Forge is a sequence-based generative model for binder design built on latent flow matching. We frame binder generation analogously to text-to-image generation, treating the problem as a target-sequence-to-binder-sequence conditional generation task. But instead of operating directly in sequence space, Forge performs generation in the latent space of Raygun, an autoencoder-based model whose embeddings capture semantic and functional relationships. Using a standard flow-matching formulation with classifier-free guidance, we generated binder sequences by conditioning on the NiV G C-terminus sequence.

TRAINING DATA Rather than relying on protein complexes from the PDB (which reflect a small and structurally biased subset of known interactions) we train Forge on a subset of the STRING protein-interaction database (~10 million interacting protein pairs). The scale and diversity of these interactions (includes interactions involving disordered or structurally uncharacterized proteins) allow Forge to learn from a broader distribution of natural binding "strategies". By using sequence data, our goal is to recover the diversity and flexibility of nature’s binders, rather than being constrained to the well-explored solved structure space.