Using BioPAX data to analyze differential measurements (e.g. micro-array)

[Oliver Ruebenacker]

Hello,

Is any one aware of the use of pathway knowledge in BioPAX to
analyze differential measurements, e.g. micro-array data?

The idea seems straight-forward and powerful, yet I am unaware of
any one working on it. Any pointers?

Thanks!

Take care
Oliver

--
Oliver Ruebenacker
Bioinformatics Consultant (http://www.knowomics.com/wiki/Oliver_Ruebenacker)
Knowomics, The Bioinformatics Network (http://www.knowomics.com)
SBPAX: Turning Bio Knowledge into Math Models (http://www.sbpax.org)

[Gary Bader]

Hi Oliver - what kind of analysis are you thinking of?

Some commons ones are:
1. pathway (gene set) enrichment analysis - convert biopax pathways to gene sets and use existing analysis tools
2. convert biopax data to SIF networks and use existing network analysis tools, such as ActiveModules or NetBox
3. work at the level of the BioPAX model itself - this is done less often, presumably because 1 and 2 are simpler e.g. http://www.ncbi.nlm.nih.gov/pubmed/20466809

Gary

http://baderlab.org
The Donnelly Centre - http://www.thedonnellycentre.utoronto.ca/
University of Toronto

[Oliver Ruebenacker]

Hello Gary,

Thanks for the response! Here is an example of what I have in mind:

Suppose a chemical reaction network of four reactions:

X -> A, Y -> B, Z -> A, Z -> B.

Suppose we measure A, B and C in a reference specimen and then see
how they differ in a trial specimen:.

- If A is larger, but B is unchanged, a possible explanation is that
X is larger
- If B is larger, but A is unchanged, a possible explanation is that
Y is larger
- If both A and B are larger, a possible explanation is that X and Y
are larger, but perhaps more plausible is that Z is larger

More typically, we will include catalysts and (negatively)
inhibitors, and reactions upstream of X, Y and Z, and the network is
much larger.

We do not need to know concentrations and kinetic parameters, but if
we have some (e.g. as SBPAX), these could be used to find that some
reactions or pathways contribute more strongly than others.

I don't know much about conversion to gene sets or SIF networks, but
wouldn't that loose much of the information?

Thanks!

Take care
Oliver

[IgorRodchenkov]

(it may be off-topic or even ad, anyway...)

coming soon are some steps in a similar direction:

paxtools-causality module

http://biopax.hg.sourceforge.net/hgweb/biopax/paxtools/file/b005495f3779/paxtools-causality/ 

(see also /paxtools-pattern/), and integrated BioPAX resources and graph queries, like the following, can help find relevant processes (you can import the result network in Cytoscape to see) 
http://awabi.cbio.mskcc.org/cpath2/graph?source=urn:miriam:uniprot:P51587&kind=commonstream&direction=upstream 

(note: it's one of dev/test cpath2 instances, so no warranty...; ask Ozgun Babur for more details; and there will be other nice features in future Pathway Commons Cytoscape 3 app and cBio portal)

IR.

[Oliver Ruebenacker]

Hello Igor,

Thanks! Do you have a documentation for this feature?

Take care
Oliver

[Gary Bader]

Hi Oliver,

On 2012-06-07, at 1:46 PM, Oliver Ruebenacker wrote:

> Hello Gary,
>
> Thanks for the response! Here is an example of what I have in mind:
>
> Suppose a chemical reaction network of four reactions:
>
> X -> A, Y -> B, Z -> A, Z -> B.
>
> Suppose we measure A, B and C in a reference specimen and then see
> how they differ in a trial specimen:.
>
> - If A is larger, but B is unchanged, a possible explanation is that
> X is larger
> - If B is larger, but A is unchanged, a possible explanation is that
> Y is larger
> - If both A and B are larger, a possible explanation is that X and Y
> are larger, but perhaps more plausible is that Z is larger
>
> More typically, we will include catalysts and (negatively)
> inhibitors, and reactions upstream of X, Y and Z, and the network is
> much larger.
>
> We do not need to know concentrations and kinetic parameters, but if
> we have some (e.g. as SBPAX), these could be used to find that some
> reactions or pathways contribute more strongly than others.

Sounds like an interesting research project - the paper I mentioned from Emek tries to evaluate this, but I don't know too many papers that tackle this.

>
> I don't know much about conversion to gene sets or SIF networks, but
> wouldn't that loose much of the information?

You lose a lot of information, but simple gene set representations of pathways likely outnumber other uses 1000 times to 1 in the literature. For instance, the GSEA software is cited over 3300 times - http://scholar.google.ca/scholar?cites=8644502642729854893&as_sdt=2005&sciodt=0,5&hl=en and that is just one tool out of dozens that works with pathway information at the gene set level.

Gary

[IgorRodchenkov]

Hi Oliver,

On Thursday, June 7, 2012 4:51:13 PM UTC-4, Oliver Ruebenacker wrote:

    Hello Igor,

  Thanks! Do you have a documentation for this feature?

Which one, the paxtools module or graph web service? 

(both are under development, but you may ask 'ozgun AT cbio.mskcc.org' regarding to algorithms and papers)

 

     Take care
     Oliver

> ...

[Ozgun Babur]

Yes I think paxtools-causality module is most relevant to Oliver's question. Aim of this module is to be able to traverse the BioPAX model in a signed and directed manner, so that we can evaluate signal flows and their effects on the downstream. This is not a specific analysis, but an infrastructure to build methods that explores cause-effect relations. It is still in development and early for advertising.

The most challenging part of developing such an infrastructure is to mine the upstream-downstream relations in the network at semantic level. Using only graph-theoretic upstream-downstream relations in the BioPAX model do not work well because current data in databases are highly fragmented and links between PhysicalEntity objects of the same EntityReference are largely missing. To complete the missing links, we need to evaluate semantics of the network. For instance we need to infer the nature of reactions, whether they have inhibiting or activating affect on their participants. This not possible most of the times. So the current implementation makes a best guess, evaluating specific patterns in the model (using paxtools-pattern module), and evaluating gained and lost modification features of molecules. The bad thing here is that these rules for mining semantics highly depend on how the data provider creates the data. For instance NCI-PID sometimes says that the inactive form of a molecule activates a reaction. This means that the active form inhibits the reaction, and its being inactivated makes the reaction active. Such relations do not exist in Reactome. If the data providers change how they model data we will need to revisit the mining rules.

When it is fully developed and tested, we plan to work on specific analysis methods for mining and integrating cause-effect relations. The good news is that not just us but anyone will be able to develop such analysis without getting lost in the details of semantics of the BioPAX model.

On Thu, Jun 7, 2012 at 4:49 PM, IgorRodchenkov <rod...@gmail.com> wrote: