Datasets:

openadmet
/

pka-modeling-data

license: cc-by-4.0
language:
  - en
tags:
  - chemistry
  - drug-discovery
  - physicochemical-properties
  - pKa
  - protonation
  - molecular-properties
pretty_name: OpenADMET pKa (SiriusT3)
size_categories:
  - 1K<n<10K
task_categories:
  - tabular-regression
annotations_creators:
  - expert-generated
source_datasets:
  - original
configs:
  - config_name: default
    data_files:
      - split: train
        path: pka_measurements.csv

OpenADMET pKa (SiriusT3)

Experimentally measured aqueous pKa values for 717 small molecules, produced through the OpenADMET-Schrodinger collaboration.

pKa describes the pH-dependent equilibrium between protonation states differing by a single proton, and is one of the most fundamental descriptors of a small molecule — it shapes expectations for solubility, permeability and broader ADME behaviour. The compound set was chosen deliberately to include chemistry where prediction is hardest: tautomerisable compounds with uncertain equilibria, electron-deficient heteroaromatics, and heterocycles with sparse literature precedent, alongside more familiar carboxylic acids, phenols and substituted pyridines. Much of it is measured here for the first time.

A reported pKa is not a direct instrument readout. It is the endpoint of a chain of measurement, modelling and interpretation: titration, extrapolation to zero cosolvent, and the choice of how many protonation transitions to fit. This release exposes that last step rather than hiding it — see Choosing a fit.

Blog post: link to be added once published
Produced by: OpenADMET and Schrodinger
Instrument: Pion SiriusT3

Dataset at a glance


Rows	7,246
Compounds	717 (unique `Enamine_ID`, one-to-one with `SMILES`)
Rows per `pKa_model`	1 pKa: 744 rows · 2 pKa: 1,466 rows · 3 pKa: 2,160 rows · 4 pKa: 2,876 rows
Assay method	UV-metric 7,222 rows; pH-metric 24 rows
Fits within the pH 2-12 titration window	79.5%
Rows with manual review (`manual_QC == 1`)	56
Rows carrying a `comment_code`	636

Choosing a fit (read this first)

Each measurement was fitted with 1-, 2-, 3- and 4-pKa models, and every fit is reported. No preferred interpretation has been selected. A compound measured once therefore contributes 1 + 2 + 3 + 4 = 10 rows, and pka_value is only meaningful once you decide how many protonation transitions the molecule actually undergoes in the titration range.

Do not train directly on the pka_value column. Select a pKa_model per compound first.

Picking that model requires an estimate of how many transitions are expected between pH 2 and 12. In the source work this came from Schrodinger's Macro-pKa workflow, which also reports tautomeric microstate populations; the predicted transition count then determined which fitting model to trust. Where several models look mathematically plausible, that external prediction is often the only way to choose correctly.

Signals in the table that help narrow it down:

pka_std — the fitted uncertainty. It climbs sharply with pKa_model as extra parameters go unconstrained (median 0.053 at 1 pKa versus 0.797 at 4).
Whether pka_value falls inside pH 2-12. Outside that range the value is extrapolated, not observed. 79.5% of all rows fall inside; the fraction drops as pKa_model rises.
comment_code and manual_QC — flags from the experimental review.

As a rough indication of coverage, 710 of the 717 compounds have at least one fit that is both inside the window and has pka_std < 1.

Repeated keys

Rows are not unique on (Enamine_ID, pKa_model, pka_index). Where a key repeats, the rows are alternative processings of the same transition — typically an original fit alongside a manually reviewed one (manual_QC == 1, often with comment_code 5), or a UV-metric screen alongside a pH-metric confirmation. Compare manual_QC, assay_method and comment_code to pick between them; the reviewed fit is generally the better one.

How to get the data

With `datasets` (HuggingFace)

from datasets import load_dataset

pka = load_dataset("openadmet/pka-modeling-data", split="train").to_pandas()

With `pandas`

import pandas as pd

pka = pd.read_csv("pka_measurements.csv")

Straight from the Hub without cloning:

import pandas as pd

pka = pd.read_csv(
    "https://huggingface.co/datasets/openadmet/pka-modeling-data/resolve/main/pka_measurements.csv"
)

cosolvent_fractions is stored as a JSON-style list in a string column; parse it if you need the individual titration fractions:

import json

pka["cosolvent_fractions"] = pka["cosolvent_fractions"].apply(
    lambda s: json.loads(s) if isinstance(s, str) else None
)

