Citation Matrix and Metadata Derived from the PKG24S4 Database

Version Upadate: we provide zipped files for faster download.

This repository contains sparse matrices and auxiliary tables describing citation and authorship relations among biomedical papers, clinical trials, and patents derived from the PKG24S4 version of PubMedKG (PKG). The source database integrates PubMed, ClinicalTrials.gov, USPTO (via PatentsView), and additional open citation resources into a unified relational schema.

We focus on the core entities (papers, clinical trials, patents) and their citation and authorship links, and reorganize them into machine‑readable matrices and TSV files suitable for large‑scale network analysis, scientometrics, and method development in bibliometrics and science‑of‑science research.

Data Statistics

We consider all papers, patents, and clinical trials in PKG24S4, consisting of 40,363,572 nodes: 38,217,186 papers, 530,522 clinical trials, and 1,615,864 patents. Nodes that appear only in citation tables (C04, C12, C16) but lack metadata (in C01, C11, or C15) are retained with missing metadata.

Files Outline

We share five TSV files, three comprising metadata information and two consisting of the citation matrix.

Metadata Files

node_id_type.zip: TSV file with one row for each node (paper, patent, clinical trial). Each row reports the node identifier (first column), the type of node (second column), and the row id in the citation matrix (third column).

node_year.zip: TSV file with one row for each node (paper, patent, clinical trial). Each row reports the node identifier (first column) and the year of publication of the node (second column). If the year is unavailable in the database, the second column is empty.

node_authors.zip: TSV file with one row for each node (paper, patent, clinical trial). Each row reports the node identifier (first column) and the list of authors of the node (second column). If author information is unavailable, the second column is empty.

Note that the first row of each file is the header information.

Node Identifiers

PKG stores PMIDs and patent identifiers as integers, while clinical trial identifiers already carry the prefix “NCT”. To avoid collisions between papers and patents with the same numeric ID, we normalize node identifiers as strings with explicit prefixes:

Papers: PMID<integer> (e.g., “PMID245”).

Clinical trials: as in the source database (e.g., “NCT00887523”).

Patents: PTID<integer> (e.g., “PTID123”).

The author identifiers are kept as in the source database (i.e., integers).

Citation Matrix

The citation matrix is shared in Compressed Sparse Row (CSR) format, using three arrays: one for non-zero values (called *data array*), one for corresponding column indices (`indices.tsv`), and one for row pointers indicating the start of each row's data in the other arrays (`indptr.tsv`). Since all edges in the citation matrix have unit weight, we can omit the separate data array. A single array of ones is sufficient. We share:

indices.zip: TSV file representing the arry with the column index for each non-zero value

indptr.zip: TSV file representing the array with information on where each row starts and ends in the indices array.

The two TSV files do not contain an header, the first row corresponds to the first element of the array.

Data Construction

Source Data and Tables

We use the PKG24S4 relational database (https://doi.org/10.1038/s41597-025-05343-8), which can be downloaded at: https://www.scidb.cn/en/detail?dataSetId=4. The documentation of PKG24S4 is available at: https://github.com/PubMedKG/PubMedKG.github.io/.

We focus on the following tables:

C01Papers: paper metadata (PubMed articles).

C02LinkPapersAuthors: links each paper to its ordered list of authors (AIDs).

C04ReferenceListPapers: paper‑to‑paper reference list (paper citing papers).

C07Authors: disambiguated author identifiers (AID) in PKG.

C11ClinicalTrials: clinical trials from ClinicalTrials.gov.

C12LinkPapersClinicaltrials: citations from papers to clinical trial records (paper citing clinical trial).

C14Investigators: investigators associated with clinical trials.

C15Patents: patents from USPTO (PatentsView).

C16LinkPatentsPapers: citations from patents to papers (patent citing paper).

C19Inventors: inventors associated with patents.

C20LinkAuthorsInventors: mapping between authors and inventors (AIDs).

Year of publication

We extract publication (or registration) years from C01 (papers), C11 (clinical trials), and C15 (patents), under the assumption that these tables contain one row per entity after deduplication. During quality checks we observed that C11 and C15 contain no duplicates, whereas C01 contains 3,159 PMIDs with multiple rows; we therefore keep only the first row per PMID when creating the year table. The resulting TSV node_year.tsv has:

node_id: normalized identifier.

year: four‑digit year, or empty if year is unavailable in the source tables.

Authors information

We extract authors or analogous contributors for each node and store them as lists of AIDs (PKG’s disambiguated author IDs). We discard names and keep only AIDs, since author names are not stored in the authors tables and the master author table A02AuthorList often contains inconsistent information for the same AID (over 5 million authors have at least two different name variants, and AID 0 aggregates over one million distinct names). In addition, AID 0 is treated as an invalid “unknown author” and is discarded.

The extraction logic is:

Papers: Use `C02LinkPapersAuthors`, selecting all authors for a given PMID and ordering them by column AuthorOrder.

Clinical trials: Use `C14Investigators`, selecting investigators with role PRINCIPAL_INVESTIGATOR (PI) for each clinical trial. For trials with multiple PIs (37,791 cases), order is arbitrary since no explicit ordering column is provided.

Patents: Use `C19Inventors`, selecting all inventors for each patent and ordering them by column InventorSequence.

The resulting TSV node_authors.tsv has:

node_id: normalized identifier.

list_auth: list of AIDs associated with that node, after removing AID 0 and null values, or empty list if authors are unavailable in the source tables.

Citation Matrix

We construct a sparse, directed adjacency matrix over all 40,363,572 nodes capturing three types of citation edges: paper→paper, paper→clinical_trial, and patent→paper. For each citing node, we ran dedicated SQL queries against C04ReferenceListPapers, C12LinkPapersClinicaltrials, and C16LinkPatentsPapers to retrieve the complete list of cited entities, rather than relying on pre‑aggregated views. This strategy ensures coverage of all references while respecting the directional semantics in the source tables. The resulting files are two: one for corresponding column indices (indices.tsv), and one for row pointers indicating the start of each row's data in the other arrays (indptr.tsv).