Dataset Construction to Explore Chemical Space with 3D Geometry and Deep Learning.

Area of Science:

• Computational chemistry
• Machine learning
• Materials science

Background:

• Deep learning model performance is contingent upon the quality and scale of training datasets.
• Developing large, high-quality datasets is crucial for advancing molecular modeling and property prediction.

Purpose of the Study:

• To introduce Frag20, a novel, large-scale fragment-based molecular dataset.
• To establish a new data preparation protocol for generating molecular datasets.
• To develop and validate deep learning models for molecular energy prediction using the Frag20 dataset.

Main Methods:

• A new data preparation protocol was implemented.
• A large fragment-based dataset (Frag20) was constructed with over half a million molecules.
• Molecular geometries were optimized using Merck molecular force field (MMFF) and density functional theory (DFT) at the B3LYP/6-31G* level.
• Simplified PhysNet architecture was employed to build molecular energy prediction models.

Main Results:

• The Frag20 dataset comprises optimized 3D geometries and calculated molecular properties for numerous molecules.
• Developed molecular energy prediction models achieved accuracy close to or exceeding chemical accuracy (1 kcal/mol).
• Model performance was validated on diverse test sets, including CSD20 and Plati20, which utilize experimental crystal structures.

Conclusions:

• The Frag20 dataset and the developed data preparation protocol facilitate robust deep learning model creation.
• The PhysNet-based models demonstrate high accuracy in predicting molecular energies for both MMFF and DFT optimized geometries.
• This work advances the application of deep learning in computational chemistry and molecular property prediction.