An advanced LLM-powered article extraction pipeline delivers a database of buffers for protein studies and protein expression conditions

Abstract

Optimizing protein expression and buffer conditions is a major challenge in life-science research and biotechnology. We present an LLM-powered extraction pipeline that systematically converts unstructured scientific literature into a structured, queryable database. Leveraging advanced optical character recognition, custom reranker models, and large language models, our system achieves ~97% extraction accuracy. By processing 100 353 structural biology publications associated with the RCSB Protein Data Bank, we constructed a database containing 947 153 distinct buffer entries and 73 210 expression protocols for 35 690 unique proteins, establishing the largest resource of its kind. The ProBufEx (https://www.probufex.org/) web platform enables rapid exploration of optimal conditions by protein properties or functions. Case studies reveal distinct patterns in chemical usage on proteins with different post-translational modifications, and function-dependent expression strategies for proteins with different enzymatic activities. This resource transforms vast structural biology literature into actionable knowledge, accelerating evidence-based protocol design, reducing trial-and-error, and streamlining experiments in protein science.