Literature-Based Discovery¶

Generate testable hypotheses by linking complementary relations stated in separate scholarly literatures—often A–B and B–C—to propose an unstated A–C connection.

Version: v1 · 2026-08-30 · History
Domain-specific #: 2196
Origin domain: information science
Subdomain: knowledge discovery from literature
Aliases: Literature-related discovery, LBD, Swanson linking

Core Idea¶

Literature-based discovery (LBD) is the information-science practice of generating hypotheses by connecting claims that are explicit in separate scholarly literatures but whose joint implication has not been explicitly investigated. Its canonical ABC pattern begins with an established relation between concept A and intermediary B and another between B and C. If the A and C literatures are substantially disconnected, the shared B suggests a potentially novel A–C relation for expert assessment and empirical testing.

Don Swanson pioneered the method in the 1980s. His best-known case joined literature about Raynaud disease with literature about fish oil through intermediate concepts such as blood viscosity and platelet aggregation. The proposed therapeutic relation was later investigated prospectively. This history demonstrates the intended epistemic status: LBD surfaces a plausible candidate from public knowledge; it does not convert transitive co-occurrence into established causation.^[1]

Modern systems combine information retrieval, named-entity recognition, word-sense disambiguation, ontology normalization, relation extraction, knowledge graphs, embeddings, ranking, and interactive visualization. Open discovery starts from A and searches through Bs for candidate Cs. Closed discovery starts with A and C and searches for bridging Bs that could explain or test their connection. Both forms end with ranked hypotheses, not automatic discoveries.

Structural Signature¶

scholarly corpus — dated papers, abstracts, databases, or curated resources;
disjoint or weakly connected literatures — communities discuss complementary relations without making the target connection;
normalized concepts — synonyms and ambiguous mentions map to controlled identities;
explicit source relations — A–B and B–C links are extracted with provenance and context;
bridging concept B — an intermediary supplies a possible mechanism or association path;
unstated candidate relation A–C — the path proposes a connection not explicit in the eligible corpus;
open or closed search mode — endpoints given determine exploration or explanation;
filters and semantic types — admissible path patterns constrain combinatorial explosion;
ranking model — novelty, strength, specificity, plausibility, and diversity order candidates;
traceable evidence chain — users can inspect the publications supporting every edge;
expert adjudication — domain knowledge rejects artifacts and prioritizes tests;
external validation — later experiments, observations, or publications assess the hypothesis;
temporal cutoff — evaluation prevents future knowledge from leaking into retrospective discovery.

The invariant is hypothesis generation from complementary, provenance-bearing relations distributed across previously disconnected literature.

What It Is Not¶

Not ordinary literature review. Review synthesizes a known topic; LBD deliberately searches for implicit cross-literature connections.
Not deductive proof. A–B and B–C do not generally entail A–C without relational and causal qualifications.
Not co-occurrence mining alone. Co-occurrence can supply candidates, but LBD requires novelty framing, linking logic, provenance, and validation.
Not a knowledge graph generally. Graphs can implement LBD but also represent already explicit knowledge.
Not drug repurposing only. Biomedicine dominates historically, while the method can support other sciences.
Not automatic scientific discovery. System output is a ranked hypothesis requiring critical and empirical work.
Not transitive closure over all predicates. Relation types, direction, negation, conditions, and evidence quality matter.

Scope of Application¶

LBD is used in biomedical informatics, drug repurposing, adverse-event detection, gene–disease association, biomarker discovery, disease mechanism studies, research-policy analysis, materials and environmental research, and interdisciplinary collaboration discovery. It is strongest where a large indexed literature has good entity normalization and partially complementary research communities.

The method can incorporate curated databases alongside text if provenance distinguishes sources. Semantic typing can restrict paths—for example disease–process–drug rather than arbitrary word chains. Contextualized relations improve on raw co-occurrence by preserving direction, negation, species, experimental setting, and causal role.

Scope should be bounded by corpus date and coverage. “Not published” may mean absent from the selected database, missed by extraction, phrased under another term, or genuinely unknown. Claims of novelty must therefore be search-audited rather than absolute.

Clarity¶

In open discovery, choose A, retrieve its associated Bs, then expand each B to candidate Cs outside the starting literature. Rank Cs and display the strongest paths. In closed discovery, choose A and C and search for Bs that connect them, producing possible mechanisms or evidentiary bridges.

The ABC labels describe roles, not fixed biomedical entity types. B can be a physiological process, material property, gene, method, or social mechanism. Multiple independent Bs usually strengthen plausibility more than one ambiguous bridge.

A valid result must preserve sentences and citations supporting each edge. If “A inhibits B” and “C increases B,” the naive A–C direction may oppose the desired effect. Negation, population, dose, time, and modality can reverse interpretation. Expert review is not an optional cosmetic step.

Manages Complexity¶

Scientific specialization distributes relevant facts across journals, vocabularies, and communities. No researcher can read every adjacent literature. LBD turns this fragmentation into a search space: normalized concepts become nodes, explicit relations become edges, and unexamined paths become candidates.

The ABC abstraction is a powerful compression. It reduces millions of documents to interpretable bridge patterns, while filters and rankings control path explosion. Evidence chains retain human auditability that an opaque similarity score lacks.

