Enhancing Molecular Contrastive Prediction by Recognizing Metric Distribution Separation.

Area of Science:

• Computational chemistry and cheminformatics
• Machine learning in drug discovery
• Artificial intelligence for molecular property prediction

Background:

• Molecular machine learning is crucial for predicting molecular properties and accelerating AI-aided drug discovery.
• A key challenge is effectively transferring knowledge from known molecules to large sets of unseen candidates.
• Existing methods face limitations in knowledge transfer efficiency for novel molecular entities.

Purpose of the Study:

• To develop a novel metric distribution contrastive prediction method for molecular property prediction.
• To enhance the transfer of knowledge from known molecules to unseen candidates in drug discovery.
• To improve the accuracy and efficiency of predicting molecular properties.

Main Methods:

• Established metric distributions among ordered categorical molecule populations.
• Constructed metric distributions, extracted class prototypes, and employed Bhattacharyya distance for distinguishing distributions.
• Integrated pre-training and prediction phases into a unified neural computation framework.

Main Results:

• The proposed method successfully minimizes intra-class sample distances while maximizing inter-class sample distances.
• Demonstrated significant improvements in prediction accuracy across diverse benchmark molecular datasets.
• Outperformed existing models in molecular property prediction tasks.

Conclusions:

• The metric distribution contrastive prediction method offers a robust framework for molecular property prediction.
• This approach effectively addresses the challenge of knowledge transfer in AI-aided drug discovery.
• The validated improvements in prediction accuracy highlight the method's potential for accelerating drug candidate identification.