Benchmarking%20Orthology%20in%20Eukaryotes

1
Benchmarking Orthology in Eukaryotes

• 12-01-2004 Nijmegen
• Tim Hulsen

2
Summary

• (1) An introduction to orthology
• (2) Orthology determination methods
• (3) Benchmarking
• co-expression
• conservation of co-expression
• SwissProt name
• (4) Conclusions

3
An introduction to orthology
(from http//www.ncbi.nlm.nih.gov/Education/BLASTi
nfo/Orthology.html)
4
Orthology determination methods

• Orthology databases/methods
• COG/KOG
• Inparanoid
• OrthoMCL
• Inclusiveness
• one-to-one/one-to-many/many-to-many
• organisms
• Best bidirectional hit/Phylogenetic trees

5
Benchmarking orthology

• Quality of orthology difficult to test no golden
  standard
• Orthologs should have highly similar functions
• Measuring conservation of function
• functional annotation
• co-expression
• domain structure

6
Benchmarked orthology determination methods

• BBH Best Bidirectional Hit
• KOG euKaryotic Orthologous Groups
• INP INPARANOID
• MCL OrthoMCL
• Z1H All pairs with Z gt 100
• COM Comics Phylogenetic Tree Method
• EQN Equal SwissProt Names

7
Data set used

• Protein World all proteins in all available
  (SPTREMBL) proteomes compared to each other
• Smith-Waterman with Z-value statistics
• 100 randomized shuffles to test significance of
  SW score

rnd ori 5SD ? Z 5
O. MFTGQEYHSV shuffle 1. GQHMSVFTEY 2.
YMSHQFTVGE etc.
seqs
SW score
8
Data set used

• Z-value compensates for
• bias in amino acid composition
• sequence length
• Proteomes used
• Human 28,508 proteins
• Mouse 20,877 proteins
• ? 595,161,516 pairs

9
BBH method

• Easiest method best bidirectional hit
• Human protein (1) ? SW ? best hit in mouse (2)
• Mouse protein (2) ? SW ? best hit in human (3)
• If 3 equals 1, the human and mouse protein are
  considered to be orthologs
• 12,817 human-mouse orthologous pairs (12,817
  human, 12,817 mouse proteins)

10
KOG method

• KOG euKaryotic Orthologous Groups
• Eukaryotic version of COG, Clusters of
  Orthologous Groups
• COG method
• All-vs-all seq. comparison (BLAST)
• Detect and collapse obvious paralogs

Sp1-Sp1 Sp2-Sp2 Sp1-Sp2
etc. for other species ? determine BBHs
EHs-Hs lt EBBH ? paralogs EMm-Mm lt EBBH ? paralogs
11
KOG method

• Detect triangles of best hits
• Merge triangles with a common side to form COGs
• Case-by-case manual analysis, examination of
  large COGs (might be split up)

12
KOG method

• KOG method mainly the same as COG method special
  attention for eukaryotic multidomain structure
• Group orthologies many-to-many
• Cognitor assign a KOG to each protein
• (mouse not yet in KOG)
• 810,697 human-mouse orthologous pairs (20,478
  human, 15,640 mouse proteins)

Tatusov et al., The COG database an updated
version includes eukaryotes, BMC Bioinformatics.
2003 Sep 114(1)41
13
INP method

• All-vs-all followed by a number of extra steps to
  add in-paralogs ? many-to-many relations
  possible
• 54,553 human-mouse orthologous pairs (19,504
  human, 17,030 mouse proteins)

Remm et al., Automatic clustering of orthologs
and in-paralogs from pairwise species
comparisons, J Mol Biol. 2001 Dec
14 314(5)1041-52
14
MCL method

• All-vs-all BLASTP ? determine orthologs
  recent paralogs ? use Markov clustering to
  determine ortholog groups
• 7,322 human-mouse orthologous pairs (human 6,332,
  mouse 6,115 proteins)

