evaluating hypotheses using sparql-dl as an abstract workflow language to choreograph sadi semantic...

“Shopping for data should be as easy

as shopping for shoes!!”

Carole Goble

Evaluating Hypotheses using SPARQL-DL as an

abstract workflow language to choreograph

SADI Semantic Web Services

Mark Wilkinson, Isaac Peral Senior Researcher in Biological InformaticsCentro de Biotecnología y Genómica de Plantas, UPM.

Start at the end...

We wanted to duplicatea real, peer-reviewed, bioinformatics analysis

simply by providing a model describingwhat the answer

(if one exists)would look like...

...the machine had to make every other decision

on it’s own

Gordon, P.M.K., Soliman, M.A., Bose, P., Trinh, Q., Sensen, C.W., Riabowol, K.

Interspecies data mining to predict novel ING-protein interactions in human.

BMC Genomics 9, 426 (2008).

This is the study we chose:

Gordon, P.M.K., Soliman, M.A., Bose, P., Trinh, Q., Sensen, C.W., Riabowol, K.: Interspecies data mining to predict novel ING-protein interactions in human. BMC genomics. 9, 426 (2008).

Original Study - simplified and abstracted:

Using what is known about interactions in other species, predict new interactions in your species of interest

Given a protein P in Species X

Find proteins similar to P in Species Y Retrieve interactors in Species Y Sequence-compare Y-interactors with Species X genome (1) Keep only those with homologue in X

Find proteins similar to P in Species Z Retrieve interactors in Species Z Sequence-compare Z-interactors with (1)

Putative interactors in Species X

For our prototype study, we simplified this further to:

X 2

Then intersect

The tricky part is...

In the abstract, the two workflowsare identical

but in reality they will be different because they call for information

from different species

Modeling the result – Step 1

OWL

Web Ontology Language (OWL) is the language approved by the W3C

for representing knowledge on the Web

ProbableInteractor:

is homologous to (

protein from ModelOrganism1…) # Potential Interactor in previous slide

and

protein from ModelOrganism2…) # Potential Interactor in previous slide

Modeling the result – Step 2

Probable Interactor is defined in OWL as a subclass of Potential Interactor that requires homologous pairs of interacting proteins to exist in both

comparator model organisms.

(Effectively, an intersection)

We then publish our OWL model of a Probable Interactor on the Web

In a local data-file we provide the protein we are interested in, and the two species we wish to use in our comparison

taxon:9606 a i:OrganismOfInterest . # humanuniprot:Q9UK53 a i:ProteinOfInterest . # ING1taxon:4932 a i:ModelOrganism1 . # yeasttaxon:7227 a i:ModelOrganism2 . # fly

These four lines represent all of the data provided to the query I am about to show you...

PREFIX i: <http://sadiframework.org/ontologies/InteractingProteins.owl#>

SELECT ?proteinFROM <file:/local/workflow.input.n3>WHERE {

?protein a i:ProbableInteractor .}

This is the question we ask:

The reference to our OWL model of the answer

Our system then derives (and executes) the following workflow automatically

These are differentWeb services!

...selected at run-time based on the same model

There are two very cool things about what you just saw...

There are two very cool things about what you just saw...

The system was able to create a workflow based on

an OWL model

There are two very cool things about what you just saw...

The workflow it created (i.e. the services chosen)

differed depending on context

We got the answer

“simply” by designing a model of the answer!

Start at the beginning...

Semantic Automated Discovery and Integration

http://sadiframework.org MicrosoftResearch

A Semantic Web-focused Web Services specification

Web Services

vs.

Semantic Web

Web ServicesXML + XML Schema

Semantic WebRDF + OWL

Web ServicesPOST of SOAP

Semantic WebGET of RDF

Web ServicesNo (rigorous) semantics

Semantic WebRich, flexible semantics

Web Services&

Semantic Web

Fundamentally different Web technologies

A design-practice for Web Service provision on the

Semantic Web

100% standards-compliantwith no “invented” standards

Lightweight(only 2 inter-related “rules”)

Rules come from observations:

Web Services in Bioinformatics create implicit biological relationships

between their input and output

SADI Observation #1:

SADI Observation #1:

SeqRet

SADI Design Practice #1

Make the implicit explicit…

A Web Service should create “triples” linking the input data to the output data thus

explicitly describing the semantic relationship between them

SADI Best Practice #1

This is what bioinformatics Web Services implicitly do anyway!

