LeFlur: A Biomolecular Design Model with Latent Structure Tokens

Protein design pipelines today are fragmented, separating backbone generation, sequence design, and structure prediction into bespoke models. We present LeFlur, a unified discrete-token model that integrates these tasks into a single standard text transformer. We offload Cartesian-space modeling to LatentGenerator, a Vision Transformer autoencoder that discretizes arbitrary 3D structures (proteins and small molecules alike) into a compact, shared-vocabulary token stream (with modality-specific quantizers $\mathcal{Q}^{(p)}$ and $\mathcal{Q}^{(\ell)}$ feeding a single combined transformer), using a Simple Linear Quantizer (SLQ), an application of Gumbel-Softmax categorical bottlenecks to per-residue 3D-coordinate latents,as a minimal-vocabulary alternative to Finite Scalar Quantization. On top of this representation, LeFlur supports both protein-only and ligand-conditioned design, a first for structure-token models. Because LeFlur is trained as a masked any-order model, it admits a sequence/structure/joint pseudo-likelihood (PLL) that is closely approximated by a single masked forward pass and made unbiased with $K$ stratified Monte-Carlo draws. This PLL correlates strongly with downstream designability and structure-prediction accuracy and serves as a best-of-$N$ ranker that re-uses the model itself. We also introduce Self-Reflection, an inference-time refinement loop that iteratively redesigns outputs against an internal forward-fold consistency check, lifting designability without any retraining or external scorer. Despite its architectural simplicity, LeFlur is competitive with specialized baselines across folding and generation tasks.