1. Comparative Data Analysis Ontology (CDAO)
Arlin Stoltzfus
Center for Advanced Research in Biotechnology
9600 Gudelsky Drive, Rockville, MD
Biochemical Science Division
National Institute for Standards and Technology
University of Maryland Biotechnology Institute

2. Computational genome analysis
New Genome Sequence
Useful information ?
Human genes
Does it vary in humans?
Is it implicated in disease?
Computer-based analysis of genome sequences yields useful knowledge about species. This information provides researchers with a framework for addressing practical questions such as whether an organism is pathogenic (if so, how is it pathogenic? What are the most likely drug targets?), what are its synthetic capabilities, what are its ecological habits, and so on. In the case of humans and their animal relatives, the genome sequence provides a framework for addressing health-related questions such as what genes are implicated in specific diseases, what are the disease risks for a particular individual, what are the possible strategies for treatment.
However, the genome does not speak by itself, or at least does not speak directly to us. Instead it must be “annotated”.
Potential pathogens
Does it make a toxin?
Will UV sterilization work?
Any organism
Does it synthesize ascorbic acid?
Will it grow at high temperatures?

3. Example: SIFT
What do we do precisely, with these comparisons? One example is SIFT. Here the object is to predict whether an observed amino acid variant in the human population is likely to be deleterious. The method is to look for homologs and judge how well the position is “conserved”.
You can read the logic right here in this paragraph.
Ng, P. C., and S. Henikoff SIFT: Predicting amino acid changes that affect protein function. Nucleic Acids Res 31:

4. Genome analysis is comparative analysis
New Genome Sequence
Useful information ?
Database with annotated genomes of other species
Comparative Analysis
. . . and comparative analysis is evolutionary biology
So, the entire process is dependent on comparisons. The input data for the process is not a genome sequence, but a set of genome sequences. The process is not to feed the data into a biophysical model, instead it is to interpret evolved differences.
The power of computational genome analysis comes from comparisons.
Comparison is an evolutionary problem. That is, if we want a theoretical framework for the process of comparison, this framework is evolutionary theory, because evolution is the process that generates similarities and differences.
This is why the editors of Science chose “Evolution” as the “breakthrough of the year” last year; and this why NIH spends such an enormous amount of money sequencing the genomes of organisms that are not humans, organisms such as chimpanzees and dogs. The comparative analysis of these sequences unlocks a treasure trove of useful information.

5. The problem, restated Power comes from comparative analysis
Genome sequences
Useful inferences
Comparative analysis
99.99 % accurate
Far less accurate
Power comes from comparative analysis
Comparative analysis is an evolutionary problem
Depends on a tree describing relationships
Depends on representing dynamics of evolution
Requires attention to uncertainty
Genome analysis can be improved by
Facilitating tree-based analysis with better informatics
Improving models of evolutionary change
Incorporating prior knowledge
Now, we can restate the problem and identify a strategy. The problem of computational genome analysis is to make more accurate, more reliable “functional” inferences.
The prevailing approach to this problem is comparative analysis.
However, this approach, as currently implemented, often does not take advantage of evolutionary theory. Computer scientists in bioinformatics often develop methods of analysis that do not assume or even imply that the things they are comparing have evolved by a branching process.
This leads to a strategy for improving genome analysis. Evolutionary methods are possible and practicable, and will yield better and more accurate inferences. The practical importance of evolutionary comparisons is now widely recognized.
More specifically, in our attempts to improve and facilitate comparative analysis, we have focused on two things: developing software to facilitate evolutionary approaches, and developing better models of the dynamics of evolutionary change. I will give an example of each.

6. A bold generalization What are these principles?
"It matters not at all whether you work with genetic elements, with viruses, bacteria, fungi, animals, or plants. The same principles apply if your subject is molecular evolution, the diversity of genetic systems, comparative morphology, physiology, ecology, or behaviour." (p. 7)
Harvey, P. H., and M. D. Pagel The Comparative Method in Evolutionary Biology. Oxford University Press, Oxford.
What are these principles?

