Probabilistic generative transformer language models for generative design of molecules.

Area of Science:

• Computational Chemistry and Cheminformatics
• Artificial Intelligence in Drug Discovery
• Machine Learning for Molecular Design

Background:

• Self-supervised neural language models are increasingly used for generative design in chemistry and biology.
• Existing deep learning models for molecular design often require large datasets and lack interpretability.
• Challenges include understanding the design logic and improving data efficiency in generative models.

Purpose of the Study:

• To develop a novel probabilistic neural network model for the generative design of molecules.
• To create a model that is interpretable, data-efficient, and capable of high-quality molecule generation.
• To enable guided molecule modification with explanations based on learned chemical principles.

Main Methods:

• Developed the Generative Molecular Transformer (GMTransformer), a probabilistic neural network.
• Adapted a blank-filling language model, originally for text processing, to learn 'molecular grammars'.
• Evaluated model performance on the MOSES datasets for novelty and scaffold diversity.

Main Results:

• The GMTransformer achieved high novelty and scaffold (Scaf) metrics compared to existing baselines.
• Demonstrated advantages in data efficiency and interpretability compared to black-box models.
• Probabilistic generation steps allow for guided molecule modification with explanations.

Conclusions:

• The GMTransformer offers a powerful and interpretable approach to generative molecular design.
• The model effectively learns underlying molecular grammar, enabling high-quality and novel molecule generation.
• Its ability to provide explanations for modifications opens new avenues for molecular tinkering and discovery.