Easy to implement this as a best-practice

...and makes SADI Services a source of Linked Data

SADI Observation #2:HTTP GET and POST

GET guarantees the response relates to the request URI

in a very precise and predictable way

POST does not…

SADI Observation #2: HTTP GET and POST

That’s why Web Services have a fundamentally different behaviour than the Semantic Web

We can fix that!

(without breaking any existing rules or standards!)

SADI Observation #2:GET and POST

SADI Best Practice #2

SUBJECT URI of the output graph (triples)

is the same as

SUBJECT URI of the input graph (triples)

(the output is “about” the input... Now explicitly!)

SADI Best Practice #2

SeqRet

BRCA1

GeneID

rdf:type

BRCA1

GeneID

rdf:type

AGCTTA...

hasDNASequence

Consequence

Web Services now exhibit a very similar behavior to the Web itself

POST “behaves like” GET

SADI Interfaces

Service Interfaces can be described by two OWL classes

(this is 100% compatible with the SAWSDL standard)

SADI Interfaces

OWL Class #1: My Input Class

SADI Interfaces

OWL Class #2: My Output Class

SADI Service Functionality

Consumes OWL Individuals of Class #1

Returns OWL Individuals of Class #2

SeqRet

BRCA1

GeneID

rdf:type

BRCA1

GeneID

rdf:type

AGCTTA...

hasDNASequence

but the URI of those two individuals is the same; they are the same individual, just now a member of a new class.

In practice, of course, you don’t return the input data

Strip it and add the new data provided by the Service

But since the output is still “rooted” in the input node,

Input and output are easily merged client-side

(just concatenate the output with the input)

Service Description

INPUT OWL ClassNamedIndividual: things with a “name” property from “foaf” ontology

OUTPUT OWL ClassGreetedIndividual: things with a “greeting” property from “hello” ontology

person:1

hello:Greeted Individual

rdf:type

Hello, Guy Incognito!

hello:greeting

POST http://example.org/myservice

person:1

hello:NamedIndividual

rdf:type

Guy Incognito

foaf:name

How do we discover services?

Input and output are about the same “thing”

Therefore, to describe what a service does simply compare (“diff”) the

Input and Output OWL classes

This is not prescriptive! Just how we use it

Service Description

INPUT OWL ClassNamedIndividual: things with a “name” property from “foaf” ontology

OUTPUT OWL ClassGreetedIndividual: things with a “greeting” property from “hello” ontology

person:1

hello:Greeted Individual

rdf:type

Hello, Guy Incognito!

hello:greeting

POST http://example.org/myservice

person:1

hello:NamedIndividual

rdf:type

Guy Incognito

foaf:name

The service providesa “greeting” to a Named Individual based on its “name”

Index all of the propertiesadded by all of the services

under all circumstances

Service Discovery

Real-world Example

Input Data: BRCA1 rdf:type Gene ID

Output Data: BRCA1 hasDNASequence AGCTTAGCCA…

Registry Index: Service provides “hasDNASequence” property to Gene IDs

Simply search for the property of interestbased on the data in-hand

Service Discovery

e.g. The question:

“what is the DNA sequence of BRCA1?”

Discover a SADI Web Service that generates the DNA Sequence property for gene identifiers

DEMOKnowledge Explorer

Plug-in

For more information about the Knowledge Explorer surf to:http://io-informatics.com

SADI has just filled-in “Encodes” property for the three genes from the output of discovered Web Service(s)

Discover services that provide the hasGOTerm property for Protein Sequence datatype

This kind of “Web Service Surfing” is veryintuitive for the Biologist!

No need to describe the algorithm or the databasejust describe the properties that will be added

Semantic Health And Research Environment

SHARE answers arbitrary SPARQL queries by finding and executing SADI Services

Example #1

What is the phenotype of every allele of the Antirrhinum majus DEFICIENS gene

SELECT ?allele ?image ?desc

WHERE { locus:DEF genetics:hasVariant ?allele . ?allele info:visualizedByImage ?image .

?image info:hasDescription ?desc }

Example #1

What is the phenotype of every allele of the Antirrhinum majus DEFICIENS gene

SELECT ?allele ?image ?desc

WHERE { locus:DEF genetics:hasVariant ?allele . ?allele info:visualizedByImage ?image .

?image info:hasDescription ?desc }

Note that there is no “FROM” clause!We don’t tell it where it should get the information, The machine has to figure that out by itself...

Enter that query into SHARE

Click “Submit”...

...and in a few seconds you get your answer.

Based on predicates in your query, SHARE utilized SADIto automatically discover the resources required to answer your question.

