coleslaw481 / ud985

UD-985

This repository contains scripts and data needed for the experiments described below.

EVERYTHING IN THIS REPO EXPERIMENTAL AND MAY CONTAIN ERRORS. USE AT YOUR OWN RISK.

The Problem

Currently the Community Detection App (CDAPS) allows users to run four community detection apps wrapped in containers (OSLOM, CliXO, Infomap, & Louvain) but the app requires the user to invest significant effort to decide on optimal parameters to use.

The Ultimate Goal

The ultimate goal is to empower CDAPS with features that makes it easier for users to generate "useful" clusters/hierarchies.

A “useful” cluster/hierarchy is defined by a set of attributes possessed in the resulting cluster/hierarchy. What those attributes are is still up for debate.

The Intermediate Goal

Given a set of attributes about a network try to predict the number of clusters that a specific algorithm will generate when run on the network.

Desired outcome

We want to gain insight as to whether a neural network can deduce a pattern for predicting outcomes from clustering algorithms. If neural networks are successful prediction, then experiments with more advanced predictive neural networks can be performed.

Experiment

1. Take Bioplex and generate thousands of subnetworks in two ways
   1. By generating random sub networks generate_subgraph.py
   2. By generating sub networks using GO terms generate_gosubgraph.py
2. Create a neural network in PyTorch that takes a set of basic network attributes (# nodes, # edges, avg node degree etc..) and outputs a single number denoting number of clusters that will be generated
3. Generated training data by running Infomap on networks from step 1
4. Need to verify there is not wild variability when feeding same input to Infomap repeatedly.
5. Use 80% of data from step 3 to train the neural network
6. Run prediction on remaining 20%
7. Assess predictive power
8. If successful repeat for CliXO, Louvain, OSLOM … could also vary parameters and see if prediction holds for a new model trained with alternate parameters and whether the old models do well

Note

For Infomap use some default set of parameters that tend to give good hierarchies.

Design of Neural Network

Description of the inputs

• # edges / # nodes ^2

  The number of edges in the graph divided by the number of nodes squared

• density

  Value from this equation described by networkx: https://networkx.github.io/documentation/stable/reference/generated/networkx.classes.function.density.html

• degree mean / # nodes

  Average of degree of all nodes divided by number of nodes in graph

• degree stddev

  Standard deviation of degree of all nodes in graph

Scripts Requirements

• ndex2 client > 3.3.1 & < 4.0.0
• pandas
• pytorch
• networkx > 2.3
• matplotlib

Scripts and what they do

• create_trainingdata.py

  Takes networks created from generate_subgraph.py or generate_gosubgraph.py along with output from a community detection algorithm and generates training data usable by predict_clusters.py

• extract_trainingdata.py

  NOT IMPLEMENTED This script will take training data generated by create_trainingdata.py and attempt to normalize the data by extracting a subset of the data where the number of clusters is evenly distributed

• generate_subgraph.py

  Creates networks from an input CX network by randomly picking nodes or selecting nodes matching genes for GO terms

• plot_trainindata.py

  Creates plots from training data generated by create_trainingdata.py

• predict_clusters.py

  Runs training and prediction