/solubility

Primary LanguageJupyter NotebookMIT LicenseMIT

Solubility prediction project

The results generated using the code in this repository have been published in Journal of Chemometrics: https://analyticalsciencejournals.onlinelibrary.wiley.com/doi/full/10.1002/cem.3349

If you use the code/library in (other) projects, please consider citing our paper.

Project structure

.
├── data                 # Data for modelling
├── results              # Results 
├── experiments          # Automated tests
├── src                  # Source, models, tools, utilities
├── LICENSE
└── README.md            # Brief repo description and installation recommendation

The repository

There are 3 important data files, all sharing the same index (canonical SMILES). The data is also published at http://doi.org/10.5281/zenodo.4008331

  • data/descriptors.csv | Descriptors file
  • data/fingerprints.csv | Fingerprints file
  • data/solubility_data.csv | Predictive target and data splits (random, picking, pca split)

The /src directory has all relevant modules and functions for modelling and preprocessing. The /results directory is the data drop from trained models.

Best regressor

This scripts creates a pickle file with model parameters and results. Should be run as:

python best_regressor.py

Running the winning models

The two winning models in our work (LASSO and Random Forest). Model parameters are included in the files. Should be run as

python run_indi_model_lasso.py

or

python run_indi_model_rf.py

The code is set up as follows:

src has all the modules necessary for modelling

src/configs.py | Parameter space definitions for ML models src/models.py | Optimization and modelling modules src/model_support.py | Preprocessing routines src/utils.py | Aux