kawow 0.1.6

Models for partitioning coefficients (logKow, logKoa and logKaw) from molecular structure, including model based on [Naef & Acree 2024](https://doi.org/10.3390/liquids4010011)

Kawow (under development)

Group-additivity prediction of logKow, logKoa, and logKaw from molecular structure.

Kawow implements some models to predict partitioning coefficients (logKoa, logKow and logKaw), in particular the Naef & Acree (2024) group-additivity scheme using RDKit SMARTS pattern matching. Two model families are available depending on how much transparency or accuracy is required.

Flagging criteria used in outputs

run_models(...) returns B/vB, M/vM, and regulatory-gap flags derived from predicted partition values.

Bioaccumulation (following doi:10.1126/science.1138275):

Flag	Condition
B	logKoa ≥ 6 and logKow ≥ 2
vB	logKoa ≥ 6 and logKow ≥ 5

Mobility — estimated via logKoc_est = logKow − 0.4 (UBA drinking-water guidance, link):

Flag	Condition on logKoc_est	Equivalent logKow
M	logKoc_est ≤ 4.5	logKow ≤ 4.9
vM	logKoc_est ≤ 3.5	logKow ≤ 3.9

Regulatory gaps (returned as in_gap1, in_gap2, in_gap3 booleans and gap_labels list):

Gap	Condition
Gap 1	3.5 < logKow < 5.0 (between vM and vB on Kow axis)
Gap 2	logKow > 4.5 and logKoa < 6 (non-M, non-B)
Gap 3	4.5 < logKow < 5.0 and logKoa < 6 (intersection of Gap 1 and Gap 2)

---

Installation

pip install kawow

Or from source (requires RDKit ≥ 2022.9):

git clone https://github.com/LucMiaz/kawow.git
cd kawow
pip install -e ".[dev]"

---

Models at a glance

logKow and logKoa R² values are from 5-fold cross-validation on the shared S01∩S02 benchmark (n = 3 319 for logKow; n = 1 956 for logKoa), except smarts which applies fixed Naef & Acree (2024) parameters directly (no re-fitting). logKaw R² is from external validation on the full S03 dataset (Naef & Acree 2024; n = 2 130–2 150; S03 was not used for training).

Model key	Class	Approach	logKow R²	logKoa R²	logKaw R²
kawow	PartitionCalculator	Ridge regression on Crippen atom-type counts + Naef special-group features	0.898 (cv)	0.937 (cv)	—
smarts	NaefAcreePartitionCalculator	Pure Naef & Acree 2024 group-additivity (no re-fitting, tabulated parameters only)	0.857	0.785	0.654
smarts_mixed	NaefAcreeCrippenMixedPartitionCalculator	Naef & Acree SMARTS contributions + Crippen atom-type Ridge hybrid	0.938 (cv)	0.943 (cv)	0.912
mqg	MQGPartitionCalculator	Ridge regression ensemble: Naef group contributions + Crippen atom types + Molecular Quantum Graph fingerprints	0.940 (cv)	0.942 (cv)	0.913
pfasgroups	PFASGroupsPartitionCalculator	Ridge regression on 77-dim PFASGroups halogenated-group descriptor	— (cv)	— (cv)	—
pfasgroups_mixed	PFASGroupsPartitionCalculator	Ridge regression on PFASGroups (77-dim) + Crippen atom-type (77-dim) concatenated features	— (cv)	— (cv)	—

Use run_models() to run several models at once and get per-molecule B/vB and M/vM flags:

import kawow

results = kawow.run_models(
    ["CCCCO", "c1ccccc1", "OC(=O)c1ccccc1"],
    models=["kawow", "smarts_mixed"],
)
for row in results:
    print(row["smiles"], row["models"]["kawow"]["logKow"],
          row["models"]["kawow"]["b_class"])

Each element of the returned list is a dict with:

Key	Description
smiles	canonical SMILES
name	molecule name from input
models	dict keyed by model name; each value contains logKow, logKoa, logKaw, b_class, m_class, ok
ok	True if at least one model succeeded

---

1 — PartitionCalculator (recommended for most uses)

