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.
- 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.
Tip
Start working with Google Colab or Jupyter Notebook!
- First, install the following packages using pip3 or pip:
numpypandassklearnpymc3theano.tensorpicklescipymatplotlibPILpyplotrdkitseabornscikit-learn
Note
- 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.
Important
For any questions please contact: 👉 Ashok K. Sharma; ashoks773@gmail.com or compbiosharma@gmail.com