Tox21 QSAR

About

This tox_21_qsar repository contains cheminformatics and machine/deep learning code for descriptive analysis and QSAR modelling for tox_21 datasets.

Several predictive models were built using various classical machine learning and deep learning algorithms, such as:

1. Random Forests
2. XGBoost
3. Feed Forward Neural Networks (FFNs)
4. Long Short-Term Memory Neural Networks (LSTMs)
5. Graph Convolutional Networks (GCNs)
6. Meta Classifiers, including hard and soft voting.

The projects explore various data splitting algorithms, such as MaxMin (diversity) Picking using various fingerprints, as well as various hyperparameter tuning approaches such as [Grid CV] (https://scikit-learn.org/dev/modules/generated/sklearn.model_selection.GridSearchCV.html) and Randomized search.

Several methods are also used for explainability, including SHAP and LIME.