Designing Protein Binders Using the Generative Model Proteina-Complexa

NVIDIA has released Proteina-Complexa, a generative model for designing de novo protein binders and enzymes. Built on the La-Proteina model, it uses a partially latent flow-matching framework to co-design fully atomistic protein structures and amino acid sequences simultaneously. The model was trained on over 1 million curated structures and incorporates inference-time compute scaling with search algorithms like Beam Search and Best-of-N to iteratively refine candidates. Experimental validation involved testing ~1 million binder candidates against 133 distinct protein targets, achieving nano- and picomolar affinities on several targets including Activin Receptor Type-2A. The model also demonstrated success in designing binders for small molecule targets and carbohydrate surfaces on red blood cells. The post includes a step-by-step CLI guide for setting up and running the binder design pipeline. Source code, model checkpoints, and datasets are released under open licenses.