Investigating effective and novel techniques to conduct pairwise alignment of very large genomic seq

1
Investigating effective and novel techniques to
conduct pairwise alignment of very large genomic
sequences. William D. Kenworthy Semester 2, 2003
2
Overview

• Aims of the project
• Background to genomic sequence alignment process
• Literature Review
• Proposed new approach
• Feature Based Sequence Alignment algorithm (FBSA)
• Results and discussion
• Conclusions and future directions

3
Aims

• The aims of this project
• Investigate the application of existing
  algorithms to very large genomic sequence
  alignment
• Propose a novel technique aligning very large
  sequences
• To evaluate the performance of the proposed system

4
Background
Simple Representation of the Human Genome
1
X
Y
22
Chromosome 6
3000 million nucleotides
Chromosome 6p21
170 Million nucleotides
19,894,959 nucleotides
3,246 nucleotides
TCAGGACCGCGACAAGCCAGCCCAGATCCGCTTCAGCAACATTTCCGCCG
CCAAAGCGGT...
1,082 amino acids
MTDDKDVLRDVWFGRIPTCFTLYQDEITEREAE...
1 or more proteins
5
Background
Genomic Features

• Various features exist in genomic sequences
• Features such as genes are often fragmented and
  scattered across a local area of the sequence

One gene
19,894,959 nucleotides
3,246 nucleotides
TCAGGACCGCGACAAGCCAGCCCAGATCCGCTTCAGCAACATTTCCGCCG
CCAAAGCGGT...

• Nucleotide or DNA (deoxyribonucleic acid)
  molecules are each represented by one of the
  letters A, C, G or T

6
Background
Alignment

• Sequence alignment involves comparing two or more
  sequences alongside each other in order to
  examine them for differences and similarities.

Individual 1
TCAGGACCGCGACAAGCCAGCCCAGATCCGCTTCAGCAACATTTCCGCCG
CCAAAGCGGT...
TCAGGACCGCGACATGCCCGC--AGATCCGCTTCAGCAACATTTCCGCCG
CCAAAGCGGT...
Individual 2

• Common algorithmic approaches are concerned with
  attempting to insert gaps in order to align the
  symbols into their correct evolutionary position

7
Literature Review

• A number of algorithms are currently in use for
  aligning small to medium genomic sequences (up to
  a few thousand nucleotides in size)
• Smith and Waterman considered the most accurate
• Also improvements by Gotoh, Miller and Webb have
  further improved the basic algorithm
• BLAST or Basic Local Alignment and Search tool is
  a fast, heuristic local alignment algorithm
• FASTA, another heuristic local alignment
  algorithm
• These are all local alignment algorithms and are
  concerned with finding the best local matches
  between areas of high similarity

8
Many possible alignments

• It has calculated there are a large number of
  possible alignments between any two sequences of
  length n
• Problem Long sequences (gt1000 symbols and above)
  can take a very long time and a large amount of
  resources to align (accurate or not!).

9
Literature Review
Aligning very large sequences

• proposals based on patterns or motifs
• a group of heuristic algorithms have been
  developed.
• SHAHA, MUMmer, pipmaker/blastz, ...
• These involve various methods of calculating
  unique (often overlapping) motifs to characterise
  the sequences
• These motifs are then assembled to create the
  final alignment
• In some cases, a secondary alignment is done to
  improve the accuracy.

10
Literature Review
Assemble the motifs
Individual 1
19,894,959 nucleotides
TCAGGACCGCGACAAGCCAGCCCAGATCCGCTTCAGCAACATTTCCGCCG
CCAAAGCGGT...
TCAGGACCGCGACAAGCCAGCCCAGATCCGCTTCAGCAACATTTCCGCCG
CCAAAGCGGT...
19,894,959 nucleotides
Individual 2
11
Literature Review
disadvantages of motif based approaches

• although fast and efficient in resource usage
• but
• miss-alignment can and does occur
• motifs are often small, and overlapping so gains
  are sometimes small
• Some are very heuristic in their approach -
  accuracy suffers so much a second stage of
  alignment is required.
• Not aware of biological features in a sequence so
  can distort them (don't ignore features)

12
Proposed new approach to aligning very large
sequences

• Genomic sequences already contain features such
  as repetitive elements, genes, low complexity
  regions etc.
• This is the basis of the new approach "Feature
  Based Sequence Alignment process" or "FBSA"
• These features characterise the sequence, and
  intuitively, can be seen to be conserved across
  two related sequences.
• Instead of discarding them as the existing
  alignment process dictates, we include them in
  the process!

