Sampling out-of-distribution chemical spaces via Bayesian flow.

Area of Science:

• Computational chemistry
• Drug discovery
• Machine learning

Background:

• Generating novel molecules with improved properties (out-of-distribution generation) is crucial for de novo drug design.
• Distribution learning models, like diffusion models, struggle with out-of-distribution generation as they prioritize fitting training data distributions.

Purpose of the Study:

• To demonstrate the capability of Bayesian flow networks, specifically the ChemBFN model, in generating high-quality out-of-distribution samples.
• To enhance ChemBFN's performance and sampling speed for generative tasks.

Main Methods:

• Utilized a Bayesian flow network (ChemBFN) model.
• Integrated a reinforcement learning strategy with ChemBFN.
• Employed a controllable ordinary differential equation solver-like generation process.
• Introduced a semi-autoregressive strategy during training and inference.

Main Results:

• ChemBFN demonstrated intrinsic capability for high-quality out-of-distribution sample generation.
• The integrated strategies accelerated sampling processes.
• The semi-autoregressive approach enhanced model performance, surpassing state-of-the-art models.
• ChemBFN showed outstanding out-of-distribution performance on small molecule and protein generation tasks.

Conclusions:

• ChemBFN effectively generates molecules outside the training data distribution without complex modifications.
• The model proves valuable for exploring novel chemical spaces and accelerating drug design and materials discovery.