Published October 19, 2026 | Version v2

Verified external pharmacokinetic data for drugs approved from 2017, and the AstraZeneca assay provenance behind TDC's clearance and plasma protein binding endpoints

Description

Supporting data for two manuscripts: "Two undeclared data defects in the pharmacokinetic endpoints of the TDC ADMET benchmark", and "A plasma protein binding benchmark that cannot support the use its endpoint is for".

What this contains

Three things, kept together because the papers use them together.

1. A temporal external set. Plasma protein binding (n = 129) and terminal half-life (n = 76) for small molecules first approved by the FDA from 2017 onward, after every TDC pharmacokinetic source dataset was compiled. Values were extracted from FDA product labels, then read individually against their own source sentence by hand; half-life was read twice. Compounds appearing in TDC, or in PKSmart's released external sets, were removed. Every accept, correct, reject and withhold judgement from the hand review is included with its reason, so the curation can be audited rather than taken on trust.

2. Assay provenance. The four AstraZeneca assays as deposited in ChEMBL (document CHEMBL3301361), retaining two fields that the benchmark's conversion drops: the species of each assay, and the "<" or ">" qualifier marking a value at a detection limit. These are the files behind the first paper's two principal findings, that Clearance_Hepatocyte_AZ merges a human and a rat assay, and that a quarter to a third of the clearance labels sit exactly on the assays' declared detection limits.

3. New in this version. The species composition of the plasma protein binding endpoint. The same deposition contains five binding assays, one per species: human, rat, dog, mouse and guinea pig, holding 2828 records over 1797 distinct molecules. TDC's PPBR_AZ holds 2790 rows over exactly those 1797 molecules. Matching each benchmark row against all five assays by canonical SMILES and value, 97.5 percent match exactly one assay, and of those 57.6 percent are human, 25.0 percent rat, 8.6 percent dog, 5.6 percent mouse and 3.2 percent guinea pig. So 42.4 percent of the endpoint is not human plasma. The duplication this creates also puts a floor under the benchmark's metric: the test partition is 559 rows over 343 molecules, 140 of which carry labels that disagree by up to 49.3 percentage points, so 1.964 percentage points of mean absolute error cannot be removed by any model.

Cautions on use

Roughly a fifth of the binding values in the external set are censored, reported as "> 99%" or similar rather than measured. The censoring column says which. Treating those as point measurements is the error the first paper is about.

Half-life values marked is_apparent come from oral studies and are not comparable with intravenous terminal half-life. They are supplied but excluded from the verified set.

PPBR_AZ rows are per measurement, not per molecule. A molecule measured in several species appears several times. No molecule appears in both the train and test partitions, so this is not a leakage finding.

Files

Each file ships with a column description in README.md. The three added in this version are ppb_all_species_upstream.csv (all five assays as fetched), ppb_species_composition.csv (the row-level attribution) and ppb_metric_floor.csv (the floor computation).

Licence

CC BY 4.0. The underlying FDA label text is a US government work; the ChEMBL deposition is CC BY-SA 3.0. This deposit covers the curation, verification and provenance annotation.

Code

https://github.com/MinhNhat-2504/ivive-pk-multitask

Files

az_assays.csv

Files (3.6 MB)

Name Size
md5:40d3567de89b2fb57284d71871f00b49
548.4 kB Preview Download
md5:ff0459eea8c252d48e9677f9e5236cbb
2.4 MB Preview Download
md5:c4b56b523bea60d13d20819f2c8a5028
21.2 kB Preview Download
md5:470a6a92f9d3d384231b97be28e5bb83
1.1 kB Preview Download
md5:9962f6951fcf9bd4eba9e82e973fe96b
30.7 kB Preview Download
md5:6b613b2bd4abe4ec3e28cc10a9a4c1d3
209.0 kB Preview Download
md5:196641efaaec9b3ff2573c6e3783d261
340.4 kB Preview Download
md5:085c4ed6bca1ead4cfc79602a4244aea
35.3 kB Preview Download
md5:d87c078c38eb3228f00e22fabfb76122
1.6 kB Preview Download
md5:d74a2c1a1db29c32fc63e91c240e3706
164 Bytes Preview Download
md5:56f884057e91eb57135dd2164012871a
205 Bytes Preview Download
md5:9f7d84f21fa6f8e9bf4627adeb7d398c
6.2 kB Preview Download

Additional details

Related works

Is derived from
Dataset: 10.6019/CHEMBL3301361 (DOI)
Is supplemented by
Software: https://github.com/MinhNhat-2504/ivive-pk-multitask (URL)

Dates

Collected
2017-01-01/2026-10-19
FDA approval dates of the included compounds