Lumen
Lumen retrieves literature and extracts methodology consensus from full-text Methods sections. Each extracted parameter carries a citation count and a strength indicator, with retracted papers flagged rather than silently counted.
Inputs
- Research question or topic
- Optional corpus filters (date, journal, organism, method)
- Tunable ranking weights: semantic, citation authority, recency, journal quality, retraction status
Outputs
- Ranked literature set with retraction-status flags
- Methodology consensus blocks (per-parameter value, citation count, strength indicator)
- Exportable citations (BibTeX, RIS, EndNote XML, Zotero RDF, CSL)
How the AI suggests, and where you decide
Lumen surfaces consensus values and ranked evidence. It reports “no literature evidence available” where coverage is zero rather than fabricating a value. The researcher selects which consensus to carry forward.
Illustrative
Lumen in practice
Libraries were sequenced to a depth of approximately 30 million reads per sample. Differential expression was assessed with DESeq2; genes with absolute log2 fold-change ≥ 1.5 and Benjamini–Hochberg adjusted p < 0.05 were considered significant. Each condition comprised six biological replicates.
| Parameter | Consensus value | Evidence |
|---|---|---|
| Sequencing depth | 30M reads / sample | 28 citations· Strong |
| Fold-change threshold | |log2FC| ≥ 1.5 | 19 citations· Moderate |
| Multiple-testing correction | Benjamini–Hochberg FDR 0.05 | 34 citations· Strong |
| Sample size | n = 6 per group | 12 citations· Moderate |
| Batch-effect model | No literature evidence available | No literature evidence |
Citation grounding
Every consensus value links to its supporting papers with author, year, journal, and URL. Outputs trace back to retrieved sources, never invented from model knowledge.
For the technical reader
Technical details
- Full-text Methods-section parsing across PubMed Central and Europe PMC open-access subsets.
- Hybrid ranking with user-tunable weights (semantic similarity, citation authority, recency, journal quality, retraction status).
- Methodology fields aggregated across the corpus; methods used in under 10% of papers flagged as outliers.
- Multilingual corpora: SciELO (Spanish/Portuguese), HAL (French), CNKI (Chinese), J-STAGE (Japanese), KCI (Korean).
Researchers, Clinical research, Public health, Pharma enterprise. The decision-support boundary is surfaced consistently: the platform suggests; the user decides.
Each capability is one link in the closed loop. See the whole chain end to end.