Learning machine learning through a familiar chemical problem
pKa-ML Education introduces decision trees, feature engineering and generalization through organic acid pKa prediction. A graphical interface supports the main teaching workflow without programming.
Five datasets, progressively refined descriptors
- Start with simple carbon and carboxyl counts.
- Introduce halogen counts and more specific substituent descriptors.
- Add electron-withdrawing group positions and strengths.
- Compare training, validation and fixed external test performance.
Teaching focus
Students explore overfitting by changing tree depth, interpret tree splits and feature importance in chemical terms, and examine how chemical knowledge improves feature engineering.
Getting started
Windows users can obtain an executable from figshare. Python users can run python DTcode/predictpka260324.py; consult the repository for dependencies.
The repository contains the source and dataset1.csv through dataset5.csv. The associated teaching manuscript is listed as in preparation; consult the repository for updated citation information.