Transcript NCBIO-Berkeley

Some thoughts on PATO
Chris Mungall
BBOP
Hinxton
May 2006
Outline
 Motivation revisited
 The Ontology: PATO
 OBD & using PATO for annotation
Who should use PATO?
 Originally:
 model organism mutant phenotypes
 But also:
 ontology-based evolutionary systematics
 neuroscience; BIRN
 clinical uses
 OMIM
 clinical records
 to define terms in other ontologies
 e.g. diploid cell; invasive tumor, engineered gene,
condensed chromosome
Unifying goal: integration
 Integrating data
 within and across these domains
 across levels of granularity
 across different perspectives
 Requires
 Rigorous formal definitions in both
ontologies and annotation schemas
Some thoughts on the
ontology itself
 Outline
 Definitions
 how do we define PATO terms?
 what exactly is it we’re defining?
 is_a hierarchy
 what are the top-level distinctions?
 what are the finer grained distinctions?
 shapes and colors
It’s all about the definitions
 Everything is doomed to failure without
rigorous definitions
 even more so with PATO than other ontologies
 OBO Foundry Principle
 Definitions should describe things in reality, not
how terms are used
 def should not use the word ‘describing’
 Should we come up with a policy for
definitions in PATO
 currently: 19 defs (2.5 are circular)
 proposed breakout session: examine all these
consistency: the property of holding together and retaining shape
amplitude: The size of the maximum displacement from the 'normal' position, when
periodic motion is taking place
placement: The spatial property of the way in which something is placed
pointed value: A sharp or tapered end
epinastic value: A downward bending of leaves or other plantnparts
oblong value: Having a somewhat elongated form withnapproximately parallel sides
elliptic value: Elliptic shapen
hearted value: Heart shaped
fasciated value: Abnormally flattened or coalescedn
opacity: The property of not permitting the passage of electromagnetic radiatio
opaque value: Not clear; not transmitting or reflecting light or radiant energy
undulate value: Having a sinuate margin and rippled surface
permeability: The property of something that can be pervaded by a liquid (as by osmosis
or diffusion)
porosity: The property of being porous; being able to absorb fluids
porous value: able to absorb fluids
viscosity: a property of fluids describing their internal resistance to flow
viscous value: a relatively high resistance to flow.
latency: The time that elapses between a stimulus and the response to it
power: The rate at which work is done
Proposal: genus-differentia
definitions
 An S is a G which D
 Each def should refine the is_a parent
 Single is_a parent
 Example: (non-PATO)
 binucleate cell def= a cell which has two nuclei
 Example (proposed PATO def):
 convex shape def= a shape which has no indentations
 opacity def= an optical quality which exists by virtue
of the bearer’s capacity to block the passage of
electromagnetic radiation
 v similar to existing def
This policy will reap benefits
 Advantages:
 Helps avoid circularity
 Ensures precision
 Consistency in wording user-friendly
 Considerations:
 Sometimes leads to awkward phrasing
 -ity suffix - “an opacity which…”
 Solution:
 allow shortened gerund form
 having…, being…., ….
 most of the existing defs conform already
 implicit prefix “A G which exists by virtue of the bearer…”
From the top down
 First, the fake term ‘pato’ must be removed
 How do we define ‘attribute’?
 Note: I prefer the term ‘quality’ or ‘property’
 attribute implies attribution
 length_in_centimetres is an attribute
 we can of course continue to say ‘attribute’ but I
use ‘quality’ in these slides
 most of new new pato defs are phrased as ‘a
property of…’ which I like, but inconsistent with
calling the root ‘attribute’
 Well then, what is a quality/property?