13
FBSA
Genomic sequences already contain natural motifs
...

• Genes and other features are already present, and
  can be used to guide the alignment process

Individual 1
19,894,959 nucleotides
TCAGGACCGCGACAAGCCAGCCCAGATCCGCTTCAGCAACATTTCCGCCG
CCAAAGCGGT...
TCAGGACCGCGACAAGCCAGCCCAGATCCGCTTCAGCAACATTTCCGCCG
CCAAAGCGGT...
19,894,959 nucleotides
Individual 2
14
Hypothesis
The hypothesis for this project is that "it is
possible to devise an approach for aligning very
large genetic sequences in a fast, efficient and
intuitive manner by making use of existing
information which is held within the genetic
structure evidenced by the sequences under
examination, to guide the alignment process."
15
The Plan

• Investigate the features that best characterise
  the physical structure of a large genetic
  sequence.
• Investigate methods to extract and use the most
  suitable features to guide the alignment process.
• Design and test new algorithms based on features
• Examine the results and report on findings

16
Why repetitive elements?

• It became apparent that the class of features
  known as "repetitive elements" were ideal
  subjects.
• Very common in many genomes (Human, Mouse, Rice)
• Size is generally a few hundred base pairs
• Class includes "Low Complexity Repeats"
• Software to identify these elements is part of
  any standard alignment process (which normally
  results in the removal of the identified
  elements)

17
Low Complexity Regions

• Areas of simple repetitive patterns create
  difficulties for algorithmic approaches to
  alignment.
• AAAAAAATAAAAAAATAAAAAAAT
• Problem Common algorithms require such regions
  to be removed before attempting alignment to
  improve performance and accuracy.
• FBSA can make use of this information

18
FBSA algorithm
Identify features (e.g., using repeatmasker)
in both original sequences

• repeatmasker is a software program used to
  identify features by using a stored lookup table
• The output files are used by most standard
  algorithms to "mask" out and remove repetitive
  elements and loq complexity regions

19
FBSA algorithm
Identify features (e.g., using repeatmasker)
in both original sequences
Create a symbol table from the identified
features

• A symbol table is created to map symbols to
  different types of feature
• Computers manipulate symbols - not molecules!
• The symbols used are integers
• easy and simple to manipulate

20
FBSA algorithm
Identify features (e.g., using repeatmasker)
in both original sequences
Create a symbol table from the identified
features
Create feature sequence files for repeats
identified

• A vector is created for each sequence with
  symbols representing each of the features
  identified in the original sequences, in the
  correct order of appearance.

21
FBSA algorithm
Identify features (e.g., using repeatmasker)
in both original sequences
Create a symbol table from the identified
features
Create feature sequence files for repeats
identified
Align feature files using an adapted Smith and
Waterman algorithm.

• The core of the new algorithm
• The Smith and Waterman algorithm has been adapted
  to use integer symbols instead of alpha characters

22
FBSA algorithm
Identify features (e.g., using repeatmasker)
in both original sequences
Create a symbol table from the identified
features
Create feature sequence files for repeats
identified
Align feature files using an adapted Smith and
Waterman algorithm.
Subdivide both original input sequences on
identified feature boundaries

• Each matched pair of features AND the
  corresponding areas between features is generated
  as individual pairs of sub-sequences.

