NVIDIA and Baker Lab Release RFD4-Proteína for Biomolecular Design

NVIDIA and Baker Lab have released RFD4-Proteína, a flow-matching model designed for biomolecular design tasks. This model can generate protein sequences and all-atom structures simultaneously, conditioned on specific molecular contexts such as binding targets or enzyme motifs. It features a 147-class residue vocabulary and a 32-dimensional latent space, resulting from a LoRA fine-tune of the pretrained base model. The model is available via the Hugging Face Diffusers library, with weights stored in FP32 and recommended to run in BF16 on a CUDA GPU. According to the model card, RFD4-Proteína supports various tasks, including protein binder design and motif scaffolding, with the ability to handle up to 1024-token designs on a 16 GB GPU. The model and weights are licensed under the BSD 3-Clause, and users can install the necessary dependencies from GitHub until the next Diffusers release.