Ridge regression fitted on the same S01/S02 datasets. Coefficients are stored in kawow/data/*.json so no re-fitting is needed at import time.

from kawow import PartitionCalculator

calc = PartitionCalculator()           # Ridge (default)

# Single molecule from SMILES
result = calc.predict("CCCCO")        # 1-butanol
print(result)
# {'logKow': 0.88, 'logKoa': 4.12, 'logKaw': -3.24, 'status': 'ok'}

# Batch prediction
smiles = ["c1ccccc1", "CCCCCCCCCC", "OC(=O)c1ccccc1"]
for r in calc.predict_batch(smiles):
    print(r["smiles"], r["logKow"], r["logKoa"], r["logKaw"])

Predict from an InChI string or SDF file:

r = calc.predict("InChI=1S/C4H10O/c1-2-3-4-5/h5H,2-4H2,1H3")
results = calc.predict("compounds.sdf")   # returns list[dict]

Inspect model metadata:

info = calc.model_info
print(info["logKow"])
# {'target': 'logKow', 'n_train': 3234, 'alpha': 51.8,
#  'r2_cv': 0.8980, 'rmse_cv': 0.6643,
#  'ccc_cv': 0.9470,          # Lin's concordance correlation coefficient
#  'nrmse_sd_cv': 0.2770,     # RMSE / σ(logKow_train)
#  'nrmse_range_cv': 0.0580,  # RMSE / range(logKow_train)
#  'bf10_log10_cv': '>10^300',# log₁₀ Bayes factor (H₁: ρ ≠ 0)
#  'r_ci95_cv': [0.945, 0.951],  # 95 % CI on Pearson ρ
#  ...}

Re-fit on your own training data:

import kawow
kawow.fit(
    sdf_logkow="my_logkow.sdf",
    sdf_logkoa="my_logkoa.sdf",
    logkow_prop="logP",
    logkoa_prop="logKoa",
)
calc = kawow.PartitionCalculator()   # reload after fitting

Performance (Naef & Acree benchmark datasets)

All values are from 5-fold cross-validation on the shared S01∩S02 intersection, except smarts (no fitting; evaluated on full S01/S02/S03) and the logKaw rows (external validation on full S03 — S03 was never used for training).

Model	Property	n	R²	RMSE	Note
kawow (Crippen Ridge)	logKow	3 319	0.898	0.664	5-fold CV
kawow (Crippen Ridge)	logKoa	1 956	0.937	0.740	5-fold CV
smarts (Naef & Acree)	logKow	3 344	0.857	0.786	external (S01 full)
smarts (Naef & Acree)	logKoa	1 983	0.785	1.387	external (S02 full)
smarts (Naef & Acree)	logKaw	2 150	0.654	1.758	external (S03 full)
smarts_mixed (hybrid)	logKow	3 319	0.938	0.518	5-fold CV
smarts_mixed (hybrid)	logKoa	1 956	0.943	0.702	5-fold CV
smarts_mixed (hybrid)	logKaw	2 150	0.912	0.886	external (S03 full)
mqg (ensemble)	logKow	3 319	0.940	0.510	5-fold CV
mqg (ensemble)	logKoa	1 956	0.942	0.705	5-fold CV
mqg (ensemble)	logKaw	2 130	0.913	0.882	external (S03 full)

Regulatory classification performance (F1 scores)-

Binary classification F1 scores on the shared S01∩S02 benchmark (1 083–1 102 molecules with paired experimental logKow and logKoa). Flags are applied to predicted values using the same thresholds as run_models(). naef_mqg and crippen_mqg are available via EnsemblePartitionCalculator.

