AI models for drug metabolism & pharmacokinetics
Predict ADME, PK, PK/PD, CYP interactions, clearance, exposure, and other DMPK properties using production-ready machine-learning models — accessible from a single API.
input · smiles
output · endpoints
Platform
Machine-learning models purpose-built for DMPK
DMPK.AI provides a unified suite of machine-learning models for drug metabolism and pharmacokinetic prediction — from early discovery screening through translational PK/PD modeling — delivered as a single, consistent API.
Full DMPK coverage
One platform spanning ADME, CYP interactions, clearance, and PK/PD — instead of stitching together point tools.
Production-ready inference
Models are versioned, benchmarked, and served behind a stable API built for high-throughput screening.
Calibrated, not overconfident
Every endpoint reports uncertainty alongside the prediction, validated against held-out data.
atoms
27
rings
2
embedding
512-d
Model catalog
Eleven endpoints across the DMPK stack
Every model is trained on curated, harmonized data and shipped with calibrated uncertainty — from early ADME triage to PK/PD translation.
CYP
CYP Inhibition
Predicts inhibition risk across major CYP450 isoforms.
- Input
- SMILES / molecular structure
- Output
- Per-isoform inhibition probability + confidence interval
CYP
CYP Substrate Classification
Classifies whether a compound is a substrate of major CYP enzymes.
- Input
- SMILES / molecular structure
- Output
- Multi-label substrate classification (per isoform)
ADME
Microsomal Stability
Predicts metabolic stability in liver microsome assays.
- Input
- SMILES + species
- Output
- % remaining at 30 min / predicted half-life class
Clearance
Intrinsic Clearance
Predicts hepatic intrinsic clearance (CLint).
- Input
- SMILES + assay system
- Output
- CLint (µL/min/mg protein) with uncertainty band
ADME
Plasma Protein Binding
Predicts fraction unbound in plasma across species.
- Input
- SMILES + species
- Output
- Fraction unbound (fu) with calibrated interval
ADME
Permeability
Predicts passive membrane permeability (Caco-2 / PAMPA-like).
- Input
- SMILES / molecular structure
- Output
- Papp class + continuous estimate (10⁻⁶ cm/s)
PK
Half-Life
Predicts terminal elimination half-life.
- Input
- SMILES + species + route
- Output
- t½ (hours) with uncertainty band
PK
Volume of Distribution
Predicts steady-state volume of distribution.
- Input
- SMILES + species
- Output
- Vss (L/kg) with confidence interval
PK
Oral Bioavailability
Predicts fraction of oral dose reaching systemic circulation.
- Input
- SMILES + species + formulation notes
- Output
- %F estimate with calibrated interval
PK
PK Parameter Prediction
Joint prediction of core PK parameters from structure.
- Input
- SMILES + species + dose route
- Output
- CL, Vss, t½, Cmax, AUC — joint posterior estimates
PK/PD
PK/PD Modeling
Links exposure predictions to pharmacodynamic response.
- Input
- PK profile + target/assay context
- Output
- Simulated exposure–response curves
Developer & API
Predict from a single call
The Python SDK and REST API expose every model behind one consistent interface — built for pipelines, ELNs, and internal screening tools.
from dmpk import DMPK
client = DMPK(api_key="YOUR_API_KEY")
result = client.predict(
model="cyp3a4-inhibition",
smiles="CC(=O)Oc1ccccc1C(=O)O",
)
print(result){
"model": "cyp3a4-inhibition-v2",
"input": {
"smiles": "CC(=O)Oc1ccccc1C(=O)O"
},
"prediction": {
"inhibition_probability": 0.12,
"confidence_interval": [0.07, 0.19]
},
"latency_ms": 142
}AI & ML technology
Modeling approach built for chemistry
A shared modeling foundation underlies every endpoint — designed to represent molecular structure faithfully and report predictions you can trust.
Graph neural networks
Molecules represented as atom-bond graphs, preserving structural and stereochemical detail.
Transformers
Attention-based architectures capture long-range substructure interactions and context.
Molecular representations
Learned embeddings combine graph, descriptor, and fingerprint views of each compound.
Uncertainty estimation
Every prediction ships with a calibrated confidence interval, not a bare point estimate.
Ensemble learning
Model committees reduce variance and flag out-of-distribution inputs.
Multi-task learning
Shared representations across related endpoints improve data efficiency and consistency.
Explainable AI
Atom- and fragment-level attributions surface the structural drivers behind a prediction.
Calibrated predictions
Confidence intervals are validated against held-out data, not just reported.
Workflow
One pipeline, from structure to decision
Each stage runs on dedicated models and feeds the next — so a single molecule input produces a coherent DMPK profile.
Molecule
Input structure via SMILES, batch file, or API call.
ADME
Absorption, permeability, and stability predicted in parallel.
CYP
Inhibition and substrate liability screened across isoforms.
Clearance
Intrinsic and hepatic clearance estimated with uncertainty.
PK
Core pharmacokinetic parameters projected across species.
PK/PD
Exposure linked to pharmacodynamic response models.
Decision
Ranked, explainable output supports go/no-go calls.
Data
Model-ready data, curated end to end
Models are trained and validated on curated public, licensed, and proprietary datasets — standardized and quality-controlled before a single parameter is fit.
Data standardization
Structures, units, and assay conditions are normalized to a consistent schema before modeling.
Assay harmonization
Results from heterogeneous assay formats and protocols are reconciled onto comparable scales.
Quality control
Automated and expert review filters outliers, duplicates, and low-confidence measurements.
Train / validation / test separation
Scaffold-aware splitting prevents structural leakage and overstated performance.
Model-ready datasets
Curated datasets are versioned and packaged for direct use in training and benchmarking.
Solutions
Built for every stage of drug development
Early discovery
Screen virtual libraries for ADME and CYP liability before synthesis.
Lead optimization
Rank analogs by predicted clearance, exposure, and interaction risk.
DMPK screening
Triage compound sets across the full DMPK endpoint panel in one pass.
Candidate selection
Compare candidates on calibrated, explainable PK and safety-relevant signals.
Translational modeling
Bridge in vitro predictions to projected human PK and PK/PD behavior.
API integration
Embed predictions directly into internal pipelines, ELNs, and screening platforms.
Custom model development
Train endpoint-specific models on your proprietary assay data.
Validation & metrics
Performance reported the way scientists expect
Held-out performance is tracked with standard regression and classification metrics, plus calibration coverage — not marketing figures.
Figures below are illustrative example metrics for demonstration purposes only — they are not production performance claims or validated regulatory results.
plasma concentration · time
Example predicted vs. observed profile (demo)
CYP3A4 Inhibition
AUROC
0.91
Intrinsic Clearance
RMSE (log)
0.38
Plasma Protein Binding
R²
0.85
Permeability
MAE (log)
0.44
Half-Life
Coverage @ 95% CI
92%
Microsomal Stability
R²
0.79
Enterprise & custom models
Bring your own data. Own the model.
Organizations can request custom model development, private deployment, or training on proprietary assay data — extending the DMPK.AI platform to endpoints and chemical spaces unique to your pipeline.
Talk to our teamPrivate deployment
Run inference in your own VPC or on-prem environment.
Proprietary training
Fine-tune or train new endpoints on your internal assay data.
Custom models
Commission bespoke models for endpoints outside the standard catalog.