Whole Genome Phylogenetic Analysis

1
Whole Genome Phylogenetic Analysis

• Yifeng Liu and Reihaneh Rabbanyk Khorasgani
• April 8th, 2009

2
Agenda

• Introduction
• Our method proposals
• Datasets and experiments
• Results
• Discussion
• Future work
• Conclusion

3
Whole Genome Phylogeny Motivations

• Currently the dominant method for phylogenetic
  analysis is based on a single gene or protein.
• However different gene tells a different story
• Recently more genomic sequences became available
• We hope to resolve the above inconsistency by
  using the entire genome (or proteome) to
  reconstruct phylogenetic tree.

4
Whole Genome Phylogeny Methods

• Major categories of methods are based on
• Shared gene (ortholog) content
• Nucleotide and amino acid (string) composition
• Genome Compression
• Gene order
• In our study, we focus on string composition and
  compression methods

5
Complete Composition Vector (CCV)

• The observed occurrence probability for a
  k-string
• The estimated background occurrence probability
  based on the Markov assumption is

6
Complete Composition Vector (CCV)

• The occurrence probability due to selective
  pressureThe k-th composition vectorThe
  Complete Composition Vector (CCV)

7
Compression Methods

• Kolmogorov Complexity
• Lempel-Ziv complexity

8
Agenda

• Introduction
• Our method proposals
• Datasets and experiments
• Results
• Discussion
• Future work
• Conclusion

9
A new term weighting scheme

• CCV uses S() to weight each k-string, which
• Utilizes only local information available within
  a single sequence
• Estimates random background based on Markov
  model
• Can we have a measure that use both local and
  global information without making the Markov
  assumption?

10
Term and Document Frequency

• Genomes are documents written in a language of
  four alphabets A,T,C,G similarly, proteomes
  are documents written in a language of twenty
  alphabets.
• Each k-string can be viewed as a word within a
  gnome (or proteome) document.
• The collection of all genomes in the dataset is
  therefore a corpus.

11
Term and Document Frequency

• In statistical Natural Language Processing, a
  well-known term weighting scheme TF-IDF combines
  both term frequency and document frequency into a
  single weight.

12
CCV meets Document Frequency

• We can also combine the occurrence probability
  due to selection S() with the inverse document
  frequency into a single weight called CCV-IDF.
  
• S() provides local information and dfi provides
  global information.

13
Ensemble Measures

• Normalizing distances to same range
• Combining distance matrixes
• These parameters should be adjusted

14
Tree Evaluation

• We propose a new evaluation method for evaluating
  phylogenetic trees
• A numeric measure
• Shows how compatible the tree is with the given
  taxonomy

15
Tree Evaluation (Cont.)

• Labeling the inner nodes in the tree
• For each species
• A path in the tree
• ? sequence of inner node labels
• A taxonomy description
• ? taxonomy sequence
• There should be a many to many alignment between
  these two sequences

16
Tree Evaluation (Cont.)

• Finding alignment between these sequences for all
  the species
• Using Bayesian Network
• Finding the most probable alignments
• Measuring the Log likelihood of these alignment
• How probable is this tree given this taxonomy

17
Tree Evaluation (Example)

• Phylogenetic tree
• Taxonomy
• T1T2 A
• T1T3 B
• T1T3 C
• T1T3 D

1
2
3
1
ltT1T2,1gt
P1
12
ltT1,1gt ltT3,2gt
P2
123
ltT1,1gtltT3, 23gt
P3
123
ltT1,1gtltT3, 23gt
P4
18
Agenda

• Introduction
• Our method proposals
• Datasets and experiments
• Results
• Discussion
• Future work
• Conclusion

19
Dataset influenza virus

• Influenza virus genomes (flu)
• 44 influenza A genomes (3 for H1-H13, 2 for H16)
• 3 influenza B genomes
• 1 influenza C genome (out group)
• Coding gene sequences only
• Collected and joined from individual gene
  sequences according to the following order HA,
  NA, NP, M, NS, PA, PB1, PB2

20
Dataset Prokaryotes

• Prokaryote genomes (bac)
• 88 bacterial genomes
• 11 archaean genomes
• Uses Nanoarchaeum equitans as the out group.
• Collected from NCBI according to the accession
  number provided in the CCV paper.
• Genomeic DNA sequence including intergenic
  regions.