Li et al., OrthoMCL identification of ortholog
groups for eukaryotic genomes, Genome Res. 2003
Sep13(9)2178-89
15
Z1H method

• All human-mouse pairs with Z gt 100 in Protein
  World set are considered to be orthologs
• 290,176 human-mouse orthologous pairs (19,055
  human, 16,149 mouse proteins)

16
COM method

• Human
• 
  All 9 eukaryotic proteomes in
  Protein World
• Zgt20, RHgt0.5QL
• 24,263
  groups

PROTEOME
Hs-Mm 85,848 pairs Hs-Dm 55,934 pairs etc.
TREE SCANNING
17
COM method

• Example BMP6 (Bone Morphogenetic Protein 6)
• ? 5 Hs-Mm orthologous relations defined

18
EQN method

• Consider all Hs-Mm pairs with equal SwissProt
  names to be orthologous
• e.g. ANDR_HUMAN??ANDR_MOUSE
• Used as benchmark later on
• 5,214 Hs-Mm orthologous pairs (5,214 human, 5,214
  mouse proteins)

19
Benchmarkingthrough co-expression

• Comparison of expression profiles of each
  orthologous gene pair
• Using GeneLogic Expressor data set

organism samples fragments tissue categories SNOMED tissue categories
human 3269 44792 115 15
mouse 859 36701 25 12
20
Expression tissue categories
HUMAN MOUSE
1 Blood vessel 1 Blood vessel
2 Cardiovascular system 2 Cardiovascular system
3 Digestive organs 3 Digestive organs
4 Digestive system 4 Digestive system
5 Endocrine gland -
6 Female genital system 5 Female genital system
7 Hematopoietic system 6 Hematopoietic system
8 Integumentary system 7 Integumentary system
HUMAN MOUSE
9 Male genital system 8 Male genital system
10 Musculoskeletal system 9 Musculoskeletal system
11 Nervous system 10 Nervous system
12 Product of conception -
13 Respiratory system 11 Respiratory system
14 Topographic region -
15 Urinary tract 12 Urinary tract
21
Co-expression calculation

• Calculation of the correlation coefficient
• N?xy (?x)(?y)
• r ----------------------------
• sqrt( (N?x2 - (?x)2)(N?y2 (?y)2))
• Measured over the 12 corresponding SNOMED tissue
  categories

22
Co-expression example 1
High correlation 0.914167
23
Co-expression example 2
Low correlation -0.935731
24
Benchmarking throughco-expression

-
25
Benchmarking through conservation of co-expression
Human
Gene A Gene B
Co-expression Cab (-1ltcorr.lt1)
(Co-expression calculated over 115 tissues in
human, 25 in mouse)
All-vs-all Human 40,678 chip fragments Mouse
29,910 chip fragments
Mouse
Gene A Gene B
Cab gt Cab
? Increases probability that A and B are involved
in the same process
26
Benchmarking through conservation of co-expression

• Gene Ontology (GO) database hierarchical system
  of function and location descriptions
• Orthologs are in same functional category when
  they are in the same 4th level GO
• Biological Process class

27
Benchmarking through conservation of co-expression
28
Benchmarking through SwissProt name

• How many of the predicted orthologous relations
  have equal SwissProt names (EQN set in other
  benchmarks)
• reliable because checked by hand
• - assumes only one-to-one relationships are
  possible

29
Benchmarking through SwissProt name
(ALL if all possible human-mouse pairs (or
random fraction) would be orthologs)
30
Conclusions

• Hard to point out the best orthology
  determination method
• In most cases lessbetter, moreworse
• Method that should be used depends on research
  question do you need few reliable orthologies or
  many less reliable orthologies?
• Future directions look at conservation of domain
  structure as a benchmark

31
Credits

• Martijn Huynen
• Peter Groenen
• Comics Group
• Gert Vriend
• Rest of CMBI
• Organon Bioinf. Group