RITA: a Study on Scaling Up Generative Protein Sequence Models
This work addresses the need for scalable protein design tools for researchers, though it is incremental as it focuses on scaling existing methods.
The authors tackled the problem of scaling generative models for protein sequences by introducing RITA, a suite of autoregressive models with up to 1.2 billion parameters trained on 280 million sequences, and found that increased model size improves performance in tasks like next amino acid prediction and enzyme function prediction.
In this work we introduce RITA: a suite of autoregressive generative models for protein sequences, with up to 1.2 billion parameters, trained on over 280 million protein sequences belonging to the UniRef-100 database. Such generative models hold the promise of greatly accelerating protein design. We conduct the first systematic study of how capabilities evolve with model size for autoregressive transformers in the protein domain: we evaluate RITA models in next amino acid prediction, zero-shot fitness, and enzyme function prediction, showing benefits from increased scale. We release the RITA models openly, to the benefit of the research community.