ProteinCrow: A Language Model Agent That Can Design Proteins

Manvitha Ponnapati, Sam Cox, Cade W Gordon, Michael J Hammerling, Siddharth Narayanan, Jon M Laurent, James D. Braza, Michaela M. Hinks, Michael D Skarlinski, Samuel G Rodriques, Andrew White

Keywords: LLM, Agent, Protein Design, Tool

TL;DR: A Language Model Agent That Can Design Proteins

Abstract: Recent breakthroughs in deep learning have revolutionized protein structure and sequence modeling, enabling the design of proteins with novel functions through tools like AlphaFold, RFdiffusion, BindCraft and ProteinMPNN. Successful in silico protein engineering pipelines integrate multiple of these specialized models, incorporating biochemical insights, and iteratively optimizing sequences. Here, we present ProteinCrow, an agentic LLM-based protein design assistant that consolidates multimodal information, including structural data, deep learning models,scientific literature and biochemical data expressed in natural language, to automate protein design tasks by using 36 expert-curated tools. We evaluated its performance in designing binder libraries tailored to specific secondary‐structure motifs; redesigning protein backbones to improve stability and optimizing binders to eliminate predicted MHC Class I epitopes.

Submission Number: 54