Efficient protein structure generation with sparse denoising models

Proteins play diverse roles in all domains of life and are extensively harnessed as biomolecules in biotechnology, with applications spanning from fundamental research to biomedicine. Therefore, there is considerable interest in computationally designing proteins with specified properties. Protein structure generative models provide a means to design protein structures in a controllable manner and have been successfully applied to address various protein design tasks. Such models are paired with protein sequence and structure predictors to produce and select protein sequences for experimental testing. However, current protein structure generators face important limitations for proteins with more than 400 amino acids and require retraining for protein design tasks unseen during model training. To address the first issue, we introduce salad, a family of sparse all-atom denoising models for protein structure generation. Our models are smaller and faster than the state of the art and matching or improving design quality, successfully generating structures for protein lengths up to 1,000 amino acids. To address the second issue, we combine salad with structure editing, a sampling strategy for expanding the capability of protein denoising models to unseen tasks. We apply our approach to a variety of challenging protein design tasks, from generating protein scaffolds containing functional protein motifs (motif scaffolding) to designing proteins capable of adopting multiple distinct folds under different conditions (multi-state protein design), demonstrating the flexibility of salad and structure editing.