Whole Genome Alignment using Multithreaded Parallel Implementation

1
Whole Genome Alignment using Multithreaded
Parallel Implementation

• Hyma S Murthy
• CMSC 838 Presentation

2
Talk Overview

• Organization of the paper
• Motivation
• Technique
• Pairwise Sequence Comparison using Dynamic
  Programming
• EARTH Execution Model
• Evaluation
• Result Graphs
• Conclusions
• Related Work (MUMmer)

3
Motivation

• Importance of Genome Alignment
• Identify important matched and mismatched regions
• matches represent homolog pairs, conserved
  regions or long repeats
• mismatchesrepresent foreign fragments inserted
  by transposition, sequence reversal or lateral
  transfer
• Detect functional differences between pathogenic/
  non-pathogenic strains, evolutionary distance,
  mutations leading to disease, phenotypes, etc.
• Problems
• Large computational power, memory and execution
  time
• Existing algorithms apply dynamic programming
  only to subsequences
• Computationally intensive to apply to whole
  sequences (O(n2))
• Thus applicable only to closely related genomes

4
Solution..

• Multithreaded parallel implementation of sequence
  alignment algorithm to align whole genomes
• Parallel implementation of dynamic programming
  technique
• Uses collective memory of several nodes
• Uses multithreading to overlap computation and
  communication
• Applicable to closely related as well as less
  similar genomes
• Reliable output in reasonable time

5
Pairwise Sequence Comparison using Dynamic
Programming

• Basic Idea
• Quantify the similarity between pairs of symbols
  of target sequences
• Associate score for each possible arrangement
• Similarity is given by the highest score
• Example
• sequence x A T A A G T
• sequence y A T G C A G T
• SCORE 1 1 -1 1 1 1 1
  TOTAL -3
• sequence x A T A - A G T
• sequence y A T G C A G T
• SCORE 1 1 -1 2 1 1 1
  TOTAL 2
• Model mutation by gaps (gaps indicate evolution
  of one sequence into another)
• 
  
  

6
Dynamic Programming

• Smith and Waterman approach
• Aligns subsequences of given sequences
• Involves (a) calculation of scores indicating
  similarity
• (b) identification of alignment(s)
  corresponding to the score
• Build solution using previous solutions for
  smaller subsequences
• Construct a two-dimensional array Similarity
  Matrix to store scores corresponding to partial
  results
• Matrix represents all possible alignments of the
  input sequences
• Recurrence equation

SMi, j-1 gp SMi-1, j-1 ss SMi-1, j gp
0
SMi, j
7
Contd.

• Each element of the matrix is the max of the foll
  four values
• Left element gap, upper-left element score of
  replacing vertical
• with horizontal symbol, upper element gap, 0.
• Consider the foll example
• T G A T G G A G
  G T

G
2 max0 (-2), 1 (1), 0 (-2), 0
A
T
A
G
G
8
Identifying alignments

• Alignments with score above a given threshold are
  reported
• Start at end of the alignment and move backwards
  to the beginning
• T G A T G G A G G T

T G A T G G A G G T G A T A G G
G
A
T G A T G G A G G T G A T A G G
T
T G A T G G A G G T G A T A G G
A
G
T G A T G G A G G T G A T A G G
G
9
EARTH Execution Model

• Program is viewed as a collection of threads
• execution order determined by data and control
  dependencies
• Threads further divided into fibers
• fibers are non-preemptive and
• all data is ready before their execution
• Each node in EARTH has
• an execution unit
• synchronization unit
• queues linking the two (RQ and EQ)
• local memory
• interface to interconnection network

10
EARTH Architecture
Memory bus
node

PE
PE
PE
. .
node
To EQ
Inter connection Network
From RQ
node
EU
Local Memory
EQ
RQ
SU
11
Multithreaded parallel implementation

• Divide scoring matrix as follows
• horizontal strips (each element of input sequence
  X)
• strips into rectangular blocks
• Blocks are calculated by two fibers within a
  thread
• only one fiber is active at any given time
• Each thread is assigned to one horizontal strip
• the computation is done by even/ odd fibers
  within the thread
• Initialization delay of reading sequences from
  server is minimized
• Each thread needs only the piece of input
  sequence it grabs and not the whole of sequence X
• After computing a block, fiber sends to fiber
  beneath a piece of sequence Y
• among other information
• The computation of the anti-diagonal elements of
  the matrix is as shown

12
Computation of similarity matrix on EARTH
P1 P2 P3
Thread A
Thread B
Inactive fiber
E fibers O
E fibers O
Active fiber
Ack
Sync
Data
P1
P2
P3
P4
P1
P2
P3
P4
13
Evaluation

• Experimental environment
• Beowulf implementation of EARTH
• Uses Beowulf machine consisting of 64 nodes, each
  containing two 200MHz Pentium Pro processors (a
  total of 128 processors and 128MB of memory)
• Sequences of lengths ranging from 30K to 900K
  were tested
• Execution times for sequential and parallel
  implementation of Smith and Waterman algorithm is
  given below

14
Evaluation

• The multithreaded parallel implementation is
  named ATGC Another Tool for Genomic Comparison
• Experiment alignes
• human and mice mitochondrial genomes
• human and drosophila mitochondrial genomes
• Reason for selection
• human and mice are closely related and the other
  pair are less similar
• The results were confirmed with MUMmer another
  whole genome alignment tool
• Result graphs show that ATGC is more accurate
  than MUMmer
• (verified by using NCBI Blast)

15
Result Graphs
16
Contd.
17
Conclusions

• Comparison of whole genomes requires high
  computation and memory
• Made convenient by using a multithreaded parallel
  implementation of dynamic programming on a
  cluster of PCs
• Accurate results obtained in reasonable amount of
  time
• Aligns closely related as well as less similar
  genomes
• Slower, but plays important role where high
  accuracy is needed
• ( as seen in comparison with MUMmer for human
  and drosophila mitochondrial genome)

18
Related work MUMmer(Maximal Unique Match)

• given genomes A and B
• find all maximal, unique, matching subsequences
  (MUMs)
• extract the longest possible set of matches that
  occur in the same order in both genomes
• close the gaps
• output the alignment
• maximal unique match (MUM)
• occurs exactly once in both genomes A and B
• not contained in any longer MUM
• key idea in identifying MUMs is to build a suffix
  tree for genomes A and B