Parallel Detection of Regulatory Elements with gMP

1
Parallel Detection of Regulatory Elements with gMP
Bertil Schmidt, Lin Feng, Amey Laud, Yusdi Santoso
Damayanti Gupta CMSC 838 Presentation
2
Motivation

• Fundamental question
• How are expression levels of thousands of genes
  regulated ?
• Very important
• Understanding of gene function
• Response to environment
• Understand genetic causes of diseases
• Evaluate effects of drus
• Detect mutations
• Remember
• Sets of genes -gt Pathways -gt Genetic Networks
• Gene regulation
• Control decisions turn genes on/off
• Gene Regulation Network

3
Talk Overview

• Overview of talk
• Motivation
• Technique
• Experiment
• Related work
• Conclusions

4
Technique

• Motifs upstream of genes regulate gene expression
• Motifs are sites of regulatory activity
• Identify regulatory motifs by combining
• Gene expression data
• Detect common motifs occuring upstream of genes
• Huge datasets
• Utilise parallel computing

5
Technique

• gRNA
• Java development framework
• gMP
• Java communication library
• REDUCE
• Algorithm to identify regulatory motifs
• REDUCE parallelised with gMP
• Increase computing power
• Get motifs ranked in statistical significance

6
gRNA framework

• Consists of APIs

7
gRNA - APIs

• Interact with data sources
• Provide functionality from biology
• Pipelines tasks into unified process
• Repository of resources
• Distributed programming

8
gRNA environment

• gRNA Grid
• Clustered computing environment
• Application written for gRNA
• Multiple-tier application
• Applications operate from client computer
• Communicates with cluster through single computer
• Hosts EJB server
• Server identifies processing nodes
• each of these perform tasks

9
gRNA Grid
10
gMP

• Java based message passing tool
• Built on top of sockets
• Manages virtual processors to run on available
  machines
• Scalable
• Machines added/removed easily

11
gMP

• Processes are grouped
• Communication primitives provided for sending and
  receiving data
• Collective communication to several nodes enabled
  modularly and efficiently
• Enables functions to be implemented on data

12
REDUCE algorithm

• Based on model
• Upstream motifs contribute additively to
  expression level of each gene
• Quantify the extent to which these motifs
  contribute to expression data
• Fit log of expression ratio to sum of activating
  and inhibitory terms
• Find stastically most significant motifs
• Plots of fitting parameters suggest biological
  function

13
REDUCE algorithm

• Terms
• Occurence vector
• Measure of how often a motif is found
• Expression vector
• Measure of gene expression

14
REDUCE method

• Consists of
• 1) Motif frequency counter
• counts occurrences of DNA motifs upstream of each
  ORF
• motifs are about 711 nucleotides in length
• get occurence vectors

15
REDUCE algorithm

• 2) Significant motif finder
• Use
• i) Normalised occurrence vector made for each
  motif nµ
• ii) Normalised vector of logs of gene expression
  ratio vectors- a
• Take dot product of these (a . nµ) ,and square.
• Can be considered as frequency of occurence X
  expressive power of regulatory motif
• It is squared to get rid of negatives
• Correlate gene expression with occurence of motif
• Largest dot product is most significant motif

16
....

• a is modified to remove effect of this motif
• residual gene expression vector
• Process repeated until motifs are ranked

17
Table Finding significant motifs

• Uses a - (.5816,.2522,.2886,-.5947, -.1595,
  -.3683)

18
REDUCE parallelised with gMP...

• Parallel motif frequency counter
• Split set of ORFs equally
• Distribute across available nodes
• Each node calculates in parallel to get occurence
  vectors
• Matrix transposition
• Occurence vectors scattered across nodes
• Advantageous to store each vector in single node
• Transpose motif frequency matrix
• For each ORF can only calculate fraction of
  occurence frequencies for all motifs
• But the entire occurence frequency is needed

19
...

• Parallel significant motif finder
• Normalises occurence vector within each node
• At each node, most significant motif calculated
• Global most significant motif calculated
• Process iterated to rank occurence vectors
• Interface in gRNA allows ease of implementation

20
Experiment

• Use Compaq Alpha system
• Consists of cluster of 8 AlphaServer SC/ES45
• Connected by high-speed Alpha SC 16-Port switch
  and ELAN PCI adapter cards.
• Each server contains 4 Alpha EV68 processors

21
Results

• Use 7090 gene expressions of yeast
• ORFs of length 600
• Motifs upto length 7
• Throughput (in MBytes/s) also shown
• 20 most significant motifs computed.

22
Analysis

• Runtime scales well with number of processing
  nodes
• Frequency counter scales perfectly
• Motif finder also scales
• Cannot achieve perfect scaling because of
  communication overhead.

23
Related work

• DiscoveryLink
• Provides configurable wrappers as interfaces to
  multiple data sources
• Kleisli system
• Systematically manages and integrates external
  databases
• Uses functional query language to perform
  correlation across databases
• Toolkits designed with functionality for
  specialised areas
• BioJava, BioPerl, PAL
• Sequence Analysis
• Ensembl initiative, DAS
• provide extensible approach to issue of
  annotating genomic data

24
Related work

• Previous approaches using Java for high
  performance computing
• Bindings into native message-passing
  APIs(e.g.MPI)
• Does not allow easy integration into larger Java
  applications
• Pure Java message passing interfaces
• JMPI, CCJ
• Both implemented on top of Java RMI
• Slower than using raw sockets
• CCJ tries to overcome
• optimised RMI implementation
• not portable
• Both cannot handle integration

25
Comparison

• According to authors ...
• gRNA distinguishes itself
• Uses whole range of requirements for applications
  in computational biology
• Provides decoupled, yet inter-related subsystems
• Ease of 3rd party implementation

26
Observations

• REDUCE surpasses traditional clustering approach
• REDUCE algorithm has high runtime
• Complexity depends on product of number possible
  motifs and that of genes.
• Grows exponentially with length of sequences
• So length of motif is restricted
• REDUCE algorithm is greedy
• suboptimal
• REDUCE is simplistic
• lacks parameters for interactions between motifs
• does not consider impact of other biological
  knowledge

27
...

• Not clear that results of REDUCE are biologically
  significant
• Experiment does not effectively show how higher
  computation power helps results
• Only analysis from 9 to 16 processors, is this
  sufficient to determine good scaling?

28
Conclusions

• Finally...
• gRNA demonstrates efficient mechanism for
  development of genome-centric applications
• Further...
• Extensions to REDUCE have been proposed
• require higher computing power
• more specialised programming interfaces required
• Identifying communication patterns
• Use of data structures e.g. sequences, trees,
  matrices