umwpl-cardiotoxicity-prediction

Overview

Repository for Machine Learning in Drug Design course. The goal of the project is to predict cardiotoxicity of given substances and to determine features that are most responsible for toxicity.

Goals

As part of the project, I wanted to achieve the following goals:

perform a comprehensive data analysis of all fingerprint files.
perform regression on IC50 values.
perform classification to determine if given substance is toxic or not.
test classification models with different IC50 thresholds and compare results.
compare results of approaches with regression and classification.
figure out which fingerprint type gives the best results and test whether mix of different fingerprint types gives better results.
validate how well models behave with limited number of features and test different methods for feature selection.
determine if it's possible to find features that are essential for determining cardiotoxicity.

Installation

Simply download it with Download ZIP option on Github or clone it with git clone <repository>.

After that you need to create new conda environment - this can be done with

conda env create --file environment.yml -n <name>

After this just run jupyter notebook in the main folder.

Data

The initial data is in ready_sets.zip, as size of .csv files was too large for github. You need to extract it into ready_sets folder. After extraction it contains three different fingerprint types:

Extended-Connectivity Fingerprint
Klekhota-Roth Fingerprint
Molecular Access System Fingerprint

Project structure

ready_sets.zip : contains data used to predict cardiotoxicity
data_analysis.ipynb : fingerprint data analysis and visualization
data_preparation.ipynb : all required preparation to tranform .csv files into data frames for models.
regression.ipynb : notebook with all regression models and results
classification.ipynb : notebook with all classification models and results
class_regr_comp.ipynb : noetbook to perform comparison of best regression model with best classification model.
utils.py : contain code from data_preparation.ipynb notebook in format that is easier to import.
presentations : both entry and final presentations in .pdf format.
png : some .png files used for results section in README.

Results

For detailed results, please refer to the respective notebooks and final presentation. Here I will just show most important charts.

Regression

Regression results for different fingerprints and different amounts of used features. Results were compared with R^2 score. The best results were obtained for mixed fingerprints, with R^2 score of about 0.68

Classification

Classification results for IC50 = 10 000 threshold and different amounts of used features. Results were compared with F1 score and accuracy. The best model had accuracy and F1 score of about 0.79

mateuszpoleski/cardiotoxicity-prediction