Greetings, I am currently going through the initial set of test networks, and noticed some strange behavior in the PPI networks (1_ppi and 2_ppi). When visualizing the distribution of edge weights, both of the networks appear to have commonly concurring values which appears as peaks in the density plots, or bands in a scatter plot. For example, when the edge weights for 1_ppi and 2_ppi are plotted against one another, you get: ![scatterplot of 1_ppi vs. 2_ppi](https://s32.postimg.org/z2i6v4ddx/1_ppi_vs_2_ppi.png) - For 1_ppi (x-axis) the values are spread across the range [0,1], but there are several prominent bands present. - For 2_ppi (y-axis) the values also go between 0 and 1, however, there is a large gap between ~0.63 and 1. For the second PPI network, it seems like the data may have been thresholded such that also values above ~0.63 were set to 1 and everything else was left as-is. I'm not sure where the bands are coming from in the first PPI network. I realize that the details of the network preparation have been intentionally hidden as part of the challenge, but since these seem like they may be technical artifacts, I thought I would check to make sure this is something that is expected, and see if it would be possible to learn anything about what these are caused by? Thanks! All the best, Keith

Created by Keith Hughitt keith
a bayes/log likelihood method with several discrete inputs will result in this distribution.
Hi Daniel, Thanks for the response and clarifications! So in that case, the bands at one and zero could be explained by an unweighted PPI being combined with a weighted one. What is still unclear though is the gap between ~0.63 and 1. The fact that there is not a more gradual decreases in edge values approaching that value, and instead is a gap suggests that some threshold has been used. Would it be possible for someone to explain what is meant by the gap? Keith
FYI: One of the network providers writes: Your answer to the question is correct. The strange density pattern is a result of combining different interaction sources, each of which with its own score distribution. --daniel
Dear Keith I'll check with the data providers to get some additional details on how the weights were derived. I'm not too surprised by the bands though. These PPIs integrate data from different sources, a set of edges with the same weight could correspond to all interactions that are supported by two different databases, for example. As the weights in the networks generally do not have a well-defined physical meaning, I think it will be worth exploring different transformations of the edge weights during the leaderboard phase (e.g., unweighted, rank-transformed, ...). We haven't yet explored this ourselves so I can't say how informative the weights are for module identification in the different networks. Best, Daniel

Artifacts in Protein-protein interaction networks? page is loading…