---
title: ADMET-AI
description: Predict 41 TDC ADMET endpoints and 8 RDKit descriptors from SMILES with Chemprop-RDKit graph neural networks — hERG, AMES, DILI, CYP450 and clearance.
tags:
  - admet-ai
---

ADMET-AI predicts absorption, distribution, metabolism, excretion, and toxicity liabilities across a broad panel of endpoints straight from SMILES. Reach for it for a fast, broad first screen — early hit triage and library-scale prioritization — when you want to flag safety and pharmacokinetic risks before committing synthesis. ADMETLab3, which offered a wider endpoint set, is no longer available, so ADMET-AI is the platform's ADMET predictor. See [what you can achieve](/docs/capabilities#admet-and-safety) for the full set.

## Try it

**Try this prompt** — [open in a guest chat](/chat/new?guest=true&prompt=Compare%20gefitinib%2C%20erlotinib%2C%20and%20osimertinib%20head-to-head%20on%20safety%20%E2%80%94%20CYP%20inhibition%2C%20hERG%20cardiotoxicity%2C%20oral%20bioavailability%2C%20the%20kinome-wide%20off-target%20panel%20for%20each%2C%20and%20any%20real-world%20adverse-event%20signals.%20Tell%20me%20which%20compound%20has%20the%20cleanest%20safety%20profile%2C%20and%20whether%20the%20predicted%20ADMET%20explains%20the%20differential%20clinical%20toxicity%20we%20actually%20see.).

> Compare gefitinib, erlotinib, and osimertinib head-to-head on safety — CYP inhibition, hERG cardiotoxicity, oral bioavailability, the kinome-wide off-target panel for each, and any real-world adverse-event signals. Tell me which compound has the cleanest safety profile, and whether the predicted ADMET explains the differential clinical toxicity we actually see.

_Screens all three with ADMET-AI, then chains DeepPurpose off-targets, FAERS signals, and pharmacogenomic guidelines._

## ADMET-AI integration in .bloom* platform

ADMET-AI is a state-of-the-art machine learning platform for fast and accurate ADMET property prediction. Developed by researchers at Stanford University, it uses Chemprop-RDKit graph neural networks trained on 41 ADMET datasets from the Therapeutics Data Commons (TDC). Through .bloom*'s webapp, you can screen compound libraries for pharmacokinetic and safety properties without complex setup.

### Citations

> Swanson K, Walther P, Leitz J, Mukherjee S, Wu JC, Shivnaraine RV, Zou J. ADMET-AI: a machine learning ADMET platform for evaluation of large-scale chemical libraries. *Bioinformatics*. 2024 Jul 1;40(7):btae416. doi: [10.1093/bioinformatics/btae416](https://doi.org/10.1093/bioinformatics/btae416)

> Huang K, Fu T, Gao W, Zhao Y, Roohani Y, Leskovec J, Coley CW, Xiao C, Sun J, Zitnik M. Therapeutics Data Commons: Machine Learning Datasets and Tasks for Drug Discovery and Development. *NeurIPS Datasets and Benchmarks*. 2021. arXiv: [2102.09548](https://arxiv.org/abs/2102.09548) · [https://tdcommons.ai](https://tdcommons.ai)

### Technical foundation

### Model architecture
ADMET-AI uses **Chemprop-RDKit**, a graph neural network architecture that combines:
- Message passing neural networks for molecular graph representation
- RDKit-computed molecular descriptors as additional features
- Trained on the Therapeutics Data Commons (TDC) ADMET Benchmark datasets

### Performance benchmarks

Every figure below is author-reported, carries the dataset and sample size it was measured on, and comes from TDC's **scaffold-split** test sets — the setting where the model must generalise to chemical series it did not train on. Read [how ADMET-AI results compare to the bench](/docs/admet-ai/translatability) before acting on any of them.

