Improving molecular representation learning with metric learning-enhanced optimal transport.

Area of Science:

• Computational chemistry
• Machine learning
• Materials science

Background:

• Machine learning models in chemistry and materials science often struggle with limited or heterogeneous training data.
• Generalizing beyond the initial training domain remains a significant challenge for existing models.

Purpose of the Study:

• To develop a novel algorithm, MROT (Machine learning with Optimal transport and Regression), to enhance the generalization capability of molecular regression models.
• To address the domain gap in chemical data by learning continuous data labels and incorporating posterior variance regularization.

Main Methods:

• Developed MROT, an optimal transport-based algorithm for molecular regression.
• Introduced a new metric for measuring domain distances.
• Implemented posterior variance regularization over the transport plan to bridge domain gaps.

Main Results:

• MROT demonstrated significant performance improvements over state-of-the-art models in unsupervised and semi-supervised settings.
• The algorithm effectively handled chemical property prediction and materials adsorption selection tasks.
• MROT successfully learned continuous data labels and bridged domain gaps.

Conclusions:

• MROT offers a powerful approach to enhance generalization in molecular regression, particularly with limited or heterogeneous data.
• The algorithm shows significant potential for accelerating the discovery of novel substances with specific properties.
• This work provides a promising direction for developing more robust and versatile machine learning models in chemical and biological applications.