DG-GL: Differential geometry-based geometric learning of molecular datasets

Area of Science:

• Computational chemistry
• Machine learning
• Geometric deep learning

Background:

• Differential geometry (DG) has potential for analyzing complex molecular datasets.
• Its application in dimensionality reduction and encoding chemical/biological information is underexplored.

Purpose of the Study:

• To propose a novel differential geometry-based geometric learning (DG-GL) hypothesis.
• To demonstrate DG-GL's capability in analyzing large, diverse molecular and biomolecular datasets.

Main Methods:

• Encoding chemical, physical, and biological information into 2D element interactive manifolds.
• Utilizing multiscale discrete-to-continuum mapping with differentiable density estimators.
• Applying differential geometry to construct analytical element interactive curvatures.

Main Results:

• DG-GL represents 3D molecular structures on low-dimensional manifolds.
• Paired with machine learning, DG-GL shows strong descriptive and predictive power.
• Outperformed advanced methods in predicting protein-ligand binding affinity, drug toxicity, and solvation free energy.

Conclusions:

• DG-GL offers a powerful new strategy for molecular and biomolecular data analysis.
• The method effectively captures intrinsic physics and essential information from complex structures.
• DG-GL shows significant promise for applications in drug discovery and computational chemistry.