Bayesian Flow Network Framework for Chemistry Tasks.

Area of Science:

• Computational chemistry
• Machine learning for drug discovery

Background:

• Developing advanced language models for chemical tasks is crucial for accelerating molecular design and discovery.
• Bayesian flow networks offer a probabilistic approach to generative modeling, suitable for discrete data like molecular structures.

Purpose of the Study:

• Introduce ChemBFN, a language model leveraging Bayesian flow networks for chemistry.
• Enhance molecular generation quality and diversity using a novel accuracy schedule and classifier-free guidance.
• Demonstrate the model's versatility through fine-tuning for regression and classification tasks.

Main Methods:

• Utilized Bayesian flow networks for discrete data processing in chemistry.
• Implemented a new accuracy schedule to minimize reconstruction loss and improve sampling.
• Adapted classifier-free guidance for conditional molecular generation.

Main Results:

• ChemBFN demonstrated high-quality molecule generation with satisfied diversity, even with fewer sampling steps.
• The model achieved state-of-the-art performance when fine-tuned on regression and classification tasks.
• Reduced reconstruction loss significantly through the proposed accuracy schedule.

Conclusions:

• ChemBFN is an effective all-in-one model for diverse chemistry tasks, including generation, classification, and regression.
• The proposed methods enhance generative model performance and applicability in computational chemistry.
• Open-sourcing the model facilitates further research and development in AI-driven chemistry.