Label	Condition	n (+)	kawow	smarts	smarts_mixed	naef_mqg	crippen_mqg	mqg
G1	3.5 < logKow < 5.0	178	0.67	0.74	0.77	0.77	0.69	0.55
G2	logKow > 4.5 and logKoa < 6	24	0.56	0.54	0.62	0.63	0.58	0.15
G3	4.5 < logKow < 5.0 and logKoa < 6	11	0.00	0.27	0.13	0.13	0.00	—
M	logKoc_est ≤ 4.5	797	0.97	0.98	0.98	0.98	0.97	0.95
vM	logKoc_est ≤ 3.5	677	0.95	0.96	0.97	0.97	0.95	0.95
B	logKow ≥ 2 and logKoa ≥ 6	503	0.94	0.94	0.96	0.95	0.95	0.94
vB	logKow ≥ 5 and logKoa ≥ 6	266	0.92	0.93	0.93	0.94	0.93	0.79

n (+): true-positive molecule count. — = model makes 0 positive predictions (precision undefined). G3 has only 11 true-positive molecules; most models do not recover this rare class.

---

2 — NaefAcreePartitionCalculator (SMARTS additivity, full transparency)

Implements the Naef & Acree 2024 method exactly: each SMARTS pattern from the paper's supplementary tables is matched against the molecule and its tabulated contribution added. No matrix regression — every contribution is directly interpretable.

from kawow.smarts_model import NaefAcreePartitionCalculator

calc = NaefAcreePartitionCalculator(smiles="c1ccccc1")
result = calc.predict("c1ccccc1")
# {'logKow': 2.13, 'logKoa': 2.80, 'logKaw': -0.67, 'in_coverage': True}

# Or pass a pre-built RDKit mol:
from rdkit import Chem
mol = Chem.MolFromSmiles("CCCCCCCCCC")
result = calc.predict(mol)

# Batch via constructor:
calc_batch = NaefAcreePartitionCalculator(
    smiles=["c1ccccc1", "CCCCCCCCCC", "OC(=O)c1ccccc1"]
)
for mol, coeffs in calc_batch.results.items():
    print(coeffs)

Performance (Naef & Acree tabulated parameters, evaluated on benchmark sets)

The smarts model applies the published Naef & Acree (2024) parameters without any re-fitting. Performance is evaluated on the full individual datasets and on the shared S01∩S02 benchmark intersection.

Dataset	Property	n	R²	RMSE
S01 (Naef 2024, full)	logKow	3 344	0.857	0.786
S02 (Naef 2024, full)	logKoa	1 983	0.785	1.387
S03 (Naef 2024, full)	logKaw	2 150	0.654	1.758
S01∩S02 intersection	logKow	3 319	0.857	0.785
S01∩S02 intersection	logKoa	1 956	0.777	1.387

The remaining error is concentrated in specific chemotypes (notably highly heteroatom-rich agrochemical scaffolds), while the broad SMARTS generalization and pi-environment fixes substantially improved overall logKoa performance on S02.

Correlation plots

logKow vs Naef S01	logKoa vs Naef S02	logKow vs Arp & Hale

---

Feature engineering

Each molecule is represented by counts of SMARTS atom-type groups from the Naef & Acree parameter tables, plus five special-group descriptors:

• pi-neighbour moieties — the number of conjugated systems adjacent to a centre atom (controls which entry in a pi-stratified table applies; computed by count_conjugated_neighbor_moieties)
• H-acceptor binary presence — 1 if any intramolecular H-bond donor/acceptor pair is within 5 bonds
• Alkane flag — 1 if the molecule is a pure saturated hydrocarbon
• Unsaturated HC flag — 1 if the molecule is a pure unsaturated hydrocarbon
• Extra −COOH count — number of carboxylic acid groups beyond the first
• Endocyclic C−C single bond count

The PartitionCalculator additionally uses 72 Crippen atom-type features (from RDKit's Crippen.txt) on top of the 5 Naef special groups.

Extended metrics

In addition to R² and RMSE, kawow reports the following metrics for each trained endpoint:

Metric	Symbol	Description
CCC	—	Lin's concordance correlation coefficient (Lin 1989): combines precision and accuracy in [−1, 1].
NRMSE (σ)	nrmse_sd_cv	RMSE / σ(y_train): scale-free error relative to training set spread.
NRMSE (range)	nrmse_range_cv	RMSE / (max − min)(y_train): error as a fraction of the data range.
log₁₀(BF₁₀)	bf10_log10_cv	BIC-approximated Bayes factor for H₁: ρ ≠ 0 (Ly et al. 2015). Values > 1 decisively support the correlation hypothesis. At n > 3000, all models give ">10^300".
95 % CI on ρ	r_ci95_cv	Fisher z-transform confidence interval on the cross-validated Pearson correlation.

