Evidence Types:
Confidence:
High
Medium
Low
Review

Recent Articles

PMIDTitleYearJournalEligibilityEvidence

Articles

PMIDTitleYearJournalEligibilityConfidenceEvidenceProteinsPDB

Evidence Browser

Evidence Confidence: High = direct experimental method / clear quote | Medium = partial method, needs verification | Low = weak/incomplete, requires review | Review = flagged for curator
IDPMIDProteinUniProtPDBTypeConfidenceQuote

Protein Directory

Structure Directory

PDB IDProteinUniProtResolutionMethodEvidenceArticles

Manual Review Queue

IDPMIDIssue TypeDescriptionNext StepStatus

About Human_PPCA

What is Human_PPCA?

Human_PPCA (Human Protein Production and Crystallization Atlas) is a curated knowledgebase of experimental evidence for human protein expression, purification, construct design, crystallization, and structure determination. It integrates literature evidence with UniProt and PDB cross-references, providing a searchable, confidence-graded resource for structural biology researchers.

Current seed release covers 100 articles, 797 evidence records, 122 proteins, and 132 unique PDB structures.

Evidence Confidence Grading

High Confidence
Direct experimental method explicitly described in the full text. Clear evidence quote available. Ready for production-level use without further review.
Medium Confidence
Partially extracted or inferred from context. Method described but key parameters missing. Recommended for manual verification.
Low Confidence
Weak, incomplete, or indirectly referenced evidence. Requires full-text re-review before inclusion in production database.

needs_manual_check flag: Regardless of confidence, any evidence record where extraction is ambiguous or contradictory is flagged with review. These appear in the Review Queue.

Methods: LLM-Assisted Full-Text Extraction Workflow

  • Source Selection: 100 structure-related articles identified from PubMed using targeted queries for human protein expression, purification, crystallization, and structure determination.
  • Full-Text Retrieval: Article full texts obtained from PubMed Central and publisher APIs.
  • LLM-Assisted Extraction: A structured extraction pipeline using large language models extracted evidence records following a predefined schema covering 6 evidence types.
  • Quality Assurance: Each extracted record includes an evidence quote (verbatim text from source article), enabling traceability. Records graded for confidence and flagged for manual review.
  • Cross-Referencing: UniProt accessions and PDB IDs cross-referenced against UniProt KB and RCSB PDB.

Data Release & Download

The following data files are available in the current demo package. Production version (v0.4+) will provide full downloads via Cloudflare R2 with versioned archives.

↓ article_records_final.tsv
100 articles with PMID, title, journal, year, DOI, extraction eligibility, confidence, curator summary.
TSV · 100 rows
Download TSV
↓ evidence_records_final.tsv
797 evidence records covering 6 evidence types with confidence grading.
TSV · 797 rows
Download TSV
↓ unresolved_items_final.tsv
84 unresolved/manual review items with issue type and recommended next steps.
TSV · 84 rows
Download TSV
↓ human_ppca_seed_release.sqlite
Complete normalized database dump with full schema, relations, and indexes.
SQLite · Full database
Download Database

Note: Download buttons are placeholders in this deployment. Production version will provide real file downloads via R2.

Known Limitations (v0.3)

  • Scale: This seed release covers 100 articles. Production database will cover thousands of structural biology publications.
  • Extraction Eligibility: 10 articles have extraction_eligibility set to "not classified" — early pilot artifacts, will be re-reviewed in v0.4.
  • Compound PDB IDs: 3 evidence records contain comma-separated PDB IDs (1GP0/1GP3, 1K1A/1K1B, 1LKK/1LKL). These are split into individual sub-IDs for structure detail views.
  • No 3D Structure Viewer: Molstar visualization planned for v0.4.
  • No Authentication: Admin actions (review resolution) require JWT auth, planned for production backend.
  • LIKE-based Search: Full-text search uses SQLite LIKE. Meilisearch/Elasticsearch/Vectorize planned for production.

Citation & Contact

Citation (placeholder): Zhang, R. et al. "Human_PPCA: A Knowledgebase of Protein Production and Crystallization Evidence." (Publication pending)

Contact: For questions or collaboration, contact the development team via the project repository.

Search Results