Data

Sources, filters, and exact composition of our biomedical corpus

Every domain in Robertium is built from open-access biomedical literature, retrieved via OpenAlex API, filtered through PubTator for biomedical relevance, and processed through the same pipeline. This page documents the exact sources and parameters for full reproducibility.

Six therapeutic domains

Each domain is a separate corpus, retrieved using a specific OpenAlex concept ID. We targeted high-impact areas where drug repurposing has clinical urgency.

Domain OpenAlex Concept ID Target size Year filter
Glioblastoma C2776194525 18,000 ≥2018
Epilepsy C2778186239 18,000 ≥2018
ALS (Amyotrophic Lateral Sclerosis) C2780596555 15,000 ≥2015
Alzheimer's disease C502032728 18,000 ≥2018
Pancreatic cancer C2780210213 15,000 ≥2015
Major Depressive Disorder C2780051608 18,000 ≥2018

All concepts were verified via the OpenAlex API to confirm display name and works count before ingest. The exact YAML configuration files are available in the GitHub repository under config/domains/.

Biomedical relevance filter

After ingest, every paper passes through NCBI's PubTator service, which tags biomedical entities (genes, chemicals, diseases). Papers that don't contain at least two recognized biomedical entities are excluded from downstream processing.

Domain Papers ingested L1 passed Pass rate
Glioblastoma 17,831 6,268 35.2%
Epilepsy 17,803 8,629 48.5%
ALS 14,109 5,559 39.4%
Alzheimer's 17,408 11,724 67.3%
Pancreatic cancer 14,648 10,369 70.8%
Depression (MDD) 17,333 4,706 27.2%
Multiple sclerosis 16,881 8,546 50.6%
Type 2 diabetes 17,712 13,275 74.9%
Rheumatoid arthritis 16,210 10,565 65.2%
Inflammatory bowel disease 17,210 10,525 61.2%
Total 167,145 90,166 54.0%

Pass rates vary by domain. Oncology and neurodegenerative literature is densely biomedical (high pass rate). Psychiatric literature like depression has more behavioral/sociological papers without specific molecular entities (lower pass rate). This is expected and documented for transparency.

Structured claim extraction

Each filtered abstract is processed by an instruction-tuned LLM with a structured prompt. The model returns claims as triples: (subject, predicate, object) with entity types, polarity, and confidence.

Domain Filtered papers Claims extracted Avg claims/paper
Glioblastoma 6,268 17,379 2.8
Epilepsy 8,629 25,058 2.9
ALS 5,559 15,447 2.8
Alzheimer's 11,724 35,534 3.0
Pancreatic cancer 10,369 33,457 3.2
Depression (MDD) 4,706 13,303 2.8
Multiple sclerosis 8,546 23,794 2.8
Type 2 diabetes 13,275 42,048 3.2
Rheumatoid arthritis 10,565 31,934 3.0
Inflammatory bowel disease 10,525 32,837 3.1
Total 90,166 270,791 3.0

Variation in claims-per-paper reflects literature density: review papers and detailed mechanistic studies yield more claims than short clinical correspondences. Total: 270,791 structured claims across 10 domains, all linked to source PMIDs.

Knowledge graph composition

All claims are imported into a Kuzu graph database. Each unique entity (after lexical normalization) becomes a node; each claim becomes an edge with the predicate as label.

160,955
total
nodes
270,791
claim
edges
10
domains
in graph
24,285
cross-domain
hypotheses

Cross-domain shared entities — entities mentioned in two or more domains — are the seeds of repurposing hypotheses. A protein appearing in both glioblastoma and epilepsy literature creates a bridge that the ABC model can traverse.

Controlled vocabulary

To ensure consistency, the LLM extracts only predicates and entity types from a fixed vocabulary.

Predicates (21)

inhibits activates induces suppresses upregulates downregulates phosphorylates binds_to increases decreases correlates_with associated_with predicts prevents causes treats improves reverses exacerbates sensitizes_to confers_resistance_to

Entity types (9)

drug gene protein pathway disease phenotype cell_type tissue biomarker

This controlled vocabulary makes claims comparable across domains and amenable to graph queries. Adding new predicates requires careful evaluation — we keep the vocabulary small intentionally.

Open data release plan

All data — claims, knowledge graph, and configuration — will be released under CC-BY-4.0 alongside the first preprint (planned for Q3 2026). Currently:

  • GitHub repository with source code (MIT) and config files: github.com/routewise96/robertium
  • Hypothesis catalog publicly available at /hypotheses — cross-domain candidates with full evidence chains
  • Full claim dataset and knowledge graph dump will be released as a Zenodo deposit with persistent DOI on preprint publication
  • API access for programmatic queries — planned for late 2026

Data updates

This corpus is a snapshot. The biomedical literature publishes ~4,000 papers per day, and our domains will be re-ingested periodically:

  • Quarterly re-ingest of all 6 domains to capture new publications
  • Monthly cross-domain hypothesis recomputation to incorporate new data
  • New domains added when there is a clear scientific case (request via daniel@robertium.com)

The knowledge graph version will be tagged in the repository at each re-ingest. Old versions remain accessible for reproducibility.

Read the methodology → Browse hypotheses → View on GitHub