What a quality is NOT
 Qualities are not measurements
 Instances of qualities exist independently of their
measurements
 Qualities can have zero or more measurements
 These are not the names of qualities:




percentage
process
abnormal
high
Some examples of qualities
 The particular redness of the left eye of a
single individual fly
 An instance of a quality type
 The color ‘red’
 A quality type
 Note: the eye does not instantiate ‘red’
 PATO represents quality types
 PATO definitions can be used to classify quality
instances by the types they instantiate
the type “red”
instantiates
the particular case of
redness (of a particular
fly eye)
the type “eye”
instantiates
an instance of an eye
inheres
(in a particular fly)
in (is a
quality of,
has_bearer)
Qualities are dependent
entities
 Qualities require bearers
 Bearers can be physical objects or processes
 Example:
 A shape requires a physical object to bear it
 If the physical object ceases to exist (e.g. it
decomposes), then the shape ceases to exist
 Some qualities are relational
 they relate a bearer with other entities
 e.g. sensitivity (to)
 Compare with: functions
The PATO hierarchy
 Proposal for a new top level division
 Proposal for granular divisions
Proposal 1: top level division
 Spatial quality
 Definition: A quality which has a physical object as
bearer
 Examples: color, shape, temperature, velocity,
ploidy, furriness, composition, texture
 Spatiotemporal quality
 Definition: A quality which has a process as bearer
 Examples: rate, periodicity, regularity, duration
Proposal 2: subsequent
divisions
 Based on granularity (i.e. size scale)
 a good account of granularity is vital for inferences
from molecular (gene) level to organismal
(disease) level
 How do we partition the levels?
 Some qualities are realised at certain levels
of granularity
 Others can be realised across levels
 shape, porosity
 Sum-of-parts vs emergent
Scale
Bearer
Quality
Definition (proposed)
Physical
Cont.
Mass
Equivalent to the sum of the mass of
the parts of the bearer (mass at the
particle level is primitive/outwith
PATO)
Physical
Cont.
Opacity
An optical quality manifest by the
capacity of the bearer to block light
Phys/Che
m
Liquid
Concentration
A compositional relational quality
manifest by the relative quantity of
some chemical type contained by the
bearer
Molecular
Gene
splicing quality manifest by the splicing processes
undergone by the bearer
Cellular
Cell
ploidy
A cellular quality manifest by the
number of genomes that are part of
the bearer
Cellular
Cell
transformative
potency??
A cellular quality manifest by the
capacity of the bearer cell to
differentiate to different cell types
Scale
Bearer
Quality
Cont.
morphology
_ shape
__ 2D shape
__ 3D shape
Definition (proposed)
A morphological quality which is
manifest
Granular hierarchy
 quality
 spatial quality
 spatial physical and physico-chemical quality
 mass, concentration
 spatial biological quality
 spatial molecular quality
 spatial cellular quality
 spatial organismal quality
 spatial quality, multiple scales
 morphology/form
 optical quality
 color, opacity, fluorescence
Advantages of dividing by
granularity
 Modular
 strategic question
 should we focus on biological qualities and work with
others on morphology, physics-based qualities etc?
 Good for annotation
 easy to constrain at high level
 e.g. organismal qualities cannot be borne by molecules
 Mirrors GO and OBO Foundry divisions
 Easier to find terms
 to be proved, but I believe so
Considerations
 Possible objection:
 The upper level of an ontology is what the
user sees first
 terms such as “cross-granular quality” may
be perceived as undesirable and/or
abstruse by some users
 Counter-argument
 Solvable using ontology views
 aka subsets, slims
Relative and absolute
 Currently PATO terms often come in 3s:
 e.g. mass, relative mass, absolute mass
 Why do we need these?
PATO: One or two
hierarchies?
 Currently two hierarchies
 attribute
 value
 My position:
 there should be one hierarchy of qualities
 My compromise:
 it should be possible to transform PATO
automatically into a single hierarchy
attribute
Current
PATO
value
color
colorV
hue
sat.
var.
hueV
sat.V
var.V
blueV
darkV
paleV
is_a
…
range
blackV
attribute
Proposed
change
attribute
color
color
hue
sat.
var.
hue
sat.
var.
blue
dark
pale
is_a
…
black
Arguments for a single
hierarchy
 Practical
 elimination of redundancy
 no clear line for deciding what should be A
and what should be V
 shape, bumpy vs bumpiness
 Ontological
 what kind of thing is a ‘value’?
Diederich 1997: [quote here]
Arguments against
 Two hierarchies reflect cognitive and linguistic
structures
 e.g. the color of the rose changed from red to
brown
 3 cognitive artifacts
 we want to present data in a way that is natural to
users
 …but this can be solved with a single collapsed hierarchy
 Two are useful for cross-products
 see later - distinguish modifiers from values
 EAV is common database pattern
 so…?
