- Trained machine learning model to predict inhibition of SARS-2 Coronavisus by small molecules
- Created descriptors of each molecule relevant to drug discovery
- Optimized random forest model for prediction
Python Version: 3.7
Packages: pandas, numpy, sklearn, matplotlib, seaborn, pickle
Organic Chemistry: rdkit
For Requirements: conda env create -f drugdiscovery.yml
Data for 1669 small molecules was taken from the CHEMBL databse. Each molecule was assigned a Hit score between 0 and 1 basd on ability to inhibit Coronavirus infection of Human Renal Cortical Epithelial Cells. Scores greater than 0.6 are considered hits.
The rdkit package was used to generate relevant physical descriptors of a molecule based on their molecular formula. The following descriptors were calculated for each molecule:
| Feature | Description |
|---|---|
| Number of Atoms | Number of Atoms in the Molecule |
| Formal Charge | Electric charge on the Molecule |
| Heavy Atoms | Number of Heavy Atoms |
| Molar Refractivity | Polarizability of Molecule |
| Rotatable Bonds | Number of Rotatable Bonds |
| MW | Molecular Weight of Molecule |
| LogP | Partition Coefficient of Molecule |
| NumHDonors | Number of Hydrogen Bond Donors |
| NumHAcceptors | Number of Hydrogen Bond Acceptors |
| SAmapping | Topological Surface Area |
| Number of Rings | Number of Rings in Molecule |
Almost all the predictors are approximately normally distributed with some righ skewed outliers.
The Hit scores are also normally distributed with only 13 drugs exceeding the cutoff to be considered a hit.
A number of features are highly correlated in ways that make sense based on their physical interpretation. Therefore a number of the most highly correlated features were dropped from the final model.
I fit a multivariable linear model as a baseline for comparison. I then fit a random forest regression model with hyperparameter tuning using GridSearchCV with 3 fold cross validation. Overall, the best predictor was the random forest with n_estimators = 120, max_depth = 8 and max_features = log2
The model correctly places Remdesivir, a antiviral known to be effective above the hit cutoff. This method uses primarily chemical descriptors of a molecule and thus has difficulty predicting how a drug will perform in the complex biological interactions involved in viral inhibition. Future work could focus on using known molecule - protein motif interactions or known moleculal functions in biological pathways to predict efficacy as a viral inhibitor.
Inspiration