21
Dataset Mammal mitochondria

• Mammal mitochondria (mito)
• 425 mammal mitochondria
• 1 Arabidopsis mitochondrion (out group)
• Collected from the Organelle Genome
  Megasequencing Program website.
• converted from NCBI format to fasta format.
• Contains many duplicated entries for
• Bos taurus (cattle)
• Sus scrofa (wild Boar)
• Mus musculus (mouse)
• Rattus norvegicus (rat)

22
Experiments

• We built a multiple sequence alignment tree for
  flu
• We ran CCV, TF-IDF and CCV-IDF on all three
  datasets with the following k-string length (we
  fixed K1 1 and only vary K2, L K2 - K1 1
  K2 )
• Flu L 7, L 15
• Bac and mito L 7 and L 9
• Each run generates a pairwise distance matrix.

23
Experiments

• We ran GenCompress and LZ compression programs on
  flu and mito and calculate pairwise distance
• We tried ensembling different measures Reihaneh

24
Experiments

• We converted pairwise distance matrices into
  phylogenetic trees using the Neighbor-Joining
  program in PHYLIP
• We visualized resulting trees using DRAWGRAM and
  TreeView.

25
Agenda

• Introduction
• Our method proposals
• Datasets and experiments
• Results
• Discussion
• Future work
• Conclusion

26
MSA trees versus HA tree
HA tree by Suzuki et.al.
27
MSA versus Compression
GenCompress
1, 2, 3
10, 12, 8
7
13, 16
15, 4, 5, 6, 9
B
28
MSA versus CCV
MSA tree
H1, 2, 3
H5, 6, 9

• H4, 15, 16, 13

7, 8, 10, 12
B
29
MSA vs TF-IDF
MSA tree
H1, 2, 3
H5, 6, 9

• H4, 15, 16, 13

H7, 10, 12, 8
B
30
MSA vs CCV-IDF
CCV-IDF L15 cos
MSA tree
H1, 2, 3
H1, 2, 3
H5, 6, 9
8, 10, 12
7

• H4, 15, 16, 13

7, 8, 10, 12
4, 5, 6, 9
15

• 13, 16

B
B
31
CCV vs TF-IDF
32
CCV vs CCV-IDF
33
Observations

• All methods (MSA, CCV, GenCompress, TF-IDF,
  CCV-IDF) generate similar results.
• Our results are significantly different from
  previous studies.
• Most clades are intact while some are scattered
  around.
• Most clades are pure while some are mixed with
  species from nearby clades.
• CCV and CCV-IDF results are highly similar.

34
AA versus DNA
CCV k13, k27 protein
CCV k11, k27 DNA
35
CCV L7 and L9
CCV k11, k27 DNA
CCV k11, k29 DNA
36
Observations

• Most clades are intact.
• For similar CCV length, the DNA tree is worse
  than the protein tree and unable to recognize
  Archaea as a distinctive clade.
• CCV trees are similar for length 7 and length 9.
• Similarly the L7, L15 and L21 tree for flu are
  almost identical

37
Mito results

• For the mito dataset, we have similar
  observations.
• All methods failed to resolve fine branches of
  the tree by mixing in distant species.

38
Mito primates
TF-IDF L9 cos

• CCV L9 cos

39
Agenda

• Introduction
• Our method proposals
• Datasets and experiments
• Results
• Discussion
• Future work
• Conclusion

40
DNA versus AA Sequence

• There are more k-strings for protein sequence
  than DNA sequence for the same length.
• We need longer k-strings for DNA to achieve the
  same resolution as amino acid (AA) sequence.
• Due to the redundant nature of the genetic code,
  different DNA k-strings may correspond to the
  same AA k-string.
• AA k-strings can share information even though
  their DNA sequence might be different
• DNA sequence may contain intergenic regions which
  do not response to selection pressure
• Intergenic region may not contribute much to the
  resolution of the tree they might even reduce
  such resolution.

