The All-Route Terminal Half-Life Dataset is a curated collection of terminal pharmacokinetic half-life observations reported in the biomedical literature. It contains 19,952 records from 3,270 PubMed Central articles and covers oral, intravenous, subcutaneous, and other administration routes.

Each row represents a paper-, analyte-, regimen-, cohort-, occasion-, and statistic-specific terminal half-life observation. The normalized numeric value is expressed in hours, while the source-facing reported value and its uncertainty or dispersion remain available as text.

This dataset was created using Paperclip Datasets, GXL's AI-powered tool for transforming scientific literature into structured datasets. You can create your own datasets or read our announcement to learn more about how it works.

Dataset summary

Property	Value
Records	19,952
Source papers	3,270
Columns	14
Oral records	10,002
Intravenous records	7,705
Subcutaneous records	820
Other-route records	1,425
Records with paired SMILES and InChIKey	12,434
Records with a normalized drug name	16,874
Records with a molecule class	19,952

Release version: 2026-10-26.2
Release date: 2026-10-26
Release SHA-256: 89584232b5fbb52bf08a14c3c8fecabc818646b33e5fc7600a8447778ae21a90

Data format

The repository contains one CSV file:

data/all_route_half_life_release.csv

Hugging Face exposes the table as the train split for compatibility with the datasets library. This split name is a storage convention; the dataset was not designed as a predefined machine-learning training set and does not include official validation or test partitions.

from datasets import load_dataset

dataset = load_dataset("gxl-ai/paperclip-half-life")
records = dataset["train"]
print(records)

Schema

Column	Type	Description
`row_id`	string	Stable opaque identifier for the curated record.
`pmcid`	string	PubMed Central identifier for the source article.
`drug_name`	string	Analyte or drug label used for the extracted record. This may include metabolites, total antibody, radioactivity, transgene, or other measured entities.
`normalized_drug_name`	string, nullable	Normalized analyte or compound name when an identity could be resolved.
`molecule_class`	enum	Single controlled class for the measured analyte or presentation. See the vocabulary and interpretation notes below.
`smiles`	string, nullable	Validated SMILES representation when an appropriate structure could be assigned.
`inchikey`	string, nullable	InChIKey corresponding to the assigned SMILES. SMILES and InChIKey are either both populated or both blank.
`administration_route`	enum	Normalized route: `oral`, `iv`, `subcutaneous`, or `other`.
`terminal_half_life`	float	Terminal half-life in hours.
`reported_half_life`	string	Source-facing value, units, statistic, and available dispersion or interval information.
`summary_statistic`	enum	Controlled classification of the statistic represented by `terminal_half_life`.
`drug_information`	string	Regimen, formulation, dose, analyte, assay, sampling, and PK-occasion context when available.
`cohort_information`	string	Population, subgroup, sample size, demographic, disease, and eligibility context when available.
`notes`	string	Source interpretation, unit conversion, terminal-phase qualification, and provenance caveats.

Blank strings represent unavailable values. They should not be interpreted as zero or as evidence that a compound lacks a structure.

Molecule-class vocabulary

Every record has one molecule_class value:

Value	Records	Meaning
`small_molecule`	14,904	Discrete small-molecule analyte.
`antibody`	2,564	Antibody or antibody-like therapeutic without a classified drug conjugate.
`protein`	854	Protein therapeutic or protein analyte not classified more specifically as an antibody or enzyme.
`peptide`	620	Peptide or peptide-like analyte.
`other`	401	Analyte or presentation not represented by the other controlled classes, including some total-radioactivity, elemental, gas, cellular, polymer-bound, or formulation-level measurements.
`antibody_drug_conjugate`	175	Antibody–drug conjugate or conjugated-antibody analyte.
`enzyme`	172	Enzyme therapeutic or measured enzyme activity.
`oligonucleotide`	157	DNA-, RNA-, aptamer-, or related oligonucleotide analyte.
`oligosaccharide`	105	Oligosaccharide, cyclodextrin, or related carbohydrate analyte.

molecule_class describes the measured analyte or presentation in that record, not necessarily the normalized parent structure alone. Consequently, records with the same normalized_drug_name or InChIKey can legitimately have different classes—for example, a free small molecule versus its polymer-conjugated or carrier-bound measurement.

Summary-statistic vocabulary

summary_statistic uses the following controlled values:

Value	Meaning
`mean`	The source says mean or average without establishing whether it is arithmetic or geometric.
`arithmetic_mean`	Explicitly identified as an arithmetic mean.
`geometric_mean`	Explicitly identified as a geometric or back-transformed log-scale mean.
`harmonic_mean`	Explicitly identified as a harmonic mean.
`median`	Explicitly identified as a median.
`individual_value`	Value for one participant or patient.
`model_estimate`	Model-derived point estimate without a more specific reported statistic.
`pooled_estimate`	Pooled, profile, or group estimate without a more specific reported statistic.
`typical_value`	Explicit typical-population or typical-participant estimate.
`unspecified`	The source does not identify the governing summary statistic.

The release contains 317 unspecified records and no range_endpoint records. Consult reported_half_life and notes before treating unspecified values as conventional point estimates.