7. Example: Residue “conservation”
Principle 1: hierarchically structured data demand appropriate statistics
The “entropy”
Example: Residue “conservation”
Valdar, W. S Scoring residue conservation. Proteins 48:
Seq_1 D D
Seq_2 D E
Seq_3 D D
Seq_4 D E
Seq_5 E D
Seq_6 E E
Seq_7 E D
Seq_8 E E
S = 1 bit
To see the importance of a tree, lets go back to the issue raised by SIFT, which is about the use of “conservation”.
Valdar considers a variety of schemes. One of them is “entropy”. But look at how entropy makes a mistake by not taking into account the tree-- these two positions have the same numbers of Ds and Es, thus the same entropies, but clearly they do not have the same amount of “conservation”.
Figure 1. Some example columns from different multiple alignments. Each labeled column represents a residue position in a multiple-sequence alignment . . .

8. Statistics for tree-related data
How can one characterize a set of data collected from different biological species, or indeed any set of data related by an evolutionary tree? The structure imposed by the tree implies that the data are not independent, and for most applications this should be taken into account. We describe strategies for weighting the data that circumvent some of the problems of dependency.
Altschul, S. F., R. J. Carroll, and D. J. Lipman Weights for data related by a tree. J Mol Biol 207:
Seq_1 D D
Seq_2 D E
Seq_3 D D
Seq_4 D E
Seq_5 E D
Seq_6 E E
Seq_7 E D
Seq_8 E E
But if we want to use a tree, how do we use the tree? I won’t tell you about a debate on this issue that raged for some years in the systematics community between those who advocated a principle of “parsimony” and those who opted for probabilistic models based on evolutionary mechanisms.
In the end, the modelers one. The answer to how do we use the tree is: we run an evolutionary model along the branches

9. Principle 2: evolution is the generating process
Because the non-independence arises via descent with modification, the proper framework for addressing hierarchy is as to interpret it as an evolved pattern
Seq_1 D D
Seq_2 D E
Seq_3 D D
Seq_4 D E
Seq_5 E D
Seq_6 E E
Seq_7 E D
Seq_8 E E
To:
From: D E -  
How to run the model is actually not that hard to understand. To analyze these models numerically can get very complicated, but the basic idea is very simple. Given some rate matrix, we can treat change as a markov transition process, and this will give us approximate formulas to convert the rates into probabilities over a given time (branch length). This means that we can write down an equation to describe the probability of each scenario of possible evolutionary change, and we can even find parameter values that will maximize the probability of the observed data given the tree-- this is the “maximum likelihood” approach to analysis.
Let r = +, then
P(DE,t)=(/r)(1-e-rt) t

10. Example: intron “loss vs. gain” problem1 -  
Example: intron “loss vs. gain” problem
Possibilities 1
Probability of presence
max
Distance from root
(Prob) B C D E F A
Probabilities
intron
A 1
B 1
C 0
D 0
E 0
F 0
max
Distance from root
gain
intron
A 1
B 1
C 0
D 0
E 0
F 0
max
Distance from root
gain
loss
intron
A 1
B 1
C 0
D 0
E 0
F 0
max
Distance from root
gain
loss
The old problem: “loss vs. gain”. There are different scenarios. We can’t apply “parsimony” because it’s a dubious principle.
How to answer? Use a probabilistic model.
Note, this is really just the same 2-state model we used in the previous slide, I have just re-labeled it with 1 and 0 instead of D and E.
intron
A 1
B 1
C 0
D 0
E 0
F 0
max
Distance from root
present
loss

11. Principle 3: the result is an inference with uncertainty that should be treated explicitly
assign uncertainties to inferences
provide explicit probability distribution
Example from Huelsenbeck 1 -  
And this brings us to our third principle. Notice that once we stop fighting about what is the “best” or “most parsimonious” history, we are left with uncertainty .
This principle is still evolving.
Note that here we have yet another 2-state character, which is the presence of a soldier caste in social ants.
“The phylogeny is usually treated as known without error; this assumption is problematic because inferred phylogenies are subject to both stochastic and systematic errors.”
Huelsenbeck, J. P., B. Rannala, and J. P. Masly Science 288:

12. Principle 3: Explicit treatment of uncertainty
This principle is not followed in non-evolutionary approaches:
“tree-based weighting schemes require more assumptions than those based on only the alignment. After all, many plausible trees can describe a single alignment. Choosing one, even if it is the most probable, introduces additional uncertainty and thus hidden complexity”.
Valdar, W. S Scoring residue conservation. Proteins 48:
Let us return one more time to the treatise on “conservation”by Valdar He suggests there is no “generally accepted standard” for defining or evaluating “conservation”. Fom my perspective, this is odd. Conservation is clearly an evolutionary concept-- it means a lack of change, or a reduced rate of change relative to some background expectation.
At the end of the paper, we see why Valdar does not opt for an evolutionary method. He refers to such methods, but says that “tree-based weighting schemes require more assumptions than those based on only the alignment. After all, many plausible trees can describe a single alignment. Choosing one, even if it is the most probable, introduces additional uncertainty and thus hidden complexity”.
In other words, the true answer involves uncertainty, therefore we will use a measure that, though inaccurate, is reproducible and precise. One of the principles of evolutionary analysis is that we are not going to bury our heads in the sand. We are going to confront uncertainty and deal with it explicitly.

13. Example: functional inference
presence
A 1
B 1
C 0
D 0
E 0
F 0 t
Let r = +, then
P(01,t)=(/r)(1-e-rt)
Note that we can treat “functional assignment” as an inference problem just like the 2-state problem we addressed earlier for introns, the soldier ant caste, and the D/E variability in sequence alignment.
functional
attribute
A 1
B 1
C ?
D ?
E 0
F 0

14. Evolutionary analysis in practice
Homologize characters
Discriminate character states
Assume or infer a phylogeny
Carry out tree-based analysis
Parameter estimation
State reconstruction
Model comparison
Correlation analysis
Now let us put this character-state data model into the context of a flow of operations in evolutionary analysis. Here is an example from Eisen.

15. Character-state data model
OTU: Operational Taxonomic Unit
Character Data
Tree 13 Q E
The “state” is Q (Glutamine)
for “character” 13 (column 13)
of “OTU” H_sapiens_
The start is the “character-state” data model. We take observable features of organisms (or genes, or protein, or whatever are the “operational taxonomic units”) and organize them into “states” of “characters”.

16. NEXUS
#NEXUS
BEGIN TAXA;
DIMENSIONS ntax=26;
TAXLABELS O_volvulus_AAB O_volvulus_AAB C_elegans_AAF C_elegans_AAA
S_cerevisiae_CAA C_albicans_AAC S_pombe_CAB N_crassa_AAA M_musculus_AAA
C_capitata_AAA D_virilis_CAA D_erecta_AAF D_orena_AAF D_teissieri_AAF
D_yakuba_AAF D_melanogaster_AAF D_mauritiana_AAF D_sechellia_AAF
D_simulans_CAA Z_mays_AAB O_sativa_AAC O_sativa_AAC A_thaliana_AAF
P_tremuloides_AAD A_thaliana_BAB A_thaliana_AAD ;
END;
BEGIN CHARACTERS;
DIMENSIONS nchar=30;
FORMAT datatype=protein gap=- missing=?;
CHARLABELS ;
MATRIX
M_musculus_AAA QGTIHFEQKASGE--PVVLSGQITGLTE-G
C_capitata_AAA KGTVHFEQQDAKS--PVLVTGEVNGLAK-G
N_crassa_AAA KGTVIFEQESESA--PTTITYDISGNDPNA
--stuff deleted here--
D_simulans_CAA KGTVFFEQESSGT--PVKVSGEVCGLAK-G
S_cerevisiae_CAA SGVVKFEQASESE--PTTVSYEIAGNSPNA
S_pombe_CAB SGVVTFEQVDQNS--QVSVIVDLVGNDANA;
BEGIN ASSUMPTIONS;
WTSET MySoapWeights (VECTOR) = ;
BEGIN TREES;
TREE "Cu-Zn Superoxide Dismutase" = (((((O_volvulus_AAB : ,O_volvulus_AAB : ):
[1],(C_elegans_AAF : ,C_elegans_AAA : ):0.2533[1]): [0.98],
((S_cerevisiae_CAA : ,C_albicans_AAC : ): [0.91],(S_pombe_CAB :
0.3159,N_crassa_AAA :0.1635): [0.97]): [1]): [0.77],(M_musculus_AAA :
,(C_capitata_AAA : ,(D_virilis_CAA : ,(((D_erecta_AAF : ,
D_orena_AAF : ): [0.92],(D_teissieri_AAF :0.004,D_yakuba_AAF : ):
0.0073[0.87]): [0.88],(((D_melanogaster_AAF : ,D_mauritiana_AAF : ):
[0.28],D_sechellia_AAF : ): [0.7],D_simulans_CAA : ): [0.75]):
[1]): [1]): [1]): [1]):0.0712[0.9],(((((Z_mays_AAB : ,
O_sativa_AAC : ): [0.98],O_sativa_AAC : ): [0.99],
(A_thaliana_AAF : ,P_tremuloides_AAD :0.1075): [1]): [1],
A_thaliana_BAB : ): [0.75],A_thaliana_AAD : ): [0.94]);
NEXUS is a file format designed to store character-state data and trees. We will say more about this format later.

