Creating Free Form Feedforward Network Models

Robert Ness
Jan 27, 2015

Motivation

  • We can reconstruct signaling network topology from single cell data
  • However, we need to choose which proteins to measure
  • To do this, we model how signal flows through feedforward topology

Approach: 3 inputs

  1. A feedforward topology of a signaling pathway
  2. A set of input signals on upstream nodes (eg. drug treatments)
  3. A set out output values on downstream nodes (eg. transcription factor activity level)

Model how signal flows through the network using basic assumptions. Identify nodes of high entropy as priorities for measurement.

Toy Example: Gaussian Model

A set of variables with a joint normal distribution factorizes as follows:

plot of chunk unnamed-chunk-2

Application of Model on Gaussian Data

  • Intermediate node estimates depend on variance in inputs and outputs

plot of chunk unnamed-chunk-7

Interpretation to general case

  • Intermediate nodes are estimated as a function of input nudes and minimizing error of prediction of output nodes
  • Suggests importance of a node is function of network connectivity from inputs and to outputs.

Ongoing work

  • Currently model only accepts one output, need to generalize to multiple outputs.
  • Need to change activation function to simple biological system kinetics.
  • Need to adjust optimization so weights cannot be negative.