FAQ

Is this just another GPT wrapper?

No. Robertium uses LLMs for one specific task: extracting structured claims from abstracts according to a controlled vocabulary. The LLM doesn't generate hypotheses — it only converts unstructured text into a graph format.

The hypothesis generation uses Don R. Swanson's literature-based discovery model from the 1980s, applied to a knowledge graph. This is a deterministic graph algorithm, not LLM inference.

How do I know the hypotheses are real and not hallucinated?

Every hypothesis is a chain of evidence with PMIDs to peer-reviewed papers. You can click any PMID on the /hypotheses page and verify the claim against the original abstract.

If a claim doesn't match its source paper, that's a bug — please report it via GitHub or email. We do regular manual spot-checks of top-scored hypotheses and the majority capture meaningful relationships, but precision and recall have not been systematically evaluated yet — that work is planned alongside the first preprint. Errors exist and we want to know about them.

How is this different from existing drug repurposing tools?

Most repurposing tools work in one of three ways: (1) molecular similarity to known drugs, (2) gene expression signature matching, or (3) clinical trial data mining. Robertium uses literature-based discovery (LBD) — surfacing connections that exist in published research but haven't been explicitly studied.

Specifically, our cross-domain approach (drug from one disease literature, target/mediator shared with another disease literature) is rarely automated systematically. Most LBD tools work in single domains.

How is Robertium different from existing literature mining tools (PubTator, BERN2, SemMedDB)?

Robertium operates at a higher abstraction level. Tools like PubTator and BERN2 perform named-entity recognition; SemMedDB extracts semantic predications via SemRep. Robertium consumes these as inputs (we use PubTator for the L1 biomedical-relevance filter) and adds two layers on top: (1) LLM-based structured claim extraction with explicit subject–predicate–object triples and confidence scores, and (2) cross-domain Swanson ABC chains across multiple disease graphs simultaneously.

The output is not entity tags or single-sentence relations, but candidate repurposing hypotheses with full evidence chains spanning two literatures.

Can I use this in my research?

Yes. The /hypotheses catalog is freely browsable. The full code is MIT licensed. The dataset will be released under CC-BY-4.0 alongside the first preprint.

If you want to:

• Validate a hypothesis experimentally — please cite the project and let us know the results (positive or negative).
• Use the methodology in your own corpus — clone the repo, configure a new domain, run the pipeline.
• Collaborate on improving the methodology — open an issue on GitHub or email daniel@robertium.com.

What's the validation rate of generated hypotheses?

This is an open research question. We are early-stage and don't yet have systematic experimental validation of top hypotheses.

Anecdotally, several top hypotheses match published findings — for example, vortioxetine (antidepressant) → glioblastoma efficacy was published in Nature Medicine 2024 (Lee et al.) and is automatically surfaced by Robertium from epilepsy ↔ glioblastoma analysis. This suggests the methodology surfaces real signal, not noise.

Systematic validation requires lab partnerships, which we are actively seeking. Top hypotheses prioritized for validation will be flagged on /hypotheses.

What are the known limitations and false-positive sources?

Three main sources of noise:

• Entity resolution is lexical. "EGFR" and "Epidermal Growth Factor Receptor" remain separate nodes in the current graph, pending UMLS-based normalization.
• The LLM occasionally extracts secondary mentions — control-arm drugs, citation references — as primary claims, especially in densely written abstracts.
• Some mediators are highly generic (TNF-α, IL-6, NF-κB, autophagy, oxidative stress) and connect any inflammatory drug to any inflammation-associated disease. These are filtered out from outreach-quality hypotheses but remain in the raw graph for downstream research uses.

Systematic precision evaluation is planned for the first preprint. In the meantime, every hypothesis links to its source PMIDs so reviewers can verify claims independently.

Why these specific therapeutic domains?

The current ten domains were chosen for diversity and tractability: four neurological (epilepsy, ALS, Alzheimer's, multiple sclerosis), one psychiatric (major depressive disorder), two oncological (glioblastoma, pancreatic cancer), one metabolic (type 2 diabetes), and two autoimmune/inflammatory (rheumatoid arthritis, inflammatory bowel disease). They share enough molecular biology to produce meaningful cross-domain bridges (EGFR signaling, BDNF, autophagy, mitochondrial dysfunction, AMPK, TNF-α) while being distinct enough that cross-pollination of literature is non-trivial.

Practical considerations: each domain in the current corpus has 14,000–18,000 papers — enough signal without overwhelming the L1 filter and downstream extraction. Adding or re-ingesting a domain takes pipeline runtime proportional to corpus size, so the catalog can be expanded based on user demand. If you have a strong argument for a specific domain, please reach out.

Can I run this on my own corpus?

Yes. The pipeline is domain-agnostic. To add a new domain:

• Find your OpenAlex concept ID
• Create a new YAML config under config/domains/
• Run scripts/run_full_pipeline.py --domains your_domain

The README has detailed setup instructions. The pipeline is model-agnostic: any modern instruction-tuned LLM with structured-output support works for the extraction step.

What about clinical trial data, drug-target databases, etc.?

Currently Robertium uses only literature. Integration with structured biomedical databases is planned for late 2026:

• DrugBank — for drug-target validation and metadata
• ChEMBL — for binding affinity data
• ClinicalTrials.gov — for known clinical investigations

These will be additional evidence layers, not replacements for literature-based discovery.

How is this funded?

Currently self-funded by the maintainer. We are actively pursuing open science grants:

• Mozilla Open Source AI Fund
• Chan Zuckerberg Initiative Open Science
• Wellcome Trust Discovery Awards

If your foundation supports open biomedical infrastructure, please reach out.

How can I contribute or collaborate?

Several entry points depending on what you want to do:

• Methodology contributions — see open issues on github.com/routewise96/robertium. Pull requests for filters, predicates, or pipeline improvements are welcome.
• Experimental validation — for collaborations on validating specific hypotheses in the lab, contact daniel@robertium.com directly.
• Catalog feedback — corrections on specific hypotheses or claims can come via email or a GitHub issue. Both work; both get acted on.
• New domain or replication — the full pipeline is open-source. Clone the repo and follow the README; happy to help debug if you hit something unclear.

Who validates the methodology?

A preprint with detailed methodology and validation will be submitted to bioRxiv in mid-2026. Pre-publication, the methodology is open for review by the research community — please open issues on GitHub for technical concerns.

We welcome critique. The project gets better when domain experts find problems.

I'm a journalist. Can I cite or quote Robertium?

Yes, with attribution. For technical accuracy please reach out before publication — a five-minute conversation prevents misrepresentation. daniel@robertium.com.

We particularly welcome stories about:

• Open science infrastructure
• Cross-domain biomedical research
• Drug repurposing as alternative to traditional drug discovery
• AI/ML in scientific research (responsibly framed)

I'm a clinician, can I make decisions based on this?

No. Robertium generates hypotheses for research investigation, not clinical guidance. Generated candidates require:

• Expert review by domain specialists
• Preclinical validation (cell lines, animal models)
• Phase I–III clinical trials before any clinical application

Please don't change your treatment decisions based on /hypotheses content. This is research infrastructure, not medical advice.