Ontology‑driven LLM Platform for Assayability Evidence from Literature

An ontology‑driven, LLM‑enabled platform for extracting and structuring assayability evidence from biomedical literature

High attrition rates in drug discovery continue to place increasing emphasis on robust early target validation. However, critical assayability evidence, detailing which assays, models and cellular systems can generate decision-enabling data, is often fragmented across biomedical literature and requires time-consuming manual review.

In this poster, we describe an LLM-enabled, ontology-driven platform designed to automatically extract, normalise and prioritize assayability evidence from large-scale literature sources. By transforming unstructured publications into a structured, target-centric resource, the platform helps teams assess target tractability faster and with greater confidence.

What the poster covers

• Automated literature mining at scale

Use of transformer-based language model to extract genes, assays, in vivo models and cell lines from millions of abstracts

• Fine-tuned biomedical NER

A staged training strategy based on Bioformer-16L, delivering robust entity recognition performance for experimentally relevant concepts that are difficult to capture with rule-based approaches

• Ontology-driven normalisation

Introduction of SygOntology, an extensible, purpose-built ontology that integrated public resources with curated terms to resolve ambiguous biomedical language and inconsistent nomenclature

• Capability-aware evidence ranking

Target-assay relationships are ranked using literature strength, recency and frequency, combined with internal experimental feasibility, surfacing the most actionable experiments rather than the most commonly reported ones

• A modular, extensible workflow

A scalable architecture that supports ongoing refinement, human oversight, and future expansion to additional entity types and full-text analysis.