Organizations: School of Computer Science and Technology, Zhejiang University of Science and Technology, Hangzhou, China · Institute of Medicinal Biotechnology, Chinese Academy of Medical Sciences & Peking Union Medical College, Beijing, China · Zhejiang Key Laboratory of Biomedical Intelligent Computing Technology, Hangzhou, China
Biomedical knowledge resources often either preserve evidence as unstructured text or compress it into flat triples that omit study design, provenance, and quantitative support. Here we present EvidenceNet, a disease-specific dataset of record-level evidence collections and corresponding graph representations derived from full-text biomedical literature. EvidenceNet uses a large language model (LLM)-assisted pipeline to extract experimentally grounded findings as structured evidence records, normalize biomedical entities, score evidence quality, and connect related records through typed semantic relations. We release EvidenceNet-HCC with 7,872 evidence records and a corresponding graph with 10,328 nodes and 49,756 edges, and EvidenceNet-CRC with 6,622 records and a corresponding graph with 8,795 nodes and 39,361 edges. Technical validation shows high component fidelity, including 98.3% field-level extraction accuracy, 100.0% high-confidence entity-link accuracy, 87.5% fusion integrity, and 90.0% semantic relation-type accuracy. Downstream analyses show that the data support retrieval-augmented question answering and graph-based tasks such as future link prediction and target prioritization. These results establish EvidenceNet as a disease-specific biomedical knowledge base dataset for evidence-aware analysis and reuse.
Figures & tables
Aspect
Article-level literature
Compact knowledge graph
EvidenceNet release
Primary unit
Article abstract
Triple typed edge
Evidence record
Provenance
Article-level metadata
Citation-level
Article metadata + source span
Study context
Free-text narrative
Sparse or compressed
Structured context fields
Quantitative detail
Reported in text
Rarely retained
Captured when reported
Typical use
Search Reading Screening
Navigation Integration Graph mining
Retrieval Synthesis Structural mining
Table 1: Positioning of EvidenceNet relative to article-level literature resources and compact biomedical knowledge graphs.
Figure 1: Dataset construction workflow. Disease-specific full-text PDF collections are parsed into text, filtered into extractable sections, converted into source-grounded evidence records, normalized to biomedical entities where possible, checked for duplicate or related records and exported as record-level JSON files with derived graph files.
Measure
HCC
CRC
SLE
Unique source PDFs linked in records
470
472
392
Evidence records retained
7872
6622
4261
Records per linked source PDF
16.8
14.0
10.9
Unique article titles
457
452
382
Table 2: Input-literature and record-construction summary.
File group
Format
Contents
records/ evidence_records_{disease}.json
JSON
Primary evidence-record files for HCC, CRC and SLE, each keyed by evidence_id .
graphs/ evidence_graph_{disease}.json
JSON
Derived NetworkX node-link graph files for the corresponding disease subsets.
schema/
JSON, TSV
Record schema, graph schema and field dictionary for released file interpretation.
validation/
JSON, CSV
Audit samples, integrity checks, external-concordance summaries, record-lookup outputs and metadata-QC records.
prompts/
TXT
Prompt templates used during extraction, aggregation and relation verification.
checksums_sha256.txt
TXT
SHA-256 checksums for release-level integrity checking.
Table 3: Main deposited file groups and their roles.
Figure 2: Representative transformation from full-text literature to a structured evidence record and its companion relation view. The left panel shows preserved article-level provenance and supporting source text. The middle panel shows the corresponding structured evidence object with provenance, context, finding, entity and quality fields. The right panel shows how the same evidence_id is linked to normalized entities and related records in the companion graph files.
Field
Description
evidence_id
Unique record identifier.
source
Article-level metadata, including DOI, title, journal, year, authors, citation count and local source path when available.
pico
Structured study context, including population, intervention, comparison and outcome metrics.
core_entities
Extracted biomedical entities with raw names, semantic classes and normalized identifiers when available.
target_disease
Disease context for the resource: HCC, CRC or SLE.
bio_mechanism
Text field describing the reported biological mechanism when available.
Table 4: Main fields in the primary evidence-record JSON files.
Figure 3: Score-filtered cross-paper semantic display views of the released disease-specific companion graph files. Each panel shows a disease-specific display subgraph derived from the corresponding released graph JSON file after restricting to cross-paper evidence-to-evidence relations above the visualization threshold. Record colour indicates study design and relation colour indicates semantic relation type.
Figure 4: Statistical overview of the three released disease subsets. Panel A shows the annual distribution of source years for released evidence records. Panel B shows the study-design composition of each disease subset. Panel C shows the distribution of composite evidence scores. Panel D shows the composition of released evidence-relation annotations associated with the companion graph layer.
Field
HCC
CRC
SLE
Comparison
4492 (57.1%)
3951 (59.7%)
2598 (61.0%)
p-value
800 (10.2%)
835 (12.6%)
571 (13.4%)
Sample size
1955 (24.8%)
1856 (28.0%)
1648 (38.7%)
Fold change
340 (4.3%)
325 (4.9%)
323 (7.6%)
Bio-mechanism
5355 (68.0%)
4585 (69.2%)
2344 (55.0%)
Phenotype
7666 (97.4%)
6480 (97.9%)
4148 (97.3%)
Table 5: Coverage of selected evidence-record fields.
Record component
Validation criterion
source metadata
DOI, title and publication year are checked for plausibility, and obvious metadata outliers are reviewed against article and repository information.
source_text
Sampled records are checked for direct support from the deposited source passage.
pico , phenotype , bio_mechanism
Sampled records are checked for reusable, source-grounded findings rather than background text alone.
study_design , clinical_stage
Study-context labels are checked against explicit wording or unambiguous cues in the source passage.
core_entities , linked_tarkg_entities
Accepted links are checked when the mention can be grounded in the record context, focusing the audit on deposited normalized identifiers.
merged_from , version
Sampled merged records are reviewed for coherence of the retained canonical record after duplicate handling.
Table 6: Validation checks used for the main record components.
Validation item
HCC
CRC
SLE
Source-text support
47/50 (94.0%)
46/50 (92.0%)
48/50 (96.0%)
Key-entity capture
42/50 (84.0%)
40/50 (80.0%)
45/50 (90.0%)
Grounded entity-link mention
50/50 (100.0%)
50/50 (100.0%)
50/50 (100.0%)
Normalized identifier validity
50/50 (100.0%)
50/50 (100.0%)
50/50 (100.0%)
Fusion record coherence
25/30 (83.3%)
29/30 (96.7%)
28/30 (93.3%)
Evidence-relation edge validity
50/50 (100.0%)
50/50 (100.0%)
50/50 (100.0%)
Table 7: Balanced manual audit of core record-quality components.