- **Toxicity classifiers:** hERG AUROC 0.840 (hERG, n = 655) · mutagenicity 0.850 (AMES, n = 7278) · drug-induced liver injury 0.887 (DILI, n = 475).
- **Strongest regression endpoints:** aqueous solubility MAE 0.761 log mol/L (AqSolDB, n = 9982) and lipophilicity MAE 0.467 log-ratio (AstraZeneca, n = 4200) — best on the leaderboard for both.
- **Weakest endpoints, and they are the pharmacokinetic ones:** half-life Spearman 0.239 (Obach, n = 667) and volume of distribution Spearman 0.389 (Lombardo, n = 1130). Do not read these as quantities.
- **Highest average rank** across the 22 TDC ADMET Benchmark Group datasets as of the paper's snapshot (4 October 2023). That is an average: on the same snapshot the model ranked 10th on volume of distribution and 11th on blood–brain barrier.
- **Fastest of the public ADMET web servers** measured in the 2024 paper — a 45% time reduction over the next fastest for 1000 molecules. That comparison is on speed only, not accuracy.

## Webapp input and processing capabilities

### Compound input methods

**Direct SMILES entry:**
- Paste individual SMILES strings with instant validation
- Real-time structure preview and property calculation
- Support for stereochemistry and complex molecules

**Batch file processing:**
- Upload CSV files with SMILES in the first column
- Support for multi-molecule SDF files
- Automatic format detection and parsing
- Progress tracking for large datasets (thousands of compounds)

**Integration with other tools:**
- Direct import from generative chemistry tools (DeepPurpose, AiZynthFinder)
- Connection to virtual screening results
- Import from compound library searches (ZINC, PubChem)

### Prediction endpoints available

ADMET-AI computes **8 physicochemical properties** using RDKit and predicts **41 ADMET properties** across the following categories:

**Absorption predictions:**
- **Caco-2 permeability**: Intestinal absorption prediction
- **HIA**: Human intestinal absorption classification
- **Pgp inhibition/substrate**: P-glycoprotein interaction predictions
- **Bioavailability**: Oral bioavailability classification

**Distribution predictions:**
- **BBB permeability**: Blood-brain barrier penetration
- **PPB**: Plasma protein binding (fraction unbound)
- **VDss**: Volume of distribution at steady state

**Metabolism predictions:**
- **CYP450 inhibition**: CYP1A2, CYP2C9, CYP2C19, CYP2D6, CYP3A4 inhibition
- **CYP450 substrate**: Substrate classification for CYP2C9, CYP2D6, CYP3A4

**Excretion predictions:**
- **Clearance**: hepatocyte and microsomal clearance, as regression values in the assay's own units. There is no renal clearance model
- **Half-life**: a regression value in hours, not a stability class — and the weakest endpoint the model reports

**Toxicity predictions:**
- **hERG inhibition**: cardiotoxicity risk (QT prolongation)
- **AMES toxicity**: mutagenicity prediction
- **DILI**: drug-induced liver injury risk
- **Skin reaction**: dermal toxicity potential (n = 404)
- **Carcinogenicity**: carcinogen classification (n = 278)
- **Clinical trial toxicity** and **acute toxicity (LD50)**: two separate endpoints — a binary ClinTox probability and an LD50 regression value
- **Tox21 panel**: 12 nuclear-receptor and stress-response assay endpoints. These are a toxicity screen, not an off-target selectivity panel

### Quality control and validation

**Structure processing:**
- Automatic SMILES canonicalization and standardization
- Salt stripping and neutralization
- Duplicate detection and removal
- Structure validation with detailed error reporting

**Prediction reliability:**
- Confidence scores for each prediction
- Applicability domain assessment based on training set similarity
- Model version tracking for reproducibility
- Uncertainty quantification via 5-model ensemble predictions

No reliability field tells you whether the model has seen chemistry resembling your molecule. That is the dominant caveat on every number here, and it is unmeasured — see [translatability](/docs/admet-ai/translatability).

