💻 computer science

GeneTech-30: A Provenance-Tracked, Cross-Domain Scholarly Knowledge Graph of 294,330 Entities across 30 Scientific Domains

GeneTech-30 is a publicly accessible, provenance-tracked scholarly knowledge graph spanning 30 scientific domains that prioritizes transparent design and honest reporting of its moderate retrieval performance and limited literature-based discovery signal over scale, positioning it as a reproducible building block for cross-domain research rather than a comprehensive index.

Original authors: Li Xing

Published 2026-10-23

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Li Xing

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Science often advances by connecting ideas that seem unrelated, like finding a hidden bridge between two distant islands. For decades, researchers have believed that new discoveries often happen not by looking deeper into a single field, but by spotting patterns between different fields that rarely talk to each other. This idea, known as literature-based discovery, suggests that if you read enough scientific papers, you can find a clue in one area that solves a problem in another, even if no one has ever written about the connection before. However, testing this idea has been difficult because scientific knowledge is scattered across thousands of separate databases, and there has been no reliable way to track exactly where every fact came from or to prove that a connection is real and not just a lucky guess.

A researcher has now built a new tool called GeneTech-30 to help solve this problem. Instead of trying to collect every scientific paper in existence, which would be an impossible task, they gathered a specific, manageable collection of 294,330 research works from 30 different scientific fields, ranging from quantum computing to agriculture and brain science. The goal was not to create a massive library, but to build a precise, trustworthy map where every single piece of information is tagged with its origin story. In this system, every claim is linked to the exact document it came from, the time it was found, and a digital seal that proves the data has not been altered. This allows anyone to trace a connection back to its source and verify it for themselves, turning a collection of facts into a transparent, auditable record.

The researcher used this map to test whether they could automatically spot the kind of hidden bridges that lead to new discoveries. They looked for pairs of fields that shared a topic but had never cited each other in the past, hoping to see if a connection would suddenly appear later. To their surprise, the data showed that sharing a topic did not actually make it more likely for two fields to connect. When they compared their findings against a random selection of fields, the specific topic shared between them offered no advantage in predicting a future link. The signal they were looking for was no stronger than chance. This result is not a failure of the tool, but a clear, honest answer to a long-standing question: simply sharing a topic label is not enough to predict where the next big discovery will come from.

Despite this negative finding on prediction, the tool proved highly effective at organizing and retrieving information. When researchers asked the system to find papers based on the text of their summaries, it succeeded in nearly half of the attempts, a solid performance for such a complex task. However, the system struggled when asked to find papers based on broad subject tags, showing that while it can match words well, it still needs more advanced methods to understand the deeper meaning behind those words. The researcher also tested the system against a list of 40 real, historical discoveries that had already happened between different fields. The tool correctly identified about two-thirds of these past breakthroughs as potential connections, provided the fields had not already been talking to each other too much. This suggests the system is good at spotting the conditions where new ideas might emerge, even if it cannot predict the exact moment they will happen.

The value of GeneTech-30 lies in its honesty and its structure. It does not claim to have solved the mystery of scientific discovery or to have found a magic formula for the future. Instead, it provides a clean, verified foundation that other scientists can use to study how knowledge moves between fields. By admitting where the system works well and where it falls short, and by providing the code and data for others to check the work, the researcher has created a resource that prioritizes truth over hype. It stands as a reminder that in the pursuit of new knowledge, knowing what does not work is just as important as knowing what does, and that the most reliable path forward is built on data that can be traced, checked, and trusted by anyone.

Technical Summary: GeneTech-30

Problem Statement
Cross-domain scholarly knowledge is currently fragmented across disciplinary silos, creating a significant barrier for Literature-Based Discovery (LBD). LBD, pioneered by Swanson, aims to surface non-obvious connections between disparate fields without new experiments. However, the field lacks reproducible, provenance-tracked, cross-domain datasets necessary to transparently validate bridging methods. Existing comprehensive indices (e.g., OpenAlex) prioritize scale over provenance guarantees, while smaller datasets often lack the rigorous auditability required for trustworthy engineering.

Methodology
GeneTech-30 is constructed as a curated, cross-domain scholarly knowledge graph with a focus on data integrity and reproducibility rather than raw scale.

