Drug-likeness scoring based on unsupervised learning.

Area of Science:

• Computational chemistry
• Machine learning in drug discovery

Background:

• Drug-likeness prediction is crucial for virtual screening but lacks definitive regression data due to complex clinical trial requirements.
• Existing binary classification models, often using graph neural networks, show performance variability dependent on the negative training set selection.

Purpose of the Study:

• To develop a novel unsupervised learning model for drug-likeness scoring that overcomes limitations of existing methods.
• To create a pragmatic tool for drug discovery utilizing only known drugs for training.

Main Methods:

• An unsupervised learning approach was implemented using a recurrent neural network-based language model.
• The model was trained exclusively on datasets of known drugs, eliminating the need for a negative set.

Main Results:

• The unsupervised model demonstrated consistent performance across various datasets, outperforming classification models.
• Drug-likeness scores generated by the unsupervised model showed meaningful separation in distributions, correlating with drug development stages.
• Classification models exhibited polarized, overconfident predictions for unseen data, unlike the proposed unsupervised method.

Conclusions:

• The novel unsupervised learning model offers a robust and pragmatic solution for drug-likeness scoring.
• This approach provides a valuable tool for virtual screening and has potential applications in broader biochemical research.