Because it is the Semantic WebThe query results are live hyperlinksto the respective Database or images

Importantly

We posed, and answered a complex database query

WITHOUT A DATABASE

(in fact, the data didn’t even have to exist... as I’ll now show you)

Example #2

Show me the latest Blood Urea Nitrogen and Creatinine levelsof patients who appear to be rejecting their transplants

SELECT ?patient ?bun ?creatFROM <http://sadiframework.org/ontologies/patients.rdf>WHERE {

?patient rdf:type patient:LikelyRejecter .?patient l:latestBUN ?bun . ?patient l:latestCreatinine ?creat .

}

Likely Rejecter:

A patient who has creatinine levelsthat are increasing over time

- - Wilkinson “MD”

Likely Rejecter:

Our database contains various blood chemistry measurements

at various time-points

Likely Rejecter:

…but there is no “likely rejecter” column or table in our database…

SemanticsSHARE determines

by itself

the need to do a Linear Regression analysis over

Creatinine blood chemistry measurements

SemanticsSHARE determines

by itself

how and where that analysiscan be done

and does it

The SHARE system utilizes Semantics (via SADI) to discover and access analytical services on the Web that do linear regression analysis

VOILA!

How does SHARE work?

Ontologies

Ontology Spectrum

Catalog/ID

SelectedLogical

Constraints(disjointness,

inverse, …)

Terms/glossary

Thesauri“narrowerterm”relation

Formalis-a

Frames(Properties)

Informalis-a

Formalinstance

Value Restrs.

GeneralLogical

constraints

Originally from AAAI 1999- Ontologies Panel by Gruninger, Lehmann, McGuinness, Uschold, Welty; – updated by McGuinness.Description in: www.ksl.stanford.edu/people/dlm/papers/ontologies-come-of-age-abstract.html

Catalog/ID

SelectedLogical

Constraints(disjointness,

inverse, …)

Terms/glossary

Thesauri“narrowerterm”relation

Formalis-a

Frames(Properties)

Informalis-a

Formalinstance

Value Restrs.

GeneralLogical

constraints

Basic SADI/SHARE functionality

Ontology Spectrum

BETTER!!

Catalog/ID

SelectedLogical

Constraints(disjointness,

inverse, …)

Terms/glossary

Thesauri“narrowerterm”relation

Formalis-a

Frames(Properties)

Informalis-a

Formalinstance

Value Restrs.

GeneralLogical

constraints

WHY?

Because I say so!

Because it fulfils XYZ

Ontology Spectrum

Why is this data a member of this OWL Class? (and therefore valid input to the service)

Catalog/ID

SelectedLogical

Constraints(disjointness,

inverse, …)

Terms/glossary

Thesauri“narrower term”relation

Formalis-a

Frames(Properties)

Informalis-a

Formalinstance

Value

Restrs.GeneralLogical

constraints

Categorization Systems – like library shelves, inflexible

Discovery systems - flexible

In the upper end of the Ontology Spectrum,

if the data has the right properties

It can be discovered to be a valid input to a Service

regardless of how it was originally classified or which

ontology was used for that classification

...and in the context of a SHARE query

those individual properties may have been aggregated

from many different places;

The data becomes a valid input as properties aggregate

In exactly the same way that the OWL property restrictions of a SADI Input Class tell SHARE what

properties a service requires as input

The property restrictions of an OWL Class in the SPARQL query tell SHARE what properties it

needs to retrieve to create members of that class

Show me the latest Blood Urea Nitrogen and Creatinine levelsof patients who appear to be rejecting their transplants

PREFIX rdf: <http://www.w3.org/1999/02/22-rdf-syntax-ns#> PREFIX patient: <http://sadiframework.org/ontologies/patients.owl#> PREFIX l: <http://sadiframework.org/ontologies/predicates.owl#> SELECT ?patient ?bun ?creatFROM <http://sadiframework.org/ontologies/patients.rdf>WHERE {

?patient rdf:type patient:LikelyRejecter .?patient l:latestBUN ?bun . ?patient l:latestCreatinine ?creat .

