NEWS · RESEARCH · #151
Fraglingo: Attachment-aware autoregressive fragment generation for molecular design
This arXiv preprint introduces Fraglingo, an autoregressive fragment-based molecular generator that jointly models fragment identity and attachment in a continuous latent space. It uses a wildcard-anchored readout to represent the growing molecule from the active attachment site, predicts an attachment-aware fragment embedding, retrieves the next fragment by latent-space nearest-neighbor search, and reports stronger joint property control and competitive validity/novelty while generalizing to fragment libraries up to 4× larger than seen in training.
KEY POINTS
- This arXiv preprint introduces Fraglingo, an autoregressive fragment-based molecular generator that jointly models fragment identity and attachment in a continuous latent space.
- It uses a wildcard-anchored readout to represent the growing molecule from the active attachment site, predicts an attachment-aware fragment embedding, retrieves the next fragment by latent-space nearest-neighbor search, and reports stronger joint property control and competitive validity/novelty while generalizing to fragment libraries up to 4× larger than seen in training.
- By jointly modeling fragment identity and attachment in a continuous embedding space, Fraglingo aligns generation with chemists' edit-style workflows and enables adding new fragments at inference time without retraining, which can make property-directed molecular design more flexible and scalable.
WHY IT MATTERS
By jointly modeling fragment identity and attachment in a continuous embedding space, Fraglingo aligns generation with chemists' edit-style workflows and enables adding new fragments at inference time without retraining, which can make property-directed molecular design more flexible and scalable.