wllgrnt / protein-networks Goto Github PK

Sarah Teichman had some project suggestions, as did Madan Babu and Eileen Nugent. It's been a while since both of those meetings, but I should write down what avenues they proposed, and at least assess feasibility.

This should probably be done by breaking each protein into a chain, then running the CD on each chain. There's potential to stitch the resulting networks back together, but its not strictly necessary.

This is needed to compare and contrast the different network generation inputs, and the resulting communities. The ability to output a network as a networkx Graph would be good. And Modularity, Conductance, Jaccard, NMI would be nice too

We should compare a broad family of proteins with known structural features and see if CD can pull these features out

After the community structure has been found for the proteins, we should perform an all-to-all comparison and identify conserved modules. We can compare these with the existing protein databases.

Changing the Network class to allow chains is going to break all my tests; fix them and reconsider my architecture.

In order to perform a comprehensive analysis of how different community detection methods break up proteins, we need lots of data for different proteins and different methods. Currently we have validation only for AFG and Infomap, and only data /gen capacity for Infomap. However, we're looking to answer the "why did you choose X?" question, so the different CD methods probably don't need integrating into the module proper.

We (and perhaps everybody else) need a weak isomorphism algorithm, but no quick ones are known.
Idea: By using community detection to generate a dendrogram, we can compare two graphs by comparing their trees.

We've shown that optimising Q gives us module fragments that don't show nearly the same compression as that of amino acids; CJP suggests plugging in the information calculator and performing a direct optimisation.

Chris currently uses steepest descent when doing energy minimisation, but with lots of local moves. Want to do group moves, but which groups? -> community detection!

Currently the module is unusable if it can't spot a MongoDB instance (and the location of said MongoDB instance is hardcoded). It would make more sense if:

• the URI and parameters were actually parameters
• the failure mode was to generate everything from scratch and store in a temp folder

Unclear how to do this at present.

Current code coverage is poor for everything except the Database class, and the unit tests don't test many of the possible networks I can generate e.g. atomic networks.