Reducing overconfident errors in molecular property classification using Posterior Network.

Area of Science:

• Computational chemistry
• Machine learning in drug discovery

Background:

• Deep learning models are vital for predicting molecular properties in drug development.
• Traditional Softmax-based models exhibit poor uncertainty estimation, leading to overconfident mispredictions for out-of-distribution data.
• These limitations pose significant risks and costs in the drug development pipeline.

Purpose of the Study:

• To improve uncertainty estimation in deep learning models for molecular property classification.
• To mitigate overconfident mispredictions in drug development applications.
• To introduce a novel approach using normalizing flows within an evidential deep learning framework.

Main Methods:

• Replaced the standard Softmax function with a normalizing flow in deep learning classification models.
• Utilized an evidential deep learning approach inspired by Posterior Network.
• Evaluated the strategy on synthetic datasets, ADMET predictions, and ligand-based virtual screening.

Main Results:

• The proposed strategy significantly reduced overconfident mispredictions compared to traditional Softmax models.
• Enhanced model reliability in identifying out-of-distribution samples.
• Demonstrated improved performance across various molecular property prediction tasks.

Conclusions:

• Evidential deep learning with normalizing flows offers a robust solution for uncertainty estimation in molecular property prediction.
• The developed framework effectively addresses critical limitations of current deep learning models in drug discovery.
• This approach provides a valuable tool for safer and more efficient drug development.