## When to use ADMET-AI in drug discovery

### Early compound evaluation

**Virtual screening triage:**
- Filter large compound libraries before synthesis
- Prioritize compounds with favorable ADMET profiles
- Flag molecules with potential safety concerns
- Order the wet-lab queue — not replace it. No endpoint here is validated to substitute for its assay

**Lead optimization:**
- Assess ADMET properties of hit compounds
- Guide medicinal chemistry optimization decisions
- Balance potency with developability parameters
- Identify compounds requiring formulation development

**Compound library design:**
- Evaluate virtual compound collections
- Design libraries with improved ADMET properties
- Assess synthetic accessibility alongside drug-like properties
- Optimize for multiple parameter objectives

### Safety triage

These are internal triage uses. ADMET-AI is not qualified under ICH M7 or any other guideline that admits a computed prediction in place of a test, so nothing here is suitable for a regulatory submission.

**Safety profiling:**
- Early flagging of toxicological liabilities
- A cardiotoxicity signal (hERG probability) to prioritize a patch-clamp assay
- A first look at metabolic liabilities

**Drug metabolism:**
- CYP inhibition and substrate probabilities — these are *inputs* to a drug–drug interaction assessment, not a DDI assessment
- Hepatocyte and microsomal clearance estimates, for ranking rather than as values
- Planning which in vitro metabolism studies to run first

It does **not** predict metabolic pathways, metabolite structures, or drug–drug interactions.

## Performance characteristics

### Batch processing efficiency

**Scalability:**
- Support for thousands of compounds per job
- Parallel processing with automatic load balancing
- GPU acceleration for large prediction batches
- Memory optimization for high-throughput screening

**Error handling:**
- Individual compound failure handling
- Detailed error reporting per molecule
- Batch continuation despite individual failures
- Retry logic for transient service issues

### Result management

**Output formats:**
- **CSV exports**: Tabular format for spreadsheet analysis
- **JSON artifacts**: Structured data for programmatic access
- **Interactive dashboards**: Real-time result visualization
- **Statistical summaries**: Distribution analysis across datasets

**Result interpretation:**
- **Scoring systems**: Standardized risk categories (Low/Medium/High)
- **Confidence intervals**: Prediction uncertainty quantification
- **Comparative analysis**: Side-by-side evaluation of compound series
- **Trend identification**: Property optimization patterns

## Integration with .bloom* ecosystem

### Upstream data sources

**Chemistry tools:**
- **DeepPurpose**: Evaluate generated molecule properties
- **AiZynthFinder**: Assess synthetic intermediates
- **Fragment libraries**: Screen building block collections
- **Virtual screening**: ADMET evaluation of docking hits

### Downstream applications

**Decision support:**
- Multi-parameter optimization workflows
- Compound advancement criteria definition
- Risk-benefit analysis frameworks
- Portfolio prioritization strategies

**Experimental planning:**
- Guide in vitro assay selection
- Plan in vivo pharmacokinetic studies
- Design formulation development strategies
- Optimize preclinical candidate selection

### Workflow automation

**High-throughput pipelines:**
- Automated batch processing workflows
- Integration with compound management systems
- Real-time result monitoring and alerting
- Quality control and data validation

**Research collaboration:**
- Shared prediction results across teams
- Integration with electronic lab notebooks
- Audit trails for regulatory compliance
- Reproducible analysis workflows

## Chat Usage

### Example Prompts

- "Run ADMET-AI on CC(=O)Oc1ccccc1C(=O)O" (aspirin)
- "Screen these 10 compounds through ADMET-AI and export results as CSV"
- "Predict drug-likeness for gefitinib and erlotinib — compare hERG profiles"

### Workflow Chaining

- **Feeds into**: Compound Passport (ADMET section), comparison tables
- **Accepts from**: PubChem, ZINC, ChEMBL (compound SMILES), REINVENT/GenMol (generated molecules)

## FAQ

### What inputs does ADMET-AI accept?

  Provide SMILES or SDF; batches are queued as jobs for processing.

### Where do results live?

  Outputs are stored as job artifacts and downloadable CSVs in the workspace.

## Read next
