Biology Atlas

Map / Concepts

concept · intermediate · partly written

How one fact becomes an edge with evidence

In INDRA a fact becomes a typed Statement with a list of Evidence objects and a belief score, so the edge can be ranked by trust without deleting anything.

Short answer

A sentence is read, its names are grounded to identifiers, it is filled into a typed Statement, each supporting sentence or database row is kept as an Evidence object, identical Statements are merged and a belief score is computed. The edge is a summary; the evidence list is the record. 12

Reported, not yet replicated

The worked example more

The published example (2017 Figure 2): MAP2K1 (MEK1), phosphorylated at serines 218 and 222, phosphorylates MAPK1 (ERK2) at threonine 185. The enzyme is HGNC:6840 and the substrate HGNC:6871. The Evidence shown is from the BEL Large Corpus, the text "c-Raf activates MEK1 by phosphorylating at serine residues 218 and 222", PubMed 8621729; that Evidence supports the enzyme's activation, the condition part of the Statement. 1345

Reported, not yet replicated

How belief is computed more

Per source type a systematic error (wrong the same way every time) and per Evidence a random error: P(correct) = 1 minus the product over sources of [systematic + product of random errors over that source's Evidences]. The defaults are priors set by the authors: readers 0.05 systematic and 0.3 random, BEL 0.01 and 0.1. With those defaults one BEL Evidence gives 0.89 and three Reach Evidences 0.923 (our own arithmetic, an illustration, not a real record). More of one reader cannot beat its systematic error; a different source moves belief more than repetition. 62

Our inference

Where it breaks more

Reader errors are correlated, so belief is overconfident wherever two readers share a mistake; the authors say so. The curated training set over-represents high-mention Statements (35% single-reader versus 81% in the benchmark corpus), so AUPRC 0.942 is not a general accuracy. The claim that belief is a calibrated probability rests on the authors' own comparison against BioGRID, assuming BioGRID entries are correct. 2

Reported, not yet replicated

Still open here

Not yet written: Plain answer, Why it matters, Equation, Real numbers.

Linked from

Sources

  1. Gyori BM, Bachman JA, Subramanian K, Muhlich JL, Galescu L, Sorger PK. From word models to executable models of signaling networks using automated assembly. Molecular Systems Biology 13:954 (2017). paper; opened 2026-11-18 all records citing this
  2. INDRA paper on automated assembly of mechanistic knowledge from text and databases (PMC10167483). PubMed Central (2023). paper; opened 2026-11-17 all records citing this
  3. HGNC. HGNC record for MAP2K1. rest.genenames.org (2026). database, not peer reviewed; opened 2026-11-18 all records citing this
  4. HGNC. HGNC record for MAPK1. rest.genenames.org (2026). database, not peer reviewed; opened 2026-11-18 all records citing this
  5. INDRA developers. INDRA documentation: Statements. indra.readthedocs.io (2026). web, not peer reviewed; opened 2026-11-18 all records citing this
  6. INDRA developers. INDRA belief module and default belief probabilities. GitHub (gyorilab/indra) (2026). database; opened 2026-11-18 all records citing this
Where it sits on the map
History of this record (6)

This record has been changed 6 times, on the dates below. The note kept against each change is written for the people who maintain the atlas, in their own shorthand, so it is not shown here; it is in the record itself.

Status
partly written
Last updated
2026-11-19
Record
content/concepts/fact-to-edge-with-evidence.json