Filter genemania results based on interaction confidence score

[nivedas...@gmail.com]

I was wondering if the raw weights or the normalized weights can be used as a confidence score to filter the number of interactions retrieved? My goal is to visualize protein-protein interaction network. Since the network is very dense, I would like to limit the number of interactions by showing only those with highest confidence and discard the rest. For example, here is a table I've retrieved from genemania :

SUID	data type	networks	normalized max weight	raw weights
432	Co-expression	Boivin-Vidal-2012	1.25E-04	0.011788654
434	Co-expression	Boivin-Vidal-2012	1.27E-04	0.012002653
436	Co-expression	Hernandez-Novoa-Kovacs-2008|Peng-Stevenson-2013	7.58E-05	0.005101466551423073|0.007797231897711754
438	Co-expression	Anderson-Neville-2007	9.03E-05	0.009969791
439	Co-expression	Lattin-Sweet-2008	9.24E-05	0.00718554
440	Co-localization	Zhang-Hughes-2004	7.25E-05	0.012013779
441	Predicted	Conservation profile-Inparanoid	4.37E-05	0.007320687

 

Should the filtering be done on the basis of raw weights or normalized max weight? How should I interpret the scale of these weights? Is lower weight indication of a more confident interaction or vice versa?

[Scooter Morris]

Overall, I would suggest using the normalized max weight if you are going to compare between networks.  My guess is that larger values are better (it is a weight, after all), but I'll confirm that with the GeneMania folks.

-- scooter

[Gary Bader]

Hello - GeneMANIA is designed to filter by node weight (e.g. give me the top 20 closest genes to my query). While edge weights are used for this computation, they are used only in the context of a composite network that contains edges that are a combination of all the source network edges, linearly weighted by the overall weight of each network.  Edge weights within each network are ranked by what you could generally and informally call confidence or strength (i.e. higher weights are better), but these have different meanings for different network types. For example, in co-expression networks, edge weight is related to Pearson correlation of gene expression profiles, but protein interaction network edges start off as binary, but change to a fraction when normalized. If you are having trouble with too many edges, one thing you could try is to turn off certain networks that contain too many edges and are more redundant and less informative than others.

Hope that helps.

Best,

Gary

[nivedas...@gmail.com]

Do you mean filtering by unchecking the interaction networks option as indicated below?

[Gary Bader]

Hello - filtering by network type works too. I thought the original question was about the weights on the edges.

Best,
Gary

> On Jan 17, 2020, at 8:29 PM, nivedas...@gmail.com wrote:
>
> Do you mean filtering by unchecking the interaction networks option as indicated below?

> --
> You received this message because you are subscribed to the Google Groups "cytoscape-helpdesk" group.
> To unsubscribe from this group and stop receiving emails from it, send an email to cytoscape-helpd...@googlegroups.com.
> To view this discussion on the web visit https://groups.google.com/d/msgid/cytoscape-helpdesk/80903b6a-f012-4d0e-8f92-b52a20eb36c6%40googlegroups.com.
> <genemania-settings.PNG>

[nivedas...@gmail.com]

Yes, it would be ideal to filter using weights. My intention is to use a specific cutoff score to filter interactions. But my understanding was that since weights have different meanings for different network types, a single cutoff score would not work. 

1) Can I, in fact, use a score to filter based on edge weight for all network types?

2) How would I numerically define edge weights of having a) Low score b) Medium score c) High score?

Thanks!

Niveda