Toward De Novo Protein Design from Natural Language

Abstract

De novo protein design (DNPD) aims to create new protein sequences from scratch, without relying on existing protein templates. However, current deep learning-based DNPD approaches are often limited by their focus on specific or narrowly defined protein designs, restricting broader exploration and the discovery of diverse, functional proteins. To address this issue, we introduce Pinal, a probabilistic sampling method that generates protein sequences using rich natural language as guidance. Unlike end-to-end text-to-sequence generation approaches, we employ a two-stage generative process. Initially, we generate structures based on given language instructions, followed by designing sequences conditioned on both the structure and the language. This approach facilitates searching within the smaller structure space rather than the vast sequence space. Experiments demonstrate that Pinal outperforms existing models, including the concurrent work ESM3, and can generalize to novel protein structures outside the training distribution when provided with appropriate instructions. This work aims to aid the biological community by advancing the design of novel proteins, and our code will be made publicly available soon.