DILI_Bayesian_Model

This repository contains example code for Bayesian Neural Networks (BNN) and other ML approaches utilizing publicly available AstraZeneca's in-vitro assay data for predicting Drug Induced Liver Injury (DILI) risk.

Objectives

• Step 1: AZ's in vitro assay data used for the development of the Bayesian Hapatic Safety model was downloaded.
• Step 2: The BNN model was re-developed and the performance of the BNN model was compared with the other Machine learning-based methods.
• Step 3: Important descriptors (based on Mutual Informative Score) were selected based on DILI classification abilities
• Step 4: A refined model using original features + top 20 descriptors was generated which shows improved performances.
• Step 5: A similar approach was used utilizing internal in vitro hepatic safety data + descriptors
• Step 6: The model was finally incorporated into the DILI de-risking strategy. It provides DILI prediction probability as well as the relative contribution of each assay.

Requirements

Tip

Start working with Google Colab or Jupyter Notebook!

• First, install the following packages using pip3 or pip: numpy pandas sklearn pymc3 theano.tensor pickle scipy matplotlib PIL pyplot rdkit seaborn scikit-learn

Steps

Note

The pipeline has been divided into multiple steps

• Step1: Use BNN_Check.ipynb This script was used to develop a BNN model using AZ's in vitro assay data.
• Step2: Use BNN_vs_otherML.ipynb This script was used to develop other ML models to compare the performance with BNN model
• Step3: Use MultiClass_ROC.ipynb and BinaryClass_ROC.ipynb scripts were used to compute ROC performance for multi-class (No-DILI; Medium-DILI; and Severe-DILI) and binary class (No-DILI vs DILI)
• Original internal assay data and plots has not been included here due to confidentially issues.

Contact: 🤚

Important

For any questions please contact: 👉 Ashok K. Sharma; ashoks773@gmail.com or compbiosharma@gmail.com