Compromise: transformations
 The Two Hierarchies approach is workable if
they can be automatically collapsed
 Prerequisite: univocity
 Each ‘value’ must be defined to mean exactly one
thing only
 i.e. Each ‘value’ must be the ‘range’ of a single attribute
 Example
 having a value ‘fast’ that could be applied to both the
spatial quality ‘velocity’ and the process quality ‘duration’
would be forbidden
attribute
Collapse on ‘ranges’
value
color
colorV
hue
sat.
var.
hueV
sat.V
var.V
blueV
darkV
paleV
is_a
…
range
blackV
 Shapes and colors
How many types of shape are
there?
 notched, T-shaped, Y-shaped,
branched, unbranched, antrose, retrose,
curled, curved, wiggly, squiggly, round,
flat, square, oblong, elliptical, ovoid,
cuboid, spherical, egg-shaped, rodshaped, heart-shaped, …
 How do we define them?
 How do we compare them?
 Is it worth the effort?
Shape types need precise
definitions to be useful
 Real shapes are not mathematical entities
 but mathematical definitions can help
 Axes of classification:
 Dimensionality
 2-4D (process “shapes”)
 concave vs convex
 angular vs non-angular
 number of
 sides
 corners
 Primitive and composed shapes
 Work with morphometrics community?
Shape likeness
 We can post-coordinate some shape types
 egg-shaped
 head-shaped
 A2-segment-shaped
 Dangers of circularity
 Only for genuine likeness (e.g. homeotic
transformation)
 not “heart-shaped leaf”
 See annotation section of this presentation
Color
 Keep PATO HSV model
 but is black a color hue?
 We should allow overlapping partitions of
color space
 different domains have ‘sub-terminologies’ of color
 Is color relational?
 Humans vs tetrachromatic UV-seeing animals
 Composition
 using has_part
Color hierarchy
 Physical quality
 Optical quality: a physical quality which exists in virtue of the
bearer interacting with visible electromagnetic radiation
 Chromatic quality: an optical quality which exists in virtue of the bearer
emitting, transmitting or reflecting visible electromagnetic radiation




Color hue
Color saturation
Color variation
Color
 Opacity: an optical quality which exists in virtue of the bearer aborbing
visible electromagnetic radiation
 opaque
 translucent
 transparent
Part 2: Annotation using PATO
 Annotation scheme desiderata
 OBD Dataflow
 Proposed annotation scheme
Annotation scheme desiderata
 Rigour
 There is a subset of the scheme which
is simple
 The entire scheme is expressive
It should have an unambiguous
mapping to real world entities
 Even if PATO is completely unambiguous, an illdefined annotation scheme may leave room for
ambiguity
 Example:
 Annotation:
 E=eye, Q=red
 What does this mean?







both eyes are red in this one fly instance
at least one eye is red in this one fly instance
a typical eye is red in this many-eyed spider
both eyes are red in this one fly at some point in time
both eyes are red in this one fly at all times
all eyes are red in all flies in this experiment
some eyes are red in some flies in this experiment
There should be a certain usable
subset that is simple
 Rationale - MODs have limited resources:
 building entry tools for simple subsets is easier
 building databases and query/search engines is easier
 curating with a less expressive formalism is easier, faster
and requires less training
 MODs primary use case is search, for which expressivity is
less useful
 Specifics
 Tools should have an (optional) simple facade
 Simple annotations should be expressible in a simple syntax
that is understood by users with relatively little training
 There should be an exchange format and/or database
schemas that use traditional technology as might be used in
a MOD
 eg XML, relational tables
The scheme must be highly
expressive
 Rationale
 May be required by other NCBCs (BIRN)
 May be required for cbio 200 gene list
 Will be required in future
 Specifics
 Expressive superset will be optional
 MODs can ‘pick and choose’ their subset
 Native exchange and storage format will be logicbased
 Details outwith scope of this presentation
Dataflow
 How will various kinds of phenotypic
data get into OBD?
 what kinds of data suppliers will use
different formalisms?
 3 scenarios… (more possible)
Example dataflow I
 generic MOD curators annotates phenotypes
using Phenote
 Annotations stored directly in MOD’s central
DB
 MOD periodically submits to OBD
 eg using Phenote to create pheno-xml
 OBD converts pheno-xml to native logicbased formalism
 Users can query MOD directly, or OBD
 OBD will allow more expressive queries and have
more data integrated
Example dataflow 2
 Non-MOD generates complex annotations
and stores them locally
 e.g. BIRN group?
