Evaluating Machine Learning Models for Drug Efflux Predictions
Author Information
Author(s): LeDonne Norman C Jr, Rissolo Kevin, Bulgarelli James, Tini Leonard
Primary Institution: AstraZeneca Pharmaceuticals
Hypothesis
The use of structure-activity landscape index curves can improve the evaluation of predictive model performance.
Conclusion
The study suggests that the choice of training sets significantly impacts model quality and that the SALI/SCI approach can enhance predictive model evaluation.
Supporting Evidence
- The study evaluated multiple machine learning models including neural networks and random forests.
- SALI curves provided insights into model performance across different prediction spaces.
- Models built with 3D descriptors showed improved predictive capabilities compared to those with 2D descriptors.
Takeaway
This study shows how scientists can use special graphs to better understand how well their computer models predict drug behavior.
Methodology
The study used various machine learning models to predict drug efflux potential and evaluated their performance using SALI curves and SCI values.
Potential Biases
The random selection of training and test sets may lead to inferior model performance.
Limitations
The training and test sets were derived from the same chemical series, which may bias the model performance assessment.
Digital Object Identifier (DOI)
Want to read the original?
Access the complete publication on the publisher's website