- Léa Goffinet (lea.goffinet@epfl.ch)
- Simon Lee (simon.lee@epfl.ch)
In an experiment on epigenetics and memory, Giulia Santoni (from the lab of Johannes Gr¨aff at EPFL) measured the gene expression levels in multiple cells of a mouse brain under three different conditions that we call KAT5, CBP and eGFP. In this challenge, the goal is to predict – as accurately as possible – for each cell the experimental condition (KAT5, CBP or eGFP) under which it was measured, given only the gene expression levels.
environment.yml - file is included in the repository. Takes dependencies from requirements.txt
To build an environment run the following command:
conda env create -f environment.yml --name [environment_name]
conda activate [environment_name]
The training data contains the normalized counts for 32285 genes in 5000 different cells together with the experimental condition under which each cell was measured. The test data contains only the normalized counts and your task is to predict the experimental condition. Data is included in this repository
./data/train.csv.gz
./data/test.csv.gz
However in our analysis we have saved a new filtered version of our train data which filtered out some genes expressed based on a threshold. In our case we decided to filter out genes that were seen across less than 10% of the cells. Filtering took extremely long so in our analysis we recommend you take advantage of our new dataset:
./data/filtered_train.csv.gz
- Description: Our linear based methods to compare the performance between different models. Multinomal logistic regression is a very basic classification algorithm and it performs the best out of all the methods. Similarly we also tested Multiple linear regressions and obtained similar results.
- Description: An ensemble learning method which has been consistently good at supervised learning tasks. We therefore took this tree based method which improves each previous model (boosting), and we obtain a model that is able to predict with 87% acurracy.
All the preprocessing, methods and analysis are contained in the ./src/cell_prediction_analysis.ipynb notebook.