This project focussed on the implementation of a machine learning model with respect to predicting kinase inhibition. The model could be used as a means for future drug discovery, which up until now has been a time-consuming process. Computer aided drug discovery (CADD), will allow researchers to find more molecular candidates for medical purposes in a more timely manner. To this end multiple approaches have been implemented. This page contains all of those, and some unfinished or unused codes that were not deemed significant enough to mention in the final product. However, some of this code could be used in further research as they might aid further accurate prediction models.
The complete code is divided into 7 main sub categories each contained in a folder
This folder contains the dataset with which the model is trained, and the dataset on which the final results are tested (which is handed in in the final assignment). Using this dataset all features are generated using the RdKit. Other files in this folder include a piece of code that generates a dataset with only a couple of of the features. This smaller dataset was occasionally used for quickly testing machine learning inputs.
This folder contains all the code and files that are associated with the exploratory data analysis. These include:
- A code that filters for outliers using an isolation forest, called Cleaning_dataset.py which produces a cleaned_dataset file.
- A code that filters out correlated variables, called Correlations.py which creates the reduced_dataset from the cleaned_dataset.
- A code that filters out variables with a low variance, called low_variance_filter.py
- A code that generates the boxplots images found the in "/Boxplots" folder, called boxplots.py
- A couple of HTML files that contain the descriptive statistics for different inhibition groups based on an analysis done using the Jamovi software.
This folder contains two main files: A code that generates Extended Connectivity Fingerprints (ECFP) for each molecule and the accompanying dataset that is created.
This folder contains the images of the boxplots that are generated using the boxplot code. It visulizes the distributions of all features.
The machine learning folder contains all the code of the machine learning. This contains one subfolder: KNN, which contains the 4 KNN codes that each use different inputs from the EDA. The main folder furthermore contains the code for the random forest for the different types of input. Finally the ML folder contains the code for the SVM model and the reduced_dataset which is used as possible input for the ML models.
This folder contains again a subfolder called Models, this consists of the acutal models that are created when run on training data and is used in the final results. The prediction folder further contains the results of the model on the untested_molecules_properties_all dataset called final_df.
Finally the visulalizations of the codes are gathered in the visulations folder. Including scatter plots, ROC-curces and confusion-matrices.
All contributors to this project are: Laura Moiana, Giorgia Barra, David Keur, Kirsten Brusse, Jaleel Dibani, Roan Jacobs, Jort Lokers and Famke Klop. Special thanks to Francesca Grisoni and Dragan Bosnacki for providing the necessary data.