Toward trustworthy LLM-based scientific information extraction: An ontology-guided, validation-aware workflow and a single-domain case study

This is the authors' abstract. We don't add key points for this paper.

This study proposes an ontology-guided, validation-aware LLM workflow for scientific information extraction and evaluates it in a single-domain case study of yeast metabolic engineering. The workflow integrates section-aware evidence targeting, context expansion, adversarial “Seed-of-Doubt” validation, and ontology-guided semantic checking. We instantiate the approach by constructing a unified schema of 37 structured features and a hybrid featurization strategy that bridges semantic interoperability with downstream machine-learning readiness. From 220 eligible research articles, the workflow extracted 2312 experimental records. Evaluation against 183 expert-curated records from 15 articles yielded an F 1 of 0.96 (95% article-clustered bootstrap CI [0.93, 0.99]), compared with 0.80 for zero-shot full-text extraction and 0.89 for self-refinement. On a shared 21-field subset, the workflow achieved F 1 = 0.95 versus 0.83 for SPIRES/OntoGPT. The rates of strain–data mismatch and genotype hallucination decreased from 22.0% to 3.0% and from 15.0% to 1.8%, respectively; numerical extraction MAE decreased from 2.45 to 0.12 g/L. The records preserved the reference correlation structure ( r = 0.94 ) and supported downstream modeling ( R 2 = 0.762 under article-grouped cross-validation; 0.805 under record-level splitting). In a separate comparison on the 183-record reference set, training on curated rather than extracted records improved R 2 by 0.027 (95% CI [−0.011, 0.064]); this interval does not establish equivalence. Ablation supports an incremental contribution from ontology-guided validation beyond structural constraints. The results support improved extraction reliability and analytical reuse within the studied domain.