La-Proteina
4 days ago
- #protein-design
- #bioinformatics
- #generative-models
- La-Proteina introduces a novel partially latent protein representation for atomistic protein design, combining explicit coarse backbone structure with per-residue latent variables for sequence and atomistic details.
- The model achieves state-of-the-art performance in all-atom co-designability, diversity, structural validity, and atomistic motif scaffolding, outperforming previous models.
- La-Proteina can generate co-designable proteins up to 800 residues, demonstrating scalability and robustness where most baselines fail.
- The framework includes an autoencoder and latent diffusion model, with specific checkpoints provided for different tasks and protein lengths.
- Detailed instructions are provided for environment setup, model training, sampling, and evaluation, including conditional and unconditional generation tasks.
- The project is released under the Apache 2.0 license for source code and NVIDIA Open Model License Agreement for model weights.