17. Visualization & editing: Nexplorer (www.molevol.org/nexplorer)
Nice tutorial
Thousands of pre-compiled data sets
Standard format allows user-supplied data
Publication-quality graphics
Intuitive user interface
The first example is the Nexplorer server, shown here with a view of the ATPase family that I mentioned previously. In this phylogenetic view, with animals, plants and fungi colored red, green and blue (respectively), it becomes immediately obvious that there are two sub-families of these ATPases, and that the mouse sequence is in one, while the human sequence is in another. We would be wary, therefore, of using this mouse sequence as a model for the human gene.
The Nexplorer server is built on a layer of software that we have been developing for several years, and that we think will be useful in analyzing various kinds of data on genes and proteins. It provides thousands of pre-compiled data sets, it implements a standard data format that allows users to upload, modify, and save their own custom data sets; it provides the capacity to generate publication-quality graphics (I.e., if you generate a nice view of the data that demonstrates an important point, the graphics are so nice that they can be inserted directly into a publication or a grant proposal); and it does all this with a convenient graphical user interface.
As an example of the possible uses of this kind of software, we have been working informally with scientists from GlaxoSmithKline on developing tools to visualize and analyze data on human kinases, which are important drug targets. We hope to develop a CRADA later this spring.

18. Character State Data (example from MacClade documentation)
#NEXUS
[!Data and tree from: Schluter, D Pp in D.B. Wake and G. Roth, eds.,
Complex organismal functions: Integration and evolution in vertebrates. Wiley, N.Y. ]
BEGIN DATA;
DIMENSIONS NTAX=14 NCHAR=5;
FORMAT MISSING=? GAP=- ;
CHARLABELS
[1] Maxillary_tomia
[2] lateral_groove
[3] posterolateral_teeth
[4] intercalary_ridge
[5] maxillary_tomia;
STATELABELS
1 thick thin,
2 deep shallow,
3 sharp reduced,
4 absent present,
5 'round-edged' 'sharp-edged';
MATRIX
presumed_ancestor
Geospiza_difficilis
Geospiza_scandens
Geospiza_conirostris
Geospiza_magnirostris
Geospiza_fortis
Geospiza_fuliginosa
Camarhynchus_pallidus
Camarhynchus_heliobates
Camarhynchus_psittacula
Camarhynchus_pauper
Camarhynchus_parvulus
Platyspiza_crassirostris
Certhidea_olivacea ;
END;
BEGIN ASSUMPTIONS;
OPTIONS DEFTYPE=unord PolyTcount=MINSTEPS ;
BEGIN TREES;
TRANSLATE
1 presumed_ancestor, 2 Geospiza_difficilis, 3 Geospiza_scandens, 4 Geospiza_conirostris,
5 Geospiza_magnirostris, 6 Geospiza_fortis, 7 Geospiza_fuliginosa, 8 Camarhynchus_pallidus,
9 Camarhynchus_heliobates, 10 Camarhynchus_psittacula, 11 Camarhynchus_pauper,
12 Camarhynchus_parvulus, 13 Platyspiza_crassirostris, 14 Certhidea_olivacea;
TREE * UNTITLED = [&R] (1,(((2,(3,4),((5,6),7)),(((8,9),((10,11),12)),13)),14));
Character State Data (example from MacClade documentation)
BEGIN DATA;
DIMENSIONS NTAX=14 NCHAR=5;
FORMAT MISSING=? GAP=- ;
CHARLABELS
[1] Maxillary_tomia
[2] lateral_groove
[3] posterolateral_teeth
[4] intercalary_ridge
[5] maxillary_tomia;
STATELABELS
1 thick thin,
2 deep shallow,
3 sharp reduced,
4 absent present,
5 'round-edged' 'sharp-edged';
MATRIX
presumed_ancestor
. . .
MATRIX
presumed_ancestor
Geospiza_difficilis
Geospiza_scandens
Geospiza_conirostris
Geospiza_magnirostris
Geospiza_fortis
Geospiza_fuliginosa
Camarhynchus_pallidus
Camarhynchus_heliobates
Camarhynchus_psittacula
Camarhynchus_pauper
Camarhynchus_parvulus
Platyspiza_crassirostris
Certhidea_olivacea ;
END;
Another example, this time from classical morphology (Darwin’s finches).

