Methodology

The pipeline runs end-to-end across ten therapeutic domains. Each stage is independently reproducible: same input, same output. Numbers below reflect the current corpus.

Each abstract is processed by an LLM with a controlled vocabulary of 18 predicates (inhibits, downregulates, associated_with, sensitizes_to, etc.) and 10 entity types (drug, gene, protein, pathway, phenotype, disease, biomarker, organism, cell_type, technique). The model returns JSON with strict schema validation.

{
  "subject": "verteporfin",
  "subject_type": "drug",
  "predicate": "downregulates",
  "object": "EGFR",
  "object_type": "gene",
  "context": "EGFR-mutant glioblastoma",
  "polarity": "positive",
  "confidence": 0.85
}

Across the 167,145 papers in the current corpus, we extract an average of 3.0 claims per abstract that passes the L1 biomedical-relevance filter. Total: 270,791 structured claims across ten therapeutic domains. All claims are linked to source PMIDs for verification.

To discover drug repurposing candidates, we apply Don R. Swanson's literature-based discovery (LBD) model. A drug A may be repurposed for disease C if there exists a mediator B such that A affects B and B is associated with C, but A and C have no direct co-mention in the literature.

Robertium currently surfaces hypotheses where:

• Direct A→C evidence is zero or minimal
• Both A→B and B→C have at least one supporting peer-reviewed paper
• Mediator B is a meaningful biological entity (gene, protein, pathway)

Each hypothesis is scored using a logarithmic combination of A→B and B→C evidence counts, with boosts for multi-evidence chains.

Robertium enforces a strict novelty constraint on the hypotheses it surfaces. Any candidate where the drug → outcome connection already appears as a direct claim in more than one paper is filtered out (max_direct_evidence = 1, see src/robertium/graph/repurposing.py:192). This is not a parameter to tune — it is the Swanson 1986 prerequisite for literature-based hidden discovery.

The catalog at /hypotheses therefore explicitly excludes the most well-known drug repurposings: ketamine → treatment-resistant depression, metformin → Alzheimer's disease, valproate → bipolar disorder, and others whose drug → outcome link has accumulated direct citation chains over decades of clinical research. These cases were filtered out at construction time.

What you see instead are connections that the published literature does not yet make explicit through direct citation — drug → mediator → outcome chains where the two halves live in separate literature pools that rarely cite each other. This is Robertium's value proposition: novelty precision, not known-repurposing reproduction. See /benchmark for empirical validation of this framing against 50 verified historical repurposings.

Each outreach-quality hypothesis is enriched with two external signals so a researcher can immediately tell what kind of candidate they are looking at.

• Clinical trial status (via the public ClinicalTrials.gov v2 API). For each drug–outcome pair we run a query and bucket the result into active, completed, preclinical_only, none, or unknown (when the drug name is a combination, identifier, or otherwise not searchable). When trials exist, the top three NCT IDs are linked from the expanded view of the hypothesis.
• Literature status (via PubMed E-utilities). Two queries are issued per hypothesis: a direct co-mention "{drug}"[tiab] AND "{outcome}"[tiab] and the same query restricted to repurposing/treatment/therapy. The result is bucketed into reproduces_known (≥5 repurposing-style mentions), partial_evidence, novel_signal (no repurposing mention and no more than two co-mentions), or unknown.

This lets the catalog distinguish hypotheses that reproduce known associations (a sanity check on the methodology) from genuinely novel signals worth investigating. The combination novel signal + no clinical trial is the most useful starting point — those are candidates that Robertium surfaced and that nobody appears to have followed up on yet.

Enrichment is rate-limited (5 req/s for ClinicalTrials.gov, ~2.5 req/s for PubMed without an API key), idempotent, and runs only against the outreach-quality hypotheses.

Robertium is research infrastructure, not a clinical decision tool. Generated hypotheses require expert review and experimental validation. Known limitations:

• Entity resolution. Lexical merging means "EGFR" and "Epidermal Growth Factor Receptor" are still separate nodes in the current graph. UMLS-based normalization is planned.
• Extraction noise. The LLM occasionally extracts secondary mentions as primary claims (for example, a control arm drug rather than the focus drug). Precision has not been systematically evaluated; manual spot-checks of top-scored hypotheses suggest the majority capture meaningful drug–target relationships, but rigorous precision/recall measurement is planned alongside the first preprint.
• Mediator and drug-vocabulary filters. Hypothesis ranking now incorporates a mediator-specificity filter that excludes generic inflammatory cytokines (TNF-α, IL-6, NF-κB) and trivial pathway mediators (inflammation, cell death, autophagy), plus a drug vocabulary filter that removes botanical extracts and undefined nano-formulations. See classify_outreach_quality.py in the public repo for the full logic.
• Self-validation extensions (May 2026). Manual review of the top-20 high-confidence hypotheses surfaced several drug-name categories that were not actionable as repurposing candidates and have since been added to the excluded set: PET imaging tracers (e.g., PBB3, 18F-PI-2620), biological samples (e.g., ALS-CSF), surgical or device interventions (e.g., deep brain stimulation, focused ultrasound), internal pharma research codes without an approved common name (e.g., WSD-0922, ABBV-221), vague natural-product mixtures (e.g., "active components", slash-joined combinations), and peptide constructs identified by Greek letters or LIR/TP53INP-style identifiers. Drugs in these categories are valid mentions in the literature but are not things a researcher can repurpose, so excluding them improves the quality of the remaining catalog.
• Direct A→C filter is heuristic. A hypothesis with one direct co-mention may still be novel if the co-mention is incidental. We surface but don't auto-exclude these.

We document these limitations openly. Methodology improvements happen in public — see GitHub for current issues.

If you reference Robertium in a manuscript, presentation, or grant proposal, please use the citation below. The hypothesis catalog is archived on Zenodo with a permanent DOI. A formal preprint with full methodology and validation is in preparation.

Source code: github.com/routewise96/robertium · Web interface: robertium.com

For collaborations, dataset access, or methodology questions: daniel@robertium.com