Unsupervised manifold embedding to encode molecular quantum information for supervised learning of chemical data

Area of Science:

• Computational chemistry
• Cheminformatics
• Machine learning

Background:

• Molecular representation is crucial for chemical machine learning, influencing model complexity and preventing overfitting.
• Understanding molecular interactions and properties requires examining local electronic information on molecular surfaces.

Purpose of the Study:

• To develop a novel, lower-dimensional molecular representation capturing surface electronic quantities.
• To assess the utility of this representation in predicting molecular interactions and properties.

Main Methods:

• Developed Manifold Embedding of Molecular Surface (MEMS) by treating molecular surfaces as manifolds and computing embeddings.
• Utilized MEMS as input features for shallow and deep neural network models.
• Evaluated MEMS performance in solubility prediction tasks.

Main Results:

• MEMS effectively embodies surface electronic quantities, retaining chemical intuition.
• Demonstrated the feasibility of using MEMS with both shallow and deep learning models.
• Achieved accurate solubility predictions, showcasing MEMS's expressiveness and robustness.

Conclusions:

• MEMS offers a powerful and chemically intuitive molecular representation for machine learning.
• The method is robust against dimensionality reduction, suitable for diverse chemical learning applications.
• MEMS advances the prediction of molecular interactions and properties in computational chemistry.