Method Profile
Molecular-graph models for absorption, distribution, metabolism and excretion (ADME)-related properties were compared with measured metabolic stability, solubility and passive permeability in selected Tox21 compounds.
Early compound prioritization when cellular bioactivity data need exposure-related context. This is a laboratory ADME benchmark, not a validated in vivo toxicity predictor.
NCATS used ADME@NCATS molecular-graph models, including directed message-passing neural networks, to predict rat liver microsomal stability, kinetic aqueous solubility, passive gastrointestinal permeability and passive blood–brain barrier permeability.
The basic ADME comparison covered 180 selected compounds; 80 existing measurements were removed from model training before evaluation and 100 compounds underwent new assays. A separate availability-selected set of 55 acetylcholinesterase-active compounds was tested for passive blood–brain barrier permeability.
Tox21 compound sets
A selected 180-compound subset for basic ADME endpoints and a separate, nonrandom 55-compound subset for the blood–brain barrier assay. The sets include drugs and other chemicals.
Laboratory comparators
Rat liver microsomal stability, kinetic aqueous solubility, PAMPA gastrointestinal permeability and PAMPA blood–brain barrier permeability, each compared with its corresponding model output.
Compare each model with its corresponding experimental classification; remove overlapping measured compounds from training; examine structural similarity and errors across chemical space. Interpret the basic ADME and blood–brain barrier sets separately.
Experimental classification of selected Tox21 compounds for three basic ADME endpoints after training-overlap removal.
180 compounds; 80 with prior NCATS measurements were removed from training and 100 were newly assayed. Endpoint measurements had some missing or inconclusive values.
Laboratory rat liver microsomal stability, PAMPA gastrointestinal permeability and aqueous solubility.
ROC AUC: 0.87 for rat liver microsomal stability, 0.92 for PAMPA gastrointestinal permeability and 0.81 for aqueous solubility. Balanced accuracy was 0.78, 0.81 and 0.75, respectively; solubility overall accuracy was 0.83, sensitivity 0.59 and specificity 0.91. Approximately 40% of selected compounds had maximum Tanimoto similarity at or below 0.4 to relevant training data.
Separate comparison of passive blood–brain barrier permeability predictions with PAMPA-BBB measurements.
55 available acetylcholinesterase-active compounds, not a random or fully representative Tox21 sample.
Corresponding PAMPA-BBB passive-permeability classification.
ROC AUC 0.77; overall and balanced accuracy both 0.76; sensitivity 0.68 and specificity 0.85.
NIH researchers compared ADME@NCATS predictions with laboratory measurements for selected Tox21 compounds. Three basic ADME models reached ROC AUCs of 0.81–0.92; a separate passive blood–brain barrier model reached 0.77. The work supports compound triage, not validated prediction of in vivo toxicity.
