 2. Flag paired reads using unambiguous long nodes unambiguous long nodes

17 Breadcrumb: resolution of repeats Extends the nodes as far as possible using flagged paired reads All nodes between A and B are paired up to either A or B

18 Experimental Results Test error removal pipeline on simulated data Simulate reads are from E. coli, S. cerevisiae, C.elegans, and H. sapiens Coverage density vs N50 for H. sapiens Limited by natural repetition of the reference genome Ideal+ Error (1%)+ SNP N50

19 Experimental Results (cont) Test error removal pipeline on experimental data 173,428 bp human BAC was sequenced using Solexa machines Reads were 35bp long, and k=31 Tour Bus increased sensitivity by correcting errors and preserved the integrity of the graph structure

20 Experimental Results (cont)

21 Experimental Results (cont)

22 Conclusions Velvet is a de Bruijn graph based sequence assembly method for short reads Errors are handled by removing tips and Tour Bus algorithm A large number of repeats are resolved by Breadcrumb algorithm Velvet was assessed using simulated and real datasets and it performed well