rohitfarmer / bittersweet

Portfolio Project for Machine Learning

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

Dataset and scripts to build machine learning models for the classification of the bitter and sweet taste of small molecules

This is an example/portfolio project in which using a publicly available real-life dataset, I have demonstrated different aspects of data science/machine learning. The dataset is based on the study from Cosy lab (https://github.com/cosylabiiit/bittersweet) on classifying small molecules into bitter and sweet classes. From the original project, I have only borrowed the list of the molecules with their SMILES and target labels. I have converted the SMILES into their canonical forms using RDKit and then used Mordred (https://github.com/mordred-descriptor) to calculate molecular descriptors (features). My intention in this project is to demonstrate steps all the way from feature generation/selection to modeling. A more elaborate discussion on the methods used is documented within the Jupyter notebooks along with the code. Steps involving feature generation require some domain knowledge of computational chemistry; however, the rest of the project can be viewed as any generic data science project. For feedback and comments please send an email at rohit.farmer@gmail.com.

Material and Methods

Folder Structure

  • data: In the repository. Original data in TSV format.
  • notebooks: Jupyter notebooks with the code used in this project.
  • output: Not in the repository; create it. Contains intermediate outputs.
    • df_canon_smiles.pkl: Pandas pickle with canonical SMILES for each parsable SMILES string.
    • bitter_sweet_2d_descriptors.[pkl, tsv].gz: Pandas pickle and TSV files with 2D descriptors for each parsable SMILES string.
    • bitter_sweet_2d_plus_3d_descriptors.[pkl, tsv].gz: Pandas pickle and TSV files with 2D and 3D descriptors for each parsable SMILES string.

Workflow Reproducibility

For reproducibility please use the latest singularity container hosted at: https://cloud.sylabs.io/library/_container/5ddc495c54b1ada33dabbed7. The Singularity recipe file that was used to compile the container can be found here.

To execute jupyter notebooks via singularity container run singularity exec bittersweet.sif jupyter notebook --no-browser. Copy and paste the URL to a web browser.

Descriptor/Feature Generation

Molecular descriptors are properties of a molecule that can be described numerically for an example, molecular weight. These molecular descriptors can be used as features in machine learning. In this project Mordred python package was used to calculate descriptors. Mordred using RDKit under the hood and can calculate both 2D and 3D descriptors that include all the descriptors from RDKit and more. feature-calculation-mordred notebook describes a protocol and the code used to calculate these descriptors.