19. CDAO Project Wiki: http://www.evolutionaryontology.org/CDAO
Artifacts:
Development team:
Enrico Pontelli (NMSU)
Brandon Chisham (NMSU)
Julie Thompson (U. Strasbourg, France)
Franciso Prosdocimi (U. Strasbourg, France)
Arlin Stoltzfus (CARB, NIST)

20. CDAO: development strategy
Ontology refinement
Specification:
Study use-cases to clarify scope
Choice of representation:
Choose language and development tools
Conceptualization:
Identify terms from use cases, artefacts
Build concept glossary
Classify key concepts and relations
Implementation:
Formalize the concepts and relations using the chosen language and tools
Evaluation:
Test the ontology for its ability to represent data called for in the use cases, and to support reasoning

21. CDAO: development & evaluation tools
OWL (Ontology Web Language)
Widely supported, emerging as a standard
Includes Description Logics concepts (OWL 1.1)
Has convenient RDF-XML file syntax
Protégé 4 alpha
Supports OWL-DL
Nice graphical interface
Improving rapidly due to active user-developer community
Integrates Reasoners (Pellet, FaCT++)
Racer (external reasoner)
Ad hoc translators from NEXUS or NeXML to CDAO
NCL (C++) or Bio::NEXUS (Perl) libraries

22. CDAO: key concepts & relations
topology
transformation
character state data matrix
tree
rooted
unrooted
edge
directed
node
character state
data matrix TU
character
datum
state
Annotation:
Alignment
procedure…
Tree
Procedure
Model…
taxonomic_link …
Length…
is_a
has
part_of
belongs_to
parent
child
parent_node
left_state
right_state
right_node
child_node
left_node
is_transformation_of
represents_TU
descendant
ancestor
connects_to

23. has_descendant min 2 Nodes
CDAO: tree concepts A B C D E
Node
Ancestral Node
Edge Transformation
has_root
Subtree
Rooted_tree
MRCA_Node
has_descendant min 2 Nodes
Directed_Edge
has_Parent_Node
has_Child_Node
Directed Edge
Lineage

24. CDAO: tree concepts, continuedB C D E F G H J K L M
connects_to
Edge
has
Right_Node
Left_Node
Annotation:
Length…
Edge Transformation
represents_TU TU
taxonomic_link …
Unrooted tree
Node b) I
Rooted Subtree
has_Root

25. CDAO: how to find out more
talk to developers
view with Protégé
browse OWLdocs on term request server

26. CDAO: plans
CDAO is intended to be useful in solving problems (its not intended as an educational tool)
Ontologies are useful for creating semantically rich computable representations, and for semantic transformation (translation) of other representations
Two projects beginning in 2009
Support for MIAPA (Minimal Information for a Phylogenetic Analysis) standard
to cover various types of data (not just sequences)
to include meta-data on sources and methods
workflow description capacity leads on to bigger and better things . . .
Interoperability of Data resources
Note in regard to MIAPA.

27. Acknowledgements Former Stoltzfus group members Lev Yampolsky
Weigang Qiu
Vivek Gopalan
Tom Hladish
Chengzhi Liang
Peter Yang
Support
CARB
NIH
NIST
National Evolutionary Synthesis Center
Collaborators on CDAO project
Enrico Pontelli (NMSU)
Brandon Chisham (NMSU)
Julie Thompson (U. Strasbourg, France)
Franciso Prosdocimi (U. Strasbourg, France)
Some of the ways to advance evolutionary analysis (previous slide) involve informatics

28. Outline Introduction: comparative analysis and evolutionary analysis
Principles underlying evolutionary analysis
Use a tree
Use a model of change
Treat uncertainty explicitly
Generalized aspects of methodology used in evolutionary analysis
The character-state data model
NEXUS files
Examples
Infrastructure needs and the Evolutionary Informatics Working Group
Ongoing projects and plans
Bio::NEXUS, Nexplorer
Comparative Data Analysis Ontology
We just finished #1, now we are moving on in the order shown here.