Adaptive language model training for molecular design.

Area of Science:

• Computational chemistry
• Artificial intelligence in drug discovery
• Machine learning for molecular design

Background:

• Vast chemical space requires computational methods for efficient molecular design.
• Genetic algorithms and masked language models (MLMs) are used for automated molecule generation and mutation.
• MLMs learn chemical sequences from large libraries for predicting molecular rearrangements.

Purpose of the Study:

• To investigate adaptive language models for improved molecule generation in drug discovery.
• To compare fixed (pre-trained) and adaptive (re-trained) MLM strategies for molecular optimization.
• To assess the impact of adaptive training on optimizing drug-likeness, synthesizability, and binding affinity.

Main Methods:

• Developed and compared two language model strategies: fixed and adaptive.
• The fixed strategy uses a pre-trained MLM for generating molecular mutations.
• The adaptive strategy retrains the MLM on newly generated molecules with desired properties.

Main Results:

• The adaptive strategy enables language models to better fit the molecular distribution of the population.
• Adaptive training significantly improved fitness optimization compared to the fixed pre-trained model.
• Demonstrated improved optimization of drug-likeness, synthesizability, and predicted protein binding affinity.

Conclusions:

• Adaptive language models offer significant advantages for molecular design and optimization tasks.
• A hybrid approach using fixed strategy initially, followed by adaptive strategy, is recommended for enhanced fitness optimization.
• This work empowers the application of language models in accelerating drug discovery pipelines.