Final ML project @EPFL (Brisken lab)

70% of breast cancers can be classified as estrogen receptor positive (ER+). Recent evidences describe tumor as a very complex and heterogeneous disease, highlighting the importance of taking into consideration inter- and intra-tumor variability. Patient-derived xenografts (PDXs) emerged as promising and clinically-relevant preclinical model able to recapitulate the clinical settings. One promising way to estimate human cells receptivity to hormones is to analyse patient genes expression. In this document, we present a set of machine learning algorithms and features selection techniques to classify human cells implanted into mice according to their hormone receptivity by using only their gene expression. According to the results of our study, we show that this method can be used to classify patient's cells receptivity to hormones.

This Github contains all the code and the notebooks used for our experiments.

Final report can be found here.

Dependencies

numpy==1.17.2
pandas==0.25.1
scikit-learn==0.21.3
matplotlib==3.1.1

Dataset

The dataset has been collected at the EPFL Brisken Laboratory, Switzerland and is part of a PhD research project. It contains information related to 15004 genes describing a total of 29 samples collected from mice which were exposed to 3 different hormonal treatments.

Notebooks

We provide a set of notebooks presenting the different algorithms tried and the associated plots and results.

Feature selection and visualization

feature_selection.ipynb : Comparison of different feature selection methods
gaussian_mixture_model.ipynb : Cluster visualization with Gaussian mixture model
pca_visualization.ipynb : 2D and 3D visualization with Principal Component Analysis (PCA)

Classification algorithms

knn.ipynb : K-Nearest Neighbours algorithm
logistic_regression.ipynb : Logistic regression
random_forest.ipynb : Random Forest
decision_trees.ipynb : Decision Trees

Dataset preprocessing

The preprocessed dataset can be generated from the raw dataset using preprocessing.py script with the following options:

python preprocessing.py --norm features: This option will normalize the data with respect to the columns (features).
python preprocessing.py --norm samples: This option will normalize the data with respect to the rows (samples).

Other options:

--saving_path: Path where you want to save preprocessed dataset CSV file.
--dataset_path: Path where the dataset CSV file is located.

Prediction on unlabeled dataset

The prediction obtained with the different models can be generated using generate_prediction.py script with the following options:

python generate_prediction.py --model all: This option will produce prediction for all algorithms implemented in different CSV files.
python generate_prediction.py --model your_model: This option will produce prediction for a specific algorithm (your_model) in a CSV files.