 Periodic submissions to OBD
 e.g. as OWL or Obo-format instance data
 OBD converts to native logic-based formalism
 Users can query OBD using more complex
queries
Example dataflow 3
 cBio MOD curates 200 genes using Phenote
 Annotations may be stored outside normal MOD schema
 schema may not be expressive enough for complicated phenotypes
 TBD - up to MOD
 Periodic submissions to OBD
 Phenote can be used to submit pheno-xml, OWL or OBO
 MOD doesn’t have to worry about format
 OBD converts to native formalism
 Users can query OBD using relatively complex queries
 Is this (should it be) different from #1?
MOD A
MOD B
pheno-detailed
XML file
OBD
MOD C
Non-MOD
Proposed annotation schema
 The schema will be described informally
using a simple syntax
 I use ‘E’ for entity and ‘Q’ for quality
 Pretend it is EAV if you like
 with implicit superfluous ‘A’
 The schema has (will have) a formal
interpretation
 aim: database exchange and removal of
ambiguities
 can be expressed using logical language
 OBD will use an internal logic-based
representation
Outline of annotation schema
 ‘EAV’ or ‘EQ’ is not enough
 Fine for (very) simple subset
 Extensions:






time
relational qualities
post-coordination of entity types
count qualities
measurements
…
Standard case: monadic
qualities
 Examples
 E=kidney, Q=hypertrophied
 autodef: a kidney which is hypertrophied
 We assume that there is more contextual
data (not shown)
 e.g. genotype, environment, number of organisms
in study that showed phenotype
 Interpretation (with the rest of the database
record):
 all fish in this experiment with a particular
genotype had a hypertrophied kidney at some
Quantification
 long thick thoracic bristles
 2 statements
 E=thoracic bristle, Q=long
 E=thoracic bristle, Q=thick
 Default interpretation
 A typical thoracic bristle is long and thick
 Optional entity quantifiers
 EQuant={some,all,most,<percentage>,<count>}
 E=thoracic bristle, Q=long, EQuant=80%
 80% of the thoracic bristles in this one individual fly
OBD internal representation
Time
 Example:
 E=brain,Q=small,during=stage
 A E which has quality that instantiates Q
during T
 E has the quality Q for some extent of time,
and that extent overlaps T
 during and other temporal relations will
come from the OBO Relations ontology
Relational qualities
 E.g. sensitivity
 E=eye, Q=sensitive, E2=red light
Post-coordinating entity types
 E=blood in head Q=pooled
 Problem:
 The E may not be pre-defined (pre-coordinated,
pre-composed) in the anatomy ontology
 We can post-compose a type representation
(aka make a cross-product)
 E=(blood  has_location(head))
 The ability to post-coordinate may not be
available in the ‘simple-subset’
 can be expressed easily in pheno-xml, obo, owl,
phenote(soon)
 OBD will handle all required reasoning
Pre-coordinating phenotypes
 Mammalian phenotype ontology has precoordinated phenotype terms
 osteoporosis
 pink fur
 OBD will be able to translate
 post-coordinated queries to annotations on predefined terms
 queries on pre-defined terms to post-coordinated
phenotypes
 Requirement
 computable logical definitions are added to MP
Count qualities
 wingless
 polydactyly
 spermatocytes devoid of asters
Absence can never be
instantiated
 wingless
 E=wing, Q=absent
 autodef “an instance of wing which is
absent”
 Proposal: restate as:
 E=mesothoracic segment, Q=missing part,
E2=wing
 This has other advantages
 works better for “spermatocyte devoid of
asters”
The quality of ‘being many’
does not inhere in a finger
 Polydactyly
 E=finger, Q=supernumerary
 autodef: “a finger which is supernumerary”
 Restate as:
 E=hand, Q=supernumerary parts, E2=finger
 “a hand which has more fingers as parts than is
typical”
 With count extension
 E=hand, Q=supernumerary parts, E2=finger,
Count=6
 could also say +1
 “a hand with 6 fingers, which is more than normal”
Proposed PATO sub-hierarchy
part count quality
lacking
parts
having normal
part count
lacking
all
lacking
some
having extra
parts
Mass count qualities
 furriness
 porosity
 Bearers possess these qualities by
virtue of the number and qualities of
their granular parts
 hairiness by virtue of: number, width,
length, spacing, orientation of hair-parts
What is the essence of hairy?
 Attempt 1:
 E=skin,Q=hairy
 but what if we do not have ‘hairy’ pre-coordinated
in PATO?
