ChemFOnt LLM Pipeline

ChemFOnt is a hierarchical, OWL-compatible ontology that provides standardized, machine-readable terms for describing the biological, environmental, and industrial functions of chemicals. It comprises four functional aspects, 12 functional super-categories, and more than 173,000 functional terms, covering biological targets, health effects, exposure routes, biological locations, sources, biological roles and processes, environmental roles and processes, industrial roles and processes, and organoleptic properties.

CF3P combines ChemFOnt with a prompt-tuned large language model (gemma-4-31B) to extract structured chemical–relation–entity triples from published scientific text. Rather than treating co-occurrence of a chemical and a concept as evidence of association, CF3P records only relationships that are explicitly stated in the source text, along with the type of relationship reported. For example, chemical–target relationships distinguish agonists, antagonists, inhibitors, activators, substrates, and binders, while health-related annotations distinguish general associations, biomarkers, therapeutic uses, and changes in chemical abundance in human biofluids.

1. PubMed Search & Document Retrieval — Publications mentioning SynCED chemicals were identified via PubMed searches and downloaded as full-text records. Over 100,000 publications were processed for SynCED.
2. Named Entity Recognition — Chemical names and associated biological, environmental, and industrial entities are identified in the text.
3. Relation Extraction — The prompt-tuned gemma-4-31B model extracts structured chemical–relation–entity triples, recording the type of relationship explicitly stated in the source text along with supporting text spans and source identifiers.
4. ChemFOnt Mapping — Extracted entities and relationships are mapped to ChemFOnt concepts, linking specific annotations to broader ontological categories.
5. Normalization & Quality Control — Extracted records undergo normalization, ontology mapping, semantic validation, and deduplication. Each record is assigned a quality outcome: PASS, PASS_UNRESOLVED_ID, REVIEW, or REJECT.
6. Database Integration — Accepted records are loaded into SynCED with provenance metadata, supporting text, and source identifiers preserved.

CF3P was evaluated against a manually annotated corpus of more than 213 documents across multiple document types and subject areas. Domain experts annotated chemical–relation–entity triples independently, and CF3P output was compared against these gold-standard annotations.

Evaluation was performed on held-out documents not used during prompt tuning. Performance may vary across document types, chemical classes, and relation categories.

In addition to structured triples, the knowledge extracted by CF3P was used to generate comprehensive compound descriptions for each chemical in SynCED. These descriptions were assembled using the gemma-4-31B LLM from the extracted knowledge triples and their supporting references. Compound descriptions in SynCED average 256 words in length and contain an average of 7.3 references each.

All CF3P-derived content in SynCED is accompanied by a disclaimer indicating its computational origin. Each annotation includes a link to the supporting source publication so that users can verify the underlying evidence.

ChemFOnt was published in 2023:

Wishart DS, Girod S, Peters H, et al. ChemFOnt: the chemical functional ontology resource. Nucleic Acids Res. 2023;51(D1):D1220–D1229. PMID: 36305829.

A manuscript describing the updated ChemFOnt LLM pipeline is currently in preparation.