Fixed-stage LLM agent with literature access designs protein binders from a single natural-language prompt
- LPT chains fixed LLM stages, each running a bounded tool-using loop, so binder design starts from one plain-language request while keeping token use and compute low.
- The stages can query a curated database of about 14,000 digested papers; a model with access answered most held-out questions drawn from it, whereas one without access answered none.
- Three end-to-end campaigns targeted CALCRL/RAMP1, PD-1/PD-L1 and PVR-TIGIT. Support is in silico only, from agreement across several folding models, with no experimental testing reported.
💡 Points Based on the abstract
The novelty in LPT is a bounded orchestration layer, not a new binder-design method, and its literature test only shows retrieval beats no retrieval, a low bar. Folding-model agreement is weak evidence without wet-lab data, since those models share training biases. On my project, hit rates against plain RFdiffusion pipelines would settle it.
LPT: Orchestrating bounded agentic loops for automated protein binder design ↗