Large property models: a new generative machine-learning formulation for molecules.

Area of Science:

• Computational chemistry
• Machine learning
• Drug discovery

Background:

• Generative models for inverse molecular design are hyped but lack significant gains over expert intuition.
• A key challenge is poor accuracy in data-scarce regimes, typical for discovering rare molecular outliers.
• Existing models struggle to learn accurate property-to-structure mappings from limited datasets (tens to hundreds of samples).

Purpose of the Study:

• To test the hypothesis that property-to-structure mapping becomes unique with sufficient training properties.
• To explore if data-scarce properties can be predicted using more accessible molecular properties.
• To investigate if generative models trained on multiple properties exhibit an accuracy phase transition.

Main Methods:

• Developed novel "large property models" (LPMs), the first transformers for the property-to-molecular-graph task.
• Supplemented model training with abundant, basic chemical property data.
• Utilized transformer architectures for property-to-structure mapping.

Main Results:

• Demonstrated that incorporating multiple properties enhances predictive accuracy in data-scarce scenarios.
• Observed an accuracy phase transition in LPMs as model size and property data increased.
• Showcased the potential for LPMs to discover prized outliers in molecular design.

Conclusions:

• The large property model paradigm offers a promising approach to overcome data limitations in inverse molecular design.
• Training on diverse and abundant chemical properties is crucial for accurate and generalizable property-to-structure mappings.
• This work lays the foundation for more effective AI-driven discovery of molecules with targeted properties.