Admetica
Admetica is an open-source ADMET prediction toolkit from Datagrok that uses Chemprop-based machine learning models for accurate pharmacokinetic property predictions. Ideal for drug design and lead optimization.
What is Admetica?
Admetica is an open-source ADMET prediction toolkit developed by Datagrok. It uses Chemprop message-passing neural networks to predict 23 pharmacokinetic and toxicity properties from SMILES strings. Unlike rule-based filters that check molecular descriptors against thresholds, Admetica learns structure-property relationships directly from experimental data.
The toolkit addresses a practical problem in drug discovery: evaluating absorption, distribution, metabolism, excretion, and toxicity (ADMET) early enough to avoid late-stage failures. Compounds that look promising in target assays often fail because they cannot reach the target, get metabolized too quickly, or cause toxicity. Predicting these properties computationally allows researchers to prioritize compounds before committing to expensive synthesis and testing.
How does Admetica work?
Admetica builds on Chemprop, a graph neural network framework that represents molecules as directed graphs. Atoms become nodes; bonds become edges. The network passes information along bonds, learning to associate local structural features with global molecular properties.
Each ADMET property has its own model trained on curated datasets from ChEMBL, AstraZeneca, and academic sources. Training data ranges from 332 compounds (P-gp substrate) to nearly 10,000 compounds (solubility), with model performance validated against held-out test sets.
The regression models output continuous values (permeability coefficients, clearance rates, LD50), while classification models output probabilities for binary outcomes (CYP inhibitor yes/no, hERG liability yes/no).
How to use Admetica online
ProteinIQ hosts Admetica on cloud infrastructure, eliminating the need to install Python environments or configure GPU compute.
Input
| Format | Description |
|---|---|
| Plain SMILES | One compound per line |
| Tab-delimited | compound_name\tSMILES format to preserve identifiers |
| SDF file | Structure-data file with multiple molecules |
| CSV file | Comma-separated with SMILES column |
| PubChem fetch | Retrieve structures by compound name or CID |
Output columns
Results appear as a spreadsheet with one row per compound. Columns are organized by ADMET category.
Molecular properties
| Column | Description |
|---|---|
molecular_weight | Molecular weight in Daltons |
logp | Octanol-water partition coefficient |
tpsa | Topological polar surface area (Ų) |
hbd | Hydrogen bond donor count |
hba | Hydrogen bond acceptor count |
qed | Quantitative estimate of drug-likeness (0–1) |
Absorption
| Column | Unit/Type | Description |
|---|---|---|
caco2_permeability | log cm/s | Permeability across intestinal epithelial cells |
lipophilicity | log ratio | Octanol-water partition (model-predicted) |
solubility | log mol/L | Aqueous solubility |
pgp_substrate | probability | Likelihood of P-glycoprotein efflux |
Distribution
| Column | Unit/Type | Description |
|---|---|---|
ppbr | percentage | Plasma protein binding rate |
Metabolism
CYP enzymes metabolize the majority of small-molecule drugs. Inhibiting them causes drug-drug interactions; being a substrate affects clearance.
| Column | Type | Description |
|---|---|---|
cyp1a2_inhibitor | probability | CYP1A2 inhibition liability |
cyp2c9_inhibitor | probability | CYP2C9 inhibition liability |
cyp2c19_inhibitor | probability | CYP2C19 inhibition liability |
cyp2d6_inhibitor | probability | CYP2D6 inhibition liability |
cyp3a4_inhibitor | probability | CYP3A4 inhibition liability |
cyp2c9_substrate | probability | Metabolized by CYP2C9 |
cyp2d6_substrate | probability | Metabolized by CYP2D6 |
cyp3a4_substrate | probability | Metabolized by CYP3A4 |
Excretion
| Column | Unit | Description |
|---|---|---|
clearance_hepatocyte | log mL/min/g | Clearance rate in hepatocyte assays |
clearance_microsome | log mL/min/g | Clearance rate in microsome assays |
Toxicity
| Column | Unit/Type | Description |
|---|---|---|
herg | probability | hERG channel inhibition (cardiac risk) |
ld50 | log mg/kg | Acute oral toxicity |
ames | probability | Mutagenicity (Ames test) |
dili | probability | Drug-induced liver injury risk |
skin_sensitization | probability | Skin sensitization potential |
carcinogenicity | probability | Carcinogenic potential |
clinical_toxicity | probability | General clinical toxicity risk |
Interpreting results
Absorption predictions
Caco-2 permeability values above −5 log cm/s suggest good intestinal absorption. Values below −6 indicate poor passive permeability, though active transport can compensate.
Solubility above −4 log mol/L is generally favorable for oral drugs. Lower values may require formulation strategies.
P-gp substrate probability above 0.5 suggests the compound may be pumped out of cells, reducing bioavailability and potentially causing drug-drug interactions with P-gp inhibitors.
Metabolism predictions
CYP inhibition probabilities above 0.7 warrant attention for drug-drug interaction potential. CYP3A4 inhibition is particularly significant because this enzyme metabolizes approximately 50% of marketed drugs.
High clearance predictions (hepatocyte or microsome) indicate rapid metabolism, which may limit exposure. Low clearance can cause accumulation.
Toxicity predictions
hERG inhibition probability above 0.5 flags potential cardiac liability. This channel controls heart rhythm; blocking it can cause fatal arrhythmias. Most drug development programs include hERG screening as a safety gate.
Ames positivity (probability > 0.5) indicates potential mutagenicity. Mutagenic compounds are typically excluded from development unless the therapeutic benefit justifies the risk (e.g., oncology).
Admetica vs ADMET-AI
Both tools use Chemprop neural networks for ADMET prediction. Key differences:
| Aspect | Admetica | ADMET-AI |
|---|---|---|
| Properties | 23 models | 41 models |
| Training data | ChEMBL, AstraZeneca | Therapeutics Data Commons |
| Developer | Datagrok | Stanford/Greenstone |
| Additional properties | Skin sensitization, carcinogenicity, clinical toxicity | BBB penetration, half-life, oral bioavailability |
The tools complement each other. Running both provides broader coverage and a second opinion on shared endpoints.
Limitations
- Chemical space: Models trained primarily on drug-like small molecules. Performance degrades for peptides, natural products, and structures distant from training data.
- No uncertainty quantification: Predictions lack confidence intervals. A 0.6 probability could reflect genuine uncertainty or training data bias.
- Endpoint-specific accuracy: Regression models (Caco-2, solubility, clearance) have published accuracy metrics; classification thresholds are less rigorously calibrated.
- Correlation not causation: Models identify structural patterns associated with properties but do not explain mechanisms.
Related tools
- ADMET-AI: Alternative Chemprop-based predictor with 41 properties including BBB penetration
- Lipinski's Rule of 5: Rule-based druglikeness filter
- Molecular Descriptors: Calculate LogP, TPSA, and other physicochemical properties
- Toxicity Prediction: Structural alert screening with PAINS and BRENK filters
- eToxPred: Deep learning toxicity prediction