De novo design of ligand binding proteins using large language models alone

Protein design has significantly progressed with the emergence of sequence- and structure-based machine learning models. However, these models have not been as effective in employing reasoned design, which incorporates physicochemical principles and rules based on sequence-structure-function relationships. This study examines the capacity of prevalent large language models (LLMs) like Claude, ChatGPT, and Gemini to generate entirely new proteins that bind metals and lipophilic small molecules without replicating existing sequences.

The analysis reveals that LLMs can accomplish two key tasks: 1) devising protein sequences to adopt a predefined structure and bind the target ligand, and 2) elucidating the reasoning behind the design choices. Following structure prediction and filtering, a limited set of designs (6 to 12 per query) was chosen for experimental validation, resulting in a 25% hit rate for metal binders in a single LLM-based design round and a 25% hit rate in the second round for perfluorooctanoic acid binders.

Crucially, the LLMs furnish comprehensive explanations for the de novo designed sequences, offering a conceptual basis for evaluating the designs. Although the successful designs deviate somewhat from the prompted parameters and the LLM's rationale, these experiments serve as a case study demonstrating the potential of LLMs in making protein design more comprehensible and accessible to users without specialized design expertise.