23
FBSA algorithm
Identify features (e.g., using repeatmasker)
in both original sequences
Create a symbol table from the identified
features
Create feature sequence files for repeats
identified
Align feature files using an adapted Smith and
Waterman algorithm.
Subdivide both original input sequences on
identified feature boundaries

• Each individual pair of sub-sequences is aligned
  using a standard algorithm (ClustalW was used in
  this project)

Align sub-sequences (e.g., using Blast, Fasta,
SSearch, or ClustalW)
24
FBSA algorithm
Identify features (e.g., using repeatmasker)
in both original sequences
Create a symbol table from the identified
features

• Using information from the symbol table and
  repeatmasker lists, reassemble the individual,
  and now aligned sub-sequences into a
  super-sequence in one of the standard formats

Create feature sequence files for repeats
identified
Align feature files using an adapted Smith and
Waterman algorithm.
Subdivide both original input sequences on
identified feature boundaries
Align sub-sequences (e.g., using Blast, Fasta,
SSearch, or ClustalW)
Assemble results and prepare for viewing or
further processing
25
FBSA
Nucleotides are now pre-aligned

• Gaps have been inserted at natural boundaries to
  put orthogonal features in the correct
  relationship

Individual 1
TCAGGACCGCGACAAGCCAGCCCAGATCCGCTTCAGCAACATTTCCGCCG
CCAAAGCGGT...
TCAGGACCGCGACAAGCCAGCCCAGATCCGCTTCAGCAACATTTCCGCCG
CCAAAGCGGT...
Individual 2
26
Characteristics

• The resulting super-alignment has a number of
  desirable properties
• major insertions and deletions are made at a high
  level, and at a biologically significant boundary
  - not in the middle of important features as
  machine generated motif based methods may do.
• The pre-aligned status of each-subsequence allows
  one of the standard dynamic algorithms to be used
  closer to its area of most efficient operation.
• Assembly of the sub-sequences is accurate because
  it is under the control of the original
  information.

27
Performance Evaluation
Quality

• Measurement of alignment quality is usually a
  combination of
• Number of gaps
• Number of matches
• Percent identity
• Measured by EMBOSS "infoalign"
• Time to produce the final alignment
• How does FBSA rate

28
Time

• Timings are for the "pre-alignment" stage
• The results reflect the fact that FBSA only needs
  to handle hundreds of features instead of
  thousands of bases
• Comparisons with ClustalW AL022723 vs AF055066
• FBSA 18 minutes (1000 pairs of sub-sequences)
• possibility to parrallelise and obtain further
  gains
• ClustalW gt 11 hours (1 pair)

29
Feature Density

• General performance is dependent on the number of
  features present
• Some areas of the genome are characterised as
  feature rich, and others as feature poor
• Index was relatively constant for test sequences
  used.
• New statistical methods for assessing sequence
  quality are required when large genetic sequences
  are involved

30
Future paths

• Many avenues for further research became
  apparent
• weighted substitutions based on feature homology
• cross-species performance
• mostly tested on human sequences
• some aribidopsis and rice sequences used.
• tuning of the type of features used
• is it appropriate to include LCRs?

31
Future paths

• Generating meaningful views of large sequence
  data
• Parrallelisation and job management
• Upper and lower size limits
• Investigate feature fragmentation

32
Publication
Bellgard, M. and Kenworthy, W. (to be published
2003) FBSA feature-based sequence alignment
technique for very large sequences Accepted for
publication by Applied Bioinformatics.
33
I would like to acknowledgment the contributions
of

• A number of people who were involved with project
  including
• Professor M. Bellgard for the original idea, and
  for his supervision of the project
• David Shibeci for his help in provision of
  computational resources, and the programming of
  some of the software modules used in this
  research
• David Dunn for his advice and patient
  explanations on matters biological
• Adam Hunter for data preparation when using the
  CMAP utility

34
Finally

• The FBSA algorithm developed during the course of
  this research improves on the performance of
  current algorithms when applied to very large
  genomic sequences.
• It has opened up many avenues for further
  research
• Publication implies that the concept is
  considered viable in the wider community of
  computational biology

35
Questions?