Implicitly perturbed Hamiltonian as a class of versatile and general-purpose molecular representations for machine...

Area of Science:

• Molecular sciences
• Computational chemistry
• Machine learning applications

Background:

• Machine learning (ML) is increasingly used to solve complex problems in science.
• Developing effective ML models requires translating molecular structures into quantitative representations.
• Current molecular representations are often suboptimal for challenging molecular science problems.

Purpose of the Study:

• Introduce Implicitly Perturbed Hamiltonian (ImPerHam) as a novel class of molecular representations.
• Enhance the efficiency of ML models for complex molecular science challenges.
• Provide versatile representations applicable to diverse computational chemistry tasks.

Main Methods:

• ImPerHam representations are defined using energy attributes of the molecular Hamiltonian.
• These representations incorporate implicit perturbations from continuum solvation models.
• ML models were developed and evaluated using ImPerHam for specific chemical prediction tasks.

Main Results:

• ImPerHam-based ML models demonstrated outstanding performance.
• Accurate prediction of CYP450 enzyme inhibition.
• High-precision, transferable evaluation of non-covalent interaction energies.
• Accurate reproduction of solvation free energies for extensive benchmark sets.

Conclusions:

• ImPerHam representations offer a versatile and efficient approach for ML in molecular sciences.
• This method addresses limitations of existing molecular representations for complex problems.
• ImPerHam shows significant potential for advancing computational chemistry and drug discovery.