Perceptromic analysis

Reference

Identification of distinct pathological signatures induced by patient-derived α-synuclein structures in nonhuman primates

Bourdenx M., Nioche A. et al. Science Advances 2020 Vol. 6, no. 20, eaaz9165

AIM

This analysis allows ranking of variables using Multiple Layer Perceptrons (MLPs).

We use MLPs with a constrained architecture (3 neurons per layer - 1 hidden layer).

A random matrix of the same size of the input matrix is used as a control of learning.

Steps

Overall description

1. All combinations of 3 variables from the input dataset will be use to train an MLP to predict 3 ouput variables. with a 50-folds cross validation (random sampling among available samples using a 80%/20% splitting)
2. All combinations are then ranked regarding their performance in perdicting the 3 output variables. The top 1% is selected for further analysis.
3. The top 1% selected from the previous step is break into single variable and variables are ranked by the frequency of appearance in the top 1% of best combinations of 3 input variables.

Data preparation

A matrix of n examples (as row) and p variables (columns) is used as input. The variables used as output of the MLPs (the variables needed to be predicted should be the last 3 columns of the matrix.

Main script description

Using combinations.py (main folder):

• The DataManager class will

  • Load the file (as .txt)
  • Split input and output variables
  • Center and normalize the data (Standard scoring : https://en.wikipedia.org/wiki/Standard_score)
  • Scale the data between -0.5 and 0.5

• The Supervisor class will then

  • prepare kwargs of the different combinations of input variables (including cross validation steps that need to be performed)
  • Compute the different combinations

Data are obtained as SQLite databases.

Prerequisites

Several packages are necessary:

• Python >3
• Cython (MLP have written in cython to increase computation speed)
• Numpy
• Scipy
• Matplotlib
• sqlite3
• tqdm