A mechanism-aware and multi-omic machine learning pipeline characterises yeast cell growth

Objective

This code uses strain-specific metabolic modelling coupled with gene expression data to produce predictive models for yeast growth.

Workflow

Step one:

The msb145172 dataset (Duibhir et al., 2014) is used to create strain-specific metabolic models through tailoring microarray-based constraints on a baseline metabolic model.

Step two:

Parsimonious flux balance analysis is utilised to simulate the metabolic activity and obtain strain-specific reaction fluxes.

Step three:

Feature selection techniques are applied. We explore three state-of-the-art techniques:

• NSGA-II - a genetic algorithm that allows for multiple objectives
• Sparse group lasso - a regularisation technique for grouped variables which promotes group sparsity
• Iterative random forest - accounting for non-linear interations among variables

Step four:

Machine learning and deep learning techniques are applied to produce predictive models for yeast growth, starting from gene expression, metabolic flux and both together. We use early, intermediate and late data integration, exploring the following baseline methods:

• Support vector regression
• Random forest
• Deep neural networks

Their multi-view counterparts are:

• Bayesian efficient multiple kernel learning
• Bagged random forest
• Multi-modal deep neural networks

Step five:

Generalisation tested using an experimentally independent test set (Sameith 2015), some feature imputation is also included to ensure the data is consistent with the MSB data.

Files and requirements

Metabolic Modelling

The metabobolic modelling code is in Code/Metabolic_Modelling and has been tested in Matlab R2015b.

To calculate the metabolic fluxes for all yeast strains, run RUN.m. This will automatically start the FBA process using the gene expression data in msbdataAltered.mat and the metabolic model in iSce926.mat.

The obtained flux levels can then be extracted and used to create a complete dataset

Approx running time: ~3 hours - this code will re-run the modelling process

Matlab dependencies

• The COBRA toolbox

Machine Learning

All the material for reproducing the machine learning tests is in Code/Machine_Learning.

To build the predictive models, all the dataset files are located in Code/Machine_Learning/data

The R script ML_Code.R trains and applies all the machine learning approaches (SVR, RF, BEMKL), except for the deep neural networks. This script has been tested with R 3.4 and can be run by simply calling Code/Machine_Learning/ML_Code.R.

Approx running time: 6 hours - this code will run utilising the features selected by the feature selection techniques (the code to re-calculate these is inside the Code/feature_selection folder) and re-train the non-deep-learning models

R dependencies

The script machineLearning.R requires that the libraries listed at the top are installed to ensure it runs to completion. In particular, implementations of the machine learning methods used are in:

• caret
• nsga2R
• SGL
• iRF
• bemkl - This file is included in the repo

Deep Learning

A separate Python script is dedicated to training and testing the deep learning models. To re-run the experiments simply call DL_Code.py. This also includes the SHAP analysis calculations for feature importance and the experimentally independent test-set tests.

Approx running time: 3 days for full training

Python dependencies

The deep learning code has been tested with the following libraries:

• Python 3.5
• Tensorflow 2.0
• Keras 2.0
• Numpy 1.5
• Pandas 0.25
• Matplotlib 3.1.1
• Seaborn 0.9

Note

In all cases we have provided the SEED inside the code files to ensure reproducability along with the train-test split indicies