Reproduce “Molecule Generation with Fragment Retrieval Augmentation” on OpenResearch

About this paper

Seul Lee, Karsten Kreis, Srimukh Prasad Veccham, Meng Liu, Danny Reidenbach, Saee Paliwal, Arash Vahdat, Weili Nie

Researchers from KAIST and NVIDIA developed Fragment Retrieval-Augmented Generation (f-RAG), a framework that enhances molecule generation by applying retrieval-augmented techniques at the fragment level. The system generates diverse, novel, and synthesizable drug candidates by better balancing the exploration of new chemical space with the exploitation of known fragments, outperforming state-of-the-art baselines on key benchmarks.

Key insights

• Applying Retrieval-Augmented Generation (RAG) at a fine-grained, fragment-level provides more nuanced and effective guidance for molecular generative models, enhancing exploration beyond predefined fragments.
• Dynamically updating the fragment vocabulary with newly generated high-quality fragments, alongside post-hoc genetic modification, is crucial for continuous exploration and discovery of novel chemical space.
• A self-supervised, target property-agnostic training objective for the fragment injection module allows the framework to be efficiently adapted to various optimization tasks without re-training the core generative model.

Summary by alphaXiv.