41
Thoughts on Document Frequency

• We did not observe significant performance
  difference by adding in document frequency
  information.
• For longer genome (e.g. bac), we need longer
  k-strings to see the effect of DF.
• All bac genomes share 87.9 9-strings and only
  0.8 11-strings

42
Compression programs

• Current compression programs are problematic
• LZ could not handle large datasets
• Kolmogorov is not applicable for large sequences
• These method should be reimplemented

43
Agenda

• Introduction
• Our method proposals
• Datasets and experiments
• Results
• Discussion
• Future work
• Conclusion

44
Future works

• Run the same experiments on protein sequence
• To investigate the effect of using AA versus DNA
  sequences.
• We expect to see better results with protein
  sequences
• New result may reveal subtle difference between
  different methods.

45
Future works

• Speed up the implementation for TF-IDF and run
  them on longer k-strings
• Computational complexity is the bottle neck for
  achieving high resolution in a reasonable amount
  of time.
• Initially the calculations for TF and IDF are
  separated slow
• We achieved significant speedup by integrating
  the calculation of TF and IDF into a two-pass
  algorithm
• We may drop k-string with low TF-IDF values to
  further speed up the program.

46
Future works

• Perform bootstrapping analysis
• We are unable to perform bootstrapping analysis
  due to time and computational resource constraints

47
Future works

• In our proposed evaluation method, we need a Many
  to many alignment which is not a trivial task
• It is well studied in Machine translation and
  Natural Language Processing and those techniques
  could help here
• This measure could also be used as a measure of
  similarity between trees

48
Agenda

• Introduction
• Our method proposals
• Datasets and experiments
• Results
• Discussion
• Future work
• Conclusion

49
Conclusion

• All string composition methods (CCV, TF-IDF,
  CCV-IDF) somewhat group most similar species
  together and produce consistent results.
• However they all failed to resolve big branches
  as well as fine branches.
• We did not observe significant improvement by
  adding document frequency.
• But we will need further experiments (with longer
  k-strings on AA sequences) to fully understand
  the effect.

50
Major Contributions

• We proposed a novel term weighting scheme which
  achieves similar performance as CCV in our
  experiments
• We proposed the notion of adding in global
  information in the form of document frequency
• We discovered that using protein sequence may
  significantly improve performance for all methods
• We proposed a novel evaluation method for
  phylogenetic trees

51
Author Contributions

• Yifeng
• Collected all three data sets
• Performed CCV experiments
• Implemented TF-IDF and CCV-IDF
• Reihaneh
• Built MSA tree for flu
• Performed Compression experiments
• Implemented ensemble and evaluation methods.

52
Special thanks to

• Professor Guohui Lin
• Dr. Zhipeng Cai
• Proteome Analyst Research Group

53
References

• 1 Xin Chen, Sam Kwong, and Ming Li. A
  compression algorithm for dna sequences and its
  applications in genome comparison. In in Genome
  Informatics, pages 52-61, 1999.
• 2 Joseph Felsenstein. Phylip - phylogeny
  inference package (version 3.2). Cladistics,
  5164-166, 1989.
• 3 M. Li, J. H. Badger, X. Chen, S. Kwong, P.
  Kearney, and H. Zhang. An information-based
  sequence distance and its application to whole
  mitochondrial genome phylogeny. Bioinformatics,
  17(2)149-154, 2001.
• 4 Yifeng Liu and Reihaneh Rabbanyk Khorasgani.
  A survey on whole genome phylogenetic analysis.
  CM- PUT 606 course survey, Feburary 2009.
• 5 Christopher D. Manning and Hinrich Schtze.
  Foundations of Statistical Natural Language
  Processing. The MIT Press, June 1999.
• 6 Hasan H. Otu and Khalid Sayood. A new
  sequence distance measure for phylogenetic tree
  construction. Bioinformatics, 19(16)2122-2130,
  2003.
• 7 N. Saitou and M. Nei. The neighbor-joining
  method a new method for reconstructing
  phylogenetic trees. Mol Biol Evol, 4(4)406-425,
  July 1987.
• 8 Yoshiyuki Suzuki and Masatoshi Nei. Origin
  and evolution of in?uenza virus hemagglutinin
  genes. Mol Biol Evol, 19(4)501-509, April 2002.
• 9 Xiaomeng Wu, Xiufeng Wan, Gang Wu, Dong Xu,
  and Guohui Lin. Phylogenetic analysis using
  complete signature information of whole genomes
  and clustered neighbor-joining method.
  International Journal on Bioinformatics Research
  and Application, 2(3)219-248, 2006.