Development & deployment of a machine learning model for predicting boiling points from chemical structures.
Note: The deployment folder necessarily includes JavaScript from ChemDoodle Web Components, which is distributed under the GNU General Public License. All other portions of this repository are original products owned by this author.
A chemical's boiling point (the temperature at which it boils) is useful in a variety of different contexts, from selecting appropriate experimental conditions to characterizing the components of a mixture.
Gas chromatography (GC) is a very common laboratory technique that separates the components of a mixture, primarily by boiling point. Without supplementing it with other techniques (such as mass spectrometry, MS), however, GC alone cannot provide information about the identity of a mixture's components -- only the total number of components, their relative ranking in boiling point (low to high), and their relative abundance in the mixture.
An example of annotated GC data for three different mixtures. The x-axis is approximately a surrogate for boiling point. Component names are shown for clarity, but are not provided directly by the GC measurement. Shared by Wikipedia user Quantumkinetics under the Creative Commons Attribution 3.0 license.
A GC sample could contain, for example, a desired reaction product (such as a useful pharmaceutical) and an undesired contaminant (such as a toxic byproduct accidentally formed while preparing the valuable drug). We may even know the chemical structure of each molecule, but which component is which in a GC analysis? Answering this question helps us understand: Is this mixture mostly the useful drug, or mostly the toxic byproduct?
Boiling point data are not available for all possible chemicals. With new molecules being made every day, there will will never be a database of every chemical's boiling point. A reliable tool for predicting unknown boiling points would simplify the interpretation of GC data.
This project aims to provide a tool for predicting unknown boiling points from a given chemical structure for organic molecules in the range of 400 - 700 K. The finished predictive tool is deployed at bp.withjosh.net.
This repository is divided into two sections:
-
The development of a machine learning model for boiling point prediction, stored within the root of the repository. This includes Jupyter notebooks showing:
1_data_collection_bp.ipynb: Collection of boiling point data (8.8k molecules) from a PDF reference textbook. Chemicals are identified by CAS number, which do not convey any structural information.2_data_collection_smiles.ipynb: Collection of a SMILES code for each identifying CAS number. SMILES codes relay chemical structures in plain text.3_data_splitting.ipynb: Splitting the data into training, test, & validation sets.4_exploratory_analysis.ipynb: Exploratory analysis and feature selection, in part using the Python library RDKit.5_regression.ipynb: Training of an XGBoost regression model.
-
A Flask-based deployment of the model for prediction of user-drawn models in the
deploymentdirectory.
No installation is necessary in order to view the Jupyter notebooks used in the development of this model.
If you wish to install a copy of the deployed model for generating predictions or if you wish to reproduce & extend this analysis, first install the required Python modules listed in requirements.txt. Then, clone this repository into a directory of your choice. The web app may be launched by executing python3 frontend.py from the deployment subdirectory.