Technical Background
FoodAtlas is an evidence-based knowledge graph linking foods, the chemicals they contain, the diseases those chemicals correlate with, and the bioactivities they express. Every edge in the graph is traceable to a public source or a peer-reviewed publication.
The graph is built from two pipelines: an information extraction pipeline that pulls food–chemical relations from PubMed / PMC literature, and a knowledge graph construction pipeline that ingests public databases, resolves entities, and stitches everything into the released graph. For the full study, see the npj Science of Food paper.
What's in the graph
A node is a Food , Chemical , Disease , or Bioactivity . An edge informs about the relationship between two nodes.
Every edge carries an attestation — the supporting evidence, source, and any measurement metadata (concentration values, assay outcomes, Hill-curve fits).
| Code | Relation | Between | Description |
|---|---|---|---|
| r1 | CONTAINS | Food → Chemical | A food contains a chemical at a measured concentration. Values normalised to mg / 100g where the source unit allows. |
| r2 | IS_A | Any → Any (same type) | Ontology hierarchy. Ties into FoodOn, ChEBI, MeSH, and the bioactivity hierarchy so children inherit parent context. |
| r3 | WORSENS | Chemical → Disease | Peer-reviewed evidence that the chemical worsens the disease's health outcomes or increases risk of onset. |
| r4 | IMPROVES | Chemical → Disease | Peer-reviewed evidence that the chemical improves the disease's health outcomes or reduces risk of onset. |
| r5 | EXHIBITS | Food → Bioactivity | A food exhibits a bioactivity — either from direct assay evidence or from a model inference over its composition. |
| r6 | MEASURED | Chemical → Bioactivity | A chemical was measured against a bioactivity in an assay. Carries the raw measurement + a Hill-curve fit when the assay reports dose–response. |
Where the data comes from
How it's built
Our pipeline uses state-of-the-art AI models to extract and quantify food connections. The two major steps are (a) knowledge extraction, i.e., converting literature into food–chemical relations, and (b) knowledge graph construction, which adds meta-information and new information to our knowledge base.
Step 1 — Information Extraction
From literature to structured triplets.
Refresh a local BioC-PMC corpus and rebuild the PMC-ID index.
Query PubMed for food terms; retrieve and fuzzy-match candidate sentences.
Binary classification with a fine-tuned BioBERT model — only sentences above a 0.99 confidence threshold pass through.
The remaining sentences are batched through OpenAI's Batch API using gpt-5.5 and structured extraction prompts.
Step 2 — Knowledge Graph Construction
From sources + triplets to the released graph.
Parse each source into standardised parquet files (nodes / edges / cross-references).
Three-pass entity resolution → stable `foodatlas_id`s that survive across releases. Ambiguities are recorded on attestations, not silently collapsed.
Build typed edges (r1–r6) from source data. Duplicates merge; ambiguous resolutions explode into candidates for later review.
Concentration parser normalises the information extraction pipeline's output to mg/100g. Chemical + food names resolved through the entity registry.
Add derived metadata: chemical/food classifications, flavor descriptors, common names, display grouping.
Per-attestation plausibility signals from a Gemini 3.1 Flash-Lite LLM judge. Emits a 0–1 score and a short justification per triplet.
Diagnostics on the finished graph — orphan detection, unclassified entities, per-source coverage.
Bioactivity in detail
A bioactivity is a biological effect a chemical or food can exhibit — antibacterial, antioxidant, hepatotoxic, and so on. The graph organises 21 bioactivity concepts in a small hierarchy so queries roll up naturally (e.g. an antibacterial hit is also an antimicrobial hit).
Chemical → bioactivity evidence ( r6 ) comes in two flavors:
- Experimental — assay measurements pulled from PubChem and ChEMBL, with outcome, potency (value + unit), and a four-parameter Hill-curve fit when the assay reports a dose response.
- Predicted — random-forest model inferences (e.g.
RF_antioxidant_v1) that score a chemical against a bioactivity based on structural features.
Food → bioactivity evidence ( r5 ) is either directly measured against the food or inferred from the bioactivities of its composed chemicals.
Trust signals
Every attestation in the graph carries a per-triplet trust signal. A Gemini 3.1 Flash-Lite LLM judge scores each (food, chemical, concentration) triplet on world-knowledge plausibility from 0 to 1, alongside a short justification. Low-trust rows aren't hidden — they're surfaced on the frontend so a curator can review them without losing the underlying evidence.
Trust signals are versioned and stored separately from the main attestations, so a rebuild of the graph doesn't lose the per-row judgements that were made against an earlier snapshot.
About AIFS
The AI Institute for Next Generation Food Systems, or AIFS aims to meet growing demands in our food supply by increasing efficiencies using Al and bioinformatics spanning the entire system–from growing crops through consumption. We are dedicated to creating AI applications for a healthier, more sustainable planet from farm to fork.
Connect with us
Subscribe to our newsletter to stay up-to-date on AIFS events, industry news, and AI research.