Using GEO data and network inference methods (GRN), I have got relationships between genes from each GRN method (example, GENIE3, CLR, LASSO). For specific genes which I am interested, I have got also the network from literature (text mining based.....Ariadne, Ingenuity, STRING). As all these relationships are purely predictions, how to effectively combine/overlap these predictions (to make an ensemble network) so that I don't loose novel relationship?
GEOdata: Mouse macrophage cell line RAW264.7 with LPS stimulation
I think I would start something like this (just some ideas for discussion):
Each of your methods probably gives a score, or even better, p-value for each predicted edge. If not, then the score would be 1 or 0 or I guess for edge or not between 2 nodes.
With all these scores, you have a data set that you can try to use to create a consensus - for each edge, add up the scores from each method (you would probably scale all scores to between 0 and 1, or use the edge ranks or something). This gives a consensus score to each interaction based on how many methods rate it, and you can rank edges based on how many methods predict them.
You can go one better than this maybe, and give each method a weight. Then, you can optimise these so that the weight of each method best matches the probability that that method will agree with the weighted consensus of all methods.
This might be too simple an approach, and might not work so well for sparse data in a small network neighbourhood, maybe it needs a genome scale network to start to make sense of it.
Many thanks Jonathan for the reply. This sounds as a good start to get some clues what is happening in the system. I will try on my predictions.
cheers,
Aravind
If you are not concerned with edge weighting, as Jonathan is describing, which would provide relative levels of confidence in each edge, but are simply interested in compiling all possible edges as a predictive network, then you could use Cytoscape to assemble them. Other than the fact that you need to have your data formatted correctly, it should be pretty straightforward to compile. You'd need to carry out some filtering of duplicate interactions in a spreadsheet prior to Cytoscape upload, unless you're not bothered by multiple edges between node pairs, but it should be fairly easy to do. Once it's in Cytoscape, there are many apps/tools that can be helpful to interrogate the final network. The one I'm most familiar with is BiNGO, which allows ontological analysis of your network. Of course, in the end, these are all simply predictions, so it is best to consider the data as a predictive tool, which in turn allows you to generate testable hypotheses that are best assessed experimentally.
Aravind, if you give Cytoscape a try and run into any problems formatting your data properly, let me know.
- Badges
- Science topic
- Similar topics
- Computer Science and Engineering
- Bioinformatics
- Bioinformatic Tools