}

The definition of a Likely Rejecter category is encoded in a machine-readable document written in the OWL Ontology language

Basically:

“the regression line over creatinine measurements should have an increasing slope”

SHARE burrows down through the ontological definition to learn about what the properties of “regression models” are

SHARE utilizes SADI to discover analytical services on the Web that do linear regression analysis then other services that can

determine the “latest” measurements from a time-series

Let’s go through that again from a different perspective

OWL Classes that include property restrictions can be “executed” as if they were workflows

Ontology = Query = Workflow

WORKFLOW

QUERY:

SELECT images of mutations from genes in organism XXX that share homology to this gene in

organism YYY

Concept:

“Homologous Mutant Image”

As OWL Axioms

HomologousMutantImage is owl:equivalentTo {

Gene Q hasImage image P

Gene Q hasSequence Sequence Q

Gene R hasSequence Sequence R

Sequence Q similarTo Sequence R

Gene R = “my gene of interest” }

Those axioms combine to create an OWL Class:

Homologous Mutant Image

QUERY:

Retrieve owl:Homologous Mutant Image for

gene XXX

SHARE:

Decomposes the owl:Homologous Mutant Image Class, discovers SADI services

relevant to that class, and pipelines them together into a

workflow

“The user experiments showed that workflow re-use… is difficult for bioinformaticians.”

- Gooderis, A. (2008) Ph.D. Thesis

Under slightly different circumstances(e.g. studying the same phenomenon in a

different organism)

this workflow

WILL NOT SOLVE THE PROBLEM

It must be edited on a case-by-case basis

This editing turns out to be extremely difficult, and can ~only be done manually

This great idea would be even better if workflows were a little bit easier to re-purpose!

With SHARE, a workflow is generated dynamically, based on all of the information

presented in the query

e.g. to create a new workflow simply specify a different organism ID in the query!

With SHARE, a workflow is generated dynamically, based on all of the information

presented in the query

Moreover, the workflow plan CHANGES dynamically based on service outputs!

With SHARE, the ontology is the workflow(not the same as “an ontology that describes workflows”)

The ontology acts as an abstraction of a workflow, which is concretized at run-

time based on circumstances of the query

Works (best) with ontologies in the “Frames+” part of the spectrum, if there are SADI services available

Catalog/ID

SelectedLogical

Constraints(disjointness,

inverse, …)

Terms/glossary

Thesauri“narrowerterm”relation

Formalis-a

Frames(Properties)

Informalis-a

Formalinstance

Value Restrs.

GeneralLogical

constraints

As far as we are aware, SHARE is the only system that exhibits this particular behaviour

...and IMO this is a pretty big deal...

Gordon, P.M.K., Soliman, M.A., Bose, P., Trinh, Q., Sensen, C.W., Riabowol, K.: Interspecies data mining to predict novel ING-protein interactions in human. BMC genomics. 9, 426 (2008).

That experiment can now be represented ONCE as an OWL Class

It becomes concretized automatically

for each individual researcher

as a distinct workflow given any starting protein and any combination of comparator species

This is far beyond simply changing the parameters entered into a workflow...

Moreover, the idea of automated workflow “individuality” is quite interesting to us...

This is an early prototype of a

Patient-driven Personalized Medicine

Web interface

Matching based on official name, compound name,

brand name, trade name, or “common name”

Still needs some work...

??!?!?

Link out to PubMed

Why the alert?

The SADI+SHARE workflow and reasoning waspersonalized to YOUR medical data

In future iterations, we will enable the workflowto be further customized through “personalized”OWL Classes (e.g. Provided by your Clinician!!)

These OWL Classes might include information about the current trajectory of your treatment for a chronic disease, for example, such that what you read on the Web is placed in the context of your expert Clinical care...

Frankly, I think it’s quite cool that “people” are creating and running personalized workflows at the touch of a button...

...as many of you know, that has been my dream since I started studying this

problem a decade ago!

Final thoughts

An experiment... based on a hypothesis

An experiment... based on a hypothesis

now modeled in OWL

. . .

. . .

Does this OWL Class represent a Hypothesis?

I think it does!

. . .

. . .

I believe that we will soon show

using SADI + SHARE

that we can model a non-trivial hypothetical biological scenario

then evaluate if that hypothesis is supported or notbased on whether the automatically-synthesized workflow

returns any individuals that conform to the model.

Ontology = Hypothesis = Query = Workflow [= Materials and Methods ]

These can be automatically derived through provenance information during workflow execution

Most of your publication is done!

All you need to do now is interpret the results!

Please join us!

SADI and SHARE are Open-Source projects

http://sadiframework.org

My New Home!

Luke McCarthy – Lead Dev.Everything...

Benjamin VanderValk SHARE & SADI & Experimental modeling & myHeath Button

Soroush Samadian Cardiovascular data modeling and queries

University of British Columbia

Edward Kawas SADI Service auto-generator

Ian WoodExperimental modeling project

U of New Brunswick

Dr. Chris BakerAlexandre Riazanov

Carleton University

Dr. Michel DumontierMarc-Alexandre NolinLeonid ChepelevSteve EtlingerNichaella KiethJose Cruz

C-BRASS Collaborators at other sites

Microsoft Research