Improving Fidelity and Diversity in Chemical Language Transformers for Inverse Molecular Design.

Area of Science:

• Computational chemistry
• Machine learning in chemistry
• Molecular modeling

Background:

• Exploring vast chemical spaces for functional molecules is challenging.
• Chemical language models (CLMs) show promise but often generate invalid or off-spec molecules.
• Inverse molecular design requires efficient methods to generate molecules with desired properties.

Purpose of the Study:

• To develop a novel CLM-based framework for efficient and reliable inverse molecular design.
• To optimize latent representations for target properties and ensure molecular validity.
• To address challenges of ill-posed inverse design problems and decoder-induced latent-space drift.

Main Methods:

• Developed a CLM-based inverse design framework optimizing latent representations.
• Introduced a round-trip fidelity metric to diagnose and mitigate latent-space drift.
• Implemented post-decoding re-ranking and minimal-edit repair for invalid structures.
• Targeted surfactant critical micelle concentration (CMC) for demonstration.

Main Results:

• The framework generated a high proportion of valid and diverse molecules (∼90%).
• Achieved a target property error close to 1% for critical micelle concentration (CMC).
• Interpretability analysis confirmed adherence to physical design rules.

Conclusions:

• The proposed framework offers an efficient and broadly applicable solution for inverse molecular design.
• It successfully generates novel functional molecules with desired properties and validity.
• The approach provides physical insights crucial for molecular design.