Data Acquisition & Scope: The graph aggregates 294,330 entities from 30 mutually near-orthogonal scientific domains (e.g., quantum computing, agritech, biomed-ai). Entities include research works and resources, each containing at minimum a DOI or PMID, title, abstract, tags, and source URL.
Entity Resolution: A two-stage deterministic strategy maps entities to canonical identifiers (genetech:<namespace>:<native-id>):
1. DOI Exact-Match: Entities sharing a DOI string are collapsed into a single canonical entry.
2. Conservative Title Matching: Entities without DOIs are blocked by normalized titles (lowercase, punctuation-stripped) and merged only on exact equality. Fuzzy matching is intentionally disabled to prevent false merges.
- Result: 276,960 canonical entries with a de-duplication rate of 0.72%.
Provenance & Signature Scheme: Every assertion (entity or edge) is sealed with an Ed25519 signature chain. Records include firstSeenAt, validFrom, validTo, and source tier. This allows third parties to audit the origin of claims and verify that data has not been tampered with.
Citation Graph Construction: Edges are extracted from reference lists and resolved against the canonical identity layer. The graph contains 25,682 resolved cross-domain citation edges.
LBD Formalism: The paper operationalizes Swanson's principle by defining a "candidate bridge" between domains A and B at a temporal cut point $T_0$ as a pair sharing a topic tag but exhibiting zero (or near-zero) historical citations before $T_0$ . The method validates whether citations "emerge" after $T_0$ .
Evaluation Protocol: Retrieval is tested against a non-homologous benchmark (300 queries across 30 sites) where queries are not copied from the answer entity's own fields. Metrics include Recall@K, MRR, and NDCG@K.

Key Results
The paper reports results with high transparency, explicitly highlighting both strengths and limitations:

Retrieval Performance:
- Overall: Mean Recall@10 is 0.444, with a Mean MRR of 0.425.
- Decomposition: Performance is highly dependent on query type.
  - Strong: Abstract-mid phrase matching (Recall@10 = 97.4%).
  - Weak: Tag-based retrieval (12.7%), cross-entity mixing (9.2%), and site-topic browsing (6.9%).
- Vector Comparison: A TF-IDF cosine-similarity stack slightly outperformed the BM25 baseline (+1.0% Recall@10, +4.1 MRR points), confirming that moving beyond pure BM25 is beneficial, though neural embeddings are still required for semantic association.
LBD Bridging Validation (Retrospective):
- Null Result: The "co-tag" bridging signal did not exceed random baselines. The lift vs. random pairs was 0.89–0.93, and the lift vs. a label-permutation null model was 1.00.
- Interpretation: Shared topical tags alone do not predict specific-topic citation emergence better than chance. The paper reports this as a transparent null result, positioning the dataset as an exploratory resource for computing such negative findings rather than a predictive discovery tool.
Known-Discovery Reproduction:
- On a benchmark of 40 historical cross-domain discoveries, the method achieved a 67.5% recall rate (27/40) after relaxing strict criteria (allowing up to 5 historical citations and fuzzy topic matching).
- Misses were primarily due to high historical citation counts or topic mismatches in the tag vocabulary.
External Validation (OpenAlex):
- Validated "zero-citation" pairs against OpenAlex. Across 4,201 references checked, zero cross-domain citations were found in the sampled "gap" pairs, supporting the label as a proxy for absence (though with a 42% coverage caveat where OpenAlex lacked reference lists).

Significance and Claims
GeneTech-30 does not claim to have automated scientific discovery. Instead, its significance lies in providing:

A Provenance-Tracked Infrastructure: A composable building block where every data point carries a cryptographic signature and source history, enabling fail-closed auditability.
Honest Evaluation Standards: A framework that reports failure modes (e.g., weak tag retrieval, null LBD signals) as explicitly as successes, establishing a baseline for trustworthy scholarly-data engineering.
Reproducibility: All ingestion, resolution, evaluation, and signing code is open-source, allowing the community to recompute metrics and validate the graph's integrity.
A Benchmark for Negative Results: By demonstrating that simple tag-based bridging does not predict emergence, the resource helps refine the hypotheses required for future LBD tooling.

The paper positions GeneTech-30 not as a replacement for massive indices like OpenAlex, but as a specialized, high-integrity resource for researchers studying cross-domain knowledge structures and developing reproducible LBD methodologies.

Technical Summary: GeneTech-30

More like this