In this repository there is the surce code for the application molecule_solubility_app which allows the user to input a SMILE formula, e.g. the aspirin CC(=O)OC1=CC=CC=C1C(=O)O, and it predicts its solubility (in log scale). We've trained a simple linear model to predict the solubility value from some features.
Here is the link at the app: https://moleculesolubilityprediction.streamlit.app. On the left hend side it is possible to input one ore more SMILE formula (divided by a newline) and then the system will provide the solubility prediction based on some molecular descriptors. Notice that at the beginning some default SMILE formula are provided.
- The dataset I've worked on
delaney_solubility_with_descriptors.csv - Configuration files
packages.txtandrequirements.txtwhich are needed to deploy the app on Streamlit - The script for the app
solubility-app.py - The Jupiter Notebook in which the linear model can be found
solubility-web-app.ipynb - The results of the trained linear model asi pickle file `solubility_model.pkl.
- A logo `solubility-logo.jpg
The dataset we used for the training process is the following: John S. Delaney. ESOL: Estimating Aqueous Solubility Directly from Molecular Structure. J. Chem. Inf. Comput. Sci. 2004, 44, 3, 1000-1005. Available here:ESOL: Estimating Aqueous Solubility Directly from Molecular Structure (J. Chem. Inf. Comput. Sci. 2004, 44, 3, 1000-1005). The dataset is in csv format with the following features:
- MolLogP (partition coefficient octane/wather) abbreviated as LogP
- MolWt (melecular weight) abbreviated as MW
- NumRotatableBonds (number of specific chemical bonds) abbreviated as RB
- AromaticProportion (aromatic proportion) abbreviated as AR
- logS (logarithm of the solubility)
The feature of interest is logS.
This project was part of a bigger project I worked on with Unilab. I couldn't provide here all the results we obtained, but I've tryied to show here some important intermediate steps which I think can be valuable on their own.