Bioinformatics Project from Scratch - Drug Discovery #6 (Deploy Model as Web App) | Streamlit #22

In previous videos of the "Bioinformatics from Scratch" series, topics covered included compiling bioactivity datasets from the ChEMBL database, performing exploratory data analysis on Lipinski descriptors, building a random forest model, and comparing model performances using the Lazy Predict library. The current video focuses on converting these machine learning models into a web application using the Streamlit library, allowing deployment in the cloud for user interaction and prediction tasks. The example application predicts bioactivity for acetylcholinesterase, a target protein for Alzheimer's disease.

The video demonstrates the setup of the application environment, starting with activating a Conda environment and locating the app files, including 'app.py'. It walks through running the application with Streamlit and shows an example where users upload an input file containing SMILES and ChEMBL ID. These identifiers aid in generating molecular fingerprints with the PaDEL-Descriptor tool, which are crucial for bioactivity predictions using the trained models. An acetylcholinesterase model is highlighted for its relevance to Alzheimer's research.

The process of building and using a random forest model is explained, beginning with feature selection and descriptor preparation from SMILES strings. The procedure involves reducing descriptors from 881 to 218 to enhance model efficiency. The video provides insight into model training using Jupyter Notebook, saving the model as a pickle file, and preparing it for integration into the web app. Performance evaluation metrics, such as R-squared and mean squared error, are calculated to assess model prediction effectiveness.