Model	Property	n	CCC	NRMSE (σ)	NRMSE (range)	log₁₀(BF₁₀)
kawow	logKow	3319	0.946	0.319	0.039	>10^300
kawow	logKoa	1956	0.968	0.252	0.043	>10^300
smarts	logKow	3319	0.925	0.378	0.046	>10^300
smarts	logKoa	1956	0.894	0.472	0.082	>10^300
smarts_mixed	logKow	3319	0.968	0.249	0.030	>10^300
smarts_mixed	logKoa	1956	0.971	0.239	0.041	>10^300
mqg	logKow	3319	0.483	0.817	0.099	>10^300
mqg	logKoa	1956	0.783	0.594	0.102	>10^300
naef_crippen_mqg	logKow	3319	0.969	0.245	0.030	>10^300
naef_crippen_mqg	logKoa	1956	0.970	0.242	0.042	>10^300
pfasgroups	logKow	3319	0.880	0.464	0.056	>10^300
pfasgroups	logKoa	1956	0.945	0.322	0.056	>10^300
pfasgroups_mixed	logKow	3319	0.954	0.298	0.036	>10^300
pfasgroups_mixed	logKoa	1956	0.970	0.242	0.042	>10^300
pfasgroups_naef	logKow	3319	0.969	0.246	0.030	>10^300
pfasgroups_naef	logKoa	1956	0.968	0.247	0.043	>10^300
pfasgroups_naef_mixed	logKow	3319	0.970	0.240	0.029	>10^300
pfasgroups_naef_mixed	logKoa	1956	0.970	0.239	0.041	>10^300

Access these via model_info (see code example above) or via the benchmark script output (benchmark_results.csv).

Y-randomization (permutation test)

Run via shared_fold_benchmark.py --y-randomization (default: 1 000 permutations, Ridge 5-fold CV). Trainable models are tested (kawow, smarts_mixed, mqg, naef_crippen_mqg, pfasgroups, pfasgroups_mixed, pfasgroups_naef, pfasgroups_naef_mixed); the pure SMARTS lookup is excluded. Results are saved to tests/out/y_randomization.csv.

A non-significant p-value (fraction of permuted R² ≥ observed R²) confirms the model captures genuine structure–property relationships rather than overfitting to label order. Use the CSV directly for publication numbers, because observed values depend on run settings (e.g., max-samples/permutation count) while permutation baselines remain near zero.

---

Reference

Naef, Rudolf, and William E. Acree, Jr. 2024. "Calculation of the Three Partition Coefficients logPow, logKoa and logKaw of Organic Molecules at Standard Conditions at Once by Means of a Generally Applicable Group-Additivity Method." Liquids 4, no. 1: 231–260. 10.3390/liquids4010011

Arp, H.P.H. and Hale, S.E. 2023. "From Measured Partition Coefficients to the Prediction of Environmental Fate." Supplementary data: vg2c00024_si_001 (ACS).

Lin, L.I.-K. 1989. "A Concordance Correlation Coefficient to Evaluate Reproducibility." Biometrics 45: 255–268. 10.2307/2532051

Ly, A., Verhagen, A.J., and Wagenmakers, E.-J. 2015. "Harold Jeffreys's Default Bayes Factor for Testing Point Null Hypotheses from a Continuous Prior Distribution." Journal of Mathematical Psychology 28: 71–84. 10.1016/j.jmp.2015.06.004

Steiger, J.H. 1980. "Tests for Comparing Elements of a Correlation Matrix." Psychological Bulletin 87: 245–251. 10.1037/0033-2909.87.2.245

Chicco, D., Tötsch, N., and Jurman, G. 2021. "The Matthews Correlation Coefficient (MCC) Is More Reliable Than Balanced Accuracy." BioData Mining 14: 13. 10.1186/s13040-021-00244-z

License

CC BY 4.0 — Luc T. Miaz, 2026