Miscellaneous notes
Plots of both transitive resolved and unresolved networks
Simulating parameters for structure:
- Fraction of known Sgenes
- Number of Sgenes in model
- Number of links among Sgenes
- Number of Egenes per Sgene
- Number of Sgenes per Egene (new parameter)
- Cycles vs. no cycles.
- Fraction of inhibitory links
Simulating effects from structure:
- Signal transmission strength (example 0.9)
- Signal emission strength
- Additive vs. Multiplicative signal transmission
Reselecting Egenes from Tier 2:
PDF version
Figure Cancer Invasiveness Predictions:
- Expression data from RNAi knockdowns, ordered by predicted effect attachment point.
- Predicted cancer signaling network with top predicted effects.
- Results of matrigel invasiveness assay.
Signed models are more accurate for typical levels of microarray noise
We generated synthetic networks. First, we built a tree of 5-15 network genes. We then added a random number of cyclic links, on average 1/4 of the number of network genes. Next, we make 1/4 of the links inhibitory. Finally, we connect an average of 20 response genes to each network gene, with an average of 1/4 of the network-to-response links as inhibitory.
- In order to generate reasonable expression values from the synthetic network, we found the average expression change between members of the ribosome and proteasome pathways across a compendium of microarrays: PNG PDF
- We generated 500 synthetic networks at various levels of observation. Precision-recall with separations of: 1.75 PDF 3.0 PDF. More available in the directory
- At each level of separation, we have confusion matrices and association plots in the directory
- When using the full model, the "likelihood win," defined as the ratio of the most likely interaction over the second most likely interaction, is indicative of the correct predictions. Plots directory
Observation: Pairwise scoring finds transitive relationships. Compare to models that explicitly require transitivity
- Measure transitivity in generated networks
- Compare fully transitive networks to non-transitive networks
- Old plots: PNG PDF *
Biological Data
Performance on yeast Rosetta compendium of knockouts
Network likelihoods
Likelihood should be reported as the per-observation likelihood to normalize for the quantity of data
Also do runs with the positive only models
Connected vs. non-connected links
- Run on all of Hughes compendium, compare the likelihoods of pairs in the same GO category to those in different categories
- Table: top 10 links in positive part here
- Table: top 10 links in "negative" set
Colon Cancer Invasiveness
Leave one out finds current S-genes
results directory
Predictions
Used differentially expressed genes to build network, then scored the likelihood of the entire genome. Results in
XLS. Here, the "posRank" columns are the rank of the prediction when confined to just genes predicted to be positively attached. Similarly for "negRank" and "unattachedRank."
Heatmaps and other also in
results directory.
Ideas for negative controls of predictions.
- Randomly chosen from list. We believe that there's a small chance that a gene chosen at random will be essential for invasiveness.
- Fit null better than predicted model. We believe that the spot does not fit the data/the spot is mostly noise/the gene is not transcribed.
- High likelihood of not being attached to the model. This is sort of equivalent to the above, but it's the spots that we are most confident are not attached to the model.
- High likelihood of negative attachment to the model. We believe that these genes are involved in the network, but that they must be down regulated in order to work.
Other stuff
KnockoutNetsInvasivenessGenes - notes on which genes and categories may be active in invasive cells
- Yeast Pheromone
- Yeast Hyperosmostic Stress
- Norm's cancer dataset
- TO BE FOUND: cyclic pathway