 Alternate representation:
 E=skin,Q=excess fine-grained parts,E2=hair
 open Q: is this equivalent to, subsumed by, or
related to representation 1?
 Another representation:
 E=hair, Q=long
 this is something different
increased brown fat cells
 “increased brown fat cells”
 Attempt 1:
 E=brown fat cell, Q=increased
 autodef: a brown fat cell which is increased
 Restate as:
 E=organism, Q=increased (granular) parts,
E2=brown fat cell
 works better for “increased brown fat cells in upper
body”
 OBD handles reasoning
 should annotations to above be returned for
queries of PATO term “fatty”?
Relativity
 PATO has terms like
 large
 increased
 Context is implicit
 strain
 species
 genus/order
 Extension to make explicit
In_comparison_to
 Bigger than average for species/genus/etc
 E=brain,Q=large,In_comparison_to=<taxon-id>
 default is same species as specified by genotype
 Comparative phenotypes
 E=brain,Q=large,In_comparison_to=<phenotypeid>
 requires recording phenotype IDs
 e.g. two experiments, same genotype, different
environment, phenotype stronger in one
Ratio & relative_to
 Use cases:
 Size of brain relative to size of skull
 Size of brain relative to size of skull in an
individual when compared to size brain
relative to size of skull in a typical
individual of that species
 E=brain,Q=large,relative_to=skull,
in_comparison_to=<taxon_id>
 defaults to: whole organism
Modifiers
 E=bone,Q=notched,Mod=mild
 Standardised qualitative modifiers
 Meaning dependent on E and Q
 Can have multiple, cross-cutting scales
 qualitative and numeric/score based
absent mildly realised
normal
strong
extreme
0
1
10
100
0.00
1
0.01 0.1
Modifiers modify meaning of Q
 Influence of Mod on Q is subjective but the direction
is objective
 Example: E=adult_human_body, during=sleep
 Q={low,high} temperature, Mod=mild,normal,moderate,extreme
abn+
abnormal
normal
abnormal abn+
absent mildly realised
normal
strong
extreme
word scale
NOT
0.00
1
1
10
100
score scale
N/A
35
37
39
temperature
37
36.5
36
35
low
temperature
37
37.5
38
39
high
temperature
0.01 0.1
Modifiers and PATO
 Modifiers are not qualities
 Modifiers should not be in a true
ontology
 But we can still give these PATO IDs
 kept separate from core PATO ontology
 Modifiers can be relational
 relatum may be implicit
 e.g. abnormal_with_respct_to
 Modifiers serve similar purposes as
Values in tripartite EAV model
 Difference:
 absent, low, high are not treated in the same
way as genuine quality types like ‘notched’,
‘large’, ‘diploid’, ‘pink’
 they are ingredients in the representation
language, and not types in an ontology
 Heterozygous flies have very short and
highly branched arista laterals.
 E=arista lateral, EQuant=all, Q=short,
Mod=extreme, in_comparison_to=Dmel
 E=arista lateral EQuant=all, Q=branched,
Mod=extreme, in_comparison_to=Dmel
Measurements
 Measurements are not qualities
 In the schema, representations of
measurements are attached to the
representations of qualities
 Separate measurement schema
 don’t need to discuss fine grained details
here
 some data providers will require more
detail than others here
 e.g. averages, error bars, …
 E=tail, Q=length, Measurement=2cm
 E=tail, Q=length, Measurement=+.1cm,
in_comparison_to=<individual-id>
Likeness
 Shape likeness
 Homeotic transformations
 E=A2
segment,Q=morphology,Similar_to=A3
segment
 Interp:
 An A2 segment with the morphological features
of an A3 segment
 but not “heart-shaped leaves”
Conditionals
 Some phenotypes are only realised under
certain conditions
 environment
 including chemical interactions, RNA interference etc
 we should separate conditionals (this phenotype
only seen in this envirotype with this genotype)
from data (on this occasion this phenotype seen in
this envirotype with this genotype)
Schema elements
 Phenotype character:











E
Q
EQuant
E2
Count
Mod
Relative_to
In_comparison_to
Similar_to
Measurment
Temporal
 Most of these elements are optional
 data providers pick and choose their level of
future extensions
 boolean combinations
 conditional statements
 eg environment
modifier
++
+
.
-
--