jlaw9 / fungcat-function-prediction Goto Github PK

Our algorithms can take a long time (sinksource one-versus-none takes longer than an hour and a half).

We have one big network for all the proteins of all these species (70K nodes, 5M edges), but if for a specific function we only give high confidence predictions for a very small subset of proteins, we don't need the algorithm to give a score to the entire network, just that small portion.

More details outlined in the group google doc

Guarav requested that every prediction method also be responsible for creating a docker-version of their method so the ensemble prediction methods can try out different inputs or cross-validation methods without relying on us to run our methods and give them the results.

Some questions to consider:

• What will be the inputs and outputs for our image? Preferred inputs will be something like:
  1. A set of annotations in a GAF file
  2. A set of GO terms to make predictions for
  3. A network containing some combination of a sequence similarity network and a STRING network. The difficulty here is so far we're the only ones who use these networks so I'm not sure how that will work into their pipeline.

The process-go tool can transfer annotations up the GO DAG, and unknown annotations down the GO DAG. NOT annotations should also be transferred down the GO DAG.

We could also probably find another tool to transfer annotations for us.

Additional algorithms to implement as well as a short description:

The Sequence Similarity (SS) network weights range from 0 to 200, while the STRING network weights range from 150 (we're currently using a cutoff of 400) to 1000.

Currently I'm using a very simple approach and just multiplying the SS weights by 5, but we should really be using a smarter method to prioritize weights based on the contribution of these networks.

A python implementation of the original GeneMANIA algorithm (2008) is available here, but we're really interested in using the Simultaneous Weights (SW) GeneMANIA method (2010) to integrate the weights as it optimizes the weights based on labels of all functions simultaneously. The matlab code is available here.

When I build GAIN on Ubuntu 16 and try to run it on Ubuntu 14, the library libboost_filesystem.so.1.58.0 is missing:

/home/jeffl/src/c++/biorithm/trunk/gain/gain: error while loading shared libraries: libboost_filesystem.so.1.58.0: cannot open shared object file: No such file or directory

and when I build on 14 and run on 16, libboost_filesystem.so.1.54.0 is missing:

/home/jeffl/src/c++/biorithm-ubuntu14/trunk/gain/gain: error while loading shared libraries: libboost_filesystem.so.1.54.0: cannot open shared object file: No such file or directory

On some of the lab computers, both of those libraries are missing.