Introducing SAFE, Sequential Attachment-based Fragment Embedding, a novel molecular line notation that redefines molecular design tasks as simple autoregressive sequence generation. 🧵

SAFE represents molecules as a sequence of interconnected fragments where each fragment is also a valid molecule. This sequential block structure of SAFE preserves the integrity of molecular scaffolds and fragments while offering flexibility for de novo molecular design.

Using 1.1B SAFE representations, we created SAFE-GPT: an 87.3M parameter GPT-like generative model. This is our first step towards a universal foundation model for generative chemistry. SAFE-GPT streamlines fragment-constrained generative tasks such as:

We also propose new evaluation standards for generative chemistry models using fragments and scaffolds from 10 existing drugs, mirroring the challenges of real-world drug discovery scenarios (such as scaffold decoration and linker design).

SAFE easily integrates into existing workflows. SAFE strings are SMILES and are thus compatible with existing chemoinformatic toolkits like @datamol_io and @RDKit_org.

In future work, we intend to scale SAFE-GPT to larger models and datasets, laying the groundwork for a new generation of foundation models in drug discovery. Read the paper for more details: Get started with SAFE-GPT today:

Looks great! Is it possible to start with a relatively large molecule, like a peptide, and use it for scaffold or motif extension?

Awesome work!