Browse workflow stages

Graph-based evidence workflow

What happens between a biological question and a cited answer?

BinfoNet keeps candidate retrieval, ranking, generation, and validation separate. Each stage produces evidence that can be inspected without treating the entire pipeline as one black box.

01

Make the biological input explicit.

The example question asks what role AREG and EREG play in the use of EGFR inhibitors for colorectal cancer. Those terms become named inputs rather than implicit assumptions.

Genes
AREG (NCBI Gene 374), EREG (2069), EGFR (1956)
Disease and treatment context
Colorectal cancer; EGFR inhibition; predictive biomarker role
Species
Not stated in the question. The user-selected protein-graph species defines the gene identities and network edges and is carried into publication retrieval as context.

Why this matters: a paper can be topically similar yet concern the wrong gene, disease, species, or model system. Those are distinct evaluation failures.

02

Let the graph define the candidate evidence space.

Gene anchorsResolved genes from the biological question
Publication candidatesGraph-linked PMIDs with candidate provenance
Semantic orderingGene coverage tiers, then question similarity

For the EGFR case, PubTator gene links plus a disjoint exact-symbol fallback produced 277,638 unique candidates.

263,702mention 1 of 3 requested genes
12,423mention 2 of 3 requested genes
1,513mention all 3 requested genes

What the graph link does—and does not mean

A gene–publication link is candidate-retrieval provenance. It does not prove a biological relationship, direct experimental evidence, or claim support. PubTator links in this runtime are document-level and their section is unspecified.

The current retrieval path uses this gene–publication bridge. Stored STRING and citation relationships are not yet used by default for protein-module expansion, citation-neighborhood retrieval, or publication-community selection.

03

Order graph candidates with interpretable signals.

  1. Requested-gene coveragePublications covering more requested genes form the earlier tiers.
  2. Semantic similarityWithin each tier, the question embedding is compared with publication embeddings.
  3. PMIDUsed only as a deterministic final tie-breaker.

Boundary: semantic similarity measures topical alignment. It is not a score for experimental directness, evidence strength, or scientific correctness.

04

Synthesize only the selected evidence.

After the user selects publications from a retrieval run, BinfoNet generates an answer only from their saved titles and abstracts. The evidence set remains fixed during generation.

  • Every claim must cite at least one selected PMID.
  • Available species and model-system context must be preserved.
  • Supporting, limiting, contradictory, and uncertain findings must be distinguished.
  • The answer must state that full-text methods, controls, statistics, and reproducibility were not assessed.
Inspect the generated EGFR claims →
05

Separate structural validity from scientific validity.

Deterministic checks can establish

  • valid output schema
  • claim-level citation presence
  • citation membership in selected evidence
  • valid stance and provenance fields

They cannot establish

  • whether the cited paper supports the claim
  • whether important evidence was omitted
  • whether the answer overstates the science
  • whether it is ready for biomedical use
06

Make every stage reviewable.

Persisted runs retain the question, resolved genes, ranking policy, displayed evidence, model and data versions, PMIDs, and trace identifiers. Phoenix/OpenTelemetry traces expose stage inputs and outputs without introducing BioASQ reference labels into runtime retrieval.