Evidence-Centric Knowledge Graphs for Liver and Colorectal Cancers

EvidenceNet provides structured knowledge graphs derived from approximately 500 full-text biomedical articles per disease, published between 2010 and 2025. Created by Chang Zong and released via Harvard Dataverse, this dataset transforms literature into evidence records with normalized entities and quality scores. The public release includes two disease-specific resources: EvidenceNet-HCC and EvidenceNet-CRC.

Use Cases

• Train relation extraction models on typed evidence-evidence relations within the evidence_graph.json files.
• Validate named entity recognition systems using normalized biomedical entities from evidence_nodes.json.
• Benchmark question-answering models on the provided derived evaluation outputs for component validation.
• Perform link prediction on the evidence-evidence relations to infer missing connections in the biomedical literature.
• Rank drug targets or biomarkers using the evidence quality scores attached to structured evidence records.

Strengths

• Contains structured evidence from around 500 articles per disease, providing a focused corpus.
• Includes provenance metadata linking graph elements back to source literature published from 2010-2025.
• Provides both record-level (evidence_nodes.json) and graph-level (evidence_graph.json) data representations.

Limitations

• Sample size is limited to two specific diseases (HCC and CRC), not a broad representation of all biomedical knowledge.
• Raw source PDFs and large third-party knowledge resources are excluded, limiting traceability and context.
• The evidence quality scores are derived and may reflect the biases of the underlying extraction framework.

Provenance

Source

Full-text biomedical literature published 2010-2025.

Collection Method

Constructed via the EvidenceNet framework for information extraction from literature.

Time Range

Source articles published from 2010 to 2025.

Freshness

Last updated March 2026, with source literature coverage up to 2025.