Data, cleaning and downloads

The complete imported inventory remains available. Cleaning assigns decisions and coordinates compatible records without discarding the archived source rows.

From the full source archive to representative records

174,082All source rows Complete imported inventory

150,994Eligible source rows After quarantine; before coordination

149,189Representative records Conservative within-source coordination

39,974Unique molecules Distinct parsed canonical structures

23,088 source rows are quarantined with recorded reasons; 1,805 eligible rows share representatives through within-source coordination. Cross-source records remain separate.

The representative layer contains 39,974 parsed structure entities, 1,612 source-specific protein-site context keys, and 10,380 records with unresolved identity. These are different units, not a combined compound count.

Full downloads, cleaning rules and verification

Download each layer

These complete CSV archives have no API row cap. Files use gzip compression; retain source-row IDs, original fields, decisions and inherited review annotations when reusing them.

Layer / audit table	Rows	Download
All source rows	174,082	all_source_rows_with_decisions.csv.gz 6.35 MB compressed SHA-256 `cabfde1be396eff6ff3ad320fc75b0af2f9caf4024f6aa246375b0aef8320c40`
Eligible source rows	150,994	eligible_source_rows.csv.gz 5.50 MB compressed SHA-256 `890d9ed7384a76c615040f1f2a34afe8c9875ece66067f5b5f3bfa14525225e2`
Representative records	149,189	pkaatlas_v3_candidate_records.csv.gz 8.39 MB compressed SHA-256 `3f933fe3edbe29bb3114f143eb363a0250999b191c2046ed034afd445b7bb366`
Quarantined source rows	23,088	quarantined_rows.csv.gz 0.78 MB compressed SHA-256 `81ff22966c13882bb28c438c9bd7e817fed01a6790d132d24815db8fa1b38322`
Source-row membership	150,994	source_row_membership.csv.gz 4.34 MB compressed SHA-256 `703b44f8fe057f234af361f10d91b0605b574f20595a0bedbfb86d8f568a0507`
Unresolved representative identities	10,380	identity_resolution_queue.csv.gz 0.29 MB compressed SHA-256 `4e2c262b14541dc523d178b6aefbfbbf77934d88ed80fae90e952da8da8bd6bc`
Source QC ambiguity evidence	See file	source_qc_ambiguous_keys.csv.gz 0.01 MB compressed SHA-256 `4f143b0e467713f0c38e37b98812b2c51f5dd52125469dd2bd6ed05650856132`

Release manifest, counts and checksums (JSON)

Representative coverage and inherited annotations

Denominator: 149,189 representative records. A populated field is not proof of correct experimental context or original-source verification.

Quantity	Representative records
Identity: structure_defined	137,111
Identity: unresolved_record	10,380
Identity: protein_site_context	1,698
Inherited label: experimental	149,106
Inherited label: predicted	83
Inherited grade C	13,031
Inherited grade A	7,289
Inherited grade B	128,786
Inherited grade D	83
Populated temperature_numeric_c	73,801
Populated reference	149,186
Populated doi	149,189
Populated coconut_id	13,831
Populated np_superclass	10,570
Populated structure_smiles	137,111
Identity/source-quality review flags	10,936

Inherited DOI-label acquisition groups

These eleven groups differ from the ten stored dataset labels and the eighteen nested pKaHub source labels. Known inherited DOI mismatches remain in row-level audit fields.

Inherited DOI label	Representative records
`10.5281/zenodo.7236452`	22,907
`10.1021/acs.jcim.6c00107`	99,257
`10.1021/jacs.5c02578`	8,233
`10.1093/nar/gkx1037`	690
`10.1142/s2737416525500164`	78
`10.5281/zenodo.11225612`	166
`10.5281/zenodo.12608876`	9,721
`10.1021/acs.jcim.5b00120`	5,628
`10.1021/ci800162h`	181
`10.1093/nar/gky1075`	708
`10.1093/nar/gkae255`	1,620

Sequential quarantine rules

Rules run in the order below. The first reason determines the exclusive quarantine count; all_qc_reasons retains overlapping reasons. Quarantine marks records requiring review; it does not establish that an original experiment was wrong.

Numerical rules quarantine missing/non-numeric pKa, pKa outside −50 to 100, water-labeled pKa above 60, and populated temperatures that are non-numeric or outside −100 to 300 °C. Missing temperature alone does not cause quarantine.

pKaHub quality decisions match nested dataset, parsed structure and pKa rounded to three decimals, preserving original row positions. Conflicting source-key decisions, unmatched keys and unresolved structural matches remain review flags instead of selecting an arbitrary source row.

Rule	Newly quarantined	Remaining	All rule hits (may overlap)
`pka_non_numeric_or_empty`	888	173,194	888
`pka_impossible`	4	173,190	4
`pka_water_impossible`	1	173,189	5
`temp_invalid`	1,536	171,653	1,561
`pkahub_solvent_unlabeled_nonwater`	11,010	160,643	11,010
`pkahub_cosolvent_mixed`	2,639	158,004	2,639
`pkahub_assignment_outlier`	5,073	152,931	5,073
`pkahub_suspicious_remark`	1,305	151,626	1,305
`pkahub_assignment_error_large`	632	150,994	637
`pkahub_value_unparseable`	0	150,994	690
`pkahub_inequality_value`	0	150,994	0

Identity and coordination standards

Structure parsing tries canonical SMILES, SMILES and then InChI. Parsed structures use isomeric canonical SMILES. No salt stripping, tautomer consolidation or inferred stereochemical correction is applied. Protein sites use source-specific context keys; unresolved records retain separate row identities.

Cross-source records are never merged. Within-source coordination requires a parsed structure and populated reference, temperature, pKa type, solvent and dataset. It also requires equality of the exact unrounded numerical pKa, data type, dataset, source detail, DOI, inherited grade, method, year annotations and review flags. Membership maps preserve every eligible source row.

How to interpret the data

149,106 representatives inherit experimental labels and 83 inherit predicted labels. These are not all experimentally validated small-molecule measurements.
Grades A–D are inherited annotations, not calibrated confidence scores or independent original-paper verification. Epik-based charge-state assignment is not an experimental measurement method.
Populated DOI/reference fields can identify compilations, contain inherited mismatches or require reconciliation. Row details retain provenance_status, review_flag and license_status.
Original-publication years are available for 22,305 representatives (1889–1973). Compilation years describe a different event.
Upstream source-specific terms apply. This website grants no blanket license over the integrated data; the frozen candidate's pending review annotations remain in all downloads.

Version alignment and checks

The earlier v2 snapshot has 87,308 rows and 42,813 distinct nonempty canonical identifier strings, including protein identifiers. Its web loader exposed 82,756 rows after skipping missing structure identifiers. Those counts are historical and do not describe this release.

Representative data SHA-256: 2836efd9facd950a29c9c2dff5e0765cb62f15da833d634c4e2574ed523720f5. Original field preservation was independently rechecked for 174,082 rows; eligible membership is unique and group-size sums agree. These checks verify processing integrity, not original experimental accuracy.