Many Molecular Properties from One Kernel in Chemical Space

Area of Science:

• Computational chemistry
• Machine learning
• cheminformatics

Background:

• Developing accurate machine learning models for molecular properties is crucial in chemistry.
• Existing methods often require property-specific models or extensive feature engineering.

Purpose of the Study:

• Introduce property-independent kernels for machine learning models.
• Enable the generation of models for multiple molecular properties using a unified approach.

Main Methods:

• Developed kernels that encode molecular structures and chemical space similarity.
• Utilized these kernels to generate machine learning models for various molecular properties.
• Trained and tested models on a dataset of 112,000 organic molecules.

Main Results:

• Demonstrated the capability of property-independent kernels for modeling diverse properties including energy, enthalpy, and electronic properties.
• Showcased instantaneous model generation and systematic improvement with added data.
• Successfully applied kernels to model internal energy, enthalpy, free energy, heat capacity, polarizability, electronic spread, zero-point vibrational energy, frontier orbital energies, HOMO-LUMO gap, and vibrational wavenumber.

Conclusions:

• Property-independent kernels offer a versatile and efficient approach for machine learning in molecular property prediction.
• The developed kernels facilitate rapid model creation and continuous refinement.
• This methodology holds promise for broader applications in cheminformatics and materials science.