Curation and processing

The release was produced through an iterative literature-curation and validation workflow:

Candidate terminal half-life observations were extracted from biomedical articles associated with PMCIDs.
Terminal-phase qualification, analyte identity, route, regimen, cohort, sampling occasion, source statistic, and units were reviewed from the available article evidence.
Numeric half-lives were converted to hours without averaging source-reported ranges or inventing unavailable point estimates.
Source-facing values, dispersion measures, intervals, and contextual qualifications were retained in text fields.
Duplicate candidates were adjudicated using analyte, dose, regimen, cohort, PK occasion, statistic, and source-value context. Equal numeric values alone were not treated as duplicates.
Chemical identifiers were checked as paired SMILES/InChIKey values, and normalized-name casing was consolidated deterministically.
Each record was assigned one controlled molecule_class describing the measured analyte or presentation.
Automated release validation checked row identifiers, PMCIDs, required fields, half-life positivity, identifier pairing and formatting, controlled vocabulary values, and duplicate keys.

The duplicate-review workflow included LLM-assisted comparison of 218 candidate pairs across 213 high-confidence groups. It removed 215 redundant rows and retained comparisons representing distinct doses or PK occasions. LLM assistance was used as a curation aid rather than as a substitute for the source-derived evidence recorded in each row.

Data quality and known limitations

terminal_half_life is standardized to hours, but pharmacokinetic methods, sampling duration, matrices, terminal-phase definitions, and model assumptions vary across papers.
Some values are apparent terminal half-lives or reflect absorption-, depot-, implant-, transgene-, cellular-kinetic-, or formulation-limited behavior. Review the analyte and regimen context before comparing records.
The dataset includes drugs, metabolites, biomarkers, total radioactivity, antibodies, proteins, transgenes, and other measured analytes. drug_name should not be assumed to represent only a parent small molecule.
3,078 records lack a normalized drug name.
7,518 records lack SMILES and InChIKey. Structures are commonly unavailable or inappropriate for mixtures, biologics, cells, polymers, radioactivity measurements, incompletely specified analytes, or unresolved identities.
317 records have an unspecified summary statistic.
Molecule classes are single-label, coarse categories and do not encode every relevant modality, formulation, conjugation state, or biochemical role. Interpret them together with drug_name, drug_information, and notes.
The notes explicitly flag 53 records curated from abstracts without full-text review and two records for which full text was unavailable. Absence of a source-scope statement should not be interpreted as a formal certification of full-text review.
Text fields are curated summaries and may retain study-specific terminology. Users should consult the cited paper before drawing scientific or clinical conclusions.
Deduplication substantially reduces redundant extraction records but cannot guarantee that no conceptually overlapping observations remain under every possible interpretation.

Intended uses

Appropriate uses include:

pharmacokinetic literature review;
exploratory analysis of reported terminal half-lives;
compound and analyte comparison with appropriate stratification;
retrieval, extraction, normalization, and data-quality research;
hypothesis generation and benchmarking of biomedical information-extraction systems.

Out-of-scope uses

This dataset is not clinical guidance and should not be used by itself for prescribing, dose selection, therapeutic monitoring, regulatory decisions, or patient-level risk assessment. Values from different matrices, populations, regimens, assays, sampling schedules, and analytical methods are not automatically comparable.

The dataset is also not an authoritative chemical registry. Chemical identifiers should be validated against an appropriate reference source before use in safety-critical or structure-dependent applications.

Personal and sensitive information

The dataset is derived from published biomedical literature. Cohort descriptions are generally aggregate. Some source articles report individual pharmacokinetic observations using study-level participant labels; the dataset does not intentionally add direct personal identifiers beyond what is present in the cited publication.

Source attribution

Each record includes a pmcid linking it to its source article. Users should cite the relevant source publications when reusing individual observations or subsets. The dataset does not replace the original articles, tables, supplements, or study reports.

Licensing

No blanket license is declared in the dataset-card metadata. The dataset combines factual measurements with curated, source-derived contextual descriptions from publications that may have different licenses and reuse terms. Rights in the underlying articles remain with their respective authors and publishers.

Before making the repository public, the maintainer should select and add an appropriate repository license after reviewing the intended distribution of the descriptive text fields. A dataset-level license cannot grant rights that the maintainer does not hold in third-party source material.

Citation

Until a DOI or archival citation is assigned, cite the versioned Hugging Face repository and include the release date and checksum. Cite the original paper identified by pmcid when referring to a specific observation.

Suggested citation text:

GXL. All-Route Terminal Half-Life Dataset. Version 2026-10-26.2. Hugging Face Datasets. SHA-256 89584232b5fbb52bf08a14c3c8fecabc818646b33e5fc7600a8447778ae21a90.

Versioning

This card documents release 2026-10-26.2. Relative to release 2026-10-26.1, it adds the fully populated molecule_class column while preserving all 19,952 row IDs and every existing field value. See CHANGELOG.md and the dated audit artifacts under audits/ for details.

Future releases should preserve stable row_id values for unchanged observations, publish a new file checksum, and describe additions, removals, or field-level revisions in the repository history.

Repository layout

README.md                                Dataset card and usage documentation
CHANGELOG.md                             Release history
data/all_route_half_life_release.csv     Canonical dataset file
metadata/release_manifest.json           Machine-readable release metadata
audits/2026-10-26/                       Range-endpoint review evidence
audits/2026-10-26-molecule-class/        Molecule-class validation evidence

Downloads last month: 195

Total file size:

34 MB