Related Experiment Videos
Prediction of physicochemical properties based on neural network modelling
Jyrki Taskinen1, Jouko Yliruusi
1Viikki Drug Discovery Technology Center, Department of Pharmacy, University of Helsinki, Helsinki, Finland. jyrki.taskinen@helsinki.fi
Advanced Drug Delivery Reviews
|September 5, 2003
Summary
Neural network models can predict organic compound properties from molecular structure, aiding pharmaceutical research. While accurate on training data, their performance on independent datasets needs improvement for reliable drug discovery.
Area of Science:
- Computational chemistry
- Medicinal chemistry
- Machine learning in drug discovery
Background:
- Predicting physicochemical properties of organic compounds is crucial for pharmaceutical research.
- Neural network modeling offers a powerful approach to correlate molecular structure with properties.
Purpose of the Study:
- To review the literature on neural network modeling for predicting physicochemical properties of organic compounds.
- To assess the current state and limitations of these models in pharmaceutical research.
Main Methods:
- Review of published literature on neural network applications in cheminformatics.
- Analysis of various molecular representation methods (fragments, topological indices, quantum chemical descriptors).
- Evaluation of prediction accuracy using internal and independent test sets.
Main Results:
- Feed-forward neural networks are the most common technique, but others are emerging.
- Successful modeling of properties like octanol-water partition coefficient, water solubility, boiling point, and vapor pressure.
- Models show high accuracy on internal test sets, comparable to experimental measurements.
- Performance significantly decreases on independent test sets.
Conclusions:
- Neural networks show promise for predicting physicochemical properties, but generalization to new chemical spaces remains a challenge.
- Further development is needed to improve model robustness and reliability for pharmaceutical applications.
- Addressing the gap between internal and external validation is critical for practical use.