Benchmarking the mutual agreement and experimental Caco-2 accuracy of three modern ADMET predictors across natural products

Most ADMET prediction platforms are trained on synthetic, drug-like compounds, and their performance on natural products is less well established. We compared ADMET-AI, ADMETlab 3.0, and admetSAR 3.0 across 201,905 natural products from NPASS 3.0. Native outputs were mapped to 33 canonical endpoints, of which 30 allowed numerical comparison between at least two platforms. Caco-2 predictions were evaluated against measurements for 134 natural products from one assay program. Pearson correlations with measured permeability ranged from 0.05 to 0.10 (all P ≥ 0.29). Identity-line R 2 values were negative, and binary classification showed little discrimination (AUROC, 0.54–0.61). Predictions were correlated between platforms (mean pairwise r = 0.74), but averaging them did not improve correlation with the measurements. Molecular-weight stratification and the nearest-neighbor applicability-domain criterion did not identify subsets with better predictive performance. Monolayer-integrity filtering increased correlations to at most 0.30, while identity-line R 2 remained negative. Library-wide Caco-2 agreement also differed between subsets defined by the submitted tautomeric form. These findings apply to the evaluated input representations and assay conditions. Experimental validation is needed before these predictions are used to prioritize natural products, and agreement between platforms does not establish accuracy. Scientific contribution This study combines a library-scale comparison of three ADMET platforms with an experimental Caco-2 benchmark for natural products. Neither agreement between platforms nor proximity to a drug-like reference set identified reliable predictions in this benchmark. The released resources include harmonized predictions, an endpoint crosswalk, experimental measurements, submitted structures, and code that recomputes a defined set of reported results.

Authors

Institutions

Publication Details

Journal
Journal of Cheminformatics
Published
2026-11-20
DOI
https://doi.org/10.1186/s13321-026-01331-1
Primary Topic
Computational Drug Discovery Methods
Type
article
Field-Weighted Citation Impact
0.00

Funders

Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

Benchmarking the mutual agreement and experimental Caco-2 accuracy of three modern ADMET predictors across natural products

Soon‐Mi Shim, Keunwan Park, Kyungsu Kang, Jung‐Seok Yang et al.
2 links
Journal of Cheminformatics
Computational Drug Discovery Methods
article

Benchmarking the mutual agreement and experimental Caco-2 accuracy of three modern ADMET predictors across natural products

Soon‐Mi Shim, Keunwan Park, Kyungsu Kang, Jung‐Seok Yang, Jin‐Soo Park, Eunjung Kim, Inha Hwang
article en

Abstract

Most ADMET prediction platforms are trained on synthetic, drug-like compounds, and their performance on natural products is less well established. We compared ADMET-AI, ADMETlab 3.0, and admetSAR 3.0 across 201,905 natural products from NPASS 3.0. Native outputs were mapped to 33 canonical endpoints, of which 30 allowed numerical comparison between at least two platforms. Caco-2 predictions were evaluated against measurements for 134 natural products from one assay program. Pearson correlations with measured permeability ranged from 0.05 to 0.10 (all P ≥ 0.29). Identity-line R 2 values were negative, and binary classification showed little discrimination (AUROC, 0.54–0.61). Predictions were correlated between platforms (mean pairwise r = 0.74), but averaging them did not improve correlation with the measurements. Molecular-weight stratification and the nearest-neighbor applicability-domain criterion did not identify subsets with better predictive performance. Monolayer-integrity filtering increased correlations to at most 0.30, while identity-line R 2 remained negative. Library-wide Caco-2 agreement also differed between subsets defined by the submitted tautomeric form. These findings apply to the evaluated input representations and assay conditions. Experimental validation is needed before these predictions are used to prioritize natural products, and agreement between platforms does not establish accuracy. Scientific contribution This study combines a library-scale comparison of three ADMET platforms with an experimental Caco-2 benchmark for natural products. Neither agreement between platforms nor proximity to a drug-like reference set identified reliable predictions in this benchmark. The released resources include harmonized predictions, an endpoint crosswalk, experimental measurements, submitted structures, and code that recomputes a defined set of reported results.

Journal of Cheminformatics
Sejong University (KR), Korea Institute of Science and Technology (KR), Korea University of Science and Technology (KR)
Korea Institute of Science and Technology
Good health and well-being
Computational Drug Discovery Methods
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon