Re: Advancing translational research with the Semantic Web

On May 18, 2007, at 3:06 AM, Eric Jain wrote:

> Alan Ruttenberg wrote:
>> If you want to say that the protein is found in some tissue,  
>> that's what should be said. However, in your email you wrote that  
>> the protein is expressed in the tissue.
>
> Sorry about that, should run a consistency checker on my outgoing  
> mail :-)

This is not a matter of consistency, it is a matter of saying what is  
meant :)

>> If it is know to be found in the tissue I would make the subclass  
>> be the subclass of the protein each instance of which  is located  
>> in some instance of the tissue. No processes involved at all.
>
> You would use different representations depending on how well it is  
> known?

All entities are involved in processes all the time. That we don't  
know the specifics doesn't mean they are not there, nor does it mean  
that the representation is different when we state the specifics.

By a reasonable  definition of process  (following, e.g.,  the BFO  
papers), if a process happens in a location, then each participant is  
located in some part of that location. So if it turns out that the  
truth is that the protein expression process happened in the tissue,  
and we had the relations appropriately encoded in our computational  
system, then the location of the protein - the fact that was stated,  
would be able to be inferred. So we would have extra information, but  
the information we have would stay true.

In fact, very few such axioms are currently encoded in the BFO and  
OBO ontologies, a problem which many people want to and will address  
and which some, including myself, are working on. For example, I  
recently encoded a bunch of axioms representing constraints on  
part_of (e.g. a 3-d spatial region can't be part of a 2-d  spatial  
region) in OWL and expect them to be added to a version of the  
relation ontology some time in the near future. Thomas Bittner is  
working on a FOL encoding of the BFO at http://www.ifomis.uni- 
saarland.de/bfo/fol which is substantially more detailed than any of  
the current OWL representations.

There are other computational systems that are candidates for doing  
such inferences. I'm particularly interested in OWL because it has  
the widest adoption and hence work in it has a higher chance, IMO, of  
being used by people.

>> I don't think we can make due with core RDF features
>
> Neither do I; just not enthusiastic about reimplementing core  
> features...

I think there is a lot of mileage we can get out of OWL, which  
extends RDF. Use of OWL has the dual benefit of saving us work, and  
helping the OWL people advance the state of their tools because they  
have realistic use cases. Use of OWL is not without problems - the  
reasoning techniques don't scale to anything near the size of the  
database we've created. OTOH there is ongoing research to address  
this and now they have a target (and they are very interested in  
tackling it). One area I am watching is the DL-Lite work, which  
offers some level of reasoning in a way that can be implemented in  
relational databases. I'm also aware (but haven't yet tried) the  
upcoming Oracle RDF store that implements a subset of OWL, the OWLIM  
system, and interest at Openlink in adding further inference  
techniques. I'm sure there are others that I'm not aware of.

In the mean time, the approach we tooks for the demo was to add some  
ability to query over inferred knowledge by precomputing specific  
pieces of information which we knew would be useful for some of the  
queries we wanted to do, for example the part_of relations in the GO.

-Alan

Received on Friday, 18 May 2007 16:02:22 UTC