LBD also separates generation from evaluation. High-recall discovery can propose many paths; semantic constraints, experts, and experiments progressively narrow them. This pipeline avoids demanding that one algorithm both imagine and prove a hypothesis.

Abstract Reasoning¶

Literature separation can hide public knowledge. Facts may be individually published yet jointly unrecognized.
Bridge diversity raises robustness. Several mechanistically distinct Bs reduce dependence on one extraction error.
Typed relations outperform blind transitivity. Valid inference depends on predicate semantics and direction.
Novelty is time-indexed. A candidate can be a genuine pre-cutoff prediction even if published later.
Ranking creates selection bias. Benchmarking only famous rediscoveries can reward systems tuned to a tiny canon.
Negative evidence matters. Contradictory, negated, or context-limited relations should lower or qualify ranks.
Hypothesis quality is multi-dimensional. Novelty, plausibility, actionability, testability, and potential value need not coincide.

Knowledge Transfer¶

The method transfers across fields when documents can be normalized into entities and relations. Its workflow—retrieve, normalize, link, rank, inspect, validate—remains stable even when domain ontologies change.

Transfer is weaker in fields where claims depend on long arguments, images, tacit practices, or concepts not captured as simple relations. Critics correctly note that science is not exhausted by ABC triples. Richer graph, embedding, and analogy systems extend the method but must preserve evidentiary traceability.

The broader portable abstraction is abductive bridge finding. Without scholarly corpora and an unstated cross-literature hypothesis, it is not literally LBD.

Examples¶

Raynaud disease–fish oil. Separate literatures connected through blood rheology suggested a treatment hypothesis.
Migraine–magnesium. Swanson linked physiological findings across disconnected publications to propose magnesium relevance.
Drug repurposing. A disease–pathway literature and drug–pathway literature yield candidate therapies.
Open discovery. Starting from a disease, a system ranks chemicals reached through typed mechanisms.
Closed discovery. Given a disease and drug, the system retrieves intermediate genes or processes.
Time-slice evaluation. Train on publications before year t and test whether later reported links rank highly.

Structural Tensions¶

Novelty vs. plausibility. Familiar paths are credible but uninteresting; distant paths are novel but noisy.
Recall vs. review burden. Broad generation catches more ideas while overwhelming experts.
Simple ABC paths vs. scientific nuance. Interpretability competes with context-rich representation.
Automation vs. provenance. Strong models can rank well yet obscure why.
Corpus scale vs. terminology ambiguity. More text adds evidence and homonyms.
Retrospective benchmarks vs. prospective value. Rediscovery is measurable but may not predict new usefulness.

Structural–Framed Character¶

The pipeline is structural; evaluation is framed. Whether a hypothesis is interesting, actionable, or sufficiently novel depends on current community knowledge, resources, and purposes. Evidence and cutoff rules keep that judgment auditable.

Structural Core vs. Domain Accent¶

The core is abductive path completion across disconnected knowledge fragments. The domain accent is scholarly publication, bibliographic disjointness, controlled concepts, citation provenance, and scientific validation.

Abductive Reasoning — bridges generate plausible explanations or hypotheses.
Search and Retrieval — relevant relations are located across corpora.
Knowledge Graph — concepts and relations form traversable evidence structures.
Normalization — aliases and senses map to stable entities.
Knowledge Transfer — findings cross disciplinary silos.
Triangulation — multiple independent paths strengthen a candidate.

The prospective DAG edge uses composition under prime:abductive_reasoning.

Relationships to Other Abstractions¶

Current abstraction Literature-Based Discovery Domain-specific

Parents (1) — more general patterns this builds on

Literature-Based Discovery is part of Abductive Reasoning Prime

multiple independent paths strengthen a candidate.

Hierarchy path (1) — routes to 1 parentless root

Literature-Based Discovery → Abductive Reasoning

Neighborhood in Abstraction Space¶

Literature-Based Discovery sits in a sparse region of the domain-specific corpus (96^th percentile for distinctiveness): few abstractions share its structure, so a faithful description tends to retrieve it precisely.

Family — Unclustered & Miscellaneous (1565 abstractions)

Nearest neighbors

Computed from structural-signature embeddings · 2026-09-08

Not to Be Confused With¶

Systematic review — evidence synthesis for a prespecified question.
Bibliometrics — measurement of publications and citations.
Text mining — broader extraction family.
Knowledge graph completion — technical prediction task that may lack literature-discovery goals.
Causal discovery — identification under causal assumptions.
Automated theorem proving — deductive derivation from formal axioms.

References¶

[1] Neil R. Smalheiser, “Rediscovering Don Swanson: the Past, Present and Future of Literature-Based Discovery,” Journal of Data and Information Science (2017), https://pmc.ncbi.nlm.nih.gov/articles/PMC5771422/. registry ↩

[2] “A systematic review on literature-based discovery workflow,” https://pmc.ncbi.nlm.nih.gov/articles/PMC7924697/. registry

[3] “A Graph-Based Recovery and Decomposition of Swanson’s Hypothesis,” https://pmc.ncbi.nlm.nih.gov/articles/PMC4031661/. registry

[4] “Literature-based discovery,” Wikipedia, frozen revision 1365851656 (2026-07-24), https://en.wikipedia.org/wiki/Literature-based_discovery. registry