LLM-Assisted versus Agentic Approaches to De Novo Minibinder Design for a KRAS G12D Neoantigen

We compare two modes of large language model (LLM) integration into de novo protein minibinder design: (1) human designs with LLM in the loop, where a researcher makes every scientific decision while using an LLM to accelerate code generation and analysis; and (2) LLM-agent designs with human in the loop, where the researcher specifies only the therapeutic target and the agent handles pipeline construction and in silico evaluation autonomously. Both target the KRAS G12D neoantigen peptide VVGADGVGK on HLA-A*11:01 starting with an existing framework. Across two rounds of design screens, the human-led campaign produced 15 high-confidence hits and the LLM agent presented top 41 hits. Evaluated by our new multi-layered evaluation suite of structure/sequence quality, Rosetta-based energy calculations, and docking verification, we conclude that the agent-led campaign is useful for exploring new design ideas and running large parallel campaigns, but domain expertise remained essential for correct binding orientation, calibrated specificity thresholds, and viable candidate selection.