LPT: Orchestrating bounded agentic loops for automated protein binder design

LPT enables de novo protein binder design from one natural-language prompt, using bounded LLM agent loops and a curated literature database to achieve 41/48 correct answers.

• Why it matters: Specialized, attributable scientific knowledge is hard for general LLMs to access, and unconstrained agent workflows can waste tokens. LPT targets both knowledge recall and efficient, fixed tool use.
• What they did: LPT orchestrates fixed LLM stages, each with bounded agentic loops, custom prompts, and controlled tool access to query a literature database of ~14000 digested papers. With database access, an LLM answered 41/48 test questions correctly versus none without access.
• The result: In three end-to-end campaigns, LPT autonomously selected targets, structures, and hotspot residues and designed binders to disrupt CALCRL/RAMP1, PD-1/PD-L1, and PVR-TIGIT interactions. Cross-validation with multiple folding models indicates a high in silico success rate with minimal token consumption and low computational costs.

AI summary of the abstract · may contain errors · not medical advice · read the paper.