Column reference

Column	Description
`Enamine_ID`	Catalog identifier used by the vendor Enamine
`SMILES`	Chemical structure of the compound
`assay_method`	`UV-metric` or `pH-metric` (see Methodology)
`pKa_model`	Total number of pKa values SiriusT3Refine modelled to generate this `pka_value`
`pka_index`	Index of this `pka_value` within its model (1..`pKa_model`)
`pka_value`	Extrapolated aqueous pKa at 0% cosolvent, via Yasuda-Shedlovsky extrapolation
`pka_std`	Standard deviation of the extrapolated `pka_value`
`pka_ionic_strength`	Average ionic strength of the assay solution mixture
`pka_temperature`	Average temperature the assay was performed at (degrees C)
`cosolvent`	Cosolvent used to help dissolve the analyte (methanol throughout)
`cosolvent_fractions`	Weight fraction of methanol in the total assay mixture for each of the three titrations. Not available for pH-metric samples
`manual_QC`	Whether the fit was manually reviewed: `0` = no, `1` = yes
`comment_code`	Experimental observations; comma-separated when several apply

`comment_code` values

Code	Meaning
`0`	No additional comments
`1`	Turbidity detected
`2`	Weak signal / noisy spectra
`3`	No detectable transition; possibly outside the titration pH window
`4`	Possible decomposition
`5`	Manual review; titration pH window clipped on inspection for better curve fitting

Methodology

pKa was measured on a Pion SiriusT3. The instrument does not observe pKa directly — methanol is added as a cosolvent to help solubilise the analyte, titrations are run at three methanol-to-water ratios, and the aqueous pKa is inferred by Yasuda-Shedlovsky extrapolation to 0% cosolvent.

Two measurement modes were used, in a two-tier strategy balancing throughput against confidence:

UV-metric (7,222 rows) — tracks changes in UV-vis spectra as a function of pH. Much faster, and needs only microliters of 10 mM DMSO stock, which suits high-throughput workflows. It requires the compound to absorb in the UV and for protonation to measurably shift that spectrum, so it is unsuitable for compounds lacking a chromophore (aliphatic amines and amino acids, for example).
pH-metric (24 rows) — potentiometric titration in triplicate. The canonical method, but lower throughput and needing neat material or concentrated DMSO stock.

Samples were screened by the UV-metric method first, with lower-confidence measurements promoted to pH-metric confirmation. Data were then processed in SiriusT3Refine, which fits the titration data with a user-specified number of pKa transitions — hence the four models reported here.

Assay parameters

Common to both modes: 3 titrations, pH 2-12, methanol cosolvent at 1.15 mL, ISA water 0.35 mL, additional water 0.15 mL at titration 2 and 0.34 mL at titration 3, with no additional cosolvent at either.

UV-metric adds a neutral linear buffer (0.025 mL). pH-metric adds a cleanup step: adjust pH to the starting pH, stir 60 s, add 0.25 mL water, stir 30 s.

Caveats

The table is unassigned. Model selection is left to the user; see Choosing a fit.
Out-of-window values. pka_value spans -25.9 to 38.9, far beyond the pH 2-12 titration range. The extreme values come from over-parameterised fits and are not measurements.
Reactive compounds. Aldehydes in particular are prone to degradation or reaction with the methanol cosolvent, and their fits can disagree with prediction. In some cases agreement was only recovered by modelling the compound as the hemiacetal formed with methanol.
Method mismatch. Compounds without a UV chromophore, or whose spectra do not shift on protonation, can mislead the UV-metric fitting algorithm. Where this was caught, compounds were retested by the pH-metric method.
No logP. A parallel logP study was run on the same instrument, but the remaining uncertainty was not well enough understood to release it; the blog post explains why.

Acknowledgement

We would like to acknowledge everyone at Octant who contributed to this collaborative effort. In particular, we thank Dr. Scott Simpkins, Dr. Steven Edgar, Dr. Robert Warneford-Thomson, Dr. Marco Falcioni, Dr. Ana Lindahl, Dr. Yuning Shen, Calvin Hawe, Danny Tagaloa, Hugh Fox, Dr. Steve Richards, and Dr. Naomi Handly for their valuable contributions and support throughout the project.

We would like to thank our funders for their support of OpenADMET, in particular, ARPA-H, Radial (part of the Astera Institute (https://ror.org/00ydx1s47)), Schrödinger Inc., the Gates Foundation, and the OpenAI Foundation. We would also like to thank our partners Enamine, HuggingFace, OpenEye, CDD Vault, Discovery Life Sciences, and the beamline staff at NSLS-II for their support.

This work is supported by the Advanced Research Projects Agency for Health (ARPA-H) under AVOID-OME, and Award Number 1AY1AX000035. The contents are those of the authors. They may not reflect the policies of the Department of Health and Human Services or the U.S. government. The content is solely the responsibility of the authors and does not necessarily represent the official views of the Advanced Research Projects Agency for Health.