ADMET-AI Usage and Endpoint Aggregation

ADMET-AI is a simple, fast, and accurate web interface for predicting the ADMET properties of molecules using machine learning models.

ADMET-AI predicts ADMET properties using a graph neural network architecture called Chemprop-RDKit (see the Chemprop package for details). ADMET-AI's Chemprop-RDKit models were trained on 41 ADMET datasets from the Therapeutics Data Commons (TDC). ADMET-AI’s Chemprop-RDKit models have the highest average rank on the TDC ADMET Benchmark Group leaderboard. ADMET-AI is also currently the fastest web-based ADMET predictor.

This file summarizes the CLI and Python-module usage needed by the ADMET genetic optimization workflow.

Installation

admet-ai depends on torch, so installation and first import may take longer than ordinary Python packages. Wait for model initialization before deciding that the process is stuck.

pip install admet-ai

Running it

Command line:

admet_predict \
    --data_path data.csv \
    --save_path preds.csv \
    --smiles_column smiles

Python module:

from admet_ai import ADMETModel

model = ADMETModel()
preds = model.predict(smiles="O(c1ccc(cc1)CCOC)CC(O)CNC(C)C")

If a single SMILES string is provided, preds is a dictionary mapping property names to values. If a list of SMILES strings is provided, preds is a pandas DataFrame indexed by SMILES with endpoint columns.

Inference can use a GPU when available, but CPU-only execution is supported. If GPU inference failed somehow, try CUDA_DEVICES=""

Suggested aggregation policy

Keep raw ADMET-AI outputs unchanged. Add derived fields instead:

admet_score
admet_risk_flags
admet_failed
admet_predictions_json

Column-name mapping:

admet:
  risk_threshold: 0.5
  positive_keywords:
    - hia
    - caco
    - bioavailability
    - solubility
  negative_keywords:
    - herg
    - ames
    - dili
    - cyp
    - pgp
    - p-gp
    - tox

• Positive endpoints: HIA, Caco2, Bioavailability, Solubility.
• Negative endpoints: hERG, AMES, DILI, CYP, Pgp/P-gp, ClinTox or other toxicity-like names.
• Ignore *_drugbank_approved_percentile columns for risk scoring. These are reference distribution percentiles, not raw endpoint predictions.
• Unknown endpoints should remain in raw predictions and be reported as ignored.

For numeric outputs in the 0-1 range, interpret values directly. For negative endpoints, high values indicate higher risk. For positive endpoints, low values indicate weakness or risk. For values outside 0-1, use a simple squashing transform only for demo scoring, and state this in the report.

Runtime behavior

ADMET-AI can print torch/lightning messages such as GPU availability, NVML warnings, progress bars, and dataloader worker hints. These are usually not fatal. Treat actual exceptions from ADMETModel().predict(...) as ADMET failure and either stop or use a clearly labeled demo fallback.

Use batch prediction where possible. Cache predictions by canonical SMILES so duplicate evaluation attempts do